A method for predicting slope deformation trend based on temporal feature deep learning

By decomposing and extracting features from the time-series data of slope surface deformation and displacement, and combining one-dimensional convolutional channels and long short-term memory networks, corrected deformation prediction values ​​and upper and lower bounds are generated. This solves the mixed problem of long-term cumulative changes, periodic fluctuations and triggering factor changes in slope monitoring, and achieves accurate correction and continuous updating of prediction values.

CN123116773APending Publication Date: 2026-11-17SANMING UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611065206.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-17
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing slope safety monitoring technologies, long-term cumulative changes, periodic fluctuations, and triggering factor changes are mixed together in the unified feature extraction process, and the prediction error states of different prediction step sizes are difficult to distinguish. This leads to the prediction results deviating from historical measurement values ​​and inconsistent upper and lower bounds of prediction. The state of missing value imputation and high-frequency fluctuation states are difficult to reflect in the prediction output.

Method used

The time series data of slope surface deformation and displacement are decomposed into trend components, periodic components and high-frequency residual components. The trend feature vector is extracted using a one-dimensional convolutional channel and generated with a long short-term memory network to generate a periodic time series feature vector. The uncorrected deformation prediction value is corrected by the prediction error window to generate the corrected deformation prediction value and prediction upper and lower bounds. The prediction upper and lower bounds are adjusted according to the data quality status and high-frequency residual components.

Benefits of technology

It effectively maintains the separate representation of trend patterns and cycle-trigger time series characteristics, realizes the correction of deformation prediction values ​​and continuous updating of prediction range at different prediction step lengths, solves the problems of distinguishing prediction error states and data quality changes, and improves the accuracy and consistency of prediction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116773A_ABST
    Figure CN123116773A_ABST
Patent Text Reader

Abstract

This invention relates to the field of slope safety monitoring and geological disaster early warning technology, and particularly to a method for predicting slope deformation trends based on temporal feature deep learning. The method acquires temporal data of slope surface deformation displacement and triggering factors, and generates trend components, periodic components, high-frequency residual components, and data quality status through alignment, outlier identification, missing value imputation, and temporal decomposition. Features are extracted through one-dimensional convolutional channels and long short-term memory network channels, and concatenated to obtain uncorrected deformation prediction values. The prediction error window is invoked according to the prediction step size. While keeping the model parameters unchanged, the prediction values ​​are corrected and upper and lower bounds are generated. The upper and lower bounds are adjusted according to the data quality status and high-frequency residual components, and an early warning command is output and the error window is updated. This invention enables the separate characterization of deformation features at varying scales, allows the invocation of corresponding historical error states at different prediction step sizes, and improves the adaptability of rolling prediction and early warning judgment.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of slope safety monitoring, monitoring time series data processing, and geological disaster early warning technology, and in particular to a method for predicting slope deformation trends based on time series feature deep learning. Background Technology

[0002] In the field of slope safety monitoring technology, existing solutions typically collect time-series data of slope surface deformation and displacement using displacement monitoring devices, and collect time-series data of rainfall, pore water pressure, or groundwater level using environmental sensors. Data alignment, outlier identification, and missing value imputation are then performed according to the sampling timestamps. Subsequently, statistical prediction models, convolutional neural networks, long short-term memory networks, or multi-scale time-series models are used to extract deformation and change features, generate predicted values ​​of deformation and displacement for future times, and then compare the predicted values ​​of deformation and displacement with early warning thresholds to form a risk level or early warning status.

[0003] Slope surface deformation and displacement time series typically encompass long-term cumulative changes, periodic fluctuations, and high-frequency disturbances simultaneously, with different components exhibiting varying time scales and patterns of change. Existing methods often directly input the original displacement time series and triggering factor time series data into the same time series prediction model, or employ the same type of feature extraction structure to process deformation components at different time scales. Furthermore, triggering factors may undergo end-point fusion after deformation feature extraction, making it difficult to maintain the correspondence between triggering factors and periodic deformation at various time locations. Consequently, long-term variation characteristics, periodic fluctuation characteristics, and triggering factor changes are easily mixed during the unified feature extraction process.

[0004] In continuous rolling forecasting, the distribution of prediction errors differs across different prediction step lengths, and monitoring data may exhibit short-term missing values, continuous missing values, outlier replacements, and high-frequency fluctuations. Existing solutions typically use historical prediction errors uniformly for model parameter adjustment, prediction result evaluation, or early warning threshold correction, without distinguishing the error states at different prediction step lengths. Furthermore, missing value imputation and high-frequency fluctuation states are often addressed only in the data preprocessing stage, failing to reflect them in the prediction range corresponding to the predicted values. When model parameters remain constant while monitoring data quality changes, the continuous shift between prediction results and historical measurements, error differences between different prediction step lengths, and inconsistencies between prediction upper and lower bounds and the current data state are difficult to handle uniformly. Therefore, it is necessary to address the issues of correcting deformation prediction values ​​and generating prediction upper and lower bounds at different prediction step lengths under conditions where slope deformation time series include different scales of change, monitoring data quality changes, and model parameters remain constant. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method for predicting slope deformation trends based on time-series feature deep learning, comprising: S100: Acquire slope surface deformation and displacement time series data, trigger factor time series data and sampling timestamps, align them according to the sampling timestamps, identify outliers and interpolate missing values, and generate equal interval monitoring time series and data quality status; decompose the surface deformation and displacement time series data in the equal interval monitoring time series to generate trend components, periodic components and high-frequency residual components. S200: Input the trend component into a one-dimensional convolutional channel to extract the trend feature vector; input the periodic component and the triggering factor time-series data associated with the same time step into the long short-term memory network channel encoding to generate a periodic time-series feature vector; concatenate and map the two feature vectors to generate the uncorrected deformation prediction value of the prediction step size; S300: Obtain a prediction error window formed by the uncorrected deformation prediction value and the measured value at the prediction target time, corresponding to the prediction step size; when the model parameters remain unchanged, correct the uncorrected deformation prediction value according to the mean of the prediction error to generate a corrected deformation prediction value; generate the upper and lower bounds of the prediction based on its standard deviation, and adjust the upper and lower bounds of the prediction based on the data quality status and high-frequency residual components. S400: Match the corrected deformation prediction value and the upper and lower bounds of the prediction with the graded early warning threshold, and output an early warning command; update the prediction error window after the measured value is reached.

[0006] Further, step S100 includes: resampling the time series data of surface deformation displacement and the time series data of triggering factors according to a uniform sampling interval; identifying outliers in the time series data of surface deformation displacement based on the moving median; imputing missing values ​​in missing segments with a missing length not greater than the maximum allowable imputation length, and recording missing segments with a missing length greater than the maximum allowable imputation length as consecutive missing segments; generating the data quality status according to the imputation position of the missing value imputation, the effective data ratio, and the consecutive missing length of the consecutive missing segments.

[0007] Further, the decomposition in step S100 includes: performing empirical mode decomposition on the surface deformation displacement time series data in the equally spaced monitoring time series to generate multiple candidate components; calculating the average period and energy percentage of each candidate component; combining the candidate components with an average period greater than the upper limit of the period range and the residual terms of the empirical mode decomposition into the trend component; combining the candidate components with an average period within the period range into the period component; and combining the candidate components with an average period less than the lower limit of the period range into the high-frequency residual component.

[0008] Furthermore, in step S200, the one-dimensional convolution channel is set with multiple convolution kernel scales; the slope change, inflection point and acceleration segment features of the trend component within the corresponding sliding window are extracted by each convolution kernel scale; the extraction results of each convolution kernel scale are concatenated to generate the trend feature vector.

[0009] Further, the triggering factor time series data is rainfall time series data or pore water pressure time series data; step S200 includes: standardizing the triggering factor time series data; merging the periodic component and the triggering factor time series data at the same time step in time order into a data matrix; inputting the data matrix into the long short-term memory network channel for time-series encoding to generate the periodic time series feature vector.

[0010] Further, step S200 includes: splicing the trend feature vector and the periodic time series feature vector in the feature dimension to generate a fused feature vector; mapping the fused feature vector to generate multiple uncorrected deformation prediction values ​​corresponding to multiple prediction step lengths; associating each uncorrected deformation prediction value with the slope monitoring point identifier, prediction generation time, prediction target time, prediction step length and model version identifier to generate a prediction result record.

[0011] Further, step S300 includes: associating the measured value with the uncorrected deformation prediction value in the prediction result record according to the slope monitoring point identifier and the prediction target time; calculating the difference between the measured value and the uncorrected deformation prediction value to generate a prediction error; grouping multiple prediction errors according to the model version identifier and prediction step size, and writing the prediction errors in the same group into the corresponding prediction error window in chronological order; calculating the mean of the prediction errors in the prediction error window to generate an offset correction amount; and adding the offset correction amount to the current uncorrected deformation prediction value to generate the corrected deformation prediction value.

[0012] Further, the upper and lower bounds of the prediction include an upper bound and a lower bound; step S300 includes: calculating the standard deviation of the prediction error within the prediction error window; taking the sum of the product of the corrected deformation prediction value and the standard deviation and the interval coefficient as the upper bound of the prediction, and taking the difference between the corrected deformation prediction value and the product as the lower bound of the prediction; adjusting the interval coefficient according to the effective data ratio, the continuous missing length, the energy ratio of the high-frequency residual component and the number of error samples; when the number of error samples is less than the error sample number threshold, recording the insufficient prediction error sample state, and increasing the difference between the upper bound and the lower bound of the prediction.

[0013] Further, step S400 includes: assessing a first risk level when the upper bound of the prediction exceeds the corresponding graded warning threshold and the corrected deformation prediction value does not exceed the graded warning threshold; assessing a second risk level when the corrected deformation prediction value exceeds the graded warning threshold and the lower bound of the prediction does not exceed the graded warning threshold; assessing a third risk level when the lower bound of the prediction exceeds the graded warning threshold, or when the number of consecutive exceedances reaches a preset number; outputting a corresponding warning instruction according to the first risk level, the second risk level, or the third risk level; and releasing the warning instruction when the upper bound of the prediction is lower than the warning release threshold and the duration reaches a preset time.

[0014] Furthermore, the update of the prediction error window includes: writing the newly generated prediction error into a prediction error window associated with the model version identifier and prediction step size; when the number of error samples in the prediction error window is greater than the sliding window length, removing the earliest written prediction error in chronological order; recording the update time of the prediction error window; and using the updated prediction error window for the correction of the next uncorrected deformation prediction value and the generation of prediction upper and lower bounds, while keeping the model parameters of the deformation prediction model unchanged.

[0015] The key innovations of this invention include:

[0016] (1) Decompose the time series data of slope surface deformation displacement into trend component, periodic component and high-frequency residual component; input the trend component into a one-dimensional convolution channel to extract trend feature vector, input the periodic component and the trigger factor time series data associated with the same time step into the long short-term memory network channel to generate periodic time series feature vector; concatenate the two feature vectors in the feature dimension and perform mapping to generate uncorrected deformation prediction value associated with the prediction step size.

[0017] (2) Maintain prediction error windows according to the slope monitoring point identifier, model version identifier and prediction step size respectively; the prediction error window is formed by the prediction error between the historical uncorrected deformation prediction value and the measurement value at the corresponding prediction target time; under the condition that the model parameters remain unchanged, correct the current uncorrected deformation prediction value according to the mean of the prediction error in the same prediction error window, and generate the upper and lower bounds of prediction according to the standard deviation of the prediction error.

[0018] (3) Transmit the data quality status formed by outlier identification and missing value imputation, as well as the high-frequency residual components formed by time decomposition, to the prediction upper and lower bound generation stage; adjust the prediction upper and lower bounds according to the effective data ratio, continuous missing length, energy ratio of high-frequency residual components and number of error samples; match the corrected deformation prediction value and the prediction upper and lower bounds with the graded early warning threshold, and update the prediction error window of the corresponding prediction step size after the measurement value at the prediction target time is reached.

[0019] The following are its main beneficial effects:

[0020] (1) To address the problem of long-term cumulative changes, periodic fluctuations and triggering factor changes being mixed in the unified feature extraction process, the trend component extracts local slope changes, inflection points and acceleration segment features through a one-dimensional convolutional channel, while the periodic component and the triggering factor time series data enter the long short-term memory network channel at the same time step; thus, the trend morphological features and the periodic-trigger time series features are represented separately before entering the mapping process, and an uncorrected deformation prediction value corresponding to the prediction step size is formed.

[0021] (2) To address the problem that it is difficult to distinguish the prediction error states of different prediction step lengths and that continuous prediction offset requires readjustment of model parameters, prediction error windows formed by historical uncorrected deformation prediction values ​​and target time measurement values ​​are maintained according to the prediction step length. The mean of the prediction error window directly affects the current uncorrected deformation prediction value, and its standard deviation is used to form the upper and lower bounds of the prediction. This allows each prediction step length to call the historical error state corresponding to its prediction span, and to complete the deformation prediction value correction and prediction range generation while keeping the model parameters unchanged.

[0022] (3) To address the problem that the missing value imputation status and high frequency fluctuation status remain in the data preprocessing stage and are difficult to reflect in the prediction output, the data quality status and high frequency residual components participate in the adjustment of the upper and lower bounds of the prediction. After correction, the deformation prediction value, the upper bound of the prediction and the lower bound of the prediction enter the hierarchical early warning threshold matching respectively. After the target time measurement value is reached, it returns to the prediction error window of the corresponding prediction step size, so that the prediction range is updated with the monitoring data status and historical error status, and a continuous prediction, early warning and error status back-write relationship is formed. Attached Figure Description

[0023] Figure 1 A flowchart illustrating a slope deformation trend prediction method based on temporal feature deep learning provided in this application embodiment;

[0024] Figure 2 The diagram shows a structural block diagram of a slope deformation trend prediction method based on temporal feature deep learning, which is provided in the embodiments of this application. Detailed Implementation

[0025] Example 1: Refer to Figure 1 This is a flowchart illustrating a slope deformation trend prediction method based on temporal feature deep learning provided in an embodiment of the present invention. The flowchart may include at least steps S100-S400: S100: Acquire slope surface deformation and displacement time series data, trigger factor time series data and sampling timestamps, align them according to the sampling timestamps, identify outliers and interpolate missing values, and generate equal interval monitoring time series and data quality status; decompose the surface deformation and displacement time series data in the equal interval monitoring time series to generate trend components, periodic components and high-frequency residual components. S200: Input the trend component into a one-dimensional convolutional channel to extract the trend feature vector; input the periodic component and the triggering factor time-series data associated with the same time step into the long short-term memory network channel encoding to generate a periodic time-series feature vector; concatenate and map the two feature vectors to generate the uncorrected deformation prediction value of the prediction step size; S300: Obtain a prediction error window formed by the uncorrected deformation prediction value and the measured value at the prediction target time, corresponding to the prediction step size; when the model parameters remain unchanged, correct the uncorrected deformation prediction value according to the mean of the prediction error to generate a corrected deformation prediction value; generate the upper and lower bounds of the prediction based on its standard deviation, and adjust the upper and lower bounds of the prediction based on the data quality status and high-frequency residual components. S400: Match the corrected deformation prediction value and the upper and lower bounds of the prediction with the graded early warning threshold, and output an early warning command; update the prediction error window after the measured value is reached.

[0026] S100: Acquire time series data of slope surface deformation and displacement, trigger factor time series data and sampling timestamps, align them according to the sampling timestamps, identify outliers and impute missing values, and generate equal-interval monitoring time series and data quality status; decompose the surface deformation and displacement time series data in the equal-interval monitoring time series to generate trend components, periodic components and high-frequency residual components.

[0027] This step is performed by the data acquisition module, data preprocessing unit, and time series decomposition unit of the slope monitoring system. The data acquisition module is connected to the displacement monitoring device and the environmental sensing device, and receives the time series data of surface deformation displacement, trigger factor time series data, and sampling timestamps of each slope monitoring point on the target slope. The data preprocessing unit organizes the monitoring records according to the slope monitoring point identifier, sensor identifier, and sampling timestamp. The time series decomposition unit receives the equally spaced monitoring time series formed by the data preprocessing unit and performs time series decomposition on the surface deformation displacement time series data therein.

[0028] The displacement monitoring device includes one of a Global Navigation Satellite System (GNSS) receiver, a total station, a surface displacement sensor, or an inclinometer; the environmental sensing device includes one of a rain gauge or a pore water pressure gauge. In one embodiment, the trigger factor time-series data is rainfall time-series data; in another embodiment, the trigger factor time-series data is pore water pressure time-series data; in yet another embodiment, rainfall time-series data and pore water pressure time-series data are associated at the same slope monitoring point, and the rainfall time-series data and the pore water pressure time-series data respectively retain data type identifiers.

[0029] A single monitoring record in the surface deformation displacement time series data includes the slope monitoring point identifier, displacement sensor identifier, sampling timestamp, displacement value, and sensor status; a single monitoring record in the trigger factor time series data includes the environmental sensing device identifier, data type, sampling timestamp, trigger factor value, and sensor status. The displacement value represents the cumulative displacement or displacement increment of the slope monitoring point relative to the monitoring reference position; the same displacement expression method is used within the same prediction task, and the cumulative displacement and displacement increment are not mixed. The sampling timestamp adopts the unified time reference of the slope monitoring system; the original timestamp of the device and the sampling timestamp after unified time stamp calibration are retained when the data is accessed.

[0030] This step is initiated when new monitoring records reach the preset acquisition cycle, or when the current input sliding window advances by one sliding step. The data acquisition module first groups the monitoring records according to the slope monitoring point identifier and data type, and then sorts the monitoring records in each group according to the sampling timestamp; for monitoring records with disordered sampling timestamps, they are rearranged according to time sequence; for cases where there are multiple monitoring records for the same slope monitoring point, the same data type, and the same sampling timestamp, the data preprocessing unit reads the sensor status and retains the monitoring records with normal sensor status; when there are multiple monitoring records with the same sensor status, the monitoring record with the later reception time is used, and the remaining monitoring records are written into the duplicate data record.

[0031] The data preprocessing unit performs time resampling on the surface deformation displacement time series data and the trigger factor time series data according to a unified sampling interval. The unified sampling interval is configured by the monitoring task and is consistent with the input time interval of the deformation prediction model. Specifically, taking the start time of the input sliding window as the first time position, multiple time positions are generated sequentially according to the unified sampling interval. Each monitoring record is mapped to the corresponding time position according to its sampling timestamp. When there are multiple valid monitoring values ​​within a time position, the median of the multiple valid monitoring values ​​is calculated, and the resulting median is written to that time position. When there are no valid monitoring values ​​within a time position, that time position is marked as a missing position.

[0032] In one field embodiment, the displacement monitoring device generates a displacement monitoring record every 30 minutes, the rain gauge generates a rainfall monitoring record every 10 minutes, and the uniform sampling interval of the deformation prediction model is set to 1 hour. The data preprocessing unit takes the median of multiple displacement values ​​within the same hour and accumulates multiple rainfall values ​​within the same hour to generate displacement and rainfall sequences with the same number of values ​​at the same time and location. For pore water pressure time series data, the median is taken when multiple pore water pressure values ​​exist within the same hour. The resulting displacement, rainfall, and pore water pressure sequences are correlated item by item according to their time and location.

[0033] When identifying outliers in time-series surface deformation and displacement data that has undergone time resampling, the data preprocessing unit establishes a sliding window centered on the current time position, calculates the sliding median of the effective displacement values ​​within the sliding window, and calculates the absolute difference between the current displacement value and the sliding median. The absolute difference is compared with an outlier threshold; if the absolute difference is greater than the outlier threshold, the current displacement value is marked as an isolated outlier. The outlier threshold is configured according to the historical displacement fluctuation range of the same slope monitoring point, or it can be given as a multiple of the absolute deviation within the sliding window. For time positions marked as isolated outliers, the original displacement value, the outlier marker, and the values ​​before and after replacement are retained. When effective displacement values ​​exist at adjacent time positions, the isolated outlier is replaced using a linear interpolation formed by adjacent effective displacement values. When effective displacement values ​​are missing at adjacent time positions, the location of the isolated outlier is merged into the missing segment.

[0034] A missing segment consists of multiple consecutive missing positions in time sequence. The data preprocessing unit counts the number of time positions contained in each missing segment and uses this number as the missing length. When the missing length is not greater than the maximum allowable interpolation length, the missing segment is recorded as a short-term missing segment, and missing values ​​are interpolated based on the valid monitoring values ​​before and after the missing segment. Specifically, let the displacement value of the previous valid time position of the missing segment be x_a, and the displacement value of the next valid time position be x_b. If the missing segment contains m time positions, then linear interpolation is performed between x_a and x_b in time position sequence, and the interpolated value is written to the corresponding time position. Each interpolated value retains the interpolation mark, interpolation method, and interpolation time.

[0035] When the missing length exceeds the maximum allowable interpolation length, the missing segment is recorded as a continuous missing segment. The time position within the continuous missing segment retains the missing marker, but this missing segment is not included in the calculation of valid data quantity as a normal measured segment. In one embodiment, when the continuous missing segment is located in the middle of the input sliding window and the continuous missing length does not exceed half the length of the input sliding window, a placeholder value is generated using the local trend before and after the missing segment, while retaining the continuous missing segment state. The placeholder value only maintains the time-series input length and is not included in the proportion of valid data. In another embodiment, when the continuous missing segment is located at the end of the input sliding window and there is no subsequent valid monitoring value, the data preprocessing unit retains the missing marker at that time position and records the data insufficiency state. The deformation prediction model does not read the input sliding window with the data insufficiency state until subsequent monitoring records fill in the corresponding time position, or the slope monitoring system calls a backup model version that matches the missing input.

[0036] The data quality status is determined by the interpolation location, effective data ratio, consecutive missing length, number of outliers, and sensor status. Specifically, the effective data ratio is the ratio of the number of displacement values ​​within the input sliding window that are not interpolated and whose sensor status is normal to the total number of time locations; the consecutive missing length is the number of time locations contained in the longest consecutive missing segment within the input sliding window; and the number of outliers is the number of displacement values ​​within the input sliding window that are marked as isolated outliers. The data preprocessing unit associates the slope monitoring point identifier, the start and end times of the input sliding window, the interpolation location, the effective data ratio, the consecutive missing length, the number of outliers, and the sensor status to generate the data quality status; this data quality status is transmitted to subsequent steps along with the equally spaced monitoring sequence.

[0037] After outlier identification and missing value imputation are completed, the data preprocessing unit encapsulates the monitoring time series at equal intervals according to time location. The equal interval monitoring time series includes slope monitoring point identifiers, unified sampling intervals, time location sequences, surface deformation displacement value sequences, trigger factor value sequences, measured markers, imputation markers, and data quality status. When short-term missing segments exist in the trigger factor time series data, imputation is performed according to the same missing length rules as for the surface deformation displacement time series data. When continuous missing segments exist in the trigger factor time series data, the trigger factor missing status is recorded, and the placeholder values ​​at the missing positions are not used as normal trigger factor values ​​in subsequent standardization.

[0038] The time series decomposition unit receives the equally spaced monitoring time series, reads the surface deformation displacement value sequence from it, and performs Empirical Mode Decomposition (EMD) on the surface deformation displacement value sequence. EMD constructs upper and lower envelopes based on local extrema, and filters the means of the upper and lower envelopes, sequentially generating multiple intrinsic mode components and a residue term. Each filtering only uses data from the start to the end of the current input sliding window, and does not read measurements after the prediction target time.

[0039] To address decomposition fluctuations at the end of the input sliding window, in one implementation, mirror extension is performed on the beginning and end of the surface deformation displacement value sequence, followed by empirical mode decomposition on the extended sequence. After decomposition, candidate components within the same time range as the original input sliding window are extracted. In another implementation, ensemble empirical mode decomposition is used. In yet another implementation, variational mode decomposition or multi-order moving average decomposition is used, and trend components, periodic components, and high-frequency residual components are formed according to the same scaling rules.

[0040] The candidate components include multiple intrinsic mode components generated by empirical mode decomposition; the temporal decomposition unit calculates the average period and energy proportion of each candidate component. The average period is calculated by the average time interval between adjacent local extrema of the candidate component; the energy proportion is calculated by the ratio of the sum of squares of the values ​​at each time position of the candidate component to the sum of squares of the values ​​at each time position of the original surface deformation displacement value sequence. The period range includes a lower limit and an upper limit, which are written into the model version record based on the sampling interval of the target slope, the change period of historical triggering factors, and the deformation prediction model training configuration.

[0041] The time-series decomposition unit adds candidate components with average periods greater than the upper limit of the period range to the residual terms of the empirical mode decomposition at the same time position to generate trend components; it adds candidate components with average periods between the lower and upper limits of the period range to generate periodic components; and it adds candidate components with average periods less than the lower limit of the period range to generate high-frequency residual components. Each candidate component is assigned to only one component category. Trend components, periodic components, and high-frequency residual components have the same number of time positions and correspond one-to-one with the sampling timestamps in the equally spaced monitoring time series.

[0042] When the average period of a candidate component is located at the classification boundary, it is classified according to the boundary assignment rules in the model version record. In one implementation, candidate components with an average period equal to the lower limit of the period range are classified as periodic components, and candidate components with an average period equal to the upper limit of the period range are classified as trend components. When the empirical mode decomposition does not generate candidate components with an average period within the period range, the candidate component with the smallest difference from the center value of the period range is classified as a periodic component, and a periodic component substitution marker is written into the decomposition status. When the empirical mode decomposition screening fails to converge, multi-order moving average decomposition is invoked, and a backup decomposition method identifier is written into the decomposition status.

[0043] In one field embodiment, the uniform sampling interval is 1 hour, and the input sliding window contains 336 time locations; the lower limit of the period range is configured as 24 hours, and the upper limit of the period range is configured as 168 hours. Candidate components with an average period greater than 168 hours and the remaining components are combined to form the trend component; candidate components with an average period between 24 hours and 168 hours are combined to form the periodic component; and candidate components with an average period less than 24 hours are combined to form the high-frequency residual component. The time-series decomposition unit associates the trend component, the periodic component, the high-frequency residual component, the average period of each candidate component, the energy percentage, and the decomposition state with the slope monitoring point identifier and the start and end times of the input sliding window.

[0044] This step ultimately generates equal-interval monitoring time series, data quality status, trend components, periodic components, and high-frequency residual components; the trend components and the periodic components serve as feature extraction inputs for S200; the trigger factor time series data in the equal-interval monitoring time series are associated with the periodic components at the same time position and then enter S200; the data quality status and the high-frequency residual components are transmitted to S300 for adjustment of upper and lower bounds; the decomposed status, outlier handling records, and missing value imputation records are written into the status recording unit and associated with and saved with the current input sliding window.

[0045] S200: Input the trend component into a one-dimensional convolutional channel to extract the trend feature vector; input the periodic component and the triggering factor time-series data associated with the same time step into the long short-term memory network channel encoding to generate a periodic time-series feature vector; concatenate and map the two feature vectors to generate the uncorrected deformation prediction value of the prediction step size.

[0046] This step is performed by the trend feature extraction unit, the periodic time series encoding unit, and the feature fusion module in the deformation prediction module; the trend feature extraction unit receives the trend component formed by S100; the periodic time series encoding unit receives the periodic component formed by S100 and the trigger factor time series data in the equally spaced monitoring time series; the feature fusion module receives the trend feature vector and the periodic time series feature vector, and calls the mapping parameters associated with the model version identifier to generate the uncorrected deformation prediction value.

[0047] This step begins when the temporal decomposition of the current input sliding window is completed in S100, and the decomposition status does not contain a stop prediction marker. The trend feature extraction unit first verifies whether the number of time positions of the trend components is consistent with the model input length; the periodic temporal encoding unit verifies whether the time positions of the periodic components and the trigger factor time series data correspond one-to-one; after verification, it reads the one-dimensional convolutional channel parameters, long short-term memory network channel parameters, normalization parameters, and mapping parameters corresponding to the current model version. When the model versions are inconsistent, the parameters of different model versions are not combined, and a model version conflict status is written in the prediction result record.

[0048] The one-dimensional convolutional channel is constructed using a one-dimensional convolutional neural network (1D-CNN). The one-dimensional convolutional channel takes the values ​​of the trend component at multiple consecutive time positions as input and extracts local morphological features through multiple convolutional kernel scales. The convolutional kernel scale represents the number of consecutive time positions covered by a single convolutional kernel in the temporal direction; different convolutional kernel scales correspond to local time windows of different lengths. Each convolutional kernel moves sequentially along the trend component, calculates the sum of the product of the convolutional kernel weight and the trend component value within the corresponding local time window, and generates a local convolution result through an activation function.

[0049] In one implementation, the one-dimensional convolutional channel is configured with three convolutional kernel scales, covering 3, 5, and 9 consecutive time positions, respectively. The convolutional kernel covering 3 time positions extracts local slope changes within a short time range; the convolutional kernel covering 5 time positions extracts the transformation of trend change direction; and the convolutional kernel covering 9 time positions extracts the morphology of continuous acceleration segments. The specific weights of the convolutional kernels are generated during the deformation prediction model training process and are associated with and stored with the model version identifier; the convolutional kernel weights are not changed during the online prediction phase.

[0050] The slope change refers to the change in the magnitude of the trend component's change within adjacent local time windows; the inflection point refers to the time position where the trend component's direction or rate of change changes within a local time window; the acceleration segment refers to the section where the displacement increment of the trend component gradually increases within multiple consecutive time positions. The one-dimensional convolutional channel does not output slope values, inflection point times, or acceleration values ​​separately, but rather characterizes the aforementioned local morphology through convolution results at different kernel scales; after pooling or feature aggregation, the convolution results are concatenated along the feature dimension to generate a trend feature vector.

[0051] When the input length of the trend component is insufficient for the model input length, the trend feature extraction unit reads the data quality status from S100. If the insufficient input length is caused by consecutive missing segments, the insufficient trend feature input status is recorded, and the formal output of the current prediction record is stopped. If the insufficient input length is caused by a short start time of the current monitoring task, in one implementation, existing trend components are copied and filled from the beginning of the input window, and a start-up period marker is written into the prediction result record. When the trend component contains non-numerical or infinite values, one-dimensional convolution calculation is not performed, and the abnormal location and decomposition status are written to the processing log.

[0052] The periodic temporal coding unit first standardizes the trigger factor temporal data. The standardization parameters are generated during the deformation prediction model training phase, including the mean and standard deviation of the corresponding data type, and are associated with the model version identifier and stored. For the trigger factor value z_t at time position t, the standardized trigger factor value is the difference between z_t and the training mean divided by the training standard deviation. When the training standard deviation is 0, the trigger factor type is recorded as a constant input, and the standardized value is written as 0. The interpolation position in the trigger factor temporal data retains the interpolation marker, which does not replace the standardized trigger factor value.

[0053] The periodic component and the trigger factor time series data at the same time step are merged into a data matrix in chronological order. In the embodiment where the trigger factor time series data is rainfall time series data, each row of the data matrix includes the periodic component value and the standardized rainfall value at the current time position. In the embodiment where the trigger factor time series data is pore water pressure time series data, each row of the data matrix includes the periodic component value and the standardized pore water pressure value at the current time position. In the embodiment where both rainfall time series data and pore water pressure time series data are simultaneously accessed, each row of the data matrix is ​​written with the periodic component value, the standardized rainfall value, and the standardized pore water pressure value in sequence.

[0054] In one implementation, the periodic time-series encoding unit further generates the previous cumulative rainfall based on the rainfall time-series data, and uses the previous cumulative rainfall as a data column in the data matrix; the previous cumulative rainfall is obtained by summing the rainfall within a preset time range before the current time position. In another implementation, the pore water pressure change is generated based on the difference between adjacent time positions in the pore water pressure time-series data, and the pore water pressure change is written into the data matrix. Both the previous cumulative rainfall and the pore water pressure change are standardized according to the standardized parameters recorded in the model version.

[0055] The data matrix is ​​input into a Long Short-Term Memory (LSTM) network channel for temporal encoding. The LSTM network channel reads each row of the data matrix sequentially according to its time position, updating the current hidden state through input gates, forget gates, output gates, and cell states. The input gate controls the proportion of data from the current time position written into the cell state; the forget gate controls the proportion of cell states from the previous time position retained; and the output gate controls the proportion of the current cell state written into the hidden state. After encoding all time positions within the input sliding window, the hidden state of the last time position is read, or features from the hidden states of multiple time positions are aggregated to generate a periodic temporal feature vector.

[0056] When there are consecutive missing segments in the trigger factor time series data, the periodic time series encoding unit reads the time position corresponding to the missing segment. In one implementation, the normalized value of the trigger factor at the time position is written as 0, and the interpolation marker is used as an independent data column in the data matrix. In another implementation, when the length of the consecutive missing segments is greater than the trigger factor missing threshold, no periodic time series feature vector is generated, and the trigger factor input insufficiency status is written into the prediction result record. When the number of time positions in the periodic components and the trigger factor time series data is inconsistent, they are re-associated according to the sampling timestamp; if inconsistencies still exist after re-association, the extra time positions are not written into the data matrix.

[0057] The feature fusion module concatenates the trend feature vector and the periodic time-series feature vector along the feature dimension to generate a fused feature vector. This feature dimension concatenation means preserving the element order of the two feature vectors, appending the periodic time-series feature vector to the trend feature vector, and not summing the feature elements at the same position. If the dimension of the trend feature vector is d_1 and the dimension of the periodic time-series feature vector is d_2, then the dimension of the fused feature vector is d_1 + d_2.

[0058] The feature fusion module inputs the fused feature vector into the mapping layer; the mapping layer consists of one or more fully connected layers, and its parameters are generated during the training process and associated with the model version identifier. The mapping layer outputs one or more uncorrected deformation prediction values; each uncorrected deformation prediction value corresponds to a prediction step size. The prediction step size represents the number of uniform sampling intervals between the prediction generation time and the prediction target time; when the prediction step size is 1, the prediction target time is the first uniform sampling time position after the prediction generation time; when the prediction step size is h, the prediction target time is the h-th uniform sampling time position after the prediction generation time.

[0059] In one field implementation, the uniform sampling interval is 1 hour, the input sliding window contains 336 time locations, and the deformation prediction model outputs three uncorrected deformation prediction values ​​with prediction step sizes of 1, 6, and 24 hours; the three outputs correspond to the surface deformation displacement values ​​after 1 hour, 6 hours, and 24 hours, respectively. The trend feature vector is formed by concatenating the convolution results at three convolutional kernel scales; the periodic temporal feature vector is formed by encoding a data matrix composed of periodic components, rainfall, and pore water pressure through a long short-term memory network channel; after the two vectors are concatenated, they are input into the mapping layer, which generates three uncorrected deformation prediction values ​​according to fixed output positions.

[0060] The deformation prediction module associates each uncorrected deformation prediction value with the slope monitoring point identifier, prediction generation time, prediction target time, prediction step size, and model version identifier to generate a prediction result record. The prediction result record also includes the start and end times of the input sliding window, data quality status identifier, decomposition status identifier, and measured matching status. The measured matching status is initially set to pending matching, indicating that the measurement value at the prediction target time has not yet been written into this prediction result record.

[0061] When the mapping layer outputs a non-numerical, infinite, or value exceeding the model's output range, the deformation prediction module does not treat this output as a normal, uncorrected deformation prediction value; non-numerical or infinite values ​​are written to the model output anomaly record; values ​​exceeding the model's output range retain the original output and are written to the state of input data exceeding the training range. In one implementation, when the current model version experiences a model output anomaly, the previous model version is read to complete the prediction of the current input, and a backup model version identifier is written to the prediction result record; prediction results generated by different model versions are respectively entered into the prediction error window corresponding to their respective model version identifiers.

[0062] This step generates a trend feature vector, a periodic time series feature vector, a fusion feature vector, uncorrected deformation prediction values ​​corresponding to multiple prediction step sizes, and prediction result records. The uncorrected deformation prediction values ​​and the prediction step sizes serve as the current prediction input for S300. The slope monitoring point identifier, prediction target time, prediction step size, and model version identifier in the prediction result records are used by S300 to read the corresponding prediction error window and by S400 to perform prediction error matching after the measured value arrives.

[0063] In one specific embodiment: In S200, the trend component extracts multi-scale local morphological features through a one-dimensional convolutional channel, and the periodic component and the trigger factor time series data are jointly encoded at the same time step and generated into a periodic time series feature vector through a long short-term memory network channel. After concatenating the two feature vectors, an uncorrected deformation prediction value bound to the prediction step size is generated through a mapping layer.

[0064] In the equally spaced monitoring time series generated by S100, the trend component, periodic component, and high-frequency residual component have the same number of time positions and correspond one-to-one with the sampling timestamps. The trend component represents the long-term trend of monotonic accumulation in slope surface deformation and displacement, while the periodic component represents the fluctuation component driven by environmental triggering factors. The triggering factor time series data are either rainfall time series data or pore water pressure time series data, or both. Each triggering factor time series data is associated with the periodic component item by item according to its time position. This step is initiated when S100 completes the time series decomposition of the current input sliding window and the decomposition status does not contain a stop prediction marker. The trend feature extraction unit in the deformation prediction module first verifies whether the number of time positions of the trend components is consistent with the model input length, and the periodic temporal encoding unit verifies whether the time positions of the periodic components and the trigger factor time series data correspond one-to-one. After the verification is passed, the trend feature extraction unit and the periodic temporal encoding unit read the one-dimensional convolutional channel parameters, long short-term memory network channel parameters, normalization parameters and mapping parameters corresponding to the current model version, respectively. If the model version identifiers are inconsistent, the parameters of different model versions are not combined and the model version conflict status is written in the prediction result record.

[0065] The trend feature extraction unit inputs the trend component into a one-dimensional convolutional channel. This one-dimensional convolutional channel is composed of a one-dimensional convolutional neural network, taking the values ​​of the trend component at multiple consecutive time positions as input, and extracting local morphological features through multiple convolutional kernel scales. The convolutional kernel scale represents the number of consecutive time positions covered by a convolutional kernel in the time direction, and different convolutional kernel scales correspond to local time windows of different lengths. Let the input sliding window contain... At each time point, the value of the trend component at each time point is denoted as... ,in This sequence originates from the trend component field of the equally spaced monitoring time series output by S100. For the first... Each convolutional kernel has a scale, and its kernel weight sequence is denoted as... ,in Indicates the first The number of temporal locations covered by each convolutional kernel scale, denoted by the bias term. In time location At that point, the convolution kernel interacts with the trend component from position... The initial local window is convolved to obtain the local convolution result. Formula ① gives the expression for the convolution operation: in, : No. The kernel scale at the start time position The local convolution result output represents the product of the convolution kernel weights and the local window values ​​of the trend component, plus the response value after activation. : The index number of the convolution kernel scale in a one-dimensional convolution channel, with a value of In this step These correspond to convolutional kernels covering 3, 5, and 9 consecutive time positions; : The starting time position index of the current convolution operation, with a value range of 1. , indicating the starting position of the convolution kernel sliding sequentially along the trend component; where Enter the total number of time positions within the sliding window; Activation function: This step uses a linear rectified unit (ReLU), defined as... , used to introduce nonlinear transformations; : No. The bias term at each convolutional kernel scale is generated during the deformation prediction model training phase and saved in association with the model version identifier; : The index of the relative time position within the convolution kernel, with a value range of , which represents the relative offset of the current multiplication operation within the convolution kernel coverage window; : No. The number of consecutive temporal locations covered by each convolutional kernel scale in the temporal direction, in this step , , Configured by the model version; : No. The convolutional kernel scale at the th convolutional kernel scale is... The weight coefficients at each relative time position are generated during the training phase of the deformation prediction model and saved in association with the model version identifier; Trend components at time positions The values ​​at that location are derived from the trend component field in the equally spaced monitoring time series formed by S100.

[0066] The convolutional result sequences generated at each kernel scale are pooled or feature-converged, and then concatenated along the feature dimension to generate a trend feature vector. Specifically, for the Each convolutional kernel scale, at all valid starting positions The convolution result generated above The significant local morphological response values ​​at this scale are extracted using max pooling, and the pooling result is denoted as... The pooling results of the three convolutional kernel scales are then processed according to... The order of these elements forms the trend feature vector. Formula ② gives the expression for the trend feature vector: in, Trend feature vector, with a dimension of 3, whose three components correspond to short-term local slope change response, inflection point turning response, and continuous acceleration segment response, respectively; : No. The significant response values ​​obtained after max pooling at each convolutional kernel scale are generated by formula ①. All valid start time positions We obtain the maximum value from the top, that is ; : Same as the definition in formula ①, with values ​​of 1, 2, and 3; superscript : The transpose symbol for a matrix or vector, indicating that the vector is in column vector form.

[0067] The periodic timing encoding unit standardizes the timing data of the triggering factors. Let the periodic component be located at the time position... The value at is The original value of the trigger factor time series data at the same time position is The standardized parameters are generated during the deformation prediction model training phase, including the mean values ​​for the corresponding data types. and standard deviation It is saved in association with the model version identifier. For time location... Trigger factor value at Standardized trigger factor value Calculated using formula ③: in, Time and location The standardized values ​​of the time series data of the triggering factor; : Time location index, value range is ; : Trigger factor time series data at time position The original value at the location comes from the trigger factor value sequence in the equally spaced monitoring time series formed by S100; if the trigger factor is rainfall; if the trigger factor is pore water pressure; The mean value of this trigger factor type, statistically analyzed during the training phase of the deformation prediction model, is saved in association with the model version identifier; The standard deviation of this trigger factor type, statistically analyzed during the deformation prediction model training phase, is saved in association with the model version identifier; when When this happens, the trigger factor is recorded as a constant input, and its normalized value is written as 0.

[0068] The periodic components and the trigger factor time series data at the same time step are merged into a data matrix in chronological order. Each row of this data matrix corresponds to a time position. The periodic component values ​​are written sequentially in the line. and the trigger factor value after standardization by formula ③ If both rainfall time-series data and pore water pressure time-series data are simultaneously input, then the periodic component values ​​are written sequentially for each row. Standardized rainfall values and standardized pore water pressure value .

[0069] In one implementation, the periodic time-series encoding unit further generates the previous cumulative rainfall based on the rainfall time-series data. This is obtained by summing the rainfall within a preset time range prior to the current time position and using it as a data column in the data matrix. In another implementation, the pore water pressure change is generated based on the difference between adjacent time positions in the pore water pressure time-series data and written into the data matrix. Both the previous cumulative rainfall and the pore water pressure change are standardized according to the standardized parameters recorded in the model version. When the number of time positions in the periodic component and the trigger factor time-series data is inconsistent, they are re-associated according to the sampling timestamp; if inconsistencies still exist after re-association, the extra time positions are not written into the data matrix.

[0070] The above data matrix is ​​input into a Long Short-Term Memory (LSTM) network channel for temporal encoding. The LSM network channel reads each row of the data matrix sequentially according to its time position, updating the current hidden state through input gates, forget gates, output gates, and cell states. Let the data matrix be at time position... The input row vector at position is This vector contains periodic component values ​​and standardized trigger factor values. The cell state update process of the Long Short-Term Memory (LSTM) network channel is described by equation ④: in, Time and location The input gate vector at the location has the same dimension as the hidden state dimension of the channel in the Long Short-Term Memory network, and is used to control the proportion of the current input written to the unit state. Time and location The output gate vector at that point, with the same dimension as above, is used to control the proportion of the current cell state written into the hidden state. Time and location The candidate unit state vector at the given location has the same dimensions as above, and is generated from the current input and the previous hidden state. Transformation generation; Time and location The updated unit state vector, with the same dimensions as above, is used to store long-range temporal memory information; Time and location The updated hidden state vector, with the same dimensions as above, serves as the encoded output for the current time position; Time and location The forget gate vector at that location, with the same dimensions as above, is used to control the state of the unit at the previous time step. The retention ratio; : S-type activation function The output range is (0,1), used for gating mechanisms; Hyperbolic tangent activation function; Time and location The input row vector at that position is composed of periodic component values. and the trigger factor value after standardization by formula ③ Composition; if both rainfall and pore water pressure data are simultaneously input, it also includes standardized previous cumulative rainfall and pore water pressure changes, with dimensions configured by the model version; The input weight matrix corresponding to the input gate has dimensions of (hidden state dimension × input vector dimension), generated during training, and its units are such that the product result is equal to the input weight matrix. match; The input weight matrix corresponding to the output gate has the same dimensions as above and is generated during training. The input weight matrix corresponding to the forget gate has the same dimensions as above and is generated during training. The input weight matrix corresponding to the candidate unit state has the same dimensions as above and is generated during training. The cyclic weight matrix corresponding to the input gate has dimensions of (hidden state dimension × hidden state dimension) and is generated during training. The output gate's corresponding cyclic weight matrix, with the same dimensions as above, is generated during training. The circular weight matrix corresponding to the forget gate has the same dimensions as above and is generated during training. The cyclic weight matrix corresponding to the candidate unit state has the same dimensions as above and is generated during training. The input gate bias vector has the same dimension as the hidden state, and its scale ensures that the gate value is within the effective range of the activation function. : The bias vector of the output gate, with the same dimensions as above; The bias vector of the forget gate, with the same dimensions as above; : The bias vector of the candidate unit state, with the same dimensions as above; Time and location The hidden state vector at position , dimension and same; Time and location The unit state vector at that location, with dimension and same; The Hadamard product symbol represents the element-wise multiplication of two vectors of the same dimension. Same as the definition in formulas ① and ③, time location index; Complete all input within the sliding window After encoding each time position, retrieve the hidden state of the last time position. As a periodic time series feature vector .

[0071] Simple numerical example: Let the input row vector be... The periodic component value mm, standardized rainfall value The weight matrix and bias terms generated during the training phase are fixed values. After calculation using the input gate, forget gate, output gate, and candidate unit states in formula ④, the unit state at the current time position is obtained. and hidden state Assume the current sliding window contains a total of After encoding all time positions, take... As a periodic time series feature vector .

[0072] The feature fusion module concatenates the trend feature vector obtained by formula ② along the feature dimension. and the periodic temporal feature vector formed by the hidden state at the last time position in formula ④ Generate fused feature vectors Feature concatenation refers to preserving the element order of two feature vectors, appending the periodic time series feature vector to the trend feature vector, and not summing the feature elements at the same position. If the dimension of the trend feature vector is... The dimension of the periodic time series feature vector is Then the dimension of the fused feature vector is In this step, , The dimension of the hidden state vector for the Long Short-Term Memory (LSTM) network channels is configured by the model version. Equation ⑤ gives the expression for the fused feature vectors: in, : Fuse feature vectors, with dimension It is generated by directly concatenating the trend feature vector and the periodic time series feature vector in the feature dimension; Trend feature vector, defined by formula ②, has a dimension of 3; : Periodic time series eigenvectors, obtained from formula ④ Formation, dimension (Hidden state dimension of long short-term memory network channels); Vector concatenation operators (square brackets and newline) indicate that the components are arranged sequentially along the feature dimension, without summing elements at the same position.

[0073] The feature fusion module will fuse feature vectors. The input mapping layer consists of one or more fully connected layers. Its parameters are generated during training and associated with the model version identifier. The mapping layer outputs one or more uncorrected deformation predictions, each corresponding to a prediction step size. The prediction step size represents the number of uniform sampling intervals between the prediction generation time and the prediction target time. Let the index of the prediction step size output by the mapping layer be... , , To predict the total step size, the mapping layer performs a linear transformation on the fused feature vectors. Equation ⑥ gives the output expression of the mapping layer: in, : No. The uncorrected deformation prediction value corresponding to the prediction step size represents the value of the nth prediction step since the current prediction generation time. The displacement prediction result at the uniform sampling interval number corresponding to each prediction step size; : Predicted step size index, with a value range of , The total number of prediction steps output by the mapping layer in this step These correspond to prediction step sizes of 1, 6, and 24 uniform sampling intervals, respectively. : No. The mapping layer weight vector corresponding to each prediction step size has dimensions of... Same (i.e.) This is generated during the training phase of the deformation prediction model and saved in association with the model version identifier; superscript : Transpose symbol, transposes a column vector Transpose to row vector so as to be compatible with Perform dot product operation; The fused feature vector, defined by formula ⑤, has a dimension of [missing value]. ; : No. The mapping layer bias scalar corresponding to each prediction step size is generated during the training phase and saved in association with the model version identifier. Output Related to slope monitoring point identification, prediction generation time, prediction target time, and prediction step size It is associated with the model version identifier to form a record of prediction results.

[0074] In one field implementation, the uniform sampling interval is 1 hour, the input sliding window contains 336 time locations, and the deformation prediction model outputs three uncorrected deformation prediction values ​​with prediction step sizes of 1, 6, and 24 hours, corresponding to the surface deformation displacement values ​​after 1 hour, 6 hours, and 24 hours, respectively. The current input sliding window for slope monitoring point P01 starts at 00:00 on June 1, 2026, ends at 23:00 on June 14, 2026, and the prediction generation time is 23:00 on June 14, 2026. The mapping layer generates three uncorrected deformation prediction values ​​at fixed output positions: mm corresponds to the predicted target time 2026-06-15 00:00. mm corresponds to the predicted target time 2026-06-15 05:00. mm corresponds to the predicted target time 2026-06-15 23:00. The deformation prediction module associates the above three uncorrected deformation prediction values ​​with the slope monitoring point identifier P01, the prediction generation time 2026-06-14 23:00, their respective prediction target time, prediction step size (1, 6, 24), and model version identifier V3 to generate three prediction result records. The measured matching status of each record is set to pending matching.

[0075] When the mapping layer outputs a non-numerical, infinite, or value exceeding the model's output range, the deformation prediction module does not treat this output as a normal, uncorrected deformation prediction value; non-numerical or infinite values ​​are written to the model output anomaly record; values ​​exceeding the model's output range retain the original output and are written to the input data exceeding the training range status. If the current model version exhibits an output anomaly, the previous model version is read to complete the prediction of the current input, and a backup model version identifier is written to the prediction result record; prediction results generated by different model versions are respectively entered into the prediction error window corresponding to their respective model version identifiers.

[0076] This step ultimately generates a trend feature vector, a periodic time series feature vector, a fused feature vector, uncorrected deformation prediction values ​​corresponding to multiple prediction step sizes, and prediction result records. The uncorrected deformation prediction values ​​and prediction step sizes serve as the current prediction input for S300; the slope monitoring point identifier, prediction target time, prediction step size, and model version identifier in the prediction result records are used by S300 to read the corresponding prediction error window and by S400 to perform prediction error matching after the measured values ​​arrive.

[0077] The periodic component and the triggering factor time series data enter the long short-term memory network channel at the same time step, realizing the dynamic coupling of environmental driving factors and periodic deformation response on the time axis; the trend component and the periodic component are respectively extracted by heterogeneous network and then spliced ​​in the feature dimension, maintaining the independent source of the two types of features and avoiding feature mixing caused by isomorphic processing or serial cascading.

[0078] Technical Effects of this Section: This step extracts short-term local slope changes, inflection point changes, and continuous acceleration segment morphology at three convolutional kernel scales (3, 5, and 9 time positions), by inputting the trend component into a one-dimensional convolutional channel and extracting these at three kernel scales (3, 5, and 9 time positions). This explicitly binds the physical semantics of the convolutional kernel to the engineering characteristics of the slope deformation trend. The convolutional weights remain unchanged during the online prediction phase, avoiding prediction uncertainties caused by online updates of model parameters. The periodic component and trigger factor time-series data are merged into a data matrix at the same time step and then input into the Long Short-Term Memory (LSTM) network channel. This allows rainfall or pore water pressure values ​​to participate in the updates of the input gate, forget gate, and unit state at each time position, rather than being statically concatenated at the end, thus forming an environmental... The driving factors and the periodic deformation response are coupled and encoded step by step. The trend feature vector and the periodic time series feature vector are concatenated in the feature dimension rather than added element by element, so that the morphological information and time series dependency information carried by the two types of features remain independent before entering the mapping layer. The mapping layer generates multiple uncorrected deformation prediction values ​​with a fixed output position and associates each prediction value with the slope monitoring point identifier, prediction generation time, prediction target time, prediction step size and model version identifier. The measured matching status in the prediction result record is initialized to be matched. This provides a grouped record structure that can be indexed by the slope monitoring point identifier, model version identifier and prediction step size for maintaining prediction error windows by prediction step size in S300.

[0079] S300: Obtain a prediction error window formed by the uncorrected deformation prediction value and the measured value at the prediction target time, corresponding to the prediction step size; when the model parameters remain unchanged, correct the uncorrected deformation prediction value according to the mean of the prediction error to generate a corrected deformation prediction value; generate the upper and lower bounds of the prediction based on its standard deviation, and adjust the upper and lower bounds of the prediction based on the data quality status and high-frequency residual components.

[0080] This step is initiated after S200 generates the current uncorrected deformation prediction value. The prediction error correction unit receives the uncorrected deformation prediction value, prediction step size, model version identifier, and prediction result record generated by S200, and reads the prediction error window associated with the slope monitoring point identifier, model version identifier, and prediction step size from the status record unit. The prediction error window is a sliding window that saves multiple prediction errors in the order of error generation time for the same slope monitoring point, the same model version identifier, and the same prediction step size; the prediction error is formed by the historical uncorrected deformation prediction value and the measured value at the corresponding prediction target time. The measured value comes from the surface deformation displacement value collected by the displacement monitoring device at the prediction target time and after time alignment, outlier identification, and missing value processing by S100; when the measured value has an interpolation mark, whether the measured value is written into the prediction error window is limited by the error sample rules. The prediction error correction unit reads the corresponding prediction error window according to the slope monitoring point identifier, model version identifier, and prediction step size. The prediction errors stored in the prediction error window are all from historical prediction results that have reached the prediction target time; prediction results that have not yet obtained measurement values ​​are not included in the prediction error window; prediction errors of different prediction step sizes are maintained separately, and prediction errors with a prediction step size of 1 are not written into the prediction error window with a prediction step size of 6 or 24.

[0081] Set up slope monitoring points Model version and prediction step size The corresponding prediction error window is The prediction error window contains data arranged in chronological order. Prediction error Each prediction error is calculated by subtracting the corresponding historical uncorrected deformation prediction value from the measured value; that is, the prediction error is positive when the measured value is greater than the historical uncorrected deformation prediction value, and negative when the measured value is less than the historical uncorrected deformation prediction value. The prediction error correction unit calculates the mean of the prediction errors within the prediction error window and generates the offset correction amount. Formula ⑦ gives the expression for calculating the offset correction amount: in, Offset correction amount, i.e., prediction error window Inside The arithmetic mean of the prediction errors represents the deformation prediction model at the slope monitoring points. Model version and prediction step size The systematic deviation of historical forecasts from measured values, including direction and magnitude. Positive values ​​indicate that historical forecasts are systematically underestimated, while negative values ​​indicate that historical forecasts are systematically overestimated. : Slope monitoring point identifier, derived from the slope monitoring point identifier field in the equally spaced monitoring time sequence formed by S100, used to uniquely identify the location of the monitoring point on the target slope, and the data type is string or code; Model version identifier, derived from the version record generated during the training phase of the deformation prediction model, is used to distinguish different sets of model parameters (including one-dimensional convolutional channel parameters, long short-term memory network channel parameters, normalized parameters, and mapping parameters). The data type is string or encoded. Prediction step size, representing the number of uniform sampling intervals between the prediction generation time and the prediction target time, is taken from the prediction step size field in the prediction result record generated by S200; in this step You can choose 1, 6, or 24, which correspond to predictions after 1 hour, 6 hours, and 24 hours, respectively. Slope monitoring point Model version and prediction step size The corresponding prediction error window is stored in the state record unit. It is a sliding window maintained in the order of error generation time. The window stores multiple prediction error samples. Each error sample is formed by the historical uncorrected deformation prediction value and the measurement value at the corresponding prediction target time. Prediction error window The total number of prediction error samples currently contained within the record is maintained by the state recording unit and updated as new errors are written; the value is a positive integer. The index of the sample within the window where the prediction error occurs, with a value range of [value missing]. This is used to iterate through each error sample within the window; The first one in the prediction error window, arranged in chronological order. Each prediction error sample is obtained by subtracting the corresponding historical uncorrected deformation prediction value from the measured value at the prediction target time; the measured value comes from the surface deformation displacement value collected by the displacement monitoring device (including one of the GNSS receiver, total station, surface displacement sensor or inclinometer) at the prediction target time and after time alignment, outlier identification and missing value processing by S100. fraction symbol The division factor in arithmetic mean calculations represents the division factor over time. The sum of the error samples is divided by the total number of samples.

[0082] The prediction error correction unit will adjust the offset correction amount. Add the current uncorrected deformation prediction value to generate the corrected deformation prediction value. Let the current uncorrected deformation prediction value be... The corrected deformation prediction value is Formula ⑧ gives the expression for the corrected deformation prediction value: in, Corrected deformation prediction value, i.e., the current uncorrected deformation prediction value. With offset correction amount The sum is the point prediction result after correction by the historical error mean; The current uncorrected deformation prediction value is output by the mapping layer of S200 (Formula ⑥) and written into the prediction result record. It represents the displacement prediction value directly output by the deformation prediction model without error correction. The offset correction amount is calculated by formula ⑦ and directly substituted into formula ⑧ for correction calculation. , , : As defined in Formula ⑦, these represent the slope monitoring point identifier, model version identifier, and prediction step size, respectively.

[0083] Simple numerical example: Continuing from the example above, the current uncorrected deformation prediction value mm, offset correction amount mm, then the corrected deformation prediction value mm, this value represents the displacement prediction result after historical error offset correction.

[0084] The uncertainty generation unit calculates the standard deviation of the prediction error within the prediction error window, and this standard deviation, together with the offset correction amount, constitutes the basis of the prediction uncertainty.

[0085] Set the prediction error window The mean of the internal prediction error is (Given by formula ⑦), standard deviation According to each prediction error and Calculate the sum of squares of the differences.

[0086] Formula 9 gives the expression for the sample standard deviation: in, Prediction error window The sample standard deviation of the internal prediction error represents the historical prediction error around the offset correction amount. The degree of dispersion is used to generate the basic width of the upper and lower bounds for prediction; As defined in formula ⑦, the total number of prediction error samples currently contained within the prediction error window, with a value that is a positive integer. When When the denominator is 0, the sample standard deviation is not calculated, and the process proceeds to the insufficient sample handling path for prediction error. When there is an abnormal error marker within the prediction error window and that abnormal error has already been excluded from the calculation, Take the number of normal error samples used in the calculation; Same as the definition in formula ⑦; As defined in formula ⑦, the first... One prediction error sample; The offset correction amount is calculated by formula ⑦ and directly substituted into the summation term of formula ⑨ to participate in the standard deviation calculation, so that the standard deviation reflects the fluctuation range of the prediction error around its mean. Square symbol Squaring the difference within the parentheses ensures that both positive and negative errors contribute positive numbers. fraction symbol The unbiased estimator of the sample standard deviation, using the denominator. Instead To correct the degrees of freedom for a finite sample; Square root symbol Take the square root of the value in parentheses and restore the units from mm² to mm to obtain the standard deviation. .

[0087] The uncertainty generation unit uses the corrected deformation prediction value, standard deviation, and interval coefficients to generate upper and lower bounds for prediction. The upper and lower bounds represent the upper and lower limits of the prediction uncertainty interval, respectively. The interval coefficients consist of basic interval coefficients and a mass adjustment factor. The basic interval coefficients are associated with the model version identifier and prediction step size and are configured using validation data from the deformation prediction model. The mass adjustment factor is generated based on the proportion of valid data, the length of consecutive missing data, the energy proportion of high-frequency residual components, and the number of error samples. Let the interval coefficients be... The upper bound of the prediction is The predicted lower bound is Formula 10 gives the expression for the upper and lower bounds of the prediction: in, The upper bound of the prediction indicates the upper limit of the prediction uncertainty interval, which is determined by the corrected deformation prediction value. Plus Double standard deviation get; The lower bound of the prediction indicates the lower limit of the prediction uncertainty interval, which is determined by the corrected deformation prediction value. minus Double standard deviation get; The corrected deformation prediction value is given by formula ⑧ and can be directly substituted into formula ⑩ to participate in the calculation of the center position of the upper and lower bounds. Interval coefficients are obtained by adding the base interval coefficients and the quality adjustment amount (see formula). This determines the expansion factor of the prediction interval in the direction of standard deviation, which directly affects the distance between the upper and lower bounds of the prediction. The sample standard deviation of the prediction error within the prediction error window is given by formula ⑨ and can be directly substituted into formula ⑩ to participate in the calculation of the interval width of the upper and lower bounds of the prediction. Product sign : Scalar multiplication operator, representing interval coefficients with standard deviation Multiply; , , Same as the definition in formula ⑦.

[0088] Interval coefficients The value of is determined by both the base interval coefficient and the quality adjustment amount. The base interval coefficient is associated with the model version identifier and prediction step size, and is configured by the deformation prediction model based on the validation data during the offline validation phase. The quality adjustment amount is dynamically generated based on the proportion of effective data transmitted by S100, the length of consecutive missing data, the energy proportion of high-frequency residual components, and the number of error samples in the prediction error window. The energy proportion of high-frequency residual components is calculated as the ratio of the sum of squares of the values ​​at each time position of the high-frequency residual components to the sum of squares of the values ​​at each time position of the surface deformation displacement time series data. This ratio is formed by S100 after time series decomposition and is transmitted to this step as part of the data quality status. The proportion of effective data is denoted as . The length of consecutive missing characters is denoted as The maximum permissible interpolation length is denoted as The energy percentage of the high-frequency residual component is denoted as... The number of error samples is denoted as The threshold for the number of error samples is denoted as The interval coefficients are calculated as the sum of the basic interval coefficients and the mass term. Formula The adjustment expression for the interval coefficient is given as follows: in, The adjusted interval coefficients are obtained by adding the basic interval coefficients to the four quality terms, and are directly substituted into formula 10 to participate in the generation of the upper and lower bounds of the prediction. Basic interval coefficients, and model version identifier and predicted step size The correlation is configured by the deformation prediction model based on the validation data during the offline validation phase; this value represents the expansion factor of the baseline interval under ideal data quality conditions. The quality item weights corresponding to the proportion of effective data are recorded and saved by the model version. They are used to adjust the degree of influence of insufficient effective data on the interval coefficients and remain unchanged during the online prediction stage. The effective data ratio, provided by the data quality status field of S100, represents the ratio of the number of uninterpolated displacement values ​​with normal sensor status within the input sliding window to the total number of time positions. Its value range is... , dimensionless. The closer to 1, the more abundant the available measured data. The smaller the value, the more data is imputed or missing. The quality term weights corresponding to consecutive missing lengths are recorded and saved by the model version. They are used to adjust the degree of influence of consecutive missing lengths on interval coefficients and remain unchanged during the online prediction phase. : Length of consecutive missing segments, provided by the data quality status field of S100, represents the number of time positions contained in the longest consecutive missing segment within the input sliding window, and the value is a non-negative integer; The maximum allowable interpolation length is determined by the monitoring task configuration of the slope monitoring system. When consecutive missing segments are not imputed, this value is used to normalize the length of consecutive missing segments; The mass term weight corresponding to the proportion of high-frequency residual component energy is recorded and saved by the model version. It is used to adjust the degree of influence of the intensity of high-frequency random disturbance on the interval coefficient and remains unchanged in the online prediction stage. The energy proportion of the high-frequency residual component is calculated by S100 as the ratio of the sum of squares of the high-frequency residual component at each time position to the sum of squares of the surface deformation displacement time series data at each time position. The value range is [value missing]. ; The larger the value, the stronger the unpredictable random disturbance within the current time period; The quality term weights corresponding to the number of error samples are recorded and saved by the model version. They are used to adjust the degree of influence of insufficient error sample quantity on interval coefficients and remain unchanged during the online prediction stage. Prediction error window The number of prediction error samples currently included is defined in formula ⑦ and is maintained by the state recording unit; Error sample number threshold, configured by the model version record. When the number of time error samples is sufficient, Insufficient number of time error samples; : The term indicating insufficient error sample size is a truncated linear function. When the value of this item is 0, This item is at that time , representing the proportion of missing error samples relative to the threshold; The percentage of missing valid data is calculated by subtracting the percentage of valid data from 1. The result represents the percentage of invalid data within the input sliding window; : Normalized value of consecutive missing length, which represents the length of consecutive missing values. Divide by the maximum permissible interpolation length , dimensionless, represents the severity of the current consecutive missing values ​​relative to the upper limit of allowed imputation; when and When all are time position numbers; fraction symbol The percentage of missing error samples indicates the current number of error samples. Relative to threshold The degree of inadequacy, when The value is 1 when... The value should be 0 at that time; In one field embodiment, slope monitoring point P01 uses model version V3. The prediction error window with a prediction step size of 6h stores the most recent 30 valid prediction errors. The information of this window stored in the status recording unit is as follows: window identifier W_P01_V3_6, number of error samples N=30, mean prediction error b=2.4mm, standard deviation of prediction error σ=1.8mm, and window update time 2026-06-14 23:00.

[0089] S200 completed feature extraction and mapping layer output for the current input sliding window at 23:00 on June 14, 2026. The generated prediction result record includes the slope monitoring point identifier P01, model version identifier V3, prediction step size 6, and uncorrected deformation prediction value. The predicted time is 23:00 on June 14, 2026, and the predicted target time is 05:00 on June 15, 2026. The prediction error correction unit reads the prediction error window from the status record unit according to the slope monitoring point identifier P01, model version identifier V3, and prediction step size 6. The readings show N=30, b=2.4mm, and σ=1.8mm.

[0090] Calculate the offset correction amount using formula ⑦ mm, calculated using formula ⑧, yields the corrected predicted deformation value. mm. The uncertainty generation unit reads the data quality status transmitted by S100: effective data ratio r=0.94, consecutive missing length L=2, high-frequency residual component energy ratio q=0.08, number of error samples N=30, error sample number threshold N_min=20; reads the basic interval coefficients. and weight of quality items , , , .

[0091] From the formula Calculate the adjusted interval coefficients The standard deviation is calculated using formula ⑨. mm, the upper bound of the prediction is calculated using formula ⑩. mm and predicted lower bound mm. The prediction error correction unit writes the offset correction amount of 2.4 mm, the corrected predicted deformation value of 50.4 mm, the upper prediction limit of 54.18 mm, the lower prediction limit of 46.62 mm, the number of error samples of 30, the prediction error sample shortage status (written as sufficient), the interval coefficient of 2.1, and the data quality status (effective data ratio of 0.94, consecutive missing length of 2, high-frequency residual energy ratio of 0.08) into the current prediction result record. The uncorrected predicted deformation value of 48.0 mm in the prediction result record remains unchanged.

[0092] The number of error samples N=30 is greater than the error sample number threshold N_min=20, so the state of insufficient prediction error samples is written as sufficient; the proportion of effective data r=0.94 is greater than the data quality threshold (this threshold is configured to 0.85 by the model version record), the consecutive missing length L=2 is less than the maximum allowable interpolation length L_max=6, and the proportion of high-frequency residual component energy q=0.08 is less than the high-frequency residual energy threshold (this threshold is configured to 0.15 by the model version record). Therefore, all quality states in the prediction result record are written as normal.

[0093] When the model version changes, the prediction error correction unit does not directly write the prediction error of the old model version into the prediction error window of the new model version. If the new model version has a corresponding offline validation error record, the prediction error sample shortage state is initialized with that record; if the new model version does not have an offline validation error record, the prediction upper and lower bounds are recorded as pending updates, and the output of warning instructions corresponding to the second and third risk levels is paused until the number of error samples in the prediction error window reaches the error sample number threshold. When the number of error samples is less than the error sample number threshold, the uncertainty generation unit records the prediction error sample shortage state: it reads the validation error standard deviation from the offline validation record associated with the current model version identifier and prediction step size as the current standard deviation, and adjusts the interval coefficient to the value corresponding to the sample shortage state; or it retains the standard deviation calculated by the current prediction error window and increases the interval coefficient. When the prediction error window is empty, no offset correction is generated, the uncorrected deformation prediction value is used as the corrected deformation prediction value, and the prediction upper and lower bounds are generated using the offline validation error standard deviation.

[0094] After this step, the uncorrected deformation prediction value in the prediction result record remains unchanged, and the following are added: offset correction amount, corrected deformation prediction value, prediction upper bound, prediction lower bound, number of error samples, insufficient prediction error samples status, interval coefficient, data quality status, and high-frequency residual energy ratio. The corrected deformation prediction value, prediction upper bound, and prediction lower bound serve as threshold matching inputs for S400; the uncorrected deformation prediction value remains in the prediction result record and is used by S400 to calculate the prediction error after the target time measurement value arrives, and the newly generated prediction error is written back to the corresponding prediction error window according to the prediction step size, for subsequent input sliding window to be read again by this step after S200. Under the condition that the model parameters remain unchanged, the prediction error correction unit and the uncertainty generation unit use the mean of the same prediction error window to generate the offset correction amount and the standard deviation of the same prediction error window to generate the prediction upper and lower bounds. Both outputs come from the same set of historical prediction error samples but perform different functions: the mean correction point prediction value and the standard deviation combined with the dynamic interval coefficient form the prediction uncertainty interval.

[0095] Technical Effects of this Section: This step maintains prediction error windows separately according to slope monitoring point identifiers, model version identifiers, and prediction step sizes. This allows the current uncorrected deformation prediction value to call upon the historical error state corresponding to its prediction span, avoiding the mixing of errors from different prediction spans in the correction process. The prediction error window is formed by the measured values ​​already reached at the prediction target time and the corresponding historical uncorrected deformation prediction values ​​in the order of error generation time. Each prediction error carries strict temporal anchoring information between the prediction generation time and the prediction target time, ensuring the causal order of error samples. The mean of the same prediction error window is used to generate the offset correction amount and correct the current uncorrected deformation prediction value, while the standard deviation of the same prediction error window is used to generate the upper and lower bounds of the prediction. The same set of historical errors simultaneously bears the responsibility... The system includes two functions: point prediction correction and interval prediction generation. The interval coefficient is composed of a base interval coefficient and a quality adjustment factor based on the proportion of effective data, the length of consecutive missing data, the energy ratio of high-frequency residual components, and the number of error samples. This allows the prediction interval width to be dynamically adjusted according to the current monitoring data quality status. The more severe the data missing, the stronger the high-frequency random disturbance, and the fewer the effective error samples, the wider the prediction interval. All quality factors are cumulatively applied rather than being covered by a single state. The offset correction factor only applies to the currently uncorrected deformation prediction value. The one-dimensional convolutional channel, long short-term memory network channel, and mapping layer parameters of the deformation prediction model remain unchanged during the online prediction stage. Only the prediction error window is updated as subsequent measurements arrive, forming a pure data-driven online adaptive correction under the condition of frozen model weights.

[0096] S400: Match the corrected deformation prediction value and the upper and lower bounds of the prediction with the graded early warning threshold, and output an early warning command; update the prediction error window after the measured value is reached.

[0097] This step is executed by the early warning determination unit, the status recording unit, and the error status update unit. The early warning determination unit receives the corrected deformation prediction value, prediction upper limit, prediction lower limit, data quality status, and prediction result record generated by S300, and reads the graded early warning threshold corresponding to the slope monitoring point identifier and prediction step size from the early warning rule configuration. The status recording unit records the risk level, early warning instruction, and changes in early warning status. After the measured value at the prediction target time is reached, the error status update unit updates the corresponding prediction error window according to the prediction result record.

[0098] The tiered early warning thresholds are configured according to the slope monitoring points and prediction step lengths; different prediction step lengths for the same slope monitoring point are recorded separately for each threshold. Each threshold record includes a threshold version, tiered early warning threshold, early warning cancellation threshold, consecutive boundary crossing threshold, and preset duration; the threshold version is associated with the prediction result record. When a threshold changes, the new threshold is used for prediction results generated after the change; already generated prediction results retain their original threshold version, and the historical risk level is not rewritten using the new threshold.

[0099] The early warning determination unit performs threshold matching on the corrected deformation prediction value, the upper prediction limit, and the lower prediction limit according to the same displacement direction. In the implementation where the displacement value increases with the increase of risk status, a value greater than the graded early warning threshold is recorded as exceeding the threshold. In the implementation where the displacement direction has positive and negative signs, the displacement value is first converted into a risk direction displacement value according to the risk direction configured at the monitoring point, and then matched with the graded early warning threshold. The upper prediction limit, the corrected deformation prediction value, and the lower prediction limit use the same displacement reference and unit.

[0100] When the upper bound of the prediction exceeds the corresponding graded early warning threshold, and the corrected deformation prediction value does not exceed the graded early warning threshold, the early warning determination unit assesses the first risk level; the first risk level indicates that a portion of the prediction uncertainty interval is outside the threshold. The early warning instruction corresponding to the first risk level includes the slope monitoring point identifier, prediction target time, prediction step size, prediction upper bound, corrected deformation prediction value, graded early warning threshold, data quality status, and threshold version.

[0101] When the corrected deformation prediction value exceeds the corresponding graded early warning threshold, but the lower bound of the prediction does not exceed the graded early warning threshold, the early warning determination unit assesses the second risk level; the early warning instruction corresponding to the second risk level is written into the corrected deformation prediction value exceeding the limit flag. When the lower bound of the prediction exceeds the corresponding graded early warning threshold, the early warning determination unit assesses the third risk level; the early warning instruction corresponding to the third risk level is written into the prediction uncertainty interval overall exceeding the limit flag.

[0102] The early warning determination unit also maintains the number of consecutive boundary crossings. Whenever the corrected deformation prediction value or prediction lower limit of the same slope monitoring point and the same prediction step length consecutively crosses the corresponding graded early warning threshold, the number of consecutive boundary crossings is incremented by 1. If it does not cross, the number of consecutive boundary crossings is reset, or decremented according to the early warning rules. When the number of consecutive boundary crossings reaches the preset number, the third risk level is assessed. When the third risk level is formed by the number of consecutive boundary crossings, the consecutive boundary crossing mark and the start time of the consecutive boundary crossings are written into the early warning instruction.

[0103] When the data quality status includes consecutive missing data, insufficient prediction error samples, or input data exceeding the training range, the early warning determination unit writes the corresponding status into the early warning instruction. In one implementation, when the data quality status does not meet the data quality threshold, a data anomaly instruction is output, and the first risk level formed based on the upper bound of the prediction is retained; a second or third risk level is not formed based on the corrected deformation prediction value. In another implementation, the data quality status is only used as a status field in the early warning instruction and does not change the risk level matching rules.

[0104] When the upper bound of the prediction falls below the warning cancellation threshold, the status recording unit records the start time of the cancellation condition. After the upper bound of the prediction remains below the warning cancellation threshold for a preset period of time, the current warning instruction is cancelled, and the warning status is updated to normal. If the upper bound of the prediction for any prediction period within the duration crosses the warning cancellation threshold again, the start time of the cancellation condition is cleared, and the current warning instruction is not cancelled. The first risk level, the second risk level, and the third risk level maintain the current status and the status update time, respectively.

[0105] The early warning command is output through the early warning interface of the slope monitoring system. The early warning command includes an early warning command identifier, slope monitoring point identifier, risk level, prediction generation time, prediction target time, prediction step size, corrected deformation prediction value, prediction upper bound, prediction lower bound, graded early warning threshold, threshold version, data quality status, model version identifier, and early warning status update time. The early warning interface writes the early warning command into the early warning record of the slope monitoring system and sends it to the display terminal or alarm device associated with the target slope. The display terminal and alarm device do not participate in the calculation of the deformation prediction value and the prediction upper and lower bounds.

[0106] In one field embodiment, the graded early warning threshold for slope monitoring point P01 is 50 mm; the current predicted lower limit is 48.6 mm, the corrected predicted deformation value is 50.4 mm, and the predicted upper limit is 52.2 mm. If the corrected predicted deformation value exceeds the graded early warning threshold, but the predicted lower limit does not, the early warning determination unit assesses the second risk level and outputs an early warning command containing the aforementioned predicted value and prediction step size. In subsequent prediction cycles, if the predicted lower limit exceeds the graded early warning threshold three consecutive times, the number of consecutive exceedances is recorded as 3, and a third risk level is assessed.

[0107] The error status update unit continuously receives the equally spaced monitoring time sequence formed by S100. When the current time position matches the prediction target time of a certain prediction result record, it retrieves the prediction result record with the measured matching status as pending matching according to the slope monitoring point identifier and the prediction target time. When a measurement value corresponds to multiple historical prediction results, it reads the prediction step size and model version identifier of each prediction result and calculates the prediction error separately. Historical prediction results generated with a prediction step size of 1 are written into the prediction error window with a prediction step size of 1, and historical prediction results generated with a prediction step size of 6 are written into the prediction error window with a prediction step size of 6.

[0108] When calculating the prediction error, the error state update unit reads the historical uncorrected deformation prediction values ​​from the prediction result record, without replacing the historical uncorrected deformation prediction values ​​with corrected deformation prediction values; it subtracts the historical uncorrected deformation prediction values ​​from the measured values ​​at the prediction target time to generate a new prediction error. The prediction error is calculated using the uncorrected deformation prediction values, making the prediction error window represent the offset of the original output of the deformation prediction model relative to the measured values; the offset correction amount is not repeatedly written into subsequent prediction errors.

[0109] The newly generated prediction error is associated with the slope monitoring point identifier, model version identifier, prediction step size, prediction generation time, prediction target time, error generation time, and measured value label. The error status update unit writes the prediction error into the corresponding prediction error window and updates the measured matching status in the prediction result record to "matched". When the measured value has an outlier label or a continuous missing state, it is not written into the prediction error window as a normal error sample, and the measured matching status is updated to "measured value outlier". When the measured value is a short-term missing segment interpolation value, it is written into the interpolation error label according to the error sample rules.

[0110] The prediction error window has a sliding window length. After a newly generated prediction error is written, if the number of error samples exceeds the sliding window length, the earliest written prediction error is removed according to the order of error generation, and the removed prediction error is retained in the historical prediction error record. After the prediction error window is updated, the window update time, the current number of error samples, the mean of the prediction error, and the standard deviation of the prediction error are recorded. The next time S300 is executed, the prediction error correction unit reads the updated prediction error window.

[0111] When the model version identifier changes, the error state update unit establishes an independent prediction error window for the new model version. Prediction results from the old model version are still written into the corresponding prediction error window after the target prediction time is reached, and are not mixed with the errors of the new model version. During model version rollback, the unit reads the prediction error window corresponding to the rolled-back model version that is not yet invalid. If the prediction error window exceeds the configured valid time range, the window's invalidation status is recorded, and upper and lower bounds for prediction are generated according to the path of insufficient prediction error samples.

[0112] When a measured value arrives late, the error status update unit retrieves historical prediction result records according to the predicted target time. While the prediction error window still retains the corresponding time range, the prediction error generated by the late measurement value is written into the window according to the error generation time. When the corresponding prediction error has already been generated by the interpolated measurement value, the prediction error is recalculated using the late measured value, the original prediction error is marked as replaced, and the mean and standard deviation within the prediction error window are updated. When a measured value arrives repeatedly, a valid measured value is selected based on the sensor status and reception time, and the replacement relationship is recorded.

[0113] This step ultimately generates the risk level, early warning command, early warning status record, historical prediction error record, and updated prediction error window. The updated prediction error window is read by S300 according to the slope monitoring point identifier, model version identifier, and prediction step size after the next input sliding window completes S200. The one-dimensional convolution channel, long short-term memory network channel, and mapping layer parameters of the deformation prediction model remain unchanged, and the mean and standard deviation of the prediction error window are updated as subsequent measurement values ​​arrive. This forms a continuous data write-back relationship from the generation of uncorrected deformation prediction values, the calculation of prediction upper and lower bounds, the output of early warning commands, the matching of measurement values, to the correction of the next deformation prediction value.

[0114] Example 2: Figure 2 A structural block diagram of a slope deformation trend prediction method based on temporal feature deep learning according to an embodiment of the present invention is shown. Figure 2 As shown, the structure may include: The monitoring time series generation module 01 is used to acquire slope surface deformation and displacement time series data, trigger factor time series data, and sampling timestamps. It aligns, identifies outliers, and imputes missing values ​​according to the sampling timestamps to generate equally spaced monitoring time series and data quality status. Specifically, the monitoring time series generation module receives the slope surface deformation and displacement time series data collected by the displacement monitoring device and the trigger factor time series data collected by the environmental sensing device. It associates the slope monitoring point identifier, sampling timestamp, displacement value, trigger factor value, and sensor status, arranging them in chronological order according to the sampling timestamps. The monitoring time series generation module maps the slope surface deformation and displacement time series data and the trigger factor time series data to the same continuous time position. When multiple monitoring values ​​exist at the same time position, the monitoring value with a normal sensor status is selected, and the median of the multiple valid monitoring values ​​is calculated. When no monitoring value exists at the same time position, that time position is recorded as a missing position. The monitoring time series generation module identifies outliers in displacement values ​​based on the moving median. When the difference between a displacement value and its corresponding moving median exceeds the outlier threshold, the displacement value is recorded as an outlier and replaced with an adjacent valid displacement value. When the number of consecutive missing positions is not greater than the maximum allowable interpolation length, missing values ​​are interpolated based on the valid monitoring values ​​before and after the missing position. When the number of consecutive missing positions is greater than the maximum allowable interpolation length, the consecutive missing length is recorded, and the corresponding missing status is retained. The data quality status includes the interpolation position, the proportion of valid data, the consecutive missing length, and the number of outliers. The monitoring time series generation module sends the equally spaced monitoring time series to the time series decomposition module and the periodic time series encoding module, sends the data quality status to the uncertainty generation module, and retains the correspondence between the sampling timestamp and each time position.

[0115] The time-series decomposition module 02, connected to the monitoring time-series generation module, is used to perform time-series decomposition on the surface deformation displacement time-series data in the equally-interval monitoring time-series, generating trend components, periodic components, and high-frequency residual components. Specifically, the time-series decomposition module receives the equally-interval monitoring time-series from the monitoring time-series generation module, reads the surface deformation displacement values ​​arranged according to time position, and performs empirical mode decomposition on the surface deformation displacement values ​​to form multiple candidate components and residual terms. The time-series decomposition module calculates the average period and energy percentage of each candidate component; superimposes candidate components with an average period greater than the upper limit of the period range with the residual terms at the same time position to form the trend component; superimposes candidate components with an average period within the period range to form the periodic component; and superimposes candidate components with an average period less than the lower limit of the period range to form the high-frequency residual component. The trend component, the periodic component, and the high-frequency residual component have the same number of time positions and are respectively associated with the sampling timestamp. When no candidate components within the periodic range are formed during the time-series decomposition process, the time-series decomposition module selects the candidate component with the smallest difference between the average period and the center of the periodic range as the periodic component and records the decomposition status. If the number of candidate components or the number of time positions do not meet the decomposition configuration, the equally spaced monitoring time series is retained, and the decomposition anomaly status is recorded. The time-series decomposition module sends the trend component to the trend feature extraction module, the periodic component to the periodic time-series encoding module, and the high-frequency residual component to the uncertainty generation module.

[0116] The trend feature extraction module 03, connected to the temporal decomposition module, is used to input the trend components into a one-dimensional convolutional channel for feature extraction, generating a trend feature vector. Specifically, the trend feature extraction module receives the trend components from the temporal decomposition module and forms the input of the one-dimensional convolutional channel according to the temporal position order of the trend components. The one-dimensional convolutional channel has multiple convolutional kernel scales, each covering a different number of continuous time positions. Each convolutional kernel moves along the temporal direction of the trend components, performing convolution calculations on the trend component values ​​within the corresponding time range to form multiple convolutional results. The trend feature extraction module activates and converges the features of the multiple convolutional results, recording the slope changes, inflection points, and acceleration segments of the trend components within a local time range. Then, the multiple convolutional results are concatenated according to the order of the convolutional kernel scales to generate the trend feature vector. When the number of time positions of the trend component is inconsistent with the input length of the one-dimensional convolutional channel, the trend feature extraction module reads the corresponding data quality status. If there are consecutive missing states, it records the insufficient input of the trend component and stops generating the current trend feature vector. If there are no consecutive missing states, it pads the input length according to the time positions and retains the padding positions. The trend feature extraction module sends the trend feature vector to the feature fusion and deformation prediction module and associates the trend feature vector with the model version identifier.

[0117] The periodic temporal encoding module 04, connected to the temporal decomposition module and the monitoring temporal generation module, is used to input the periodic components and the trigger factor temporal data associated with the same time step into the long short-term memory network channel for temporal encoding, generating a periodic temporal feature vector. Specifically, the periodic temporal encoding module receives the periodic components from the temporal decomposition module and receives the trigger factor temporal data and corresponding sampling timestamps from the monitoring temporal generation module. The periodic temporal encoding module associates the periodic component value at each time position with the trigger factor value at the same time position according to the sampling timestamp; it standardizes the trigger factor temporal data and merges the periodic component values ​​and the standardized trigger factor values ​​into a data matrix in chronological order. The long short-term memory network channel sequentially reads each row of data in the data matrix according to the time position, performs temporal encoding on the unit state of the previous time position and the input of the current time position, and generates hidden states for each time position; after completing the temporal encoding of all time positions, it performs feature aggregation on the hidden states to generate the periodic temporal feature vector. When the time position of the periodic component is inconsistent with that of the trigger factor time series data, the periodic time series encoding module re-associates it according to the sampling timestamp. If missing positions still exist after re-association, the missing marker is retained according to the data quality status, and the corresponding trigger factor value is written into the preset missing value. When the length of consecutive missing data exceeds the trigger factor missing threshold, the periodic time series input insufficiency status is recorded, and the current periodic time series feature vector is not generated. The periodic time series encoding module sends the periodic time series feature vector to the feature fusion and deformation prediction module, and maintains the association between the periodic time series feature vector and the model version identifier and sampling timestamp.

[0118] The feature fusion and deformation prediction module 05, connected to the trend feature extraction module and the periodic time-series encoding module, is used to concatenate the trend feature vector and the periodic time-series feature vector, and map the concatenation result to generate uncorrected deformation prediction values ​​and prediction result records corresponding to the prediction step length. Specifically, the feature fusion and deformation prediction module receives the trend feature vector and the periodic time-series feature vector, and verifies the slope monitoring point identifier, input time range, and model version identifier corresponding to the two feature vectors. When the verification is consistent, the trend feature vector and the periodic time-series feature vector are concatenated in the feature dimension according to the preset feature order to form a fused feature vector; when the verification is inconsistent, the feature version conflict state is recorded, and mapping is not performed. The feature fusion and deformation prediction module inputs the fused feature vector into the mapping layer, and generates uncorrected deformation prediction values ​​corresponding to multiple prediction step lengths based on the model parameters associated with the model version identifier in the mapping layer. Each uncorrected deformation prediction value is associated with the slope monitoring point identifier, prediction generation time, prediction target time, prediction step size, and model version identifier to form the prediction result record; the initial record of the measured matching status of the prediction result record is "to be matched". When the mapping result is non-numerical or exceeds the model output range, the feature fusion and deformation prediction module retains the original mapping result and records the abnormal state of the model output. The feature fusion and deformation prediction module sends the uncorrected deformation prediction value, the prediction step size, and the model version identifier to the prediction error correction module, and sends the prediction result record to the early warning judgment and error status update module.

[0119] The prediction error correction module 06, connected to the feature fusion and deformation prediction module, is used to obtain a prediction error window formed by historical uncorrected deformation prediction values ​​and the measured values ​​at the corresponding prediction target time, and corresponding to the prediction step size. While keeping the model parameters of the trend feature extraction module, the periodic time-series encoding module, and the feature fusion and deformation prediction module unchanged, the module corrects the uncorrected deformation prediction values ​​based on the mean of the prediction errors within the prediction error window, generating corrected deformation prediction values. Specifically, the prediction error correction module receives the uncorrected deformation prediction values, prediction step size, and model version identifier sent by the feature fusion and deformation prediction module, and also receives the prediction error window sent by the early warning judgment and error status update module. The prediction error correction module reads the corresponding prediction error window according to the slope monitoring point identifier, model version identifier, and prediction step size. Each prediction error in the prediction error window is arranged according to the error generation time and maintains its association with the historical uncorrected deformation prediction values ​​and the corresponding prediction target time. The prediction error correction module calculates the mean of multiple prediction errors within the prediction error window, records the mean as an offset correction amount, and adds the offset correction amount to the current uncorrected deformation prediction value to generate the corrected deformation prediction value. When the number of error samples in the prediction error window is less than the error sample number threshold, a prediction error sample shortage state is recorded; when the prediction error window is empty, the uncorrected deformation prediction value is recorded as the corrected deformation prediction value. During the correction process, the prediction error correction module keeps the model parameters of the trend feature extraction module, the periodic time series encoding module, and the feature fusion and deformation prediction module unchanged, and sends the corrected deformation prediction value, the prediction error window, and the prediction error sample shortage state to the uncertainty generation module.

[0120] Uncertainty generation module 07, connected to the prediction error correction module, the monitoring time series generation module, and the time series decomposition module, is used to generate an upper and lower bound for prediction based on the standard deviation of the prediction errors within the prediction error window, and to adjust the upper and lower bounds based on the data quality status and the high-frequency residual components. Specifically, the uncertainty generation module receives the corrected deformation prediction value and prediction error window sent by the prediction error correction module, the data quality status sent by the monitoring time series generation module, and the high-frequency residual components sent by the time series decomposition module. The uncertainty generation module calculates the standard deviation of multiple prediction errors within the prediction error window and reads the interval coefficients associated with the model version identifier and prediction step size; it then superimposes the interval widths formed by the standard deviations and the interval coefficients onto the upper and lower sides of the corrected deformation prediction value to generate the upper and lower bounds for prediction. The uncertainty generation module reads the proportion of valid data, the length of consecutive missing data, and the imputation position from the data quality status, and calculates the energy proportion of the high-frequency residual components. Based on the proportion of valid data, the length of consecutive missing data, the energy proportion of the high-frequency residual components, and the number of error samples, it adjusts the interval coefficients to regenerate the upper and lower bounds of the prediction. When there is an insufficient number of prediction error samples, the uncertainty generation module reads the historical error status associated with the model version identifier and prediction step size; if no corresponding historical error status exists, it records the upper and lower bounds of the prediction as pending updates. The uncertainty generation module sends the corrected deformation prediction value, the upper bound of the prediction, the lower bound of the prediction, and the data quality status to the early warning judgment and error status update module.

[0121] The early warning determination and error status update module 08, connected to the uncertainty generation module and the feature fusion and deformation prediction module, is used to match the corrected deformation prediction value, the upper bound of the prediction, and the lower bound of the prediction with the graded early warning threshold and output an early warning command. It is also used to, after the measured value arrives at the prediction target time, compare the measured value with the historical uncorrected deformation prediction value in the prediction result record, calculate the prediction error, update the prediction error window associated with the model version identifier and prediction step size, and send the updated prediction error window to the prediction error correction module. Specifically, the early warning determination and error status update module receives the corrected deformation prediction value, the upper bound of the prediction, the lower bound of the prediction, and the data quality status sent by the uncertainty generation module, and receives the prediction result record sent by the feature fusion and deformation prediction module. The early warning determination and error status update module sequentially matches the upper bound of the prediction, the corrected deformation prediction value, and the lower bound of the prediction with the corresponding graded early warning thresholds. When the upper bound of the prediction exceeds the graded early warning threshold but the corrected deformation prediction value does not exceed the graded early warning threshold, an early warning instruction corresponding to the first risk level is generated. When the corrected deformation prediction value exceeds the graded early warning threshold but the lower bound of the prediction does not exceed the graded early warning threshold, an early warning instruction corresponding to the second risk level is generated. When the lower bound of the prediction exceeds the graded early warning threshold, or when the number of consecutive exceedances reaches a preset number, an early warning instruction corresponding to the third risk level is generated. After the measured value at the predicted target time is reached, the early warning determination and error status update module retrieves the prediction result record with the measured matching status as pending matching according to the slope monitoring point identifier and the predicted target time. The measured value is compared with the historical uncorrected deformation prediction value in the prediction result record to generate a prediction error, which is then written into the corresponding prediction error window according to the model version identifier and prediction step size. When the number of error samples in the prediction error window exceeds the sliding window length, the earliest prediction error in time sequence is removed, and the update time of the prediction error window is recorded. Then, the updated prediction error window is sent to the prediction error correction module for use in the next correction of uncorrected deformation prediction values.

Claims

1. A method for predicting slope deformation trend based on temporal feature deep learning, characterized in that, include: S100: Acquire time series data of slope surface deformation and displacement, time series data of triggering factors and sampling timestamps, align according to sampling timestamps, identify outliers and interpolate missing values, and generate equal interval monitoring time series and data quality status. Decompose the surface deformation and displacement time series data in the equally spaced monitoring time series to generate trend components, periodic components and high-frequency residual components; S200: Input the trend component into a one-dimensional convolution channel to extract the trend feature vector; The periodic component and the triggering factor time-series data associated with the same time step are input into the long short-term memory network channel encoding to generate a periodic time-series feature vector; By concatenating and mapping the trend feature vector and the periodic time series feature vector, an uncorrected deformation prediction value for the prediction step size is generated; S300: Obtain the prediction error window formed by the uncorrected deformation prediction value and the measured value at the prediction target time, corresponding to the prediction step size; when the model parameters remain unchanged, correct the uncorrected deformation prediction value according to the mean of the prediction error therein, and generate the corrected deformation prediction value. The upper and lower bounds of the prediction are generated based on its standard deviation, and the upper and lower bounds of the prediction are adjusted according to the data quality status and high-frequency residual components. S400: Match the corrected deformation prediction value and the prediction upper and lower bounds with the graded early warning threshold, and output an early warning command; The prediction error window is updated after the measured value is reached.

2. The method according to claim 1, characterized in that, Step S100 includes: resampling the time series data of surface deformation displacement and the time series data of triggering factors according to a uniform sampling interval; identifying outliers in the time series data of surface deformation displacement based on the moving median; imputing missing values ​​in missing segments with a missing length not greater than the maximum allowable imputation length, and recording missing segments with a missing length greater than the maximum allowable imputation length as consecutive missing segments; and generating the data quality status based on the imputation position of the missing values, the proportion of effective data, and the consecutive missing length of the consecutive missing segments.

3. The method according to claim 1, characterized in that, The decomposition in step S100 includes: performing empirical mode decomposition on the surface deformation displacement time series data in the equally spaced monitoring time series to generate multiple candidate components; calculating the average period and energy percentage of each candidate component; combining the candidate components with an average period greater than the upper limit of the period range and the residual terms of the empirical mode decomposition into the trend component; combining the candidate components with an average period within the period range into the period component; and combining the candidate components with an average period less than the lower limit of the period range into the high-frequency residual component.

4. The method according to claim 1, characterized in that, In step S200, the one-dimensional convolution channel is set with multiple convolution kernel scales; the slope change, inflection point and acceleration segment features of the trend component in the corresponding sliding window are extracted by each convolution kernel scale. The extraction results from each convolution kernel scale are concatenated to generate the trend feature vector.

5. The method according to claim 1, characterized in that, The triggering factor time series data is rainfall time series data or pore water pressure time series data; step S200 includes: standardizing the triggering factor time series data; merging the periodic component and the triggering factor time series data at the same time step in time order into a data matrix; inputting the data matrix into the long short-term memory network channel for time series encoding to generate the periodic time series feature vector.

6. The method according to claim 1, characterized in that, Step S200 includes: concatenating the trend feature vector and the periodic time series feature vector in the feature dimension to generate a fused feature vector; mapping the fused feature vector to generate multiple uncorrected deformation prediction values ​​corresponding to multiple prediction step lengths; associating each uncorrected deformation prediction value with the slope monitoring point identifier, prediction generation time, prediction target time, prediction step length, and model version identifier to generate a prediction result record.

7. The method according to claim 6, characterized in that, Step S300 includes: associating the measured value with the uncorrected deformation prediction value in the prediction result record according to the slope monitoring point identifier and the prediction target time; calculating the difference between the measured value and the uncorrected deformation prediction value to generate a prediction error; grouping multiple prediction errors according to the model version identifier and prediction step size, and writing the prediction errors in the same group into the corresponding prediction error window in chronological order; calculating the mean of the prediction errors in the prediction error window to generate an offset correction amount; and adding the offset correction amount to the current uncorrected deformation prediction value to generate the corrected deformation prediction value.

8. The method according to claim 7, characterized in that, The prediction upper and lower bounds include a prediction upper bound and a prediction lower bound; step S300 includes: calculating the standard deviation of the prediction error within the prediction error window; taking the sum of the product of the corrected deformation prediction value and the standard deviation and the interval coefficient as the prediction upper bound, and taking the difference between the corrected deformation prediction value and the product as the prediction lower bound; adjusting the interval coefficient according to the effective data ratio, the continuous missing length, the energy ratio of the high-frequency residual component and the number of error samples; when the number of error samples is less than the error sample number threshold, recording the prediction error sample insufficiency state, and increasing the difference between the prediction upper bound and the prediction lower bound.

9. The method according to claim 8, characterized in that, Step S400 includes: assessing a first risk level when the upper bound of the prediction exceeds the corresponding graded warning threshold and the corrected deformation prediction value does not exceed the graded warning threshold; assessing a second risk level when the corrected deformation prediction value exceeds the graded warning threshold and the lower bound of the prediction does not exceed the graded warning threshold; assessing a third risk level when the lower bound of the prediction exceeds the graded warning threshold, or when the number of consecutive exceedances reaches a preset number; outputting a corresponding warning instruction according to the first risk level, the second risk level, or the third risk level; and releasing the warning instruction when the upper bound of the prediction is lower than the warning release threshold and the duration reaches a preset time.

10. The method according to claim 1, characterized in that, The update of the prediction error window includes: writing the newly generated prediction error into a prediction error window associated with the model version identifier and prediction step size; when the number of error samples in the prediction error window is greater than the sliding window length, removing the earliest written prediction error in chronological order; recording the update time of the prediction error window; and using the updated prediction error window for the correction of the next uncorrected deformation prediction value and the generation of prediction upper and lower bounds, while keeping the model parameters of the deformation prediction model unchanged.