A machine learning model comparison and screening method for slope displacement sequence

Abstract

The present application relates to the technical field of slope monitoring, and proposes a machine learning model comparison and screening method for slope displacement sequence, which comprises: reorganizing the original monitoring data of slope displacement in time and space section to obtain deep time series profile; rate turning anchoring is performed on the deep time series profile to obtain multi-stage evolution sequence; taking the convergence starting critical point as the segmentation boundary, the multi-stage evolution sequence is subjected to redundant steady-state stripping to obtain the effective deformation interval; the effective deformation interval is subjected to global magnitude normalization and sliding window cross-state construction to obtain cross-working condition time sequence sample pairs; based on the cross-working condition time sequence sample, heterogeneous model instances of the historical training set are enumerated; cross-validation deduction is performed on the heterogeneous model instances to obtain progressive folding residual archives; cross-folding error statistics and multi-fold generalization optimization are performed on the progressive folding residual archives to obtain the optimal generalization model of slope displacement; the present application can improve the efficiency of machine learning model comparison and screening for slope displacement sequence.

Description

Technical Field

[0001] This invention relates to the field of slope monitoring technology, and in particular to a method for comparing and screening machine learning models of slope displacement sequences. Background Technology

[0002] In the field of intelligent monitoring and deformation prediction for geotechnical slope engineering, deep displacement monitoring sequences of slopes generally exhibit a three-stage nonlinear evolution pattern of rapid growth, deceleration growth, and convergence stabilization. The displacement time series possesses both spatiotemporal distribution characteristics and rate inflection features. Existing conventional processing methods merely list the original displacement monitoring data in a simple time series manner, without carrying out spatiotemporal cross-sectional reconstruction and deep time series profile reconstruction. They also lack precise means for identifying and anchoring inflection points of deformation rates, making it impossible to accurately define the boundaries of multi-stage slope displacement evolution, accurately capture the starting critical position of the convergence stabilization stage, and effectively remove steady-state redundant data without deformation reference value. At the same time, existing methods lack global order-of-magnitude normalization and sliding window cross-state sampling processes for the effective deformation interval. The constructed time series samples cannot cover the deformation characteristics of slopes under multiple working conditions, and the data preprocessing accuracy is insufficient, with weak sample representation capabilities, which seriously restricts the training and evaluation effects of subsequent machine learning models.

[0003] Current methods for selecting machine learning models for slope displacement prediction generally employ a static evaluation approach that involves simply and randomly dividing the training and test sets. This approach fails to systematically enumerate heterogeneous models by incorporating time-series data dimensions and sample size characteristics, and it lacks rolling cross-validation and progressive folding residual archiving mechanisms. Traditional selection schemes rely solely on single training and testing errors for model optimization, lacking statistical and quantitative evaluation dimensions for errors under multiple working conditions and fluctuations. This makes it difficult to avoid the randomness of evaluation caused by small-sample time-series data, easily leading to problems such as model overfitting, misjudgment, and incomplete assessment of generalization performance at different deformation stages. Furthermore, the model selection process lacks a standardized system, resulting in low reliability and poor adaptability of the selection results, making it impossible to accurately select the optimal generalization model suitable for multi-stage slope displacement evolution. Therefore, improving the standardization, evaluation accuracy, and engineering generalization adaptability of machine learning models for slope displacement sequence comparison and selection has become an urgent problem to be solved. Summary of the Invention

[0004] This invention provides a method for comparing and selecting machine learning models of slope displacement sequences to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, this invention provides a method for comparing and selecting machine learning models of slope displacement sequences, comprising: P1. Spatiotemporal cross-sectional reconstruction of the original monitoring data of the slope displacement is performed to obtain the deep temporal profile of the slope displacement. P2. Rate transition anchoring is performed on the deep time series profile to obtain a multi-stage evolution sequence of the deep time series profile, which includes a rapid growth stage, a deceleration growth stage, and a convergence and stabilization stage. P3. Using the convergence start critical point of the convergence stable stage as the dividing boundary, perform redundant steady-state stripping on the multi-stage evolution sequence to obtain the effective deformation range of the multi-stage evolution sequence. P4. Perform global order-of-magnitude regularization on the effective deformation interval, and perform sliding window cross-state sampling on the regularized effective deformation interval to obtain cross-working-condition time series sample pairs of the effective deformation interval. P5. Split the cross-condition time series sample pair into a historical training set and a future test set, and enumerate the heterogeneous model instances of the historical training set according to the dimensionality characteristics of the historical training set. P6. Based on the historical training set and the future test set, cross-validation is performed on the heterogeneous model instances to obtain the progressive folded residual files of the heterogeneous model instances. P7. Perform cross-fold error statistics on the progressively folded residual file, and perform multi-fold generalization selection on the statistically obtained error to obtain the optimal generalization model of the slope displacement.

[0006] In a preferred embodiment, the process of reconstructing the spatiotemporal profiles of the original monitoring data of the slope displacement to obtain a deep time-series profile of the slope displacement includes: The original monitoring data of slope displacement is collected, and the original monitoring data is deconstructed into cross-sections to obtain the depth displacement cross-section queue of the original monitoring data. Peak spectral state analysis is performed on the depth displacement section queue to obtain the peak section spectrum of the depth displacement section queue; The displacement peaks in the peak cross-sectional spectrum are time-series evolved and grouped to obtain the peak displacement evolution series of the peak cross-sectional spectrum; The evolution profile of the peak displacement evolution series is reconstructed to obtain the deep time-series profile of the slope displacement.

[0007] In a preferred embodiment, the rate transition anchoring of the deep time series profile yields a multi-stage evolution sequence of the deep time series profile, the multi-stage evolution sequence including a rapid growth stage, a deceleration growth stage, and a convergence and stabilization stage, including: The deep time series profile is subjected to phase-by-phase differential mapping to obtain the instantaneous rate profile of the deep time series profile; Rate inflection point mining is performed on the instantaneous rate profile to obtain the rate inflection anchor point of the instantaneous rate profile; Based on the rate inflection anchor point, the working condition boundary is anchored for the deep time series profile to obtain the stage segmentation boundary of the deep time series profile. Based on the stage segmentation boundary, the deep time series profile is clustered into evolution intervals to obtain a multi-stage evolution sequence of the deep time series profile, wherein the multi-stage evolution sequence includes a rapid growth stage, a deceleration growth stage, and a convergence and stabilization stage.

[0008] In a preferred embodiment, the step of mining the velocity inflection point of the instantaneous velocity profile to obtain the velocity inflection anchor point of the instantaneous velocity profile includes: A sign scan of the instantaneous velocity profile is performed to obtain the positive and negative velocity boundary points of the instantaneous velocity profile; Extract the forward velocity, current velocity, and backward velocity of the neighborhood at the positive-negative velocity boundary point from the instantaneous velocity profile; The deceleration gradient is analyzed by analyzing the forward velocity, current velocity, and backward velocity of the neighborhood to obtain the deceleration inflection point depth value of the positive and negative velocity boundary. Based on the deceleration inflection point depth value, the inflection point confidence of the instantaneous velocity profile is verified to obtain the candidate turning point anchor points of the instantaneous velocity profile. Redundant anchor points among the candidate turning anchor points are filtered for local extrema to obtain the velocity turning anchor points of the instantaneous velocity profile.

[0009] In a preferred embodiment, the step of using the convergence initiation critical point of the convergence stabilization phase as the dividing boundary to perform redundant steady-state stripping on the multi-stage evolution sequence to obtain the effective deformation range of the multi-stage evolution sequence includes: Steady-state feature quantization is performed on the multi-stage evolution sequence to obtain the steady-state fluctuation amplitude of the multi-stage evolution sequence; Based on the steady-state fluctuation amplitude, the critical point of the multi-stage evolution sequence is captured to obtain the convergence starting critical point of the multi-stage evolution sequence. Using the convergence initiation critical point as the dividing boundary, the multi-stage evolution sequence is divided into deformable segments to obtain the middle active segment of the multi-stage evolution sequence; The effective deformation range of the multi-stage evolution sequence is obtained by stripping the residual segments of the steady-state residual segments in the intermediate active segments.

[0010] In a preferred embodiment, the step of performing global-scale normalization on the effective deformation range and then performing sliding-window cross-state sampling on the normalized effective deformation range to obtain cross-condition time series sample pairs of the effective deformation range includes: Global extreme value selection is performed on the effective deformation interval to obtain the global displacement extreme value of the effective deformation interval; Based on the global displacement extreme value, the effective deformation interval is subjected to extreme value calibration compression to obtain the order-of-magnitude regularization sequence of the effective deformation interval; The order-of-magnitude regular sequence is sliced ​​across states using a sliding window to obtain a cross-condition window sequence of the order-of-magnitude regular sequence; The time-series segments within the window of the cross-working-condition window sequence are sampled under working-condition linkage to obtain the cross-working-condition time-series sample pairs of the effective deformation interval.

[0011] In a preferred embodiment, the step of splitting the cross-condition time-series sample pairs into a historical training set and a future test set, and enumerating heterogeneous model instances of the historical training set based on the dimensionality characteristics of the historical training set, includes: The cross-operating condition time series sample pairs are subjected to time series segmentation and labeling to obtain the training and testing segmentation labels of the cross-operating condition time series sample pairs; Based on the training and testing segmentation markers, the cross-working condition time series sample pairs are partitioned and truncated to obtain the historical training set and future test set of the cross-working condition time series sample pairs. The historical training set is subjected to feature scale quantization to obtain the input dimension value and sample size value of the historical training set; Based on the input dimension value and the sample size value, multi-class model enumeration is performed on the historical training set to obtain heterogeneous model instances of the historical training set.

[0012] In a preferred embodiment, the step of performing cross-validation on the heterogeneous model instances based on the historical training set and the future test set to obtain the progressively folded residual archive of the heterogeneous model instances includes: Based on the historical training set, rolling cross-validation is performed on the heterogeneous model instances to obtain the validation residual records of the heterogeneous model instances; The verification residual records are concatenated in a stepwise manner to obtain the rolling residual vector of the heterogeneous model instance; Based on the future test set, static hold-out validation is performed on the heterogeneous model instance to obtain the hold-out validation residual of the heterogeneous model instance; The rolling residual vector and the reserved verification residual are bidirectionally coupled and archived to obtain the progressive folding residual archive of the heterogeneous model instance.

[0013] In a preferred embodiment, the step of performing rolling cross-validation on the heterogeneous model instances based on the historical training set to obtain the validation residual records of the heterogeneous model instances includes: The historical training set is progressively windowed to obtain a progressive training group and a progressive verification group for the historical training set. Based on the progressive training group, the heterogeneous model instance is subjected to iterative training in stages to obtain the learned model instance of the heterogeneous model instance on the progressive training group. Based on the progressive verification group, rolling extrapolation prediction is performed on the learned model instance, and the deviation of the predicted output value is compared to obtain the rolling step residual of the progressive verification group. The residuals of the rolling step are quantitatively recorded to obtain the verification residual records of the heterogeneous model instance.

[0014] In a preferred embodiment, the step of performing cross-fold error statistics on the progressively folded residual file and performing multi-fold generalization optimization on the statistically obtained error to obtain the optimal generalization model of the slope displacement includes: The progressively folded residual archive is reconstructed using error spectralization to obtain the multi-fold residual spectrum of the heterogeneous model instance; The average error of the heterogeneous model instance is obtained by performing cross-fold mean difference extraction on the multi-fold residual spectrum. The fluctuation steady-state quantization of the multi-fold residual spectrum is performed to obtain the fluctuation stability value of the heterogeneous model instance; Based on the fluctuation stability value and the average error, a two-dimensional game selection is performed on the heterogeneous model instance to obtain the optimal generalized model of the slope displacement.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention constructs a deep time-series profile by reconstructing the original slope displacement monitoring data in a spatiotemporal manner. Combined with rate inflection anchoring, it accurately divides the multi-stage displacement evolution sequence and completes the removal of redundant steady-state data based on the convergence initiation critical point, accurately locking the true effective deformation range of the slope. At the same time, it implements global order regularization and sliding window cross-state sampling on the effective deformation range, which can unify the dimensional distribution of displacement data, fully preserve the temporal evolution characteristics of the slope under different deformation conditions, realize the deep-level structured analysis of monitoring data and the construction of standardized samples, and ensure the integrity and accuracy of the time-series samples in representing the deformation trend.

[0016] 2. This invention rationally splits the historical training set and the future test set according to temporal characteristics, completes full coverage enumeration of heterogeneous model instances based on data dimensionality characteristics, establishes a progressively folded residual archive through cross-validation deduction, and conducts two-dimensional generalization optimization by combining cross-fold error statistics and error fluctuation stability quantification. It can realize multi-fold progressive performance quantification evaluation of machine learning models, complete model optimization from the dual dimensions of error amplitude and operational stability, standardize the entire process logic of model screening, improve the quantification degree of model screening, operational efficiency and adaptation stability of optimization results, and can stably output the optimal generalization model that adapts to the temporal evolution law of slope displacement. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a method for comparing and selecting machine learning models of slope displacement sequences according to an embodiment of the present invention. The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] This application provides a method for comparing and selecting machine learning models of slope displacement sequences. The execution subject of this method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the method for comparing and selecting machine learning models of slope displacement sequences can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cluster of cloud servers. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0020] Reference Figure 1 The diagram shown is a flowchart illustrating a method for comparing and selecting machine learning models of slope displacement sequences according to an embodiment of the present invention. In this embodiment, the method for comparing and selecting machine learning models of slope displacement sequences includes: P1. Spatiotemporal cross-sectional reconstruction of the original monitoring data of the slope displacement is performed to obtain the deep temporal profile of the slope displacement. In this embodiment of the invention, the process of reconstructing the spatiotemporal profiles of the original monitoring data of the slope displacement to obtain a deep time-series profile of the slope displacement includes: The original monitoring data of slope displacement is collected, and the original monitoring data is deconstructed into cross-sections to obtain the depth displacement cross-section queue of the original monitoring data. Peak spectral state analysis is performed on the depth displacement section queue to obtain the peak section spectrum of the depth displacement section queue; The displacement peaks in the peak cross-sectional spectrum are time-series evolved and grouped to obtain the peak displacement evolution series of the peak cross-sectional spectrum; The evolution profile of the peak displacement evolution series is reconstructed to obtain the deep time-series profile of the slope displacement.

[0021] All original slope displacement monitoring data generated by measuring points at different depths on the slope during a continuous monitoring period were collected. The complete and continuous original monitoring data was then broken down into segments according to the spatial depth dimension and the time monitoring dimension. The overall data was decomposed into discrete data units that were independent of each other and corresponded to fixed depth measuring points and fixed monitoring times. All discrete data units were then arranged in a fixed order from shallow to deep slope depth and from first to last monitoring time, forming a regular and continuous depth displacement cross-section queue.

[0022] One by one, all displacement values ​​contained in each independent section within the deep displacement section queue are traversed and retrieved point by point. The corresponding data information of the displacement value reaching the high point of each section is locked. All the retrieved displacement high point information is collected and organized in the order of the sections, and the distribution pattern and variation law of displacement peaks in each section are completely sorted out. The scattered peak information is integrated into a peak section spectrum with overall distribution characteristics.

[0023] Strictly following the temporal logic of monitoring time progression, the peak displacement information of all depth sections corresponding to each monitoring moment is extracted from the peak cross-sectional spectrum in chronological order. The peak displacements of different depth sections at the same monitoring moment are associated, integrated, and classified. Then, the classified peak combinations are continuously connected and arranged in chronological order, and the peak correlation data of all time-series nodes are continuously integrated to form a peak displacement evolution series that can fully show the dynamic changes of the peak displacement over time.

[0024] Based on the inherent temporal variation pattern of the peak displacement evolution series, and simultaneously integrating the cross-sectional distribution correlation characteristics of the slope spatial depth dimension, the evolution law of the displacement peak in the time dimension and the displacement distribution law in the spatial depth dimension are integrated and connected. The discrete peak evolution information is extended and built into a continuous and complete spatiotemporal integrated evolution structure, which fully restores the overall continuous change pattern of slope displacement in both spatial depth and temporal processes, and builds a deep temporal profile of slope displacement that can comprehensively characterize the spatiotemporal deformation characteristics of the slope.

[0025] The beneficial effects of this process are that it enables the layer-by-layer analysis and reconstruction of original slope displacement monitoring data from scattered raw sampling data to spatiotemporally structured data. It completes the fine sorting of data from both spatial cross-sectional distribution and temporal evolution dimensions, accurately extracts and completely preserves the core characteristics of the peak displacement at different depths of the slope over time, and filters out messy and invalid information in the original monitoring data that has no actual deformation characterization value. It provides standardized basic time series data with regular structure, complete features and dimension matching for subsequent deformation rate turning point anchoring and accurate segmentation of multi-stage evolution sequences, and adapts to the data preprocessing requirements of multi-stage nonlinear displacement evolution law of slope.

[0026] P2. Rate transition anchoring is performed on the deep time series profile to obtain a multi-stage evolution sequence of the deep time series profile, which includes a rapid growth stage, a deceleration growth stage, and a convergence and stabilization stage. In this embodiment of the invention, the rate transition anchoring of the deep time series profile yields a multi-stage evolution sequence of the deep time series profile. The multi-stage evolution sequence includes a rapid growth stage, a deceleration growth stage, and a convergence and stabilization stage, comprising: The deep time series profile is subjected to phase-by-phase differential mapping to obtain the instantaneous rate profile of the deep time series profile; Rate inflection point mining is performed on the instantaneous rate profile to obtain the rate inflection anchor point of the instantaneous rate profile; Based on the rate inflection anchor point, the working condition boundary is anchored for the deep time series profile to obtain the stage segmentation boundary of the deep time series profile. Based on the stage segmentation boundary, the deep time series profile is clustered into evolution intervals to obtain a multi-stage evolution sequence of the deep time series profile, wherein the multi-stage evolution sequence includes a rapid growth stage, a deceleration growth stage, and a convergence and stabilization stage.

[0027] The step of mining the velocity inflection points of the instantaneous velocity profile to obtain the velocity inflection anchor points of the instantaneous velocity profile includes: A sign scan of the instantaneous velocity profile is performed to obtain the positive and negative velocity boundary points of the instantaneous velocity profile; Extract the forward velocity, current velocity, and backward velocity of the neighborhood at the positive-negative velocity boundary point from the instantaneous velocity profile; The deceleration gradient is analyzed by analyzing the forward velocity, current velocity, and backward velocity of the neighborhood to obtain the deceleration inflection point depth value of the positive and negative velocity boundary. Based on the deceleration inflection point depth value, the inflection point confidence of the instantaneous velocity profile is verified to obtain the candidate turning point anchor points of the instantaneous velocity profile. Redundant anchor points among the candidate turning anchor points are filtered for local extrema to obtain the velocity turning anchor points of the instantaneous velocity profile.

[0028] Following the complete time sequence arrangement of the deep time-series profile, the deep displacement data of the slope corresponding to two consecutive monitoring times are extracted one by one. The entire time-series profile is continuously traversed, and the dynamic change of displacement between adjacent times is continuously calculated. The time-series information that originally only represented the magnitude of displacement is transformed into time-series information that represents the speed of slope deformation. The continuous change characteristics of deformation rate of each time-series node are completely preserved throughout the process. The rate change information of all time-series nodes is integrated to form an instantaneous rate profile that can reflect the speed of slope deformation at each moment throughout the process.

[0029] Following the temporal extension of the instantaneous rate profile, a comprehensive investigation was conducted node by node to track the entire evolution of the slope deformation rate from continuous increase to gradual decrease. The key temporal positions where the deformation growth trend underwent a fundamental change were accurately captured. All node positions with trend reversal characteristics were identified one by one. The continuous evolution of the rate changes at each node was systematically sorted out. The comprehensive excavation of trend change nodes within the instantaneous rate profile was completed, and rate reversal anchor points that can accurately identify the transition positions of the slope deformation stages were located.

[0030] Based on the precise temporal positions of all determined rate inflection anchor points on the time axis, these points are mapped onto the overall temporal structure of the deep temporal profile. Each rate inflection anchor point is directly used as a boundary marker for the switching between different construction conditions and deformation states of the slope. All boundary markers are sequentially connected along the temporal progression direction to delineate a clearly defined boundary range that conforms to the actual deformation condition switching pattern of the slope, thus forming a complete and closed stage segmentation boundary of the deep temporal profile.

[0031] Based on the temporal boundary defined by the stage segmentation boundary, all temporal nodes in the deep temporal profile that are temporally adjacent and have consistent deformation trends are collected and integrated. According to the actual evolution law of rapid deformation in the early stage of slope excavation and unloading, deceleration deformation after anchor cable prestressing intervention, and later rock mass stabilization and convergence, the entire temporal profile is clustered into continuous evolution intervals, dividing it into three independent and well-defined continuous temporal intervals. This forms a multi-stage evolution sequence of the deep temporal profile that includes a rapid growth stage, a deceleration growth stage, and a convergence and stabilization stage, and fully matches the actual deformation characteristics of the project.

[0032] The rate attribute is scanned across the entire domain node by node along the time arrangement sequence of the instantaneous rate profile. The trend of deformation rate change at each time node is determined one by one. The switching position of the rate trend of adjacent time nodes is continuously compared. The time nodes where the deformation rate growth attribute and decay attribute change are accurately located. The key node positions of all attribute switching are collected and uniformly organized to form the positive and negative rate boundary points of the instantaneous rate profile.

[0033] The precise temporal position of each positive and negative rate boundary point is locked. In the temporal arrangement structure of the instantaneous rate profile, the rate information corresponding to the monitoring time before the boundary point is selected as the forward rate of the neighborhood, the rate information corresponding to the monitoring time of the boundary point itself is selected as the current rate of the neighborhood, and the rate information corresponding to the monitoring time after the boundary point is selected as the backward rate of the neighborhood. For each positive and negative rate boundary point, the fixed-point complete extraction and retention of the three types of rate information of the adjacent time before and after the boundary point and the boundary point itself is completed.

[0034] The extracted forward velocity, current velocity, and backward velocity of the neighborhood are correlated and compared in chronological order. The progressive changes in velocity values ​​at three consecutive temporal positions are analyzed layer by layer, meticulously examining the decay evolution logic of the velocity smoothly transitioning from a high to a low state. The depth of the velocity transition at each positive-negative velocity boundary point is quantified, and a deceleration inflection point depth value is generated for each positive-negative velocity boundary point, characterizing the significance of the transition at each node. The formula for calculating the deceleration inflection point depth value is as follows: ; in, This indicates the depth value of the deceleration inflection point. This represents the current rate of the neighborhood at the positive-negative rate boundary point. This represents the forward velocity of the neighborhood at the monitoring time preceding the positive / negative velocity boundary point. This represents the backward velocity in the neighborhood at the monitoring time following the positive-negative velocity boundary point. It represents a very small positive number.

[0035] The forward velocity of the neighborhood at the monitoring moment before the positive-negative velocity boundary, the current velocity of the neighborhood at the boundary, and the backward velocity of the neighborhood at the monitoring moment after the boundary are extracted from the instantaneous velocity profile. A very small positive number is set as the calculation compensation term. By comparing the change states of the three types of velocities, the severity of the decay of the velocity at the positive-negative velocity boundary from continuous growth to gradual decline is quantified. This characterizes the significance of the boundary as a deceleration inflection point and the reliability of the actual deformation boundary. At the same time, the compensation term can ensure that the calculation process is stable and without abnormal interruption.

[0036] When the changes in the forward velocity and the current velocity of the neighborhood are more significant, and the changes in the current velocity and the backward velocity of the neighborhood are more gradual, the obtained deceleration inflection point depth value is closer to its theoretical upper limit, indicating that the velocity inflection feature at this position is more prominent and has the attribute of becoming the boundary point of the actual deformation stage. When the differences in the changes of the three types of velocities are not obvious, the obtained deceleration inflection point depth value is in the insignificant range, indicating that the position does not have obvious deceleration inflection features.

[0037] Using the inherent law of the three-stage evolution of slope rock cut displacement as the judgment criterion, and combining the deceleration inflection point depth value corresponding to each time node with the actual engineering deformation transition characteristics, the degree of fit between the inflection point characteristics of each candidate node and the actual slope deformation stage switching law is verified one by one. Invalid nodes with inflection point characteristics that do not have engineering practical significance or false transitions are eliminated. Only valid time nodes that can truly reflect the slope rate transition changes are retained. The candidate transition anchor points of the instantaneous rate profile are obtained by summarizing all valid node sets.

[0038] A full-domain feature screening is performed on all nodes within the candidate turning point anchor point set. Redundant anchor points with repeated turning features in adjacent time series positions are removed. Key nodes with unique trend turning point identification functions within each local time series range are selected and retained. Invalid and redundant nodes that cannot independently define the boundary of the deformation stage are eliminated. All effective turning points are streamlined and sorted to form rate turning point anchor points with orderly arrangement and accurate positioning of instantaneous rate profile.

[0039] The beneficial effects of this process are that it relies on deep time-series profiles to complete the conversion of time-by-time rate information, and combines a multi-level screening mechanism of global symbol scanning, neighborhood rate correlation comparison, attenuation hierarchy analysis, and confidence verification to accurately identify key rate turning points in the entire process of slope excavation, unloading, anchor cable tensioning, and convergence stabilization. It strictly conforms to the inherent mechanism of multi-stage nonlinear displacement evolution of rock cutting slopes and automatically divides deformation intervals, completely avoiding the subjective arbitrariness and boundary offset problems caused by manual segmentation. It accurately delineates the time-series boundaries of each deformation stage, providing a time-series segmentation basis that conforms to engineering reality for subsequent steady-state redundancy stripping, effective deformation interval extraction, and cross-condition sample construction. At the same time, it ensures that the subsequent machine learning model screening can truly assess the model's ability to capture the slope deceleration and convergence deformation trend.

[0040] P3. Using the convergence start critical point of the convergence stable stage as the dividing boundary, perform redundant steady-state stripping on the multi-stage evolution sequence to obtain the effective deformation range of the multi-stage evolution sequence. In this embodiment of the invention, the step of using the convergence initiation critical point of the convergence stabilization phase as the dividing boundary to perform redundant steady-state stripping on the multi-stage evolution sequence to obtain the effective deformation range of the multi-stage evolution sequence includes: Steady-state feature quantization is performed on the multi-stage evolution sequence to obtain the steady-state fluctuation amplitude of the multi-stage evolution sequence; Based on the steady-state fluctuation amplitude, the critical point of the multi-stage evolution sequence is captured to obtain the convergence starting critical point of the multi-stage evolution sequence. Using the convergence initiation critical point as the dividing boundary, the multi-stage evolution sequence is divided into deformable segments to obtain the middle active segment of the multi-stage evolution sequence; The effective deformation range of the multi-stage evolution sequence is obtained by stripping the residual segments of the steady-state residual segments in the intermediate active segments.

[0041] Following the temporal arrangement of the multi-stage evolution sequence, the complete time series interval of the convergence and stabilization stage is first located. The displacement value corresponding to each time series node in the interval is read point by point. The displacement change state of two adjacent time series nodes is continuously compared. The fluctuation amplitude of each displacement change is recorded without interruption. All time series nodes of the convergence and stabilization stage are fully covered, without missing any displacement fluctuation details. All recorded displacement fluctuation data are collected and organized. The overall level and distribution characteristics of fluctuation changes are systematically sorted out and integrated to form a steady-state fluctuation amplitude that can accurately represent the inherent level of displacement fluctuation in the convergence and stabilization stage.

[0042] Following the temporal extension of the multi-stage evolution sequence, starting from the end of the deceleration and growth phase, the sequence is traversed backwards node by node towards the beginning of the convergence and stabilization phase. The degree of matching between the displacement change amplitude and the steady-state fluctuation amplitude of each time node is compared point by point. The process of the displacement change amplitude gradually falling back and entering the horizontal range corresponding to the steady-state fluctuation amplitude for the first time is continuously tracked. The position of the first time node where the displacement change amplitude stably enters the horizontal range is accurately locked. This node is the dividing point where the slope deformation changes from the deceleration and growth state to the convergence and stabilization state, forming the convergence starting critical point of the multi-stage evolution sequence.

[0043] Using the position of the convergence initiation critical point on the time axis of the multi-stage evolution sequence as a fixed segmentation marker, along the time sequence arrangement direction, complete displacement data segments with all time sequence positions before the critical point are extracted. This segment contains all displacement data of the rapid growth stage and the deceleration growth stage in the multi-stage evolution sequence, but does not contain any time sequence nodes of the convergence and stabilization stage, forming the middle active segment of the multi-stage evolution sequence that completely covers the entire process of active deformation of the slope.

[0044] Following the time progression of the active intermediate segment, the displacement change status of each time-series node within the segment is examined node by node. The matching between the displacement change amplitude of each node and the steady-state fluctuation amplitude is continuously compared. The time-series nodes within the active intermediate segment whose displacement change amplitude is still within the steady-state fluctuation range are accurately located. These nodes belong to the steady-state residual data that has entered the small fluctuation state in advance. All steady-state residual data nodes are completely removed from the active intermediate segment, and only the time-series nodes whose displacement change amplitude always exceeds the steady-state fluctuation range are retained, forming an effective deformation interval that can completely characterize the entire process of the real active deformation of the slope.

[0045] The beneficial effects of this process are that it establishes a steady-state judgment benchmark by fully quantifying the displacement fluctuation characteristics of the convergence and stabilization stage, accurately captures the critical points where slope deformation enters a stable state by relying on reverse traversal and fluctuation amplitude matching, and completely removes redundant data in the convergence and stabilization stage by using the convergence initiation critical point as the boundary. At the same time, it further removes the steady-state residual data in the intermediate active segment, completely eliminates steady-state redundant information with no deformation characterization value, and fully retains all effective deformation characteristics in the rapid growth and deceleration growth stages of the slope. This provides a standardized deformation data foundation without redundancy or interference for the construction of subsequent time series samples, ensuring that the training and evaluation of machine learning models are carried out only around the real active deformation laws of the slope, avoiding interference from steady-state data on the evaluation of the model's generalization ability, and improving the reliability and engineering adaptability of subsequent model selection results.

[0046] P4. Perform global order-of-magnitude regularization on the effective deformation interval, and perform sliding window cross-state sampling on the regularized effective deformation interval to obtain cross-working-condition time series sample pairs of the effective deformation interval. In this embodiment of the invention, the step of performing global-scale normalization on the effective deformation range and then performing sliding-window cross-state sampling on the normalized effective deformation range to obtain cross-condition time series sample pairs of the effective deformation range includes: Global extreme value selection is performed on the effective deformation interval to obtain the global displacement extreme value of the effective deformation interval; Based on the global displacement extreme value, the effective deformation interval is subjected to extreme value calibration compression to obtain the order-of-magnitude regularization sequence of the effective deformation interval; The order-of-magnitude regular sequence is sliced ​​across states using a sliding window to obtain a cross-condition window sequence of the order-of-magnitude regular sequence; The time-series segments within the window of the cross-working-condition window sequence are sampled under working-condition linkage to obtain the cross-working-condition time-series sample pairs of the effective deformation interval.

[0047] The displacement values ​​corresponding to all time-series monitoring nodes within the effective deformation range are fully traversed. The actual value of each displacement data point is checked sequentially along the time sequence. All monitoring points are fully covered throughout the entire deformation process, including both rapid and decelerated growth. The extreme characteristic points of displacement values ​​within the entire range are completely locked. The extreme displacement values ​​of the entire range are integrated to form the global displacement extreme values ​​of the effective deformation range.

[0048] Using the extreme values ​​of displacement across the entire domain as a unified benchmark for magnitude calibration, the displacement data of each time-series node within the effective deformation range are processed using a unified magnitude scale. Throughout the process, the temporal trend and relative evolution of displacement changes between time-series nodes remain unchanged. The inherent magnitude differences in the original displacement data are smoothed out, and all displacement data are regularized into a unified magnitude expression system. After regularization, a magnitude regularization sequence of the effective deformation range is formed.

[0049] A fixed and uniform continuous time sequence is selected. Starting from the first time sequence node of the order-regular sequence, the sequence slides smoothly backward along the natural extension direction of time node by node. Each time the slide is completed, a continuous time sequence segment of fixed length is selected. The sliding selection process deliberately crosses the boundary positions of different deformation conditions such as slope excavation, unloading, anchor cable tensioning, deceleration transition, etc., so that each selected window unit contains the time sequence evolution information across conditions. All the window units obtained by the sliding selection are collected and arranged in sequence to form a cross-condition window sequence of the order-regular sequence.

[0050] The time sequence segments within each window of the cross-working condition window sequence are extracted and decomposed to correspond to different deformation working conditions. The time sequence segments of multiple working conditions connected successively within the same window are associated and bound according to the deformation evolution logic. The input time sequence segment and the subsequent evolution segment are paired and combined according to the corresponding relationship between the time sequence segments. The rate decay and trend turning point characteristics of cross-working conditions within the window are preserved. After all windows are linked and sampled, they are regularized to form cross-working condition time sequence sample pairs with effective deformation intervals.

[0051] The beneficial effects are that this step uses a global unified extreme value calibration method to complete the magnitude regularization instead of a segmented independent regularization mode, which fully preserves the relative correlation characteristics of displacement magnitudes between different deformation conditions of the slope. The sliding window cross-state slicing method realizes uninterrupted sample extraction of the condition boundary, so that the constructed time series samples naturally contain the deformation rate evolution law of the transition stage of multiple conditions, eliminate the interference of the difference in the dimensions of the original displacement data on the modeling, and automatically generate sample data that fits the real deformation evolution law of the slope without the need for manual annotation of condition switching points. This provides standardized and highly representative time series samples to support subsequent dataset splitting and heterogeneous model enumeration training.

[0052] P5. Split the cross-condition time series sample pair into a historical training set and a future test set, and enumerate the heterogeneous model instances of the historical training set according to the dimensionality characteristics of the historical training set. In this embodiment of the invention, the step of splitting the cross-condition time-series sample pairs into a historical training set and a future test set, and enumerating heterogeneous model instances of the historical training set based on the dimensionality characteristics of the historical training set, includes: The cross-operating condition time series sample pairs are subjected to time series segmentation and labeling to obtain the training and testing segmentation labels of the cross-operating condition time series sample pairs; Based on the training and testing segmentation markers, the cross-working condition time series sample pairs are partitioned and truncated to obtain the historical training set and future test set of the cross-working condition time series sample pairs. The historical training set is subjected to feature scale quantization to obtain the input dimension value and sample size value of the historical training set; Based on the input dimension value and the sample size value, multi-class model enumeration is performed on the historical training set to obtain heterogeneous model instances of the historical training set.

[0053] Following the inherent temporal order of cross-condition time series sample pairs, and using the temporal position from the final stage of slope deformation deceleration to the convergence transition as the dividing benchmark, fixed segmentation points are defined in the overall time series sample sequence. The time series sample segments before the segmentation points are uniformly labeled as training attributes specifically for model feature learning, while the time series sample segments after the segmentation points are uniformly labeled as test attributes specifically for testing the model's generalization ability. The attributes of all sample units are labeled one by one in strict accordance with the natural evolution of the time series, forming training and testing segmentation marks for cross-condition time series sample pairs with clear boundaries and unique attribute identification.

[0054] According to the time-series segmentation boundary defined by the training and testing segmentation marks, from the starting time-series position of the cross-condition time-series sample pair to the position of the segmentation mark, all time-series sample units labeled with training attributes are completely extracted and regularized to form a sample set encompassing the deformation characteristics of the slope under all conditions, including rapid growth and deceleration. Then, from the position of the segmentation mark towards the end of the time series, all time-series sample units labeled with testing attributes are completely extracted and independently collected to form a sample set corresponding to the deformation characteristics of the slope from the end of deceleration to convergence. The two types of sample sets are independent of each other and have complete time-series connection, thus obtaining the historical training set and future test set of the cross-condition time-series sample pair.

[0055] The inherent temporal structure hierarchy of each cross-condition time series sample pair in the historical training set is analyzed and sorted out one by one. The number of temporal feature levels contained in a single sample is counted and the feature structure scale is fully quantified. At the same time, the total number of all independent time series sample units contained in the historical training set is counted and sorted out. The internal feature structure scale and overall sample reserve scale of the historical training set are fully sorted out and solidified. The corresponding structural dimension information and sample quantity information are accurately extracted to obtain the input dimension value and sample capacity value of the historical training set.

[0056] By strictly matching the predetermined input dimension and sample size values ​​of the historical training set to the feature scale of the dataset, we select machine learning model categories with different operating mechanisms and network architectures suitable for the slope displacement time series prediction scenario. Based on the structural adaptation rules of each model, we match the data carrying capacity and feature input format of the current training set, and build complete structured instances for each selected model. This covers multiple technical architectures, including statistical extrapolation kernel methods, shallow neural networks, and deep time series networks. All the completed model instances are collected and organized to obtain heterogeneous model instances of the historical training set.

[0057] The beneficial effects are as follows: This method accurately defines the test set within the effective deformation range from the final stage of slope deceleration to the convergence transition, avoiding the problem of inflated model evaluation caused by using the fully converged steady-state segment as the test set. It can truly test the ability of each model to capture nonlinear deformation trends. At the same time, it performs full coverage enumeration of heterogeneous models based on the real dimensional characteristics and sample size of the historical training set, without the need for subjective limitation of model types. It fully covers the time series prediction model architecture with different mechanisms, providing a scientifically divided standard dataset and a complete model candidate library for subsequent cross-validation inference and multi-fold generalization selection. From the data partitioning source and the initial model selection stage, it ensures the rigor, objectivity and comprehensiveness of the comparison and selection of slope displacement machine learning models.

[0058] P6. Based on the historical training set and the future test set, cross-validation is performed on the heterogeneous model instances to obtain the progressive folded residual files of the heterogeneous model instances. In this embodiment of the invention, the step of performing cross-validation on the heterogeneous model instances based on the historical training set and the future test set to obtain the progressively folded residual archive of the heterogeneous model instances includes: Based on the historical training set, rolling cross-validation is performed on the heterogeneous model instances to obtain the validation residual records of the heterogeneous model instances; The verification residual records are concatenated in a stepwise manner to obtain the rolling residual vector of the heterogeneous model instance; Based on the future test set, static hold-out validation is performed on the heterogeneous model instance to obtain the hold-out validation residual of the heterogeneous model instance; The rolling residual vector and the reserved verification residual are bidirectionally coupled and archived to obtain the progressive folding residual archive of the heterogeneous model instance.

[0059] The step of performing rolling cross-validation on the heterogeneous model instances based on the historical training set to obtain the validation residual records of the heterogeneous model instances includes: The historical training set is progressively windowed to obtain a progressive training group and a progressive verification group for the historical training set. Based on the progressive training group, the heterogeneous model instance is subjected to iterative training in stages to obtain the learned model instance of the heterogeneous model instance on the progressive training group. Based on the progressive verification group, rolling extrapolation prediction is performed on the learned model instance, and the deviation of the predicted output value is compared to obtain the rolling step residual of the progressive verification group. The residuals of the rolling step are quantitatively recorded to obtain the verification residual records of the heterogeneous model instance.

[0060] The historical training set is strictly divided according to time sequence without random shuffling. A progressive window segmentation is carried out by gradually expanding the training sample range and fixing the number of verification samples. Multiple time sequence intervals are divided in turn to correspond to the initial, middle and final stages of rapid slope growth and deceleration. Each segmentation designates the earlier time sequence samples as the set used for model learning and the next fixed number of time sequence samples as the set used for model verification. The window boundary is pushed forward step by step to complete all levels of segmentation. All matching training and verification sets are organized and included to form the progressive training group and progressive verification group of the historical training set.

[0061] The time-series sample data containing multiple working conditions such as excavation, unloading, and anchor cable tensioning are sequentially taken from each progressive training group. The heterogeneous model instance is subjected to a batch-by-batch, layer-by-layer iterative training process, allowing the model to learn the displacement time-series change law of different deformation stages of the slope. As the sample range of the progressive training group expands, the model's ability to capture long-range time-series dependencies is continuously strengthened. After each progressive training group is completed, the internal feature mapping structure of the model is fixed, and the learned model instance of the heterogeneous model instance that is adapted to the corresponding deformation stage features is generated one by one on the progressive training group.

[0062] The time series samples of each progressive verification group are input into the corresponding learned model instance in the actual time series order to carry out time series forward extrapolation prediction. The natural deformation time series of the slope is completely followed without disordered prediction processing. The displacement prediction values ​​output by the model are compared with the actual values ​​of the deep horizontal displacement of the slope measured on site for each period and item. The deviation between the predicted values ​​and the measured values ​​is accurately measured. The deviation level of the model under each level window in the rapid growth and deceleration stage is quantified. The rolling step residual of each progressive verification group is obtained in turn.

[0063] Strictly following the sequential arrangement of the deformation stages corresponding to the progressive windows, the rolling step residuals generated by each window are registered, classified, and organized one by one. The original residual information corresponding to each deformation stage is completely preserved without any artificial modification or data merging. All residual information is systematically recorded and archived according to the standardized recording format. All deviation data generated by the rolling verification are completely collected to form the verification residual record of the heterogeneous model instance.

[0064] Following the temporal progression of slope deformation from rapid growth to deceleration and convergence, the residual information discretely distributed in each folded window in the verification residual record is sequentially connected and linked. The scattered single-window residual data is integrated into an overall sequence that can continuously change with the deformation process, fully restoring the dynamic evolution characteristics of the model prediction error in the entire deformation cycle of the slope. After coherent integration, the rolling residual vector of the heterogeneous model instance is formed.

[0065] The pre-defined future test set, located in the transition zone from the deceleration end to the convergence transition, is completely isolated and does not participate in any model training or rolling verification process. All trained heterogeneous model instances are subjected to a one-time static prediction verification in a fixed scenario using completely independent and unfamiliar deformation condition samples. The structural state after the model training is completed remains unchanged. The overall deviation between the model prediction output and the measured displacement data of the future test set is compared to objectively measure the model's ability to truly capture the nonlinear deformation trend in the convergence transition phase, and the residual verification of the heterogeneous model instances is obtained.

[0066] The rolling residual vector, which carries the dynamic error change characteristics of multiple deformation stages, is associated, matched and integrated with the leave-out verification residual, which represents the generalization ability of the model under unfamiliar working conditions. The rolling cross-validation multi-fold evaluation information and the static leave-out test evaluation information are incorporated into a unified archiving system. The temporal position and evaluation type of each deformation stage corresponding to each residual data are fully labeled. A complete data recording system is constructed that verifies the dynamic rolling evaluation and the static independent test. After regularization and integration, a progressive folding residual archive of heterogeneous model instances is formed.

[0067] The beneficial effects are that this step relies on a time-series progressive window to segment and replicate the actual working conditions of multi-stage deformation evolution of rock cut slopes, achieving full-coverage phased performance verification through multi-fold rolling cross-validation. This completely eliminates the problems of randomness and inflated accuracy caused by traditional static single-dataset partitioning and evaluation. The step-by-step iterative training can adapt to the scenario of small sample monitoring data in engineering, allowing various heterogeneous models to fully learn the displacement time-series laws under different construction conditions. The rolling extrapolation prediction can accurately expose the prediction deviation and overfitting risks of the model in the rapid growth and various deceleration stages. Combined with the independent test set from the deceleration end to the convergence transition interval, the static retention verification can truly assess the model's adaptability to nonlinear convergence trends. The bidirectional coupling and archiving completely preserve the error details of the model in multiple working conditions and multiple folds, providing complete traceability data support that fits the real deformation mechanism of the slope for subsequent cross-fold error statistics and multi-fold generalization selection, ensuring that the model selection process meets the actual application needs of engineering.

[0068] P7. Perform cross-fold error statistics on the progressively folded residual file, and perform multi-fold generalization selection on the statistically obtained error to obtain the optimal generalization model of the slope displacement.

[0069] In this embodiment of the invention, the step of performing cross-fold error statistics on the progressively folded residual file and performing multi-fold generalization optimization on the statistically obtained error to obtain the optimal generalization model of the slope displacement includes: The progressively folded residual archive is reconstructed using error spectralization to obtain the multi-fold residual spectrum of the heterogeneous model instance; The average error of the heterogeneous model instance is obtained by performing cross-fold mean difference extraction on the multi-fold residual spectrum. The fluctuation steady-state quantization of the multi-fold residual spectrum is performed to obtain the fluctuation stability value of the heterogeneous model instance; Based on the fluctuation stability value and the average error, a two-dimensional game selection is performed on the heterogeneous model instance to obtain the optimal generalized model of the slope displacement.

[0070] By traversing the internal records of the progressively folded residual archives and all residual information corresponding to each rolling fold and static test, and strictly following the deformation sequence of the initial deceleration, middle deceleration, and final convergence transition of the slope, the scattered and independent fold residual records are arranged and sorted in an orderly manner. The discrete single-point residual information is reshaped into an overall error distribution pattern that changes continuously with the deformation stage. The entire error evolution process of each heterogeneous model instance under different working conditions is fully restored, and all residual distribution characteristics are structurally integrated to form a multi-fold residual spectrum of the heterogeneous model instance.

[0071] This study analyzes the residual characteristics of all time-series folds covered in the multi-fold residual spectrum, collects the error performance of the same heterogeneous model instance under all deformation stages of folding conditions, integrates all folding error information for overall planning and summarization, analyzes the overall error performance level of the model across multiple deformation stages, extracts the characteristic information that can represent the comprehensive prediction deviation of the model throughout the entire cycle, and organizes and analyzes the average folding error of heterogeneous model instances that characterize the overall prediction accuracy level of the model.

[0072] Following the temporal folding arrangement of the multi-fold residual spectrum, we continuously track the fluctuations in model error between adjacent folding nodes, covering the entire transition process of the slope from rapid deformation to deceleration and convergence. We meticulously analyze the variation of model prediction error in the multi-fold verification process, consider the ability of the model to maintain the error state when switching between different deformation stages, quantify the stability of the model's prediction performance across working conditions and folding scenarios, and solidify the feature identifiers corresponding to the stability to form the fluctuation stability value of heterogeneous model instances.

[0073] The average error was used as the core criterion for evaluating the overall prediction accuracy of the model, while the fluctuation stability value was used as the core criterion for evaluating the generalization stability of the model across deformation stages. All heterogeneous model instances were comprehensively compared and screened from two dimensions: prediction error amplitude and multi-condition operation stability. Model instances with high overall error and drastic error fluctuation were eliminated, while model instances with comprehensive performance that adapted to the nonlinear deformation law of the slope in multiple stages were retained. Finally, the optimal generalization model of slope displacement that adapted to the displacement evolution characteristics of the slope under all conditions was selected.

[0074] The beneficial effects of this scheme are that it fully restores the error distribution patterns of each model at different deformation stages of the slope through error spectral reconstruction, accurately quantifies the overall prediction deviation level of the model based on the cross-fold average difference, and objectively characterizes the performance stability of the model under multiple working conditions and multiple fold scenarios through fluctuation steady-state quantification. It abandons the one-sided mode of selecting the best model based solely on prediction accuracy, and adopts a dual-dimensional game selection mechanism of accuracy and stability. It effectively avoids the problem of overfitting and misselection of models that is easily caused by single index selection. It fits the inherent mechanism of the three-stage displacement evolution of rock cut slopes and the actual engineering situation of small sample monitoring data. It can objectively distinguish the generalization and adaptation differences of various heterogeneous models at different deformation stages. The optimal generalization model selected has both low prediction error and high working condition adaptation stability, which fully meets the practical application needs of long-term time series prediction for intelligent slope monitoring and early warning.

[0075] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0076] This application embodiment can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0077] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for comparing and selecting machine learning models of slope displacement sequences, characterized in that, The method includes: P1. Spatiotemporal cross-section reconstruction of the original monitoring data of slope displacement is performed to obtain the deep temporal profile of the slope displacement. P2. Rate transition anchoring is performed on the deep time series profile to obtain a multi-stage evolution sequence of the deep time series profile, which includes a rapid growth stage, a deceleration growth stage, and a convergence and stabilization stage. P3. Using the convergence start critical point of the convergence stable stage as the dividing boundary, perform redundant steady-state stripping on the multi-stage evolution sequence to obtain the effective deformation range of the multi-stage evolution sequence. P4. Perform global order-of-magnitude regularization on the effective deformation interval, and perform sliding window cross-state sampling on the regularized effective deformation interval to obtain cross-working-condition time series sample pairs of the effective deformation interval. P5. Split the cross-condition time series sample pair into a historical training set and a future test set, and enumerate the heterogeneous model instances of the historical training set according to the dimensionality characteristics of the historical training set. P6. Based on the historical training set and the future test set, cross-validation is performed on the heterogeneous model instances to obtain the progressive folded residual files of the heterogeneous model instances. P7. Perform cross-fold error statistics on the progressively folded residual file, and perform multi-fold generalization selection on the statistically obtained error to obtain the optimal generalization model of the slope displacement.

2. The method for comparing and selecting machine learning models for slope displacement sequences as described in claim 1, characterized in that, The process of reconstructing the spatiotemporal profiles of the original monitoring data of the slope displacement to obtain a deep time-series profile of the slope displacement includes: The original monitoring data of slope displacement is collected, and the original monitoring data is deconstructed into cross-sections to obtain the depth displacement cross-section queue of the original monitoring data. Peak spectral state analysis is performed on the depth displacement section queue to obtain the peak section spectrum of the depth displacement section queue; The displacement peaks in the peak cross-sectional spectrum are time-series evolved and grouped to obtain the peak displacement evolution series of the peak cross-sectional spectrum; The evolution profile of the peak displacement evolution series is reconstructed to obtain the deep time-series profile of the slope displacement.

3. The method for comparing and selecting machine learning models for slope displacement sequences as described in claim 1, characterized in that, The process of rate transition anchoring of the deep time series profile yields a multi-stage evolution sequence of the deep time series profile. This multi-stage evolution sequence includes a rapid growth stage, a decelerating growth stage, and a convergent stabilization stage. The deep time series profile is subjected to phase-by-phase differential mapping to obtain the instantaneous rate profile of the deep time series profile; Rate inflection point mining is performed on the instantaneous rate profile to obtain the rate inflection anchor point of the instantaneous rate profile; Based on the rate inflection anchor point, the working condition boundary is anchored for the deep time series profile to obtain the stage segmentation boundary of the deep time series profile. Based on the stage segmentation boundary, the deep time series profile is clustered into evolution intervals to obtain a multi-stage evolution sequence of the deep time series profile, wherein the multi-stage evolution sequence includes a rapid growth stage, a deceleration growth stage, and a convergence and stabilization stage.

4. The method for comparing and selecting machine learning models for slope displacement sequences as described in claim 3, characterized in that, The step of mining the velocity inflection points of the instantaneous velocity profile to obtain the velocity inflection anchor points of the instantaneous velocity profile includes: A sign scan of the instantaneous velocity profile is performed to obtain the positive and negative velocity boundary points of the instantaneous velocity profile; Extract the forward velocity, current velocity, and backward velocity of the neighborhood at the positive-negative velocity boundary point from the instantaneous velocity profile; The deceleration gradient is analyzed by analyzing the forward velocity, current velocity, and backward velocity of the neighborhood to obtain the deceleration inflection point depth value of the positive and negative velocity boundary. Based on the deceleration inflection point depth value, the inflection point confidence of the instantaneous velocity profile is verified to obtain the candidate turning point anchor points of the instantaneous velocity profile. Redundant anchor points among the candidate turning anchor points are filtered for local extrema to obtain the velocity turning anchor points of the instantaneous velocity profile.

5. The method for comparing and selecting machine learning models for slope displacement sequences as described in claim 1, characterized in that, The step of using the convergence initiation critical point of the convergence stable phase as the dividing boundary to perform redundant steady-state stripping on the multi-stage evolution sequence, thereby obtaining the effective deformation range of the multi-stage evolution sequence, includes: Steady-state feature quantization is performed on the multi-stage evolution sequence to obtain the steady-state fluctuation amplitude of the multi-stage evolution sequence; Based on the steady-state fluctuation amplitude, the critical point of the multi-stage evolution sequence is captured to obtain the convergence starting critical point of the multi-stage evolution sequence. Using the convergence initiation critical point as the dividing boundary, the multi-stage evolution sequence is divided into deformable segments to obtain the middle active segment of the multi-stage evolution sequence; The effective deformation range of the multi-stage evolution sequence is obtained by stripping the residual segments of the steady-state residual segments in the intermediate active segments.

6. The method for comparing and selecting machine learning models for slope displacement sequences as described in claim 1, characterized in that, The process of performing global-scale normalization on the effective deformation range and then performing sliding-window cross-state sampling on the normalized effective deformation range to obtain cross-condition time series sample pairs of the effective deformation range includes: Global extreme value selection is performed on the effective deformation interval to obtain the global displacement extreme value of the effective deformation interval; Based on the global displacement extreme value, the effective deformation interval is subjected to extreme value calibration compression to obtain the order-of-magnitude regularization sequence of the effective deformation interval; The order-of-magnitude regular sequence is sliced ​​across states using a sliding window to obtain a cross-condition window sequence of the order-of-magnitude regular sequence; The time-series segments within the window of the cross-working-condition window sequence are sampled under working-condition linkage to obtain the cross-working-condition time-series sample pairs of the effective deformation interval.

7. The method for comparing and selecting machine learning models for slope displacement sequences as described in claim 1, characterized in that, The step of splitting the cross-condition time-series sample pairs into a historical training set and a future test set, and enumerating heterogeneous model instances of the historical training set based on the dimensionality characteristics of the historical training set, includes: The cross-operating condition time series sample pairs are subjected to time series segmentation and labeling to obtain the training and testing segmentation labels of the cross-operating condition time series sample pairs; Based on the training and testing segmentation markers, the cross-working condition time series sample pairs are partitioned and truncated to obtain the historical training set and future test set of the cross-working condition time series sample pairs. The historical training set is subjected to feature scale quantization to obtain the input dimension value and sample size value of the historical training set; Based on the input dimension value and the sample size value, multi-class model enumeration is performed on the historical training set to obtain heterogeneous model instances of the historical training set.

8. The method for comparing and selecting machine learning models for slope displacement sequences as described in claim 1, characterized in that, The method of cross-validating the heterogeneous model instances based on the historical training set and the future test set to obtain the progressively folded residual archive of the heterogeneous model instances includes: Based on the historical training set, rolling cross-validation is performed on the heterogeneous model instances to obtain the validation residual records of the heterogeneous model instances; The verification residual records are concatenated in a stepwise manner to obtain the rolling residual vector of the heterogeneous model instance; Based on the future test set, static hold-out validation is performed on the heterogeneous model instance to obtain the hold-out validation residual of the heterogeneous model instance; The rolling residual vector and the reserved verification residual are bidirectionally coupled and archived to obtain the progressive folding residual archive of the heterogeneous model instance.

9. The method for comparing and selecting machine learning models for slope displacement sequences as described in claim 8, characterized in that, The step of performing rolling cross-validation on the heterogeneous model instances based on the historical training set to obtain the validation residual records of the heterogeneous model instances includes: The historical training set is progressively windowed to obtain a progressive training group and a progressive verification group for the historical training set. Based on the progressive training group, the heterogeneous model instance is subjected to iterative training in stages to obtain the learned model instance of the heterogeneous model instance on the progressive training group. Based on the progressive verification group, rolling extrapolation prediction is performed on the learned model instance, and the deviation of the predicted output value is compared to obtain the rolling step residual of the progressive verification group. The residuals of the rolling step are quantitatively recorded to obtain the verification residual records of the heterogeneous model instance.

10. The method for comparing and selecting machine learning models for slope displacement sequences as described in claim 1, characterized in that, The step of performing cross-fold error statistics on the progressively folded residual file and performing multi-fold generalization optimization on the statistically obtained error to obtain the optimal generalization model of the slope displacement includes: The progressively folded residual archive is reconstructed using error spectralization to obtain the multi-fold residual spectrum of the heterogeneous model instance; The average error of the heterogeneous model instance is obtained by performing cross-fold mean difference extraction on the multi-fold residual spectrum. The fluctuation steady-state quantization of the multi-fold residual spectrum is performed to obtain the fluctuation stability value of the heterogeneous model instance; Based on the fluctuation stability value and the average error, a two-dimensional game selection is performed on the heterogeneous model instance to obtain the optimal generalized model of the slope displacement.

Images