A high-slope deformation prediction system and method based on frequency domain analysis and deep learning

Abstract

The application provides a high-slope deformation prediction system and method based on frequency domain analysis and deep learning, relates to the technical field of slope deformation prediction, and comprises a data acquisition module, a time sequence processing module, a frequency domain analysis module, a deformation prediction module and a deformation distribution module; the data acquisition module arranges monitoring points at intervals, divides a basic time window, and calculates the deformation difference of each monitoring point relative to a reference monitoring point; the time sequence processing module splices multi-source time sequence data to generate a plurality of cross-over sliding windows; the frequency domain analysis module outputs effective frequency band components and residual components; the deformation prediction module uses a deep learning model to output a deformation prediction value of a large deformation category; and the deformation distribution module distributes the total deformation and selects the maximum value as the overall deformation prediction result. The application solves the problems of single data dimension and prediction lag in the prior art, and realizes double-precision output of single-point deformation prediction values of each monitoring point and overall deformation prediction values of high slopes.

Description

Technical Field

[0001] This invention relates to the field of high slope deformation prediction technology, specifically to a high slope deformation prediction system and method based on frequency domain analysis and deep learning. Background Technology

[0002] With the rapid development of intelligent monitoring technology and deep learning algorithms, more and more technical solutions are being applied to the field of high slope deformation prediction. These solutions aim to address the pain points of traditional manual monitoring, such as high labor costs, low data collection frequency, and delayed early warnings, thereby improving the accuracy and timeliness of deformation prediction and promoting the upgrade of high slope safety management from passive response to proactive prevention. Currently, several deformation prediction methods based on data processing and deep learning have been publicly disclosed. Among them, a clustering-deep learning-based engineering structure deformation prediction method provided in publication number CN114881074A is a relatively typical technical solution.

[0003] The existing technology (CN114881074A) specifically includes the following steps: S1, real-time acquisition of deformation time-series data from different monitoring points in various monitoring projects in actual engineering; S2, preprocessing of the acquired deformation time-series data, including outlier detection, data resampling, and monitoring point clustering; S3, EMD decomposition of the preprocessed data to obtain multiple IMF components and one residual component; S4, inputting each IMF component and the residual component into the corresponding prediction model for training, obtaining the corresponding future prediction value through the trained prediction model, summing the future prediction values ​​corresponding to each component, and obtaining the final deformation prediction value of the monitoring project. This method has advantages such as good prediction effect, effective handling of various monitoring deformation data in actual engineering, reference value for practical engineering applications, accurate prediction, improvement of prediction lag, and reduction of computational load.

[0004] However, the following shortcomings still exist: the existing technology does not integrate multi-source data or use sliding window interception at the time series processing level, making it difficult to capture the real-time time series characteristics of high slopes, which easily leads to prediction lag and makes it impossible to accurately track their dynamic deformation process; at the feature processing level, EMD decomposition is used, but the monitoring points are not clustered and screened to focus on the key deformation areas, resulting in the inability to capture the core deformation.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a high slope deformation prediction system and method based on frequency domain analysis and deep learning, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A high slope deformation prediction system based on frequency domain analysis and deep learning includes:

[0009] The data acquisition module is used to set up monitoring points at equal intervals along the slope inclination direction on the surface of the high slope, define the monitoring interval and divide the basic time window, take the monitoring point at the bottom of the slope toe as the reference monitoring point, collect the slope deformation and environmental data of the other monitoring points in each basic time window, and calculate the difference in deformation of each monitoring point relative to the reference monitoring point.

[0010] The time series processing module is used to stitch together the slope deformation, environmental data, and deformation difference of the other monitoring points into multi-source time series data, and slide and cut according to the preset sliding window parameters to generate multiple overlapping sliding windows;

[0011] The frequency domain analysis module is used to cluster monitoring points in the multi-source time series data within each sliding window to obtain small, medium and large deformation categories. Data in the small and medium deformation categories are removed. Only the mean of the data in the large deformation category is calculated and analyzed in the frequency domain to output the corresponding frequency band components and residual components. At the same time, time-frequency domain feature matching is performed on the frequency band components of the large deformation category to obtain the effective frequency band components.

[0012] The deformation prediction module is used to input the effective frequency band components and residual components of the large deformation category corresponding to each sliding window into the trained deep learning model to predict the deformation value of the category in the next basic time window.

[0013] The deformation allocation module is used to sum the predicted values ​​of slope deformation for large deformation categories under the most recent sliding window to obtain the total predicted value of deformation. It allocates the total deformation according to the difference between the deformation of each monitoring point and the benchmark monitoring point within the category, obtains the predicted value of deformation for each monitoring point in the next basic time window, and selects the maximum value as the overall deformation prediction result of the high slope.

[0014] Furthermore, the environmental data includes rainfall, slope temperature, and soil moisture; the frequency band components include low-frequency components, mid-frequency components, and high-frequency components.

[0015] Furthermore, the monitoring intervals are defined and basic time windows are divided, with the following specific logic:

[0016] Using the current slope deformation monitoring time as the endpoint, a preset time period is traced back to form a deformation monitoring interval, which is then evenly divided into multiple continuous and non-overlapping basic time windows.

[0017] Furthermore, the deformation difference between each monitoring point and the benchmark monitoring point is calculated, with the following specific logic:

[0018] Set the benchmark monitoring point as The remaining monitoring points are ,in For the first The remaining monitoring points, , This represents the total number of the remaining monitoring points;

[0019] No. Other monitoring points In the The deformation within each basic time window is Benchmark monitoring points In the The deformation within each basic time window is Then the first Other monitoring points relative to benchmark monitoring points In the Deformation difference of each basic time window for:

[0020]

[0021] in, The index of the base time window, , The total number of base time windows.

[0022] Furthermore, by sliding and cropping according to preset sliding window parameters, multiple overlapping sliding windows are generated. The specific logic is as follows:

[0023] The sliding window parameters include the sliding window length and the sliding step size;

[0024] Wherein, the length of the sliding window is The sliding step size is , and All are positive integers, and ;

[0025] Multi-source time series data contains A basic time window, and satisfying ;

[0026] According to the sliding step size Slide sequentially from left to right, and extract segments of length [length missing]. Continuous data segments form multiple overlapping sliding windows, denoted as... ;

[0027] in For the first A sliding window, For the index of the sliding window, , Let be the total number of sliding windows, satisfying:

[0028]

[0029] in, To round down;

[0030] No. A sliding window The range of the basic time window index is: Furthermore, the index range falls entirely within the base time window index interval. Inside.

[0031] Furthermore, time-series-frequency domain feature matching is performed on the frequency band components of the large deformation category to obtain the effective frequency band components. The specific logic is as follows:

[0032] Extract all frequency band components corresponding to the large deformation category within each sliding window, and the time series sequence of deformation difference values ​​of each monitoring point of the large deformation category within the sliding window arranged in the order of the basic time window index;

[0033] For a time series sequence of frequency band components and deformation difference within a sliding window, calculate the changes of both within corresponding adjacent base time windows. The change of the frequency band components is:

[0034]

[0035] in, This represents the change in frequency band components within a sliding window over adjacent fundamental time windows. For the frequency band components in the first... The values ​​of a base time window, For the frequency band components in the first... The value of a basic time window;

[0036] The change in the time series sequence of deformation difference is:

[0037]

[0038] in, This represents the change in the time series of deformation difference within a sliding window over adjacent base time windows. The time series sequence of deformation difference is in the 1st... The values ​​of a base time window, The time series sequence of deformation difference is in the 1st... The value of a basic time window;

[0039] The signs of the slopes of the two changes are statistically analyzed, and the ratio of the number of slopes with the same sign to the total number of slopes is calculated. This ratio is the corresponding trend consistency coefficient, where the total number of slopes is... ;

[0040] The mean of all trend consistency coefficients is used as an adaptive threshold to filter out frequency band components with trend consistency coefficients greater than the threshold, which are the effective frequency band components.

[0041] Furthermore, the total deformation is allocated according to the difference in deformation between each monitoring point and the benchmark monitoring point within the large deformation category, thus obtaining the predicted deformation value for each monitoring point in the next basic time window. The specific logic is as follows:

[0042] The predicted total deformation value for the large deformation category under the most recent sliding window is: All monitoring points within the large deformation category were in the [number]th [year]. The sum of the deformation differences of each basic time window is ;

[0043] Calculate the first The monitoring point at the 1st Allocation ratio of each basic time window The specific formula is as follows:

[0044]

[0045] Calculate the first The predicted deformation value for each monitoring point within the next basic time window is calculated using the following formula:

[0046]

[0047] in, For the first The predicted value of deformation at each monitoring point within a basic time window.

[0048] To achieve the above objectives, the present invention also provides the following technical solution:

[0049] A method for predicting high slope deformation based on frequency domain analysis and deep learning, wherein the method is executed by any of the high slope deformation prediction systems based on frequency domain analysis and deep learning described above, and the specific steps include:

[0050] S1. Set up monitoring points at equal intervals along the slope inclination direction on the surface of the high slope, delineate the monitoring interval and divide the basic time window, take the monitoring point at the bottom of the slope toe as the reference monitoring point, collect the slope deformation and environmental data of the other monitoring points in each basic time window, and calculate the difference in deformation of each monitoring point relative to the reference monitoring point.

[0051] S2. The slope deformation, environmental data, and deformation difference of the remaining monitoring points are spliced ​​into multi-source time series data, and the data is slid-trimmed according to the preset sliding window parameters to generate multiple overlapping sliding windows;

[0052] S3. Cluster the monitoring points of the multi-source time series data in each sliding window to obtain small, medium and large deformation categories. Remove the data in the small and medium deformation categories. Perform frequency domain analysis on the mean of the data in the large deformation category and output the corresponding frequency band components and residual components. At the same time, perform time-frequency domain feature matching and filtering on the frequency band components of the large deformation category to obtain the effective frequency band components.

[0053] S4. Input the effective frequency band components and residual components of the large deformation category corresponding to each sliding window into the trained deep learning model to predict the deformation value of the category in the next basic time window.

[0054] S5. Sum the predicted values ​​of slope deformation for the large deformation category under the most recent sliding window to obtain the total predicted value of deformation. Allocate the total deformation according to the difference between the deformation of each monitoring point and the benchmark monitoring point within the category to obtain the predicted value of deformation for each monitoring point in the next basic time window. Select the maximum value as the overall deformation prediction result of the high slope.

[0055] Compared with the prior art, the beneficial effects of the present invention are:

[0056] The data acquisition module of this invention uses the monitoring point at the lowest point of the slope toe as a unified benchmark monitoring point. Monitoring points are evenly distributed along the tilt direction on the surface of the high slope, and the slope deformation and environmental data of each monitoring point are collected simultaneously. At the same time, the deformation difference between each other monitoring point and the benchmark monitoring point is calculated. This effectively solves the problems of existing technologies that only collect single deformation time series data, lack unified reference, and have a single dimension, laying a data foundation for subsequent prediction of the overall deformation of high slopes. The time series processing module splices and integrates the three types of data collected: slope deformation, environmental data, and deformation difference, to form comprehensive and complete multi-source time series data. By sliding and truncating through preset sliding window parameters, the real-time time series deformation characteristics of high slopes are accurately captured, effectively improving the shortcomings of existing technologies that are lagging in prediction and unable to accurately track the dynamic deformation process of high slopes.

[0057] The frequency domain analysis module adopts a three-order scheme of "clustering screening + frequency domain analysis + feature matching". By clustering screening, it focuses on the large deformation area with the highest deformation risk of high slope, and performs frequency domain analysis and feature matching. This solves the inherent defects of existing technologies that use EMD decomposition to process data, such as a lot of redundant interference, a large amount of invalid calculation, and poor feature targeting. The deformation allocation module adopts a closed-loop scheme of "total summation - difference allocation - maximum value evaluation". It sums the predicted values ​​of large deformation categories to obtain the total deformation, and then allocates the total deformation according to the difference in deformation between each monitoring point and the benchmark monitoring point. Finally, it achieves the dual accurate output of the single-point deformation prediction value of each monitoring point and the overall deformation prediction value of the high slope. Attached Figure Description

[0058] Figure 1 This is a block diagram of the modules of the present invention;

[0059] Figure 2 This is a schematic diagram of the overall method flow of the present invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0061] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0062] Example:

[0063] Please see Figure 1 The present invention provides a technical solution:

[0064] A high slope deformation prediction system based on frequency domain analysis and deep learning, the specific steps of which include:

[0065] The data acquisition module is used to set up monitoring points at equal intervals along the slope inclination direction on the surface of the high slope, define the monitoring interval and divide the basic time window, take the monitoring point at the bottom of the slope toe as the reference monitoring point, collect the slope deformation and environmental data of the other monitoring points in each basic time window, and calculate the difference in deformation of each monitoring point relative to the reference monitoring point.

[0066] All subsequent monitoring points refer to the remaining monitoring points.

[0067] Based on the above embodiments, monitoring points are evenly spaced along the slope inclination direction on the surface of the high slope, monitoring intervals are defined, and basic time windows are divided. The specific logic is as follows:

[0068] Starting from the lowest point of the slope toe, multiple monitoring points are sequentially set up along the upward slope direction at preset intervals, so that the monitoring points are distributed linearly and equally at intervals, forming an array of monitoring points covering the key deformation areas of the high slope.

[0069] Using the current slope deformation monitoring time as the endpoint, trace back a preset time period to form a deformation monitoring interval, which is then evenly divided into multiple continuous and non-overlapping basic time windows;

[0070] The preset spacing is adaptively set according to the height of the high slope, the slope gradient and geological conditions. Under the premise of ensuring uniform monitoring of slope deformation, the spacing between adjacent monitoring points is set to 3m to 10m. Smaller spacing is used in steep slopes and deformation-sensitive areas, and larger spacing is used in gentle slopes and deformation-stable areas.

[0071] Based on the above embodiment, using the monitoring point at the lowest point of the slope toe as the baseline monitoring point, the slope deformation of the remaining monitoring points within each basic time window is collected. The specific steps are as follows:

[0072] Displacement monitoring equipment is configured for each monitoring point. Displacement data of each monitoring point is collected in real time through the displacement monitoring equipment, and the collected displacement data is statistically analyzed in segments according to the basic time window.

[0073] For each basic time window, the displacement monitoring data of each monitoring point within the corresponding time range is extracted, and the average value is taken as the slope deformation of the monitoring point within the current basic time window.

[0074] This yields the slope deformation at the benchmark monitoring points within each basic time window, as well as the slope deformation at the other monitoring points within their respective basic time windows.

[0075] Based on the above embodiments, the environmental data includes rainfall, slope temperature, and soil moisture, and the specific collection methods and purposes are as follows:

[0076] Rainfall, temperature and humidity sensors are deployed in the high slope monitoring area to collect data synchronously with the displacement monitoring equipment. Rainfall, slope temperature and soil moisture data are statistically analyzed in segments according to the basic time window, and the average value of the data in each basic time window is taken as the environmental data of the corresponding window.

[0077] The purpose of collecting rainfall data is to capture the dynamic impact of rainfall on the soil saturation and seepage field of the slope, quantify the additional deformation of the slope induced by rainfall, provide data support for distinguishing the frequency band characteristics of "rainfall-induced deformation" and "slope creep" in subsequent frequency domain analysis, and avoid misjudgment of frequency band components caused by rainfall interference.

[0078] The purpose of collecting slope temperature data is to record the temporal variation of thermal expansion and contraction of slope soil and rock, extract the micro-deformation characteristics of slope driven by temperature, supplement the temporal correlation of multi-source time series data, improve the fitting accuracy of deep learning models to the temporal trend of slope deformation, and solve the problem of temporal bias when predicting single deformation data.

[0079] The purpose of collecting soil moisture data is to obtain real-time dynamic changes in slope soil shear strength and self-weight stress, quantify the correlation between soil moisture and slope deformation, provide a basis for subsequent genetic algorithm optimization of prediction model parameters and deformation allocation module to accurately allocate deformation, and at the same time assist in screening "moisture-dominated" deformation data in large deformation categories to improve prediction accuracy.

[0080] The above three types of environmental data are collected as a core component of multi-source time-series data. Together with slope deformation and deformation difference, they provide comprehensive and targeted input features for feature selection in the frequency domain analysis module and accurate prediction by the deep learning model, thereby improving the accuracy of high slope deformation prediction from the root.

[0081] Based on the above embodiments, the deformation difference between each monitoring point and the benchmark monitoring point is calculated. The specific logic is as follows:

[0082] Set the benchmark monitoring point as The remaining monitoring points are ,in For the first The remaining monitoring points, , This represents the total number of the remaining monitoring points;

[0083] No. Other monitoring points In the The deformation within each basic time window is Benchmark monitoring points In the The deformation within each basic time window is Then the first Other monitoring points relative to benchmark monitoring points In the Deformation difference of each basic time window for:

[0084]

[0085] in, The index of the base time window, , The total number of base time windows.

[0086] Based on the above, it should be noted that:

[0087] During the monitoring of high slopes, they are susceptible to common interference factors such as slight deformation of the overall crust in the region, uniform ground settlement, construction disturbance, and system errors of monitoring equipment. Such interference can cause synchronous and uniform displacement of all monitoring points (i.e., the baseline monitoring point). Deformation amount This displacement is not caused by structural instability of the slope itself and has no practical predictive significance. The difference in relative deformation is calculated. The synchronous displacement interference of all monitoring points is uniformly removed, and only the differential deformation of each monitoring point relative to the benchmark monitoring point is retained. This differential deformation is the true manifestation of the instability precursors such as shearing, creep, and sliding of the soil and rock inside the slope, which can effectively avoid the subsequent analysis deviation caused by invalid interference.

[0088] The core characteristic of high slope instability is the uneven deformation across different areas of the slope, i.e., the existence of a significant deformation gradient (larger deformation at the top, smaller deformation at the bottom, or abnormal deformation in localized areas). Deformation amount It can only reflect the absolute displacement of a single monitoring point, and cannot reflect the deformation differences between monitoring points, making it difficult to distinguish between small, medium, and large deformation categories. However, the relative deformation difference... It can intuitively quantify the deformation difference between each monitoring point and the benchmark monitoring point, and clearly present the distribution pattern of the deformation gradient of the slope along the tilt direction. The larger the deformation difference, the more severe the slope deformation in the area where the monitoring point is located, and the closer it is to the unstable state. It can provide core classification basis for monitoring point clustering in the subsequent frequency domain analysis module, help accurately screen out the large deformation category, and improve the pertinence and accuracy of large deformation identification.

[0089] Based on the above, it should be noted that:

[0090] Using the lowest monitoring point at the toe of the slope as a unified benchmark monitoring point, monitoring points are set up at equal intervals along the slope direction on the surface of the high slope. The slope deformation and environmental data of each monitoring point are collected simultaneously, and the difference in deformation of each other monitoring point relative to the benchmark monitoring point is calculated. This effectively solves the problems of existing technologies that only collect single deformation time series data, lack unified reference, and have a single dimension, and lays a data foundation for subsequent prediction of the overall deformation of high slopes.

[0091] The time series processing module is used to stitch together the slope deformation, environmental data, and deformation difference of the other monitoring points into multi-source time series data, and slide and cut according to the preset sliding window parameters to generate multiple overlapping sliding windows;

[0092] Based on the above embodiments, the slope deformation, environmental data, and deformation difference of the remaining monitoring points are stitched together to form multi-source time-series data. The specific logic is as follows:

[0093] According to the index order of the basic time window, the slope deformation, rainfall, slope temperature, soil moisture and deformation difference of the same monitoring point and the same basic time window are spliced ​​together to form time series data containing multi-dimensional features.

[0094] The above stitching operation is performed on all remaining monitoring points to finally obtain multi-source time series data covering all monitoring points and all basic time windows.

[0095] Based on the above embodiments, multiple overlapping sliding windows are generated by sliding and cropping according to preset sliding window parameters. The specific logic is as follows:

[0096] The sliding window parameters include the sliding window length and the sliding step size;

[0097] Wherein, the length of the sliding window is The sliding step size is , and All are positive integers, and ;

[0098] Multi-source time series data contains A basic time window, and satisfying ;

[0099] According to the sliding step size Slide sequentially from left to right, and extract segments of length [length missing]. Continuous data segments form multiple overlapping sliding windows, denoted as... ;

[0100] in For the first A sliding window, For the index of the sliding window, , Let be the total number of sliding windows, satisfying:

[0101]

[0102] in, To round down;

[0103] No. A sliding window The range of the basic time window index is: Furthermore, the index range falls entirely within the base time window index interval. Inside.

[0104] Based on the above, it should be noted that:

[0105] By extracting and generating overlapping sliding windows according to preset sliding window parameters, multi-source time series data can be integrated into standardized time series samples with fixed length and continuous overlap. While fully preserving the temporal variation law of slope deformation and avoiding feature loss and misjudgment, it effectively improves the shortcomings of existing technology in predicting lag and being unable to accurately track the dynamic deformation process of high slopes.

[0106] The frequency domain analysis module is used to cluster monitoring points in the multi-source time series data within each sliding window to obtain small, medium and large deformation categories. Data in the small and medium deformation categories are removed. Only the mean of the data in the large deformation category is calculated and analyzed in the frequency domain to output the corresponding frequency band components and residual components. At the same time, time-frequency domain feature matching is performed on the frequency band components of the large deformation category to obtain the effective frequency band components.

[0107] Based on the above embodiments, the monitoring points of the multi-source time-series data within each sliding window are clustered to obtain small, medium, and large deformation categories. The specific logic is as follows:

[0108] Based on the difference in deformation of each monitoring point within the current sliding window, the K-Means clustering algorithm (preset cluster number K=3) is used to classify all monitoring points within the same sliding window. According to the magnitude of the difference in deformation, the monitoring points are divided into small deformation category, medium deformation category, and large deformation category.

[0109] Among them, the smaller the difference in deformation, the lower the deformation category level of the corresponding monitoring point; the larger the difference in deformation, the higher the deformation category level of the corresponding monitoring point.

[0110] The current sliding window refers to the window in the cluster analysis that is currently in progress. A sliding window.

[0111] Based on the above, it should be noted that:

[0112] Small and medium deformation categories correspond to stable slope areas or areas subject to slight environmental disturbances; their deformation characteristics do not have the value of instability early warning.

[0113] The large deformation category corresponds to the area inside the slope where deformation is severe and instability is most likely to occur, and is the core area for slope prediction.

[0114] By clustering monitoring points and classifying small, medium and large deformation categories into small, medium and large deformation categories within each sliding window, the deformation degree of different areas of the slope can be distinguished, and the deformation area with the highest risk of slope instability can be screened out.

[0115] Based on the above embodiments, after calculating the mean of the data within the large deformation category, frequency domain analysis is performed to output the corresponding frequency band components and residual components. The frequency band components include low-frequency components, mid-frequency components, and high-frequency components. The specific logic is as follows:

[0116] Extract multi-source time series data of all monitoring points in the large deformation category within the current sliding window, calculate the mean of the corresponding data of all monitoring points in the same category under the same basic time window, and obtain a single time series mean sequence corresponding to the large deformation category within the current sliding window. This achieves feature aggregation of similar data and eliminates single-point monitoring noise and abnormal fluctuations.

[0117] The above time-series mean sequence is decomposed in the frequency domain using Fast Fourier Transform, which converts the continuous time-series signal in the time domain into a frequency domain signal, and decomposes it to obtain the frequency band components and residual components in different frequency ranges.

[0118] Based on the physical characteristics and frequency distribution of slope deformation, the decomposed frequency band components are divided as follows:

[0119] Low-frequency component: The corresponding frequency range is 0-0.1Hz. It mainly reflects the long-term creep deformation characteristics of the slope rock and soil and is the core temporal characteristic of the slow instability of the slope.

[0120] Mid-frequency component: The corresponding frequency range is 0.1-1Hz, which mainly reflects the periodic deformation characteristics of slopes induced by environmental factors such as rainfall and temperature changes;

[0121] High-frequency components: corresponding to a frequency range of 1-10Hz, mainly reflecting the instantaneous deformation characteristics caused by short-term sudden disturbances such as construction disturbances and equipment vibrations;

[0122] Component output: The low-frequency component, mid-frequency component, high-frequency component, and residual component generated during the frequency domain decomposition are output synchronously to provide frequency domain data for subsequent time-frequency domain feature matching and screening and deformation prediction.

[0123] The sampling frequency of the Fast Fourier Transform is matched with the time interval of the base time window.

[0124] Based on the above embodiments, time-series-frequency domain feature matching is performed on the frequency band components of the large deformation category to obtain effective frequency band components. The specific logic is as follows:

[0125] Extract all frequency band components corresponding to the large deformation category within each sliding window, and the time series sequence of deformation difference values ​​of each monitoring point of the large deformation category within the sliding window arranged in the order of the basic time window index;

[0126] For a time series sequence of frequency band components and deformation difference within a sliding window, calculate the changes of both within corresponding adjacent base time windows. The change of the frequency band components is:

[0127]

[0128] in, This represents the change in frequency band components within a sliding window over adjacent fundamental time windows. For the frequency band components in the first... The values ​​of a base time window, For the frequency band components in the first... The values ​​of a base time window;

[0129] The change in the time series sequence of deformation difference is:

[0130]

[0131] in, This represents the change in the time series of deformation difference within a sliding window over adjacent base time windows. The time series sequence of deformation difference is in the 1st... The values ​​of a base time window, The time series sequence of deformation difference is in the 1st... The values ​​of a base time window;

[0132] The signs of the slopes of the two changes are statistically analyzed, and the ratio of the number of slopes with the same sign to the total number of slopes is calculated. This ratio is the corresponding trend consistency coefficient. The total number of slopes represents the number of changes between adjacent base time windows within the sliding window, and its value is equal to the length of the sliding window. ;

[0133] The mean of all trend consistency coefficients is used as an adaptive threshold to filter out frequency band components with trend consistency coefficients greater than the threshold, which are the effective frequency band components.

[0134] Among them, all trend consistency coefficients include the trend consistency coefficients corresponding to the low-frequency component, the mid-frequency component, and the high-frequency component.

[0135] Among them, the slope refers to the change in frequency band components or deformation difference between adjacent base time windows, which is used to characterize the direction of change of time series data.

[0136] Based on the above, it should be noted that:

[0137] Among the low-frequency, mid-frequency, and high-frequency components obtained through frequency domain decomposition, some components may only correspond to environmental interference, equipment noise, or short-term random fluctuations, and are unrelated to the actual deformation trend of the slope. Directly inputting all frequency band components into a deep learning model would introduce a large amount of invalid information, reducing prediction accuracy and reliability.

[0138] Therefore, timing operations need to be performed on the frequency band components of the large deformation category. Frequency domain feature matching and filtering. In this step, only the sign of the slope of change is counted, rather than the specific value. The purpose is to determine whether the changing trend of the frequency band component and the actual deformation difference is consistent: the same sign indicates that the two change in the same direction, and opposite signs indicate that they are not synchronized.

[0139] The trend consistency coefficient is calculated based on the symbol consistency. Then, the mean of all trend consistency coefficients is used as the adaptive threshold. This can automatically identify and retain the frequency band components that are consistent with the actual deformation trend of the slope, and remove irrelevant noise components, interference components and spurious components, so as to ensure that the input to the deep learning model are all effective frequency band components that can truly reflect the deformation law of the slope.

[0140] Based on the above, it should be noted that:

[0141] The frequency domain analysis module adopts a three-order scheme of "clustering screening + frequency domain analysis + feature matching". By clustering screening, it focuses on the key areas of large deformation with the highest risk of deformation on high slopes, and performs frequency domain analysis and feature matching. This solves the inherent defects of existing technologies that use EMD decomposition to process data, such as a lot of redundant interference, a large amount of invalid calculation, and poor feature targeting.

[0142] The deformation prediction module is used to input the effective frequency band components and residual components of the large deformation category corresponding to each sliding window into the trained deep learning model to predict the deformation value of the category in the next basic time window.

[0143] The next basic time window refers to the first... A basic time window.

[0144] Based on the above embodiments, the predicted deformation value is the time-series mean deformation value corresponding to the large deformation category, which corresponds to the time-series mean sequence after the aforementioned large deformation category data mean calculation.

[0145] Based on the above embodiments, the deep learning model is constructed using a deep learning network based on a multilayer perceptron. The deep neural network of the multilayer perceptron includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, and an output layer. The first hidden layer, the second hidden layer, and the third hidden layer each have at least two neurons and all use ReLU (linear rectified unit) as the activation function.

[0146] The effective frequency band components and residual components of large deformation categories corresponding to multiple historical sliding windows, as well as the true temporal mean deformation of the large deformation category in the corresponding basic time window, are collected to construct multiple sets of samples. The samples are divided into training set, validation set and test set in a ratio of 7:2:1. The training set is used for learning model parameters; the validation set is used to adjust hyperparameters during training to prevent overfitting; and the test set is used to evaluate the generalization ability of the model after training.

[0147] The structure of a deep learning network with a multilayer perceptron is as follows:

[0148] Input layer: used to receive the effective frequency band components and residual components of large deformation categories;

[0149] The first hidden layer has 128 neurons and uses ReLU as the activation function.

[0150] The second hidden layer has 64 neurons and also uses the ReLU activation function;

[0151] The third hidden layer has 32 neurons and uses the ReLU activation function;

[0152] Output layer: has 1 neuron, used to output the deformation amount of large deformation categories within the base time window;

[0153] The process of training a deep learning model is as follows:

[0154] Using the effective frequency band components and residual components of large deformation categories from multiple historical sliding windows as model inputs, and the true temporal mean deformation of the corresponding base time window as the true supervision label, the mean absolute error loss function is selected. The difference between the temporal mean deformation predicted by the deep learning model and the true supervision label is calculated to obtain the model loss value. The deep learning model parameters are iteratively updated using the backpropagation algorithm. When the model loss value is within a certain range... If the loss does not decrease significantly within the specified interval for 20 consecutive training rounds, the deep learning model is considered to have converged and training is stopped.

[0155] Among them, the mean absolute error loss function is selected. This function is suitable for numerical regression prediction tasks, has strong robustness to outliers in slope monitoring data, and can effectively measure the absolute error between the predicted value and the true value of the time-series mean deformation, thus ensuring the stability of model training and prediction accuracy.

[0156] The deformation allocation module is used to sum the predicted values ​​of slope deformation for the largest deformation category under the most recent sliding window to obtain the total predicted value of deformation. The total deformation is allocated according to the difference between the deformation of each monitoring point and the benchmark monitoring point within the category, so as to obtain the predicted value of deformation for each monitoring point in the next basic time window. Since the overall safety of the slope is determined by the part with the most significant deformation and the highest risk, the average or overall value alone cannot reflect the true state of the most dangerous point. Therefore, the maximum value of the predicted deformation of each monitoring point is selected as the overall deformation prediction result of the high slope.

[0157] Based on the above embodiments, the total predicted deformation value is obtained by summing the predicted values ​​of slope deformation for the large deformation category under the most recent sliding window. The specific steps are as follows:

[0158] The most recent sliding window is the A sliding window ;

[0159] Extract the output of the deformation prediction module, corresponding to the first The large deformation category of the sliding window in the first Predicted deformation values ​​for each basic time window;

[0160] The predicted value of this deformation is directly used as the predicted value of the total deformation, and then used for the subsequent allocation of deformation at each monitoring point.

[0161] Based on the above embodiments, the total deformation is allocated according to the difference in deformation between each monitoring point and the benchmark monitoring point within the large deformation category, and the predicted deformation value of each monitoring point in the next basic time window is obtained. The specific logic is as follows:

[0162] No. A sliding window The predicted total deformation value corresponding to the lower large deformation category is All monitoring points within the large deformation category were in the [number]th [year]. The sum of the deformation differences of each basic time window is ;

[0163] Calculate the first The monitoring point at the 1st Allocation ratio of each basic time window The specific formula is as follows:

[0164]

[0165] Calculate the first The predicted deformation value for each monitoring point within the next basic time window is calculated using the following formula:

[0166]

[0167] in, For the first The predicted value of deformation at each monitoring point within a basic time window.

[0168] Based on the above, it should be noted that:

[0169] The time-series average deformation output by the deformation prediction module is the overall prediction result for large deformation categories. It can only reflect the overall deformation level of the category and cannot reflect the deformation differences between different monitoring points within the category.

[0170] The deformation patterns, deformation amplitudes, and deformation contributions relative to the benchmark point vary significantly among monitoring points at different locations on a high slope. Simply distributing the overall predicted values ​​equally would not match the actual deformation distribution, leading to distorted prediction results for each monitoring point.

[0171] Therefore, by using the difference in deformation between each of the remaining monitoring points and the benchmark monitoring point for allocation, the true relative deformation ratio of each of the remaining monitoring points can be preserved, and the overall predicted value can be reasonably allocated to each of the remaining monitoring points, so that the single-point prediction result is more in line with the actual deformation distribution of the slope, thereby improving the accuracy of the prediction and its applicability to engineering.

[0172] Based on the above, it should be noted that:

[0173] The deformation allocation module uses a closed-loop scheme of "total summation - difference allocation - maximum value evaluation" to sum the predicted values ​​of large deformation categories to obtain the total deformation. Then, it allocates the total deformation according to the difference in deformation between each monitoring point and the benchmark monitoring point, ultimately achieving a dual accurate output of the single-point deformation prediction value of each monitoring point and the overall deformation prediction value of the high slope.

[0174] Please see Figure 2 The present invention also provides a technical solution:

[0175] A method for predicting high slope deformation based on frequency domain analysis and deep learning, wherein the method is executed by any of the high slope deformation prediction systems based on frequency domain analysis and deep learning described above, and the specific steps include:

[0176] S1. Set up monitoring points at equal intervals along the slope inclination direction on the surface of the high slope, delineate the monitoring interval and divide the basic time window, take the monitoring point at the bottom of the slope toe as the reference monitoring point, collect the slope deformation and environmental data of the other monitoring points in each basic time window, and calculate the difference in deformation of each monitoring point relative to the reference monitoring point.

[0177] S2. The slope deformation, environmental data, and deformation difference of the remaining monitoring points are spliced ​​into multi-source time series data, and the data is slid and cut according to the preset sliding window parameters to generate multiple overlapping sliding windows;

[0178] S3. Cluster the monitoring points of the multi-source time series data in each sliding window to obtain small, medium and large deformation categories. Remove the data in the small and medium deformation categories. Perform frequency domain analysis on the mean of the data in the large deformation category and output the corresponding frequency band components and residual components. At the same time, perform time-frequency domain feature matching and filtering on the frequency band components of the large deformation category to obtain the effective frequency band components.

[0179] S4. Input the effective frequency band components and residual components of the large deformation category corresponding to each sliding window into the trained deep learning model to predict the deformation value of the category in the next basic time window.

[0180] S5. Sum the predicted values ​​of slope deformation for the large deformation category under the most recent sliding window to obtain the total predicted value of deformation. Allocate the total deformation according to the difference between the deformation of each monitoring point and the benchmark monitoring point within the category to obtain the predicted value of deformation for each monitoring point in the next basic time window. Select the maximum value as the overall deformation prediction result of the high slope.

[0181] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0182] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by software, electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0183] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0184] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A high slope deformation prediction system based on frequency domain analysis and deep learning, characterized in that, include: The data acquisition module is used to set up monitoring points at equal intervals along the slope inclination direction on the surface of the high slope, define the monitoring interval and divide the basic time window, take the monitoring point at the bottom of the slope toe as the reference monitoring point, collect the slope deformation and environmental data of the other monitoring points in each basic time window, and calculate the difference in deformation of each monitoring point relative to the reference monitoring point. The time series processing module is used to stitch together the slope deformation, environmental data, and deformation difference of the other monitoring points into multi-source time series data, and slide and cut according to the preset sliding window parameters to generate multiple overlapping sliding windows; The frequency domain analysis module is used to cluster monitoring points in the multi-source time series data within each sliding window to obtain small, medium and large deformation categories. Data in the small and medium deformation categories are removed. Only the mean of the data in the large deformation category is calculated and analyzed in the frequency domain to output the corresponding frequency band components and residual components. At the same time, time-frequency domain feature matching is performed on the frequency band components of the large deformation category to obtain the effective frequency band components. The deformation prediction module is used to input the effective frequency band components and residual components of the large deformation category corresponding to each sliding window into the trained deep learning model to predict the deformation amount of the large deformation category in the next basic time window. The deformation allocation module is used to sum the predicted values ​​of slope deformation under the largest deformation category in the most recent sliding window to obtain the total predicted value of deformation. It allocates the total deformation according to the difference between the deformation of each monitoring point and the benchmark monitoring point within the largest deformation category, obtains the predicted value of deformation of each monitoring point in the next basic time window, and selects the maximum value as the overall deformation prediction result of the high slope.

2. The high slope deformation prediction system based on frequency domain analysis and deep learning according to claim 1, characterized in that, The environmental data includes rainfall, slope temperature, and soil moisture; the frequency band components include low-frequency components, mid-frequency components, and high-frequency components.

3. The high slope deformation prediction system based on frequency domain analysis and deep learning according to claim 1, characterized in that, The monitoring interval is defined and the basic time window is divided. The specific logic is as follows: Using the current slope deformation monitoring time as the endpoint, a preset time period is traced back to form a deformation monitoring interval, which is then evenly divided into multiple continuous and non-overlapping basic time windows.

4. The high slope deformation prediction system based on frequency domain analysis and deep learning according to claim 1, characterized in that, The deformation difference between each monitoring point and the benchmark monitoring point is calculated using the following logic: Set the benchmark monitoring point as The remaining monitoring points are ,in For the first The remaining monitoring points, , This represents the total number of the remaining monitoring points; No. Other monitoring points In the The deformation within each basic time window is Benchmark monitoring points In the The deformation within each basic time window is Then the first Other monitoring points relative to benchmark monitoring points In the Deformation difference of each basic time window for: in, The index of the base time window, , The total number of base time windows.

5. The high slope deformation prediction system based on frequency domain analysis and deep learning according to claim 4, characterized in that, The sliding window is slid-trimmed according to preset parameters to generate multiple overlapping sliding windows. The specific logic is as follows: The sliding window parameters include the sliding window length and the sliding step size; Wherein, the length of the sliding window is The sliding step size is , and All are positive integers, and ; Multi-source time series data contains A basic time window, and satisfying ; According to the sliding step size Slide sequentially from left to right, and extract segments of length [length missing]. Continuous data segments form multiple overlapping sliding windows, denoted as... ; in For the first A sliding window, For the index of the sliding window, , Let be the total number of sliding windows, satisfying: in, To round down; No. A sliding window The range of the basic time window index is: Furthermore, the index range falls entirely within the base time window index interval. Inside.

6. The high slope deformation prediction system based on frequency domain analysis and deep learning according to claim 5, characterized in that, The effective frequency band components are obtained by performing time-frequency domain feature matching and filtering on the large deformation category frequency band components. The specific logic is as follows: Extract all frequency band components corresponding to the large deformation category within each sliding window, and the time series sequence of deformation difference values ​​of each monitoring point of the large deformation category within the sliding window arranged in the order of the basic time window index; For a time series sequence of frequency band components and deformation difference within a sliding window, calculate the changes of both within corresponding adjacent base time windows. The change of the frequency band components is: in, This represents the change in frequency band components within a sliding window over adjacent fundamental time windows. For the frequency band components in the first... The values ​​of a base time window, For the frequency band components in the first... The values ​​of a base time window; The change in the time series sequence of deformation difference is: in, This represents the change in the time series of deformation difference within a sliding window over adjacent base time windows. The time series sequence of deformation difference is in the 1st... The values ​​of a base time window, The time series sequence of deformation difference is in the 1st... The values ​​of a base time window; The signs of the slopes of the two changes are statistically analyzed, and the ratio of the number of slopes with the same sign to the total number of slopes is calculated. This ratio is the corresponding trend consistency coefficient, where the total number of slopes is... ; The mean of all trend consistency coefficients is used as an adaptive threshold to filter out frequency band components with trend consistency coefficients greater than the threshold, which are the effective frequency band components.

7. The high slope deformation prediction system based on frequency domain analysis and deep learning according to claim 5, characterized in that, The total deformation is allocated based on the difference in deformation between each monitoring point and the benchmark monitoring point within the large deformation category, thus obtaining the predicted deformation value for each monitoring point in the next basic time window. The specific logic is as follows: Set the predicted total deformation value for the large deformation category in the most recent sliding window to be... All monitoring points within the large deformation category were in the [number]th [year]. The sum of the deformation differences of each basic time window is ; Calculate the first The monitoring point at the 1st Allocation ratio of each basic time window The specific formula is as follows: Calculate the first The predicted deformation value for each monitoring point within the next basic time window is calculated using the following formula: in, For the first The predicted value of deformation at each monitoring point within a basic time window.

8. A method for predicting high slope deformation based on frequency domain analysis and deep learning, wherein the method is executed by a high slope deformation prediction system based on frequency domain analysis and deep learning as described in any one of claims 1-7, characterized in that, The specific steps include: S1. Set up monitoring points at equal intervals along the slope inclination direction on the surface of the high slope, delineate the monitoring interval and divide the basic time window, take the monitoring point at the bottom of the slope toe as the reference monitoring point, collect the slope deformation and environmental data of the other monitoring points in each basic time window, and calculate the difference in deformation of each monitoring point relative to the reference monitoring point. S2. The slope deformation, environmental data, and deformation difference of the remaining monitoring points are spliced ​​into multi-source time series data, and the data is slid and cut according to the preset sliding window parameters to generate multiple overlapping sliding windows; S3. Cluster the monitoring points of the multi-source time series data in each sliding window to obtain small, medium and large deformation categories. Remove the data in the small and medium deformation categories. Perform frequency domain analysis on the mean of the data in the large deformation category and output the corresponding frequency band components and residual components. At the same time, perform time-frequency domain feature matching and filtering on the frequency band components of the large deformation category to obtain the effective frequency band components. S4. Input the effective frequency band components and residual components of the large deformation category corresponding to each sliding window into the trained deep learning model to predict the deformation amount of the large deformation category in the next basic time window. S5. Sum the predicted values ​​of slope deformation under the most recent sliding window for the large deformation category to obtain the total predicted value of deformation. Allocate the total deformation according to the difference between the deformation of each monitoring point and the benchmark monitoring point within the large deformation category to obtain the predicted value of deformation of each monitoring point in the next basic time window, and select the maximum value as the overall deformation prediction result of the high slope.

Images