Slope deformation real-time early warning algorithm and system based on deep reinforcement learning

Abstract

The invention relates to the technical field of deep learning, and particularly provides a slope deformation real-time early warning algorithm and system based on deep reinforcement learning, and the method comprises the steps: generating a standardized state representation vector, and building a state quality evaluation index; receiving the state representation vector as input, constructing a dual-mode strategy architecture, generating a preliminary decision and confidence evaluation thereof, and forming a quantitative index of decision reliability; forming a final early warning instruction and a strategy updating suggestion, and generating a strategy performance evaluation report; continuously tracking the change trend of state feature distribution, constructing a strategy degradation early warning system, and triggering a model reconstruction mechanism; the updated policy network parameters are fed back to the dual-mode policy architecture, and the optimized feature extraction suggestions are fed back to the establishment of the state quality evaluation indexes. The system comprises an index establishment module, a report generation module and a reconstruction mechanism module. According to the slope deformation real-time early warning algorithm, a monitoring-decision-optimization end-to-end closed-loop control system is realized through a deep reinforcement learning framework.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a real-time early warning algorithm and system for slope deformation based on deep reinforcement learning. Background Art

[0002] With the rapid development of infrastructure construction in my country, the stability monitoring of slopes of highways, railways, water conservancy projects and other projects has become a major issue in the field of public safety. Traditional slope monitoring mainly relies on periodic manual inspection methods such as GNSS measurement and total station fixed-point observation, which has the problems of long data collection intervals and significant response delays. Especially in sudden geological disasters such as rainstorms and earthquakes, the existing system is difficult to capture the critical deformation characteristics of the slope in a timely manner, resulting in a high proportion of slope accidents caused by delayed warnings across the country. Although InSAR remote sensing technology has increased the monitoring range in recent years, its centimeter-level accuracy is still difficult to meet the early warning needs of millimeter-level deformations of high-risk slopes.

[0003] Prior art 1, Chinese patent, application number: CN202411648740.6 discloses a real-time monitoring and early warning device for deformation monitoring slopes based on millimeter wave radar, which belongs to the field of real-time monitoring and early warning technology for slopes. It is provided with a monitoring positioning seat that is easy to position and assemble, and a support seat is installed on the upper surface of the positioning seat; it includes: a control panel, which is installed on the inner surface of the support seat, and a worm is rotatably connected to the upper side of the inner surface of the support seat, and a worm wheel is meshingly connected to the outer surface of the worm, and a rotating part is installed on the upper surface of the worm wheel. A rotating monitoring mechanism is provided, which can effectively control the position of the positioning seat, and cooperate with the support seat and worm assembled by the positioning seat, and the worm wheel and the rotating part to control the angle of the monitoring rod assembled by the connecting seat, thereby controlling the angle of the monitoring head, controlling the monitoring range, and cooperating with the angle adjustment of the monitoring head and the monitoring rod. Although the convenience of using the monitoring head can be controlled by scanning the opposite slope according to the assembly height of the positioning seat, the monitoring dimension is limited by the single sensor data source (millimeter wave radar); the mechanical scanning monitoring (worm gear angle adjustment) has an inherent response delay defect; and the traditional equipment control system (positioning seat-support seat mechanical structure) lacks the bottleneck of autonomous decision-making ability.

[0004] Prior art 2, Chinese patent, application number: CN202011422338.8 discloses an open-pit mine slope deformation measurement method integrating InSAR and GNSS, including two technical means: synthetic aperture radar interferometry InSAR and global navigation satellite system GNSS, including the following steps: using the InSAR monitoring method to obtain preliminary deformation monitoring results of the slope; deploying GNSS online monitoring points at the profile position passing through the deformation center; using the least squares iteration method to obtain the corrected instantaneous deformation field; using the adaptive filtering method, Kalman filter equations and Kriging interpolation method to calculate and obtain deformation monitoring results covering the entire time and space domain. Although the advantages of InSAR and GNSS in spatial resolution and temporal resolution can be integrated to complement each other, it is possible to achieve accurate point-to-point real-time monitoring of key areas of open-pit mine slopes, as well as comprehensive surface monitoring of the entire mining area, providing support for real-time measurement and early warning of open-pit mine slope deformation; however, the monitoring gaps caused by the mismatch of spatiotemporal resolution in the fusion measurement of InSAR and GNSS, the strong dependence of traditional algorithms such as least squares iteration on historical data, and the insufficient applicability of spatial interpolation methods such as Kriging interpolation in real-time early warning scenarios.

[0005] At present, the existing technologies 1 and 2 have problems such as insufficient fusion of multi-source data, poor coordination between mechanical structure and electronic system, difficulty in matching time and space resolution, and insufficient real-time performance of traditional algorithms. Therefore, the present invention provides a real-time early warning algorithm and system for slope deformation based on deep reinforcement learning. Summary of the invention

[0006] In order to achieve the above object, the present invention adopts the following technical scheme:

[0007] In one aspect of the present invention, a real-time early warning algorithm for slope deformation based on deep reinforcement learning is provided, comprising the following steps:

[0008] Obtain multi-source monitoring data that has been screened for state characteristics, generate standardized state representation vectors, and establish state quality assessment indicators;

[0009] Receive the state representation vector as input, build a dual-mode strategy framework, generate preliminary decisions and their confidence assessments, and form quantitative indicators of decision reliability; form final warning instructions and strategy update recommendations, and generate a strategy performance evaluation report;

[0010] Based on the strategy performance evaluation report, identify the weak links of the strategy; continuously track the changing trend of the state feature distribution, build a strategy degradation warning system, and trigger the model reconstruction mechanism; the updated strategy network parameters are fed back to the dual-mode strategy architecture, and the optimized feature extraction suggestions are fed back to establish the state quality evaluation indicators.

[0011] In an optional implementation, the multi-source monitoring data of displacement, strain and groundwater level are integrated through a dynamic weighted fusion mechanism, and the state characteristics of the multi-source monitoring data are screened.

[0012] In an optional implementation, the process of establishing a state quality assessment indicator comprises the following steps:

[0013] Receive data streams of real-time multi-source monitoring data from displacement sensors, strain gauges, and groundwater level gauges; calculate preliminary weighting coefficients based on historical drift rates of displacement sensors, strain gauges, and groundwater level gauges, use moving averages to detect mutations in single-source data, and mark abnormal multi-source monitoring data;

[0014] The displacement change rate, strain increment and groundwater level change data in the data stream are synchronized in time and matched in space; a dynamic correlation coefficient matrix is ​​constructed to determine whether the coordinated change trend of the displacement sensor, strain gauge and groundwater level gauge data matches the physical mechanism;

[0015] The principal component analysis is used to compress the data stream that matches the physical mechanism and generate a low-dimensional state representation vector. The statistical dispersion of the low-dimensional state feature vector in a recent time window is calculated to monitor the distribution deviation of multi-source monitoring data in the data stream.

[0016] In an optional implementation, the process of generating a policy performance evaluation report includes the following steps:

[0017] Receive the real-time warning instruction sequence output by the dual-mode strategy architecture, and perform spatiotemporal alignment verification with the slope deformation events that actually occurred during the same period; obtain the true positive and false positive event matching rates within the instruction time window, and quantify the time series detection sensitivity; obtain the position detection accuracy through spatial overlap analysis, and form an initial detection efficiency matrix;

[0018] Based on the decision confidence distribution of 30 consecutive monitoring cycles, a strategy volatility index is constructed; at the same time, the trend of the strategy volatility index value under different geological conditions is analyzed, the environmental sensitivity parameters are identified, and the strategy robustness surface is generated;

[0019] Extract the key feature weight vectors in the state quality assessment indicators and perform reverse mapping with the decision error event set; establish a feature-error correlation matrix; obtain the main failure mode through singular value decomposition, and mark the feature combination that needs to be optimized; construct a three-dimensional evaluation coordinate system: the horizontal axis is the immediate detection efficiency, the vertical axis is the long-term stability, and the depth axis is the evolutionary ability. Use Monte Carlo sampling to predict the joint distribution of the three indicators and divide the strategy health level.

[0020] In an optional implementation, the process of receiving the real-time warning instruction sequence output by the dual-mode strategy architecture includes the following steps:

[0021] Input the generated standardized state representation vector and load the state quality evaluation index as the verification benchmark; remove abnormal feature values ​​according to the state quality evaluation index, and input the verification benchmark into the main decision module and auxiliary verification module of the dual-mode strategy architecture;

[0022] Generate primary warning instructions based on the current state characteristics of the standardized state representation vector and output the decision confidence score; verify the output consistency of the main decision module through feature space projection, and generate a correction coefficient if it deviates from the historical robust decision area; the confidence score of the main decision module verified by feature space projection and the correction coefficient of the auxiliary module together constitute the decision reliability index;

[0023] The primary warning instructions are dynamically adjusted according to the decision reliability index. If it is greater than the preset threshold, the main decision result is directly output as the final warning instruction. If it is not greater than the preset threshold, the auxiliary correction coefficient is enabled to reconstruct the warning parameters and generate conservative warning instructions. The adjusted warning instructions need to be fed back to the state quality assessment indicator to verify the coverage of historical failure cases.

[0024] In an optional implementation, the construction process of the main decision module and the auxiliary verification module of the dual-mode policy architecture includes the following steps:

[0025] The main decision module input receives the screened standardized state representation vector as the core feature input and loads the pre-trained policy network weights; the auxiliary verification module input simultaneously receives the same standardized state representation vector and additionally accesses the robust decision boundary parameters constructed by the historical case data set;

[0026] The main decision module generates primary warning instructions based on the state feature vector (, and at the same time calculates the matching degree of the current decision with the historical optimal solution in the feature space, and outputs the confidence score; the confidence score reflects the degree to which the current decision deviates from the distribution of training data; the auxiliary verification module maps the main decision output to the historical robust decision area through feature space projection, and determines whether it exceeds the credible interval. If it deviates, the correction coefficient is generated according to the preset decision robustness rules;

[0027] The confidence score output by the main decision module and the correction coefficient of the auxiliary module are weighted and fused to form a decision reliability index.

[0028] In an optional implementation, the process of marking the feature combination to be optimized comprises the following steps:

[0029] The feature weights in the original state quality assessment indicators are adjusted to a uniform scale; the strategy failure cases in the historical records are classified according to the performance characteristics, and a structured error event library is established. Each category of events is associated with a complete feature parameter snapshot at the time of the incident;

[0030] Construct a feature-error correlation matrix, and quantify the influence of each feature index on each type of error by calculating the matching degree between the feature vector and the error case. The value in the matrix directly reflects the correlation strength between a specific feature and a certain type of error. Perform spatial compression transformation on the complete feature-error correlation matrix, retain the dominant direction that explains most of the data variation, and realize the transformation from high-dimensional features to core features.

[0031] In the feature space after spatial compression and transformation, the main failure modes are screened according to the explanatory power of each dimension; the screened key dimensions are reversely mapped back to the original feature space to identify the most influential feature combination pattern that causes strategy failure.

[0032] In an optional implementation, a feature improvement priority list is established based on the failure mode analysis results, and the feature combinations that need to be optimized and their improvement order are clarified through quantitative evaluation of the impact of each mode.

[0033] In an optional implementation, the process of identifying policy weaknesses includes the following steps:

[0034] Receive the early warning strategy performance evaluation report. If the false alarm rate, confidence matching degree and feature drift index key indicators in the early warning strategy performance evaluation report exceed the preset safety threshold, the weak link analysis mechanism will be triggered;

[0035] Extract all false positive state representation vectors, divide them into potential failure modes, compare the data distribution characteristics of each failure cluster, and screen out the key feature deviation patterns that are significantly related to the current strategy decision error; calculate the cumulative deviation trend intensity of historical monitoring data based on the feature drift index, and predict the probability of future strategy failure;

[0036] According to the failure mode and the predicted probability of future strategy failure, the feature set of decision reliability drop is located. If the false alarm rate caused by a key feature changes significantly, the key feature is listed as the highest priority item for optimization.

[0037] Another aspect of the present invention provides a real-time early warning system for slope deformation based on deep reinforcement learning, comprising:

[0038] The indicator establishment module is used to integrate the multi-source monitoring data of displacement, strain and groundwater level through a dynamic weighted fusion mechanism, screen the state characteristics of the multi-source monitoring data, generate a standardized state representation vector, and establish a state quality assessment indicator;

[0039] The report generation module is used to receive the state representation vector as input, build a dual-mode strategy architecture, generate preliminary decisions and their confidence assessments, and form quantitative indicators of decision reliability; form final warning instructions and strategy update suggestions, and generate strategy performance evaluation reports;

[0040] The reconstruction mechanism module is used to identify the weak links of the strategy based on the strategy performance evaluation report; continuously track the changing trend of the state feature distribution, build a strategy degradation warning system, and trigger the model reconstruction mechanism; the updated strategy network parameters are fed back to the dual-mode strategy architecture, and the optimized feature extraction suggestions are fed back to establish the state quality evaluation indicators.

[0041] The slope deformation real-time early warning algorithm of the present invention realizes an end-to-end closed-loop control system of monitoring-decision-optimization through a deep reinforcement learning framework. Its technical effects are mainly reflected in: the coordinated optimization of dynamic perception and adaptive decision-making, the environmental perception module constructed by the dynamic weighted fusion mechanism and multi-level state feature screening, and the dual-mode strategy framework form a closed-loop feedback; the quality evaluation index of the state representation vector guides the mode switching of the strategy framework in real time, and the strategy performance evaluation reversely optimizes the feature extraction threshold, so that the system has the ability to adapt to the dynamic environment. Balanced control of real-time response and long-term evolution, in the short term, the dual-mode strategy framework realizes millisecond-level early warning decisions based on confidence evaluation; in the long term, the strategy degradation early warning system establishes a predictive maintenance mechanism by continuously tracking the distribution changes of state features. The two realize synergy through the parameter update channel. The deep integration of data-driven and mechanism constraints, the algorithm constructs a unified feature space through the standardized state representation vector, so that the monitoring data of different spatiotemporal scales can be treated as the same distribution samples; the strategy update suggestion is generated based on the feature-decision association analysis, which not only maintains the data-driven characteristics, but also avoids the risk of overfitting through quality control indicators. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0043] Figure 1 This is a flow chart of the real-time early warning algorithm for slope deformation based on deep reinforcement learning provided in Example 1 of the present invention;

[0044] Figure 2 A process diagram of establishing a state quality assessment indicator provided in Embodiment 2 of the present invention;

[0045] Figure 3 A process diagram of generating a strategy performance evaluation report provided in Embodiment 3 of the present invention;

[0046] Figure 4 A process diagram of identifying strategy weaknesses provided in Example 7 of the present invention;

[0047] Figure 5 This is a block diagram of a real-time early warning system for slope deformation based on deep reinforcement learning provided in Example 8 of the present invention; Figure 5In: 1. Indicator establishment module; 2. Report generation module; 3. Reconstruction mechanism module;

[0048] Figure 6 A block diagram of an electronic device provided by the present invention; Figure 6 In: 4. CPU / microprocessor / main control chip, etc.; 5. Storage medium; 6. Data bus; 7. Input / output bus / external bus / device bus, etc.; 8. Display; 9. Input / output device;

[0049] Figure 7 A block diagram of a computer-readable storage medium provided for the present invention; Figure 7 In: 10. Instructions; 11. Computer-readable storage medium. DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the present invention will be described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0051] In the following, the terms "first", "second", etc. are used only for convenience of description and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first", "second", etc. may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "plurality" means two or more.

[0052] In the present invention, unless otherwise clearly specified and limited, the term "connection" should be understood in a broad sense, for example, "connection" can be a fixed mechanical connection, or a detachable mechanical connection, or integrated; or, "connection" can be a direct connection, or an indirect connection through an intermediate medium. In addition, unless otherwise clearly specified and limited, the term "coupling" should be understood in a broad sense, for example, "coupling" can be a direct electrical connection, such as physical contact and electrical conduction between two components, and can also be understood as electrical connection between different components in a circuit structure through physical lines such as copper foil or wires on a printed circuit board (PCB) that can transmit electrical signals to transmit electrical signals; or, "coupling" can be an indirect electrical connection between two components through an intermediate medium; or, "coupling" can be an electrical connection between two components in an air-spaced / non-contact manner, for example, two components are electrically connected by capacitive coupling to transmit electrical signals.

[0053] In the embodiments of the present invention, directional terms such as "up", "down", "left" and "right" may be defined including but not limited to the orientation relative to the schematic placement of the components in the drawings. It should be understood that these directional terms may be relative concepts, which are used for relative description and clarification, and may change accordingly according to the change of the orientation of the components in the drawings.

[0054] Embodiment 1: like Figure 1 As shown, an embodiment of the present invention provides a real-time early warning algorithm for slope deformation based on deep reinforcement learning, comprising the following steps:

[0055] S100: Integrate multi-source monitoring data such as displacement, strain and groundwater level through a dynamic weighted fusion mechanism, screen the state characteristics of multi-source monitoring data, generate standardized state representation vectors, and establish state quality assessment indicators;

[0056] S200: receiving the state representation vector as input, constructing a dual-mode strategy framework, generating a preliminary decision and its confidence evaluation, forming a quantitative indicator of decision reliability; forming a final warning instruction and strategy update suggestion, and generating a strategy performance evaluation report;

[0057] S300: Based on the strategy performance evaluation report, identify the weak links of the strategy; continuously track the changing trend of the state feature distribution, build a strategy degradation warning system, and trigger the model reconstruction mechanism; the updated strategy network parameters are fed back to the dual-mode strategy architecture, and the optimized feature extraction suggestions are fed back to establish the state quality evaluation indicators.

[0058] In the above embodiment, the slope deformation real-time warning algorithm of this embodiment realizes the end-to-end closed-loop control system of monitoring-decision-optimization through the deep reinforcement learning framework, and its technical effects are mainly reflected in: the coordinated optimization of dynamic perception and adaptive decision-making, the environmental perception module constructed by the dynamic weighted fusion mechanism and multi-level state feature screening, and the dual-mode strategy architecture forms a closed-loop feedback; the quality evaluation index of the state representation vector guides the mode switching of the strategy architecture in real time, and the strategy performance evaluation reversely optimizes the feature extraction threshold, so that the system has the ability to adapt to the dynamic environment. Balanced control of real-time response and long-term evolution, in the short term, the dual-mode strategy architecture realizes millisecond-level warning decisions based on confidence evaluation; in the long term, the strategy degradation warning system establishes a predictive maintenance mechanism by continuously tracking the distribution changes of state features. The two realize synergy through the parameter update channel. The deep integration of data-driven and mechanism constraints, the algorithm constructs a unified feature space through the standardized state representation vector, so that the monitoring data of different spatiotemporal scales can be treated as the same distribution samples; the strategy update suggestion is generated based on the feature-decision association analysis, which not only maintains the data-driven characteristics, but also avoids the risk of overfitting through quality control indicators.

[0059] Embodiment 2: like Figure 2 As shown, based on Example 1, the process of establishing the state quality evaluation index in S100 provided in the embodiment of the present invention includes the following steps:

[0060] S101: receiving data streams of real-time multi-source monitoring data of displacement sensors, strain gauges and groundwater level gauges; calculating preliminary weighting coefficients according to historical drift rates of displacement sensors, strain gauges and groundwater level gauges, using moving averages to detect mutations in single-source data, and marking abnormal multi-source monitoring data;

[0061] S102: Time synchronization and spatial matching of the displacement change rate, strain increment and groundwater level change data in the data stream; constructing a dynamic correlation coefficient matrix to determine whether the coordinated change trend of the displacement sensor, strain gauge and groundwater level gauge data matches the physical mechanism;

[0062] S103: Compress the data stream that matches the physical mechanism through principal component analysis to generate a low-dimensional state representation vector; calculate the statistical dispersion of the low-dimensional state feature vector in a recent time window to monitor the distribution deviation of multi-source monitoring data in the data stream.

[0063] Among them, the initial weighting coefficient of S101 and the abnormal mark, the dynamic drift rate correction weighting coefficient expression is:

[0064]

[0065] In the formula, represents the weighting coefficient of sensor i at time t; and Represents the sensor i and j at the historical moment drift rate (long-term measurement error mean); represents the drift difference sensitivity parameter; represents the length of the sliding time window (the unit represents the number of sampling points); N represents the total number of sensors; indicates mutation penalty factor; represents the instantaneous rate of change of sensor i at time t (first-order difference); Display Window Moving average within ; Natural exponential function; Sign function, outputs the positive or negative value of the input value; represents sensor k at the historical moment Drift rate; t current moment;

[0066] Multi-source data mutation detection threshold expression:

[0067]

[0068] In the formula, Indicates the abnormal flag of sensor i at time t (1 means abnormal); Indicates the standard deviation multiple threshold; Represents a numerical stability constant; if represents a conditional statement; represents the data differential signal; k represents the sliding pointer in the time window;

[0069] S102 represents the verification of spatiotemporal matching and dynamic correlation coefficient, and the data resampling function expression after spatiotemporal alignment is:

[0070]

[0071] In the formula, represents the sensor i at the time and space point Interpolation data of , Represents the time and space interpolation weights; , represents the bandwidth parameter of the spatiotemporal Gaussian kernel; represents the target space coordinates, Represents the spatial coordinates of the sensor; Represents sensor i at discrete timestamp The original monitoring data value of Indicates the number of reference points in the time dimension in space-time interpolation; represents the time reference point index; m represents the spatial reference point index; M represents the total number of spatial reference points;

[0072] Dynamic correlation coefficient matrix and physical mechanism deviation expression: Indicates the spatial position of sensor i The monitoring data value of

[0073]

[0074]

[0075] In the formula, Indicates that sensors i and j are in the window Dynamic correlation coefficient within ; represents the expected correlation coefficient based on the physical mechanism; It represents the overall mechanism deviation index; represents the time gradient attenuation factor; represents the interpolated data of sensor i after spatial and temporal alignment at historical time τ; represents the interpolated data of sensor i in the time window [t−T,t] The mean of

[0076] S103 principal component compression and distribution shift monitoring, time-varying covariance matrix and principal component projection expression:

[0077]

[0078]

[0079] In the formula, represents the regularized covariance matrix; express forward The projection matrix composed of eigenvectors; represents the low-dimensional state representation vector; T represents the width of the time window;

[0080] Distribution deviation monitoring indicator expression:

[0081]

[0082] In the formula, represents the regularized covariance matrix The lth eigenvalue of ; Represents the history window Inside The mean of Indicates the history window The standard deviation of express The lth eigenvector of ; H represents the length of the historical reference window;

[0083] Explanation of logical consistency between formulas: S101 uses dynamic drift rate to correct weighting coefficients (and mutation detection to screen reliable data streams. S102 uses time series interpolation to align data and verifies physical consistency through dynamic correlation coefficients. S103 uses time-varying covariance matrix to reduce dimensions and quantifies data distribution changes through distribution offset indicators. All formulas form closed-loop feedback through parameters to ensure system adaptive optimization.

[0084] In the above embodiments, this embodiment realizes a complete set of slope deformation monitoring data optimization and characterization system: through the weighting coefficient and mutation detection of historical drift rate calculation, the reliability of input data stream is ensured, and the interference of noise or abnormal data on subsequent analysis is reduced; combined with dynamic weighting and correlation analysis, multi-source data with consistent physical mechanism is screened out to avoid calculation deviation caused by single sensor failure or environmental interference. Based on time synchronization and spatial matching, the monitoring data deviation caused by sampling interval or layout position is eliminated to ensure the effectiveness of collaborative analysis of displacement, strain and groundwater level data; the dynamic correlation coefficient matrix quantifies the correlation of sensor data and eliminates data that deviates from the overall deformation trend of the slope due to local interference (such as groundwater level fluctuation caused by short-term rainfall). After principal component analysis, high-dimensional unstructured multi-source monitoring data is converted into compact, low-dimensional state representation vectors to reduce computational complexity while retaining key physical characteristics. Distribution offset monitoring based on statistical scatter calculation can perceive the changing trend of data stream in real time to avoid model degradation caused by static modeling. Dynamic weighting adjustment or decision strategy optimization can be directly triggered.

[0085] In summary, this embodiment constructs a comprehensive state quality assessment system for adaptive dynamic weight adjustment, physically driven feature selection, and distribution-sensitive real-time monitoring: ensuring the reliability of multi-source sensor data, enhancing the availability of monitoring information, and responding to environmental changes or equipment degradation.

[0086] Embodiment 3: like Figure 3 As shown, based on Example 1, the process of generating a policy performance evaluation report in S200 provided in this embodiment of the present invention includes the following steps:

[0087] S201: Receive the real-time warning instruction sequence output by the dual-mode strategy architecture, and perform spatiotemporal alignment verification on it with the slope deformation events that actually occurred during the same period; obtain the true positive and false positive event matching rates within the instruction time window, and quantify the time series detection sensitivity; obtain the position detection accuracy through spatial overlap analysis, and form an initial detection efficiency matrix;

[0088] S202: construct a strategy volatility index based on the decision confidence distribution of 30 consecutive monitoring cycles; analyze the trend of strategy volatility index values ​​under different geological conditions, identify environmental sensitivity parameters, and generate a strategy robustness surface;

[0089] S203: Extract the key feature weight vectors in the state quality assessment indicators and perform reverse mapping with the decision error event set; establish a feature-error correlation matrix; obtain the main failure mode through singular value decomposition, and mark the feature combination that needs to be optimized; construct a three-dimensional evaluation coordinate system: the horizontal axis is the immediate detection efficiency, the vertical axis is the long-term stability, and the depth axis is the evolutionary ability. The joint distribution of the three indicators is predicted through Monte Carlo sampling to divide the strategy health level.

[0090] In the above embodiments, firstly, in the establishment of a multi-dimensional dynamic verification mechanism, through the closed-loop linkage of the spatiotemporal verification module and the state feature backtracking, the mapping relationship between the feature vector and the decision error (feature-error correlation analysis) is dynamically tracked while the early warning accuracy is verified in real time (spatiotemporal alignment verification), realizing the deep verification capability from simple result comparison to cause tracing; the verification mechanism can synchronously complete the error source location in the feature space while maintaining the accuracy of the time window. Secondly, in terms of strategy adaptive optimization, the confidence fluctuation analysis and feature failure mode recognition within the continuous monitoring cycle are used to form dual feedback, so that the system can not only perceive the macro stability changes (robustness surface), but also accurately locate the micro feature defects (main failure mode); the synergistic effect produces a dynamic optimization effect: the environmental sensitivity parameters drive the coarse-grained strategy adjustment, and the feature combination tag guides the fine-grained parameter update. The final three-dimensional evaluation system (immediate detection efficiency / long-term stability / evolutionary capability) realizes the three-dimensional evaluation indicators, in which Monte Carlo sampling transforms discrete detection data (initial matrix of S201), continuous stability parameters (robustness surface of S202) and incremental learning features (optimized feature set of S203) into a unified probability distribution model. This technical framework extends the strategy health assessment from traditional single-point static judgment to spatiotemporal evolution prediction, and the strategy life prediction error is lower than that of traditional methods.

[0091] Embodiment 4: On the basis of Example 3, the process of receiving the real-time warning instruction sequence output by the dual-mode strategy architecture in S201 provided in the embodiment of the present invention includes the following steps:

[0092] S2011: input the generated standardized state representation vector, and load the state quality evaluation index as a verification benchmark; remove abnormal feature values ​​according to the state quality evaluation index, and input the verification benchmark into the main decision module and the auxiliary verification module of the dual-mode strategy architecture;

[0093] S2012: Generate primary warning instructions based on the current state characteristics of the standardized state representation vector and output the decision confidence score; verify the output consistency of the main decision module through feature space projection, and generate a correction coefficient if it deviates from the historical robust decision area; the confidence score of the main decision module verified by feature space projection and the correction coefficient of the auxiliary module together constitute the decision reliability index;

[0094] S2013: Dynamically adjust the primary warning instructions based on the decision reliability index. If it is greater than the preset threshold, directly output the main decision result as the final warning instruction. If it is not greater than the preset threshold, enable the auxiliary correction coefficient to reconstruct the warning parameters and generate a conservative warning instruction. The adjusted warning instruction needs to be fed back to the state quality assessment indicator to verify the coverage of historical failure cases.

[0095] In the above embodiment, this embodiment drives dynamic decision optimization based on the dual verification mechanism of standardized state representation vector and evaluation index. The primary warning instruction is generated by the main decision module while the deviation detection of the auxiliary verification module is executed in parallel. Then, the instruction is adjusted according to the decision reliability index output by the dual modules to form the final warning instruction with both real-time and robustness. The joint input of the state representation vector and the quality evaluation index ensures the validity of the data; the correlation coupling of the main module confidence score and the auxiliary module correction coefficient solves the single decision risk; the reliability threshold criterion realizes the dynamic switching of the warning response mode; the adjusted warning instruction forms a closed-loop feedback through the return verification, so that the system can balance the warning sensitivity and false alarm rate under the stability constraint of the feature space; reduce the probability of instruction distortion caused by sudden anomalies, and improve the coverage of historical failure modes.

[0096] Embodiment 5: On the basis of Example 4, the construction process of the main decision module and the auxiliary verification module of the dual-mode policy architecture in S2011 provided in the embodiment of the present invention includes the following steps:

[0097] S20111: The main decision module input receives the screened standardized state representation vector as the core feature input and loads the pre-trained policy network weights; the auxiliary verification module input receives the same standardized state representation vector simultaneously and additionally accesses the robust decision boundary parameters constructed by the historical case data set;

[0098] S20112: The main decision module generates a primary warning instruction based on the state feature vector (, and at the same time calculates the matching degree of the current decision with the historical optimal solution in the feature space, and outputs a confidence score; the confidence score reflects the degree to which the current decision deviates from the distribution of the training data; the auxiliary verification module maps the main decision output to the historical robust decision area through feature space projection, and determines whether it exceeds the credible interval. If it deviates, a correction coefficient is generated according to the preset decision robustness rule;

[0099] S20113: The confidence score output by the main decision module and the correction coefficient of the auxiliary module are weighted and fused to form a decision reliability index.

[0100] In the above embodiment, the core technical function of the integration of the main decision module and the auxiliary verification module of the dual-mode strategy architecture is to realize dynamic optimization decision generation through a dual verification mechanism; the architecture is based on the input of standardized state feature vectors. While the main decision module outputs preliminary warning judgments, real-time decision verification is implemented through the auxiliary verification module, and finally a composite decision output with enhanced reliability is formed. Among them: the state representation vector is used as a unified input source to ensure the consistency of the operation basis of the two modules; the confidence score generated by the main decision module reflects the degree of deviation between the instantaneous decision and the historical optimal solution, and the spatial projection correction coefficient generated by the auxiliary verification module provides the decision boundary constraint. The reliability index formed by the weighted fusion of the two realizes the quantitative characterization of the decision credibility, so that the system can switch autonomously between the original decision and the robust adjustment scheme according to the actual working conditions. A two-layer decision system with self-verification capability is constructed, and its technical performance is as follows: on the basis of maintaining the real-time response capability of the main module, the decision robustness under abnormal conditions is significantly improved through auxiliary verification, and the adaptive optimization of system parameters is realized through the continuous feedback of the reliability index. The whole process forms a closed-loop system from feature input, dual calculation to decision optimization.

[0101] Embodiment 6: On the basis of Example 3, the process of marking the feature combination to be optimized in S203 provided in the embodiment of the present invention includes the following steps:

[0102] S2031: Adjust the feature weights in the original state quality assessment indicators to a uniform scale; classify the strategy failure cases in the historical records according to the performance characteristics, and establish a structured error event library. Each category of events is associated with a complete feature parameter snapshot at the time of the incident;

[0103] S2032: Construct a feature-error correlation matrix, and quantify the influence of each feature indicator on each type of error by calculating the matching degree between the feature vector and the error case. The value in the matrix directly reflects the correlation strength between a specific feature and a certain type of error. Perform spatial compression transformation on the complete feature-error correlation matrix, retain the dominant direction that explains most of the data variation, and realize the transformation from high-dimensional features to core features.

[0104] S2033: In the feature space after spatial compression transformation, select the main failure modes according to the explanatory power of each dimension; reversely map the selected key dimensions back to the original feature space to identify the most influential feature combination pattern that causes strategy failure; based on the failure mode analysis results, establish a feature improvement priority list, and through quantitative evaluation of the impact of each mode, clarify the feature combination that needs to be optimized and its improvement order.

[0105] In the above embodiments, feature weight normalization and error case classification are carried out to establish a standardized feature evaluation system, eliminate the scale differences of original data, build a structured error case library, and form a failure sample set that can be quantified and analyzed; association modeling and dimensionality compression are used to quantify the degree of association between features and failure types, extract the core influencing factors in the high-dimensional feature space, and reduce the computational complexity of the analysis; failure mode identification and optimization guidance are used to identify the feature combination corresponding to the main failure mechanism, establish the priority ranking of feature optimization, and provide a basis for improvement direction for strategy iteration;

[0106] In summary, this embodiment systematically completes the transformation from raw monitoring data to executable optimization solutions through the progressive processing of data standardization, association modeling, dimensionality reduction analysis, and pattern recognition. Each step collaboratively realizes the structural analysis of feature space and the location diagnosis of strategy weaknesses, and finally outputs a set of improvement targets with statistical basis.

[0107] Embodiment 7: like Figure 4 As shown, based on Example 1, the process of identifying policy weaknesses in S300 provided in this embodiment of the present invention includes the following steps:

[0108] S301: receiving an early warning strategy performance evaluation report, and triggering a weak link analysis mechanism if the false alarm rate, confidence matching degree, and feature drift index key indicators in the early warning strategy performance evaluation report exceed a preset safety threshold;

[0109] S302: Extract the state representation vectors of all false positives, divide them into potential failure modes, compare the data distribution characteristics of each failure cluster, and screen out the key feature deviation mode that is significantly related to the current strategy decision error; calculate the cumulative deviation trend strength of the historical monitoring data based on the feature drift index, and predict the probability of future strategy failure;

[0110] S303: Based on the failure mode and the predicted probability of future strategy failure, locate the feature set of decision reliability drop. If the false alarm rate caused by a key feature changes significantly, the key feature is listed as the highest priority item for optimization.

[0111] In the above embodiment, a slope deformation early warning closed-loop system with dynamic adaptability is constructed. The initial feature vector is generated by multi-source data fusion and standardized representation, and the decision credibility is evaluated while completing real-time decision-making through the dual-mode strategy architecture. Then, based on the quantitative indicators of the performance evaluation report (false alarm rate, confidence matching degree, feature drift index), the failure mode of the current strategy under a specific data distribution is systematically identified (including feature offset clustering analysis and degradation probability modeling), and finally the collaborative optimization of strategy parameters and feature screening mechanisms is triggered. The overall collaborative performance of the technical features of this process is as follows: continuously monitoring the trend of strategy performance attenuation, dynamically locating the crux of decision failure caused by changes in data distribution, and maintaining the stability of the early warning system under evolving conditions through a parameter iteration mechanism, realizing a closed-loop operation from data perception to strategy self-correction. Among them, the two-way feedback mechanism of feature screening optimization and strategy network update ensures the dynamic matching of feature representation and decision logic; effectively reduces the risk of early warning failure caused by changes in geological environment or sensor characteristic offset.

[0112] Embodiment 8: like Figure 5 As shown, based on Embodiments 1 to 7, the real-time early warning system for slope deformation based on deep reinforcement learning provided by the embodiment of the present invention comprises:

[0113] Indicator establishment module 1 is used to integrate multi-source monitoring data such as displacement, strain and groundwater level through a dynamic weighted fusion mechanism, screen the state characteristics of the multi-source monitoring data, generate a standardized state representation vector, and establish a state quality assessment indicator;

[0114] Report generation module 2 is used to receive the state representation vector as input, build a dual-mode strategy framework, generate preliminary decisions and their confidence assessments, and form quantitative indicators of decision reliability; form final warning instructions and strategy update suggestions, and generate a strategy performance evaluation report;

[0115] Reconstruction mechanism module 3 is used to identify strategy weaknesses based on strategy performance evaluation reports; continuously track the changing trend of state feature distribution, build a strategy degradation warning system, and trigger the model reconstruction mechanism; the updated strategy network parameters are fed back to the dual-mode strategy architecture, and the optimized feature extraction suggestions are fed back to establish state quality evaluation indicators.

[0116] In the above embodiment, the indicator establishment module converts heterogeneous monitoring data such as displacement, strain and groundwater level into standardized state representation vectors through a dynamic weighted fusion mechanism, eliminates data dimension differences, and simultaneously screens key state features; combines state quality evaluation indicators to improve data validity and the pertinence of feature expression. The report generation module adopts a dual-mode strategy architecture to simultaneously generate confidence assessments when outputting warning instructions to form a quantitative basis for decision reliability; drives strategy updates through strategy performance evaluation reports to achieve a balance between real-time warning and long-term performance optimization in the decision-making system. The reconstruction mechanism module continuously tracks changes in state feature distribution, detects strategy degradation trends and triggers model reconstruction; the updated strategy parameters and feature extraction suggestions are fed back to the decision layer and data processing layer respectively, forming a closed-loop optimization link from feature extraction, decision generation to model update, ensuring the system's continuous adaptability to dynamic changes in the environment. Through the coupling of data fusion, dynamic decision-making and self-correction mechanisms, the real-time, accuracy and long-term stability of slope deformation warning are enhanced, while reducing the risk of false alarms and missed alarms caused by monitoring data noise, environmental changes or model aging.

[0117] Figure 6 A block diagram of an exemplary electronic device suitable for implementing embodiments of the present invention is shown.

[0118] The electronic device may include a central processing unit / microprocessor / main control chip, etc. 4; a storage medium 5, coupled to the central processing unit / microprocessor / main control chip, etc. 4, and storing computer executable instructions therein, for performing the steps of each method of an embodiment of the present invention when executed by the processor.

[0119] The central processing unit / microprocessor / main control chip 4 may include but is not limited to, for example, one or more processors or microprocessors.

[0120] The storage medium 5 may include, but is not limited to, for example, random access memory (RAM), read-only memory (ROM), flash memory, EPROM memory, EEPROM memory, registers, computer storage media (such as hard disk, floppy disk, solid state drive, removable disk, CD-ROM, DVD-ROM, Blu-ray disc, etc.).

[0121] In addition, the electronic device may also include (but not limited to) a data bus 6, an input / output bus / external bus / device bus 7, a display 8, and input / output devices 9 (eg, keyboard, mouse, speaker, etc.).

[0122] The central processing unit / microprocessor / main control chip etc. 4 can communicate with external devices ( 8 , 9 etc.) through the I / O bus 7 via a wired or wireless network (not shown).

[0123] The storage medium 5 may also store at least one computer executable instruction for executing the various functions and / or method steps in the embodiments described in the present technology when executed by the central processing unit / microprocessor / main control chip 4.

[0124] In one embodiment, the at least one computer executable instruction may also be compiled into or constitute a software product, wherein one or more computer executable instructions are executed by a processor to perform the various functions and / or method steps in the embodiments described in the present technology.

[0125] Figure 7 A schematic diagram of a computer-readable storage medium according to an embodiment of the present invention is shown.

[0126] like Figure 7 As shown, instructions are stored on a non-transitory computer-readable storage medium 11, and the instructions are, for example, computer-readable instructions 10. When the computer-readable instructions 10 are executed by the processor, the various methods described above can be executed. Non-transitory computer-readable storage media include, but are not limited to, for example, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory (cache), etc. Non-transitory non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. For example, the non-transitory computer-readable storage medium 11 can be connected to a computing device such as a computer, and then, when the computing device runs the computer-readable instructions 10 stored on the computer-readable storage medium 11, the various methods described above can be performed.

[0127] In the several embodiments provided by the present invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of units is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0128] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0129] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0130] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for executing all or part of the steps of the various embodiments of the present invention through a computer device (which can be a personal computer, server, or network device, etc.). The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (full name in English: Read-Only Memory, English abbreviation: ROM), random access memory (full name in English: Random Access Memory, English abbreviation: RAM), disk or optical disk, etc. Various media that can store program codes.

[0131] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A real-time early warning algorithm for slope deformation based on deep reinforcement learning, characterized in that: The following steps are involved: Obtain multi-source monitoring data that has been screened for state characteristics, generate standardized state representation vectors, and establish state quality assessment indicators; Receiving the state representation vector as input, constructing a dual-mode policy framework, generating a preliminary decision and its confidence evaluation, which constitutes a quantitative indicator of decision reliability; Form final warning instructions and strategy update suggestions, and generate strategy performance evaluation reports; Identify strategy weaknesses based on strategy performance evaluation reports; Continuously track the changing trend of state feature distribution, build a strategy degradation warning system, and trigger the model reconstruction mechanism; the updated strategy network parameters are fed back to the dual-mode strategy architecture, and the optimized feature extraction suggestions are fed back to establish state quality evaluation indicators.

2. The real-time early warning algorithm for slope deformation based on deep reinforcement learning according to claim 1 is characterized in that: The multi-source monitoring data of displacement, strain and groundwater level are integrated through a dynamic weighted fusion mechanism, and the state characteristics of the multi-source monitoring data are screened.

3. The real-time early warning algorithm for slope deformation based on deep reinforcement learning according to claim 1 is characterized in that: The process of establishing status quality assessment indicators includes the following steps: Receive data streams of real-time multi-source monitoring data from displacement sensors, strain gauges, and groundwater level gauges; calculate preliminary weighting coefficients based on historical drift rates of displacement sensors, strain gauges, and groundwater level gauges, use moving averages to detect mutations in single-source data, and mark abnormal multi-source monitoring data; The displacement change rate, strain increment and groundwater level change data in the data stream are synchronized in time and matched in space; a dynamic correlation coefficient matrix is ​​constructed to determine whether the coordinated change trend of the displacement sensor, strain gauge and groundwater level gauge data matches the physical mechanism; The principal component analysis is used to compress the data stream that matches the physical mechanism and generate a low-dimensional state representation vector. The statistical dispersion of the low-dimensional state feature vector in a recent time window is calculated to monitor the distribution deviation of multi-source monitoring data in the data stream.

4. The real-time early warning algorithm for slope deformation based on deep reinforcement learning according to claim 1 is characterized in that: The process of generating a policy performance evaluation report includes the following steps: Receive the real-time warning instruction sequence output by the dual-mode strategy architecture and perform spatiotemporal alignment verification with the slope deformation events that actually occurred during the same period; The matching rates of true positive and false positive events within the instruction time window are obtained to quantify the sensitivity of timing detection; the position detection accuracy is obtained through spatial overlap analysis to form an initial detection efficiency matrix; Based on the decision confidence distribution of 30 consecutive monitoring cycles, a strategy volatility index is constructed; at the same time, the trend of the strategy volatility index value under different geological conditions is analyzed, the environmental sensitivity parameters are identified, and the strategy robustness surface is generated; Extract the key feature weight vector in the state quality assessment index and perform reverse mapping with the decision error event set; Establish a feature-error correlation matrix; obtain the main failure mode through singular value decomposition and mark the feature combination that needs to be optimized; construct a three-dimensional evaluation coordinate system: the horizontal axis is the immediate detection efficiency, the vertical axis is the long-term stability, and the depth axis is the evolutionary ability. Use Monte Carlo sampling to predict the joint distribution of the three indicators and divide the strategy health level.

5. The real-time early warning algorithm for slope deformation based on deep reinforcement learning as claimed in claim 4 is characterized in that: The process of receiving the real-time warning instruction sequence output by the dual-mode strategy architecture includes the following steps: Input the generated standardized state representation vector and load the state quality evaluation index as the verification benchmark; remove abnormal feature values ​​according to the state quality evaluation index, and input the verification benchmark into the main decision module and auxiliary verification module of the dual-mode strategy architecture; Generate primary warning instructions based on the current state characteristics of the standardized state representation vector and output the decision confidence score; The output consistency of the main decision module is verified by feature space projection. If it deviates from the historical robust decision area, a correction coefficient is generated. The confidence score of the main decision module verified by feature space projection and the correction coefficient of the auxiliary module together constitute the decision reliability index. Dynamically adjust the primary warning instructions based on the decision reliability index. If it is greater than the preset threshold, directly output the main decision result as the final warning instruction; When the value is not greater than the preset threshold, the auxiliary correction coefficient is enabled to reconstruct the warning parameters and generate conservative warning instructions. The adjusted warning instructions need to be sent back to the status quality assessment indicator to verify the coverage of historical failure cases.

6. The real-time early warning algorithm for slope deformation based on deep reinforcement learning as claimed in claim 5, characterized in that: The construction process of the main decision module and the auxiliary verification module of the dual-mode strategy architecture includes the following steps: The main decision module input receives the screened standardized state representation vector as the core feature input and loads the pre-trained policy network weights; the auxiliary verification module input simultaneously receives the same standardized state representation vector and additionally accesses the robust decision boundary parameters constructed by the historical case data set; The main decision module generates primary warning instructions based on the state feature vector, and calculates the matching degree of the current decision with the historical optimal solution in the feature space, and outputs a confidence score; the confidence score reflects the degree to which the current decision deviates from the distribution of training data; the auxiliary verification module maps the main decision output to the historical robust decision area through feature space projection, and determines whether it exceeds the credible interval. If it deviates, a correction coefficient is generated according to the preset decision robustness rules; The confidence score output by the main decision module and the correction coefficient of the auxiliary module are weighted and fused to form a decision reliability index.

7. The real-time early warning algorithm for slope deformation based on deep reinforcement learning according to claim 4 is characterized in that: The process of marking the feature combinations to be optimized includes the following steps: The feature weights in the original state quality assessment indicators are adjusted to a uniform scale; the strategy failure cases in the historical records are classified according to the performance characteristics, and a structured error event library is established. Each category of events is associated with a complete feature parameter snapshot at the time of the incident; Construct a feature-error correlation matrix, and quantify the influence of each feature index on each type of error by calculating the matching degree between the feature vector and the error case. The value in the matrix directly reflects the correlation strength between a specific feature and a certain type of error. Perform spatial compression transformation on the complete feature-error correlation matrix, retain the dominant direction that explains most of the data variation, and realize the transformation from high-dimensional features to core features. In the feature space after spatial compression and transformation, the main failure modes are screened according to the explanatory power of each dimension; the screened key dimensions are reversely mapped back to the original feature space to identify the most influential feature combination pattern that causes strategy failure.

8. The real-time early warning algorithm for slope deformation based on deep reinforcement learning according to claim 7 is characterized in that: Based on the failure mode analysis results, a feature improvement priority list is established. Through quantitative evaluation of the impact of each mode, the feature combinations that need to be optimized and their improvement order are clarified.

9. The real-time early warning algorithm for slope deformation based on deep reinforcement learning according to claim 1, characterized in that: The process of identifying strategic weaknesses includes the following steps: Receive the early warning strategy performance evaluation report. If the false alarm rate, confidence matching degree and feature drift index key indicators in the early warning strategy performance evaluation report exceed the preset safety threshold, the weak link analysis mechanism will be triggered; Extract all false positive state representation vectors, divide them into potential failure modes, compare the data distribution characteristics of each failure cluster, and screen out the key feature deviation modes that are significantly related to the current strategy decision error; Calculate the cumulative drift trend strength of historical monitoring data based on the characteristic drift index and predict the probability of future strategy failure; According to the failure mode and the predicted probability of future strategy failure, the feature set of decision reliability drop is located. If the false alarm rate caused by a key feature changes significantly, the key feature is listed as the highest priority item for optimization.

10. A real-time early warning system for slope deformation based on deep reinforcement learning, characterized in that: Include: The indicator establishment module is used to integrate the multi-source monitoring data of displacement, strain and groundwater level through a dynamic weighted fusion mechanism, screen the state characteristics of the multi-source monitoring data, generate a standardized state representation vector, and establish a state quality assessment indicator; A report generation module, which receives the state representation vector as input, builds a dual-mode policy architecture, generates a preliminary decision and its confidence evaluation, and constitutes a quantitative indicator of decision reliability; Form final warning instructions and strategy update suggestions, and generate strategy performance evaluation reports; Reconstruction mechanism module, used to identify strategy weaknesses based on strategy performance evaluation reports; Continuously track the changing trend of state feature distribution, build a strategy degradation warning system, and trigger the model reconstruction mechanism; the updated strategy network parameters are fed back to the dual-mode strategy architecture, and the optimized feature extraction suggestions are fed back to establish state quality evaluation indicators.

Images