A method for identifying displacement trends on open-pit mine slopes

Abstract

This invention relates to the field of engineering technology and discloses a method for identifying displacement trends on open-pit mine slopes. The method involves collecting monitoring data of open-pit mine slopes, labeling the collected data, constructing an original monitoring matrix containing time-series features, and preprocessing the original monitoring matrix. Adaptive fractional-order decoupling processing is performed on each original monitoring matrix to obtain a decoupling feature matrix. A physical correlation input tensor is constructed based on the physical coupling relationship between slope physical quantities. An accelerated deformation identification model is constructed. The physical correlation input tensor is input into the accelerated deformation identification model, and the accelerated deformation prediction probability is output to obtain an early warning result. The beneficial effects of this invention are: by employing an adaptive fractional-order decoupling method, the fractional-order order is dynamically adjusted through the mutation ratio and trend continuity coefficient, applying low-order suppression to instantaneous disturbances such as blasting pulses and high-order preservation to continuous deformation trends, thus solving the problem that fixed filtering strategies are difficult to process heterogeneous signals simultaneously.

Description

Technical Field

[0001] This invention relates to the field of engineering technology, specifically a method for identifying the displacement trend of open-pit mine slopes. Background Technology

[0002] Open-pit mine slope stability monitoring is a core component of ensuring safe mine production. With increasing mining depth and slope height, the need to accurately identify displacement trends and provide early warnings of accelerated deformation risks is becoming increasingly urgent. Currently, open-pit mine slopes are typically equipped with various monitoring devices that continuously acquire data on multiple physical quantities, including daily cumulative displacement, daily displacement rate, rainfall, blasting vibration, and groundwater level. These quantities reflect the slope's stability changes from different perspectives, such as deformation results, deformation rate, hydrological infiltration, blasting disturbance, and groundwater activity. However, effectively identifying true precursors of accelerated deformation from this multi-source, multi-physical-field monitoring data remains a challenge in the field of slope engineering. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a method for identifying the displacement trend of open-pit mine slopes, thereby resolving the problems mentioned in the background section.

[0004] To achieve the above objectives, this invention provides the following technical solution: A method for identifying the displacement trend of an open-pit mine slope, comprising the following steps: Data on open-pit mine slope monitoring was collected, including daily cumulative displacement, daily displacement rate, daily rainfall, daily peak velocity of blasting vibration, and groundwater level depth. The collected monitoring data were labeled, and an original monitoring matrix containing time series features was constructed. The original monitoring matrix was then preprocessed. An adaptive fractional-order decoupling process is performed on each original monitoring matrix to obtain a decoupling feature matrix. A physical correlation input tensor is constructed based on the physical coupling relationship between slope physical quantities. An accelerated deformation recognition model is constructed based on a grouped spatiotemporal convolutional encoder and a physical activity gating state network. The physical correlation input tensor is input into the accelerated deformation recognition model, and the accelerated deformation prediction probability is output to obtain the early warning result.

[0005] Furthermore, using a continuous 60-day observation window, training samples were extracted from historical open-pit mine slope monitoring data. The training samples were organized into the original monitoring matrix. , Indicates the first Each training sample corresponds to 60 days of open-pit mine slope monitoring data, with a size of [missing information]. .

[0006] Furthermore, an adaptive fractional-order decoupling process is performed on each original monitoring matrix to obtain the decoupling feature matrix, specifically: For the first The original monitoring matrix of each training sample was read column by column to obtain 5 one-dimensional time series; For the first The first-order backward difference is used to calculate the physical quantities of slope-like structures to obtain local variations. The mutation ratio is calculated based on the local change amount on the current date and the background change level in the local neighborhood; According to the Calculate the trend continuity coefficient of the changes in physical quantities of slope-like structures over the past 3 days. An adaptive fractional order is generated based on the mutation ratio and the trend continuity coefficient. The fractional-order memory weights are generated based on the adaptive fractional-order order, and the original monitoring values ​​of the current date and historical dates are back-weighted to obtain the fractional-order memory results. Calculate the first Short-term trend terms of physical quantities similar to slopes; By fusing the fractional memory results and the short-term trend term, we obtain the eigenvalues ​​before decoupling. Robust scaling was applied to the undecoupled eigenvalues ​​of the five types of slope physical quantities to obtain the undecoupled eigenvalues. The first The decoupling feature values ​​of 60 days and 5 types of slope physical quantities in the training samples are arranged in order of date and physical quantity to obtain the decoupling feature matrix.

[0007] Furthermore, a physically correlated input tensor is constructed based on the physical coupling relationship between the physical quantities of the slope, specifically: For the first The training sample at the th ... The five decoupling eigenvalues ​​of a day are arranged in a fixed order to form a decoupling eigenvector; Based on the physical mechanism of slope, three physical correlation kernel matrices are constructed. The physical correlation kernel matrices are used to assign different weights to different interaction positions in the instantaneous interaction matrix. Calculate the statistical correlation strength among five types of slope physical quantities within the training set, and perform a light correction on the three physical correlation kernel matrices; The first The instantaneous interaction matrix of day is multiplied element-wise by each physical correlation kernel matrix to obtain the first... Heavenly Interactive feature maps under each core channel; Stack all interactive feature maps from the past 60 days in the order of date, physical quantity row, physical quantity column, and kernel channel to obtain the physical correlation input tensor.

[0008] Furthermore, the method to accelerate the training of the deformation recognition model is as follows: The grouped spatiotemporal convolutional encoder groups channels according to the meaning of the physical association kernel channels, extracts hydrological cumulative features and blasting disturbance features respectively, and then performs cross-group fusion. The physical activity gating state network is based on a gating loop structure. The physical activity intensity vector is added to the update gate and reset gate, so that the gating process can perceive which type of slope physical quantity dominates the current date. The total loss function is composed of a weighted sum of the weighted binary cross-entropy loss and the trend smoothing regularization term. A deformation recognition model was trained using labeled historical open-pit mine slope monitoring samples.

[0009] Furthermore, the grouped spatiotemporal convolutional encoder groups channels according to the meaning of the physically associated kernel channels, extracts hydrological cumulative features and blasting disturbance features separately, and then performs cross-group fusion, specifically as follows: From the Extract the th training sample's physical association input tensor. A slice of the day was obtained. Sky interaction feature map; According to the function of the physically associated nuclear channels, the first The interactive feature map is divided into two channel groups: the first channel group is the cumulative hydrological channel group, and the second channel group is the short-term disturbance channel group. Spatial convolution is performed on the cumulative hydrological channel group and the short-term disturbance channel group respectively. The output feature maps of the two channel groups are concatenated along the channel dimension to obtain the intermediate interactive feature map. Perform cross-group fusion on intermediate interaction feature maps; For the first The above processing is performed on the 60-day interaction feature maps of each training sample day by day to obtain the spatial coding sequence.

[0010] Furthermore, the physical activity gating state network, based on a gating loop structure, incorporates physical activity intensity vectors into the update and reset gates, enabling the gating process to perceive which type of slope physical quantity dominates the current date. Specifically: Calculate the first based on the physically associated input tensor. The physical activity intensity vector of the day; The first The spatial encoding vector of the day, the hidden state of the previous day, and the physical activity intensity vector are input into the gating calculation to obtain the update gate and the reset gate. The previous day's hidden state is multiplied by the reset gate element, and then fed together with the spatial encoding vector into the candidate state linear layer and the hyperbolic tangent activation function to obtain the candidate hidden state; Based on the hidden state and candidate hidden state of the day before the update gate fusion, we obtain the... Hidden state; The hidden state is updated sequentially from day 1 to day 60 to obtain the hidden state on day 60. An auxiliary daily risk score is constructed on the hidden state of each day. The auxiliary daily risk score is obtained by passing the hidden state through a fully connected layer and a sigmoid activation function. It is not used as the final alarm result, but is used to constrain the risk evolution within the model.

[0011] Furthermore, the total loss function is constructed by weighted summation of the weighted binary cross-entropy loss and the trend smoothing regularization term, as follows: Calculate the weighted binary cross-entropy loss based on the accelerated deformation prediction probability and the true category label; The difference between the auxiliary daily risk scores of two consecutive days within each training sample is squared, and then averaged over the period from day 2 to day 60 to obtain the trend smoothing regularization term for that training sample; then the average of all training samples within a training batch is obtained to obtain the batch trend smoothing regularization term. The total loss function is obtained by weighting and summing the weighted binary cross-entropy loss with the trend smoothing regularization term.

[0012] Furthermore, the warning results are as follows: when the predicted probability is less than 0.5, a stable state is output; when the predicted probability is greater than or equal to 0.5 and less than 0.7, a state of concern is output, prompting technical personnel to strengthen the review; when the predicted probability is greater than or equal to 0.7, an accelerated deformation warning is output.

[0013] Furthermore, while outputting the early warning results, the average value of the physical activity intensity vector over the past 7 days is calculated to obtain the dominant physical factor score: if the scores corresponding to daily rainfall and groundwater level depth are high, it indicates that the current risk is mainly related to hydrological accumulation; if the daily peak velocity score of blasting vibration is high but the daily displacement rate score is not high, it indicates that the current risk may be a short-term blasting disturbance, which needs to be reviewed in conjunction with subsequent displacement development; if the scores of daily displacement rate and daily cumulative displacement continue to rise, it indicates that the current risk has manifested as enhanced deformation response, and the corresponding slope area needs to be inspected first.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention adopts an adaptive fractional order decoupling method, which dynamically adjusts the fractional order by adjusting the mutation ratio and the trend continuity coefficient. It applies low-order suppression to instantaneous disturbances such as blasting pulses and high-order preservation to continuous deformation trends, thus solving the problem that fixed filtering strategies are difficult to process heterogeneous signals at the same time.

[0015] 2. This invention adopts a physical correlation input tensor and a multi-core channel structure, designs hydrological accumulation kernels and blasting disturbance isolation kernels based on prior physical mechanisms, and constructs pairwise interaction relationships between physical quantities in the model input layer to guide the model to learn along the chain influence path of "rainfall-groundwater-displacement", avoiding mechanism confusion caused by black box feature engineering.

[0016] 3. This invention employs a grouped spatiotemporal convolutional encoder and a physical activity gating state network. First, in the shallow layer, slow-varying and short-term features are extracted independently by grouping according to physical channels and then fused across groups. Then, in the gating recursion, the normalized physical activity intensity vector is written into the update gate and the reset gate, so that the network can adopt a differentiated memory strategy for the activity level of different physical factors.

[0017] 4. This invention adopts a multi-threshold early warning and cause tracing mechanism that integrates engineering inspection information. It uses the trend smoothing regularization term of the auxiliary daily risk score to constrain the steady rise of the risk curve, and combines the physical activity intensity score output online to identify the current risk-dominant factors, aligning the model prediction with the on-site verification requirements. Attached Figure Description

[0018] Figure 1 This is a flowchart of the present invention; Figure 2 Schematic diagram of local changes in displacement rate on that day; Figure 3 Schematic diagram of the daily displacement rate change ratio; Figure 4 A schematic diagram of the continuity coefficient of the daily displacement rate trend. Detailed Implementation

[0019] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined in this application.

[0020] like Figure 1 As shown, a method for identifying the displacement trend of open-pit mine slopes is disclosed, including the following steps: S1. Sample Collection and Label Construction for Multiphysics Field Monitoring of Open-Pit Mine Slopes This invention first collects open-pit mine slope monitoring data, which includes five types of slope physical quantities: daily cumulative displacement, daily displacement rate, daily rainfall, daily peak velocity of blasting vibration, and groundwater level depth. These five types of slope physical quantities reflect slope deformation results, deformation rate, hydrological infiltration, blasting disturbance, and groundwater activity, respectively, and can describe slope stability changes from different physical perspectives.

[0021] In one implementation, the daily cumulative displacement is in millimeters (mm), the daily displacement rate is in millimeters per day (mm / d), the daily rainfall is in millimeters (mm), the daily peak velocity of blasting vibration is in millimeters per second (mm / s), and the groundwater level depth is in meters (m).

[0022] Conventional single displacement threshold methods struggle to distinguish between short-term blasting disturbances and sustained accelerated deformation, and are also inadequate for identifying slow deterioration processes caused by continuous rainfall and groundwater changes. This invention uses a continuous 60-day observation window as a single identification sample, organizing five types of slope physical quantities into a unified matrix in a fixed order. It then combines manual inspection records, displacement trends, and historical disaster records to label stable or accelerated deformation, providing structured input for subsequent model training.

[0023] In one data acquisition method, displacement monitoring equipment, rain gauges, blasting vibration monitoring equipment, and groundwater level monitoring equipment are deployed in the open-pit mine slope monitoring area. Daily-scale monitoring records are generated with a sampling cycle of one day. Among them, the daily cumulative displacement is used to represent the cumulative deformation of the slope monitoring point relative to the initial moment, the daily displacement rate is used to represent the rate of change of the cumulative displacement between two adjacent days, the daily rainfall is used to represent the cumulative rainfall value on that day, the daily blasting vibration peak velocity is used to represent the maximum particle vibration velocity generated by blasting on that day, and the groundwater level depth is used to represent the vertical distance from the groundwater level to the ground surface or monitoring reference surface.

[0024] Furthermore, using a continuous 60-day observation window, training samples are extracted from historical open-pit mine slope monitoring data through a sliding step, which can be 1 day, 3 days, or 7 days. For the first... The training samples were organized into the original monitoring matrix. , Indicates the first Each training sample corresponds to 60 days of open-pit mine slope monitoring data, with a size of [missing information]. ,in, This represents the training sample index, with a value ranging from 1 to... , Indicates the total number of training samples; This represents the date index within the observation window, with a value range from 1 to 60; This represents the index of physical quantities of the slope. Corresponding to the daily cumulative displacement, Corresponding daily displacement rate, Corresponding to daily rainfall, Corresponding to the peak velocity of daily blasting vibration, Corresponding groundwater level depth; Indicates the first In the training samples, the th Heavenly The original monitoring values ​​of physical quantities of slope-like structures; Furthermore, the original monitoring matrix Missing values ​​and obvious outliers are handled. When the number of consecutive missing days for a single slope physical quantity does not exceed two days, linear interpolation of adjacent valid dates is used to fill in the gaps. When the number of consecutive missing days exceeds two days, the 60-day observation window is removed to avoid artificial filling and the formation of false trends. When a raw monitoring value exceeds the sensor's range or is obviously inconsistent with the field record, it is first verified based on the equipment log. If verification is not possible, the 1% quantile and 99% quantile of the slope physical quantity in the training set are used for truncation.

[0025] Furthermore, for each original monitoring matrix Configure window-level labels ,in, Indicates the first The true class labels of each training sample Indicates a stable state. This indicates an accelerated deformation state. A stable state corresponds to a situation where the daily cumulative displacement and daily displacement rate do not show a continuous upward trend within the 60-day observation window, and no tensile cracks, slippage, or significant deformation expansion are found during on-site inspections. An accelerated deformation state corresponds to a situation where the daily displacement rate increases for several consecutive days in the latter part of the 60-day observation window, the slope of the daily cumulative displacement curve increases, or on-site inspection records show signs of instability such as slope crack expansion, intensified step deformation, and localized landslides.

[0026] In one embodiment, for example, if a large daily peak blasting velocity occurs only on day 32 within a 60-day observation window, but the daily displacement rate falls back to the normal fluctuation range from day 33 to day 40, and the daily cumulative displacement curve does not continue to steepen, then this 60-day observation window is marked as a stable state. If continuous rainfall occurs from day 20 to day 35 within another 60-day observation window, and the groundwater level gradually becomes shallower after day 28, with the daily displacement rate continuously increasing from day 35 to day 60, then this 60-day observation window is marked as an accelerated deformation state.

[0027] S2. Adaptive Fractional Decoupling and Physically Associated Input Tensor Construction Open-pit mine slope monitoring data exhibits multi-timescale coupling characteristics. Daily blasting vibration peak velocity typically manifests as instantaneous pulses on a single day or a few days, while continuous rainfall and groundwater level changes typically represent cumulative disturbances over multiple days. Daily cumulative displacement and daily displacement rate reflect the slope deformation results.

[0028] This invention first performs adaptive fractional-order decoupling processing on each original monitoring matrix to obtain a decoupled feature matrix; then, it constructs a physical correlation input tensor based on the physical coupling relationship between slope physical quantities, enabling the subsequent accelerated deformation identification model to simultaneously receive trend information of a single slope physical quantity and interactive information between multiple physical fields. The specific steps are as follows: S201, Adaptive Fractional Decoupling of One-Dimensional Sequences Adaptive fractional-order decoupling is used to distinguish between instantaneous abrupt changes and continuous trends within the same slope physical quantity. For short-term disturbances such as blasting pulses and sensor spikes, adaptive fractional-order decoupling reduces the control intensity of the current date on the output, making the output closer to the smooth memory result. For trend changes such as continuous increases in displacement rate and shallowing of groundwater level over multiple days, adaptive fractional-order decoupling increases the contribution of recent changes to the output, preserving early signs of accelerated deformation. The specific steps are as follows: 1> For the first The original monitoring matrix of each training sample Reading column by column, we obtained 5 one-dimensional time series.

[0029] Correspondingly, each one-dimensional time series corresponds to the first... Physical quantities of slope types, read in chronological order. This sequence represents the first... The original changes in physical quantities of slope-like structures within a 60-day observation window.

[0030] 2> Regarding the first The first-order backward difference was used to calculate the physical quantities of slope-like structures, yielding the local variations. ,in, Indicates the first In the training samples, the th Heavenly The magnitude of change in physical quantities of slope-like structures relative to the previous day is used to reflect the intensity of local abrupt changes on the current date.

[0031] In practical implementation, when season ;when season .

[0032] Furthermore, with the first A local neighborhood index set is established with Tian as the center. , Indicates the first Within 5 days before and after the first day, except for the day before and after. The set of valid dates outside of 10 days is truncated to the range of 1 to 60 at the boundary.

[0033] In one embodiment, as an example, when At that time, local neighborhood index set It can include day 1, day 2, day 4 through day 8; when At that time, local neighborhood index set It can include days 25 to 29 and days 31 to 35.

[0034] 3> Calculate the mutation ratio based on the local change amount at the current date and the background change level within the local neighborhood. .

[0035] In the specific implementation, the local neighborhood index set is first retrieved. The median of the absolute values ​​of local changes over each date is used as the background change level; then, the absolute value of the local change on the current date is divided by the sum of the background change level and the stability constant to obtain the mutation ratio, which is calculated as follows: ; in, Indicates the first In the training samples, the th Heavenly The abrupt change ratio of slope-like physical quantities is used to measure the degree of amplification of the change magnitude on the current date relative to the change magnitude of the local background. Indicates the first In the training samples, the th Heavenly The magnitude of change in physical quantities of slope-like structures relative to the previous day; Represents the local neighborhood index set Date index within; Indicates the first The stability constant corresponding to the physical quantity of the slope is used to prevent the denominator from being zero; for example, it can be taken as... ; This indicates retrieving the local neighborhood index set. The median of the values ​​corresponding to each date.

[0036] In one embodiment, for example, if the peak daily blasting vibration velocity pulses on the 25th day, the absolute value of the local change on the current date is 6.0, the local neighborhood background change level is 0.5, and the stability constant is taken as... Then the mutation ratio The value is approximately 12, indicating that day 25 is more likely to be a transient blasting disturbance or a sampling spike. If the daily displacement rate increases slowly over several consecutive days, and the absolute value of the local change on the current date is 0.25, while the local neighborhood background change level is 0.20, then the abrupt change ratio is approximately 1.25, indicating that the change is closer to a continuous trend than an isolated pulse.

[0037] 4> According to the first Calculation of the trend continuity coefficient of the physical quantities of slope-like structures over the past 3 days ,in, Indicates the first In the training samples, the th Heavenly The degree of continuity of the recent trend of slope-like physical quantities, with a value range from 0 to 1.

[0038] In the specific implementation, read the first... Heaven, the First Tianhe Di The local change over the day is counted, and the number of local changes with the same sign as the current date and whose absolute value is not lower than the local neighborhood background change level is counted. Then, the number is divided by 3 to obtain the trend continuity coefficient. When there are less than 3 available historical days, only the existing dates are used and normalized according to the actual number.

[0039] In one embodiment, as an example, suppose the first , , If all local changes in a certain physical quantity are positive and their absolute values ​​are not lower than the local background change level, then the trend continuity coefficient is... If only the first , The conditions are met on the 1st day. If the symbol for Heaven is reversed, then... ;like Only 1 day of historical data is available, and the conditions are met. .

[0040] 5> Based on the mutation ratio and trend continuity coefficient Generate adaptive fractional order ,in, Indicates the first In the training samples, the th Heavenly The fractional order of slope-like physical quantities, ranging from 0 to 1, is used to control whether the output is more biased towards historical smooth memory or more biased towards recent trend changes.

[0041] In practical implementation, the mutation ratio is first mapped to a mutation intensity within the range of 0 to 1; the larger the mutation ratio, the greater the mutation intensity. Then, the fractional order is reduced based on the mutation intensity and increased based on the trend continuity coefficient. Preferably, the minimum fractional order can be... Set it to 0.05, and set the order of the largest fractional order. Set it to 0.85; when the mutation ratio is high and the trend continuity coefficient is low, make near When the mutation ratio is moderate and the trend continuity coefficient is high, it makes near .

[0042] In one implementation, the mutation strength can be calculated first. ,in, Indicates the first In the training samples, the th Heavenly The normalized abrupt change intensity of slope-like physical quantities is calculated; then, the fractional order is set to decrease with the abrupt change intensity and increase with the trend continuity coefficient, and the results are restricted to... arrive Between. For the daily peak blasting vibration velocity, the trend continuous correction intensity can be set to a smaller value, such as 0.05; for the daily cumulative displacement and daily displacement rate, the trend continuous correction intensity can be set to a larger value, such as 0.20, in order to better preserve the continuous deformation trend.

[0043] In one embodiment, as an example, suppose the abrupt change ratio of the peak daily blasting vibration velocity on day 25 of a training sample is... Trend continuity coefficient Then the normalized mutation strength Taking the trend continuous correction strength as 0.05, the adaptive fractional order is... The order is close to The output approximates a smooth memory result. Let the abrupt change ratio of the daily displacement rate on day 40 of another training sample be... Trend continuity coefficient ,but Taking the trend continuous correction strength as 0.20, then... The order is moderate, and the output can retain recent trend changes. Among them, This indicates the operation of retrieving the maximum value. This indicates the operation of taking the minimum value.

[0044] 6> Based on the adaptive fractional order A fractional-order memory weight is generated, and the original monitoring values ​​for the current date and historical dates are back-weighted to obtain the fractional-order memory result. ,in, Indicates the first In the training samples, the th Heavenly The adaptive smoothing result of the slope-like physical quantity has the same dimensions as the slope physical quantity.

[0045] In practical implementation, the backtracking length can be set to 30 days. Let... Indicates the first The maximum available backtrack length per day ;set up Indicates the number of days that the history is lagging behind. Indicates the current date. This represents the previous day, and so on. First, the unnormalized weights are calculated using a power-law decreasing method, maximizing the weight of the current date and gradually decreasing the weight of historical dates with the lag period. Then, all unnormalized weights are summed and normalized to a sum of 1. Finally, the original monitoring values ​​of the current date and historical dates are summed according to the normalized weights to obtain the fractional-order memory result. The smaller the adaptive fractional-order number, the slower the historical date weight decays, resulting in a smoother output; the larger the adaptive fractional-order number, the faster the historical date weight decays, and the closer the output is to recent changes.

[0046] In one embodiment, as an example, suppose the adaptive fractional order of a certain physical quantity is [number] in the nth training sample on day t. Backtracking length Regarding the number of days of lag Unnormalized weights are ;right Unnormalized weights are ;right Unnormalized weights are Summing all 30 unnormalized weights gives approximately 6.15. The normalized weights are approximately Fractional order memory results It is the weighted sum of the original monitoring values ​​for the current date and the previous 29 days, adjusted for normalization.

[0047] 7> Calculate the first Short-term trend terms of slope-like physical quantities ,in, Indicates the first In the training samples, the th Heavenly The average direction and magnitude of change of physical quantities of slope-like structures over the past 3 days.

[0048] In the specific implementation, take the first... The local changes over the day and the two days prior are truncated for individual difference values ​​that significantly exceed five times the local background change level, and then the mean is calculated to obtain the short-term trend term. This truncation process prevents single-day explosive pulses from directly entering the short-term trend term.

[0049] In one embodiment, as an example, suppose the local neighborhood background change level of a certain physical quantity on day t of the nth training sample is 0.20, and the upper limit of the cutoff is... The local changes on days t-2, t-1, and t are 0.15, 1.80, and 0.22, respectively. Since 1.80 exceeds the truncation limit, it is truncated to 1.0. The three truncated values ​​are 0.15, 1.0, and 0.22, representing the short-term trend. .

[0050] 8> Memorize the fractional order results and short-term trend items The eigenvalues ​​are then fused to obtain the eigenvalues ​​before decoupling. During fusion, the adaptive fractional order is used. The strength of short-term trend term retention is controlled by adjusting the adaptive fractional order; the higher the adaptive fractional order, the greater the contribution of the short-term trend term, and vice versa. Based on this, the output corresponding to the burst pulse mainly comes from the fractional order memory result, and the output corresponding to the continuous deformation trend can retain the recent change direction.

[0051] In one implementation, the eigenvalues ​​before decoupling are denoted as... The fusion method is represented as: ,in, Indicates the first In the training samples, the th Heavenly The eigenvalues ​​of slope-like physical quantities before decoupling. For example, let... , For the blasting pulse day, ,but The contribution of short-term trend terms is weak, and the output mainly comes from smoothing memory; for continuously deformed trend days, ,but The short-term trend component makes a significant contribution, and the output retains the direction of recent changes.

[0052] 9> Robust scaling was applied to the undecoupled eigenvalues ​​of the five types of slope physical quantities to obtain the undecoupled eigenvalues. ,in, Indicates the first In the training samples, the th Heavenly Characteristic values ​​of slope-like physical quantities after pulse suppression, trend preservation, and dimension unification.

[0053] In its implementation, the robust scaling operation employs a standard robust standardization operation. Specifically, it uses the median and interquartile range of the same slope physical quantity in the training set. This is achieved by subtracting the bit depth of the corresponding slope physical quantity in the training set from the current eigenvalue before decoupling, and then dividing by the sum of the interquartile range of the corresponding slope physical quantity and the minimum constant. The minimum constant can be taken as... .

[0054] It should be noted that a smaller groundwater level depth usually indicates a higher groundwater level and a stronger risk of hydrological activity. Therefore, when applying robust scaling to the physical quantities of Type 5 slopes, directional unification can be performed first. This involves subtracting the undecoupled eigenvalue corresponding to the current groundwater level depth from the median groundwater level depth in the training set, so that a larger directional unification value indicates stronger groundwater activity. The directional unification result is still used as the undecoupled eigenvalue corresponding to the groundwater level depth in subsequent processing.

[0055] 10> will the first The decoupling feature values ​​of 60 days and 5 types of slope physical quantities in the training samples are arranged in order of date and physical quantity to obtain the decoupling feature matrix. , Indicates the first The decoupling feature matrix corresponding to each training sample has a size of . It characterizes the multiphysics field evolution characteristics within a 60-day observation window after pulse suppression, trend preservation, and dimensional unification.

[0056] It should be noted that blasting pulses are usually short in duration and should not be accumulated in a hidden state for a long time; early accelerated deformation usually manifests as continuous enhancement over multiple days and should not be excessively smoothed. The adaptive fractional-order decoupling used in this invention does not simply filter out all high-frequency changes, but uses both the mutation ratio and the trend continuity coefficient to distinguish between "single-day spikes" and "continuous trends". Based on this, misjudgments caused by blasting disturbances and sensor spikes can be reduced, while retaining the continuous risk signals in the daily cumulative displacement, daily displacement rate and hydrological factors.

[0057] In one embodiment, such as Figure 2 As shown, the local variation of daily displacement rate is analyzed, and the first-order backward difference result of daily displacement rate in the nth training sample is displayed, that is, the magnitude of displacement rate change of each day relative to the previous day. The horizontal axis represents the number of days; the vertical axis represents the displacement rate in millimeters per day (mm / d). The change in displacement rate shows obvious positive fluctuations during periods of increased rainfall and groundwater activity, while isolated spikes may appear at the moment of blasting pulses. This demonstrates that the local variation can capture both the starting point of the acceleration trend and the instantaneous intensity of short-term disturbances, and is a fundamental signal for calculating the abrupt change ratio and the trend continuity coefficient.

[0058] In one embodiment, such as Figure 3As shown, the daily displacement rate abrupt change ratio is analyzed, displaying the ratio calculated based on local variation and local neighborhood background level. The horizontal axis represents the number of days; the vertical axis represents the abrupt change ratio, which is dimensionless. This figure demonstrates that on the date of the blasting disturbance, the abrupt change ratio is significantly higher than the surrounding background level, reflecting the suddenness and isolation of the daily change; while during the continuous acceleration phase, the abrupt change ratio remains at a lower level, indicating that the change is closer to a continuous trend rather than an isolated pulse.

[0059] In one embodiment, such as Figure 4 As shown, the daily displacement rate trend continuity coefficient is analyzed, displaying the coefficient calculated based on the direction and magnitude of change over the past three days. The horizontal axis represents the number of days; the vertical axis represents the trend continuity coefficient, which is dimensionless. In the later stage of continuous acceleration and deformation, the trend continuity coefficient is close to 1.0, indicating that the direction of change is consistent and the magnitude is reliable; while during the period of data fluctuation, the trend continuity coefficient decreases, indicating that the recent changes lack continuity.

[0060] S202, Construction of Physically Related Input Tensors After obtaining the decoupling characteristic matrix Subsequently, the physical quantities of the five types of slopes on the same day were combined in pairs to construct physical interaction relationships, thereby expressing the chain effect of "continuous rainfall - shallower groundwater level - increased daily displacement rate - accelerated daily cumulative displacement". At the same time, the instantaneous effects of "blasting vibration - short-term displacement rate fluctuations" were distinguished from the cumulative hydrological effects. The specific steps are as follows: 1> For the first The training sample at the th ... The five decoupling eigenvalues ​​of the day are arranged in a fixed order to form a decoupling eigenvector. ,in, Indicates the first In the training samples, the th The 5-dimensional decoupling feature vector for each day consists of five components, which are the decoupling feature values ​​corresponding to the daily cumulative displacement, daily displacement rate, daily rainfall, daily peak velocity of blasting vibration, and groundwater level depth, respectively.

[0061] Furthermore, the decoupling eigenvectors are calculated. The instantaneous interaction matrix is ​​obtained by multiplying it by itself. ,Right now , Indicates the first In the training samples, the th The pairwise interaction strengths between the physical quantities of five types of slopes, with dimensions of... , express The transpose of . Instantaneous interaction matrix The Middle Line number Column element equals the first Decoupling eigenvalues ​​of physical quantities of slope type and the first The product of the decoupled eigenvalues ​​of slope-like physical quantities, where... and All of these represent slope physical quantity indices, with values ​​ranging from 1 to 5.

[0062] In one embodiment, for example, if the first If the decoupling eigenvalue corresponding to daily rainfall is 1.2 and the decoupling eigenvalue corresponding to groundwater level depth is 1.0, then the interaction value at the corresponding position of daily rainfall and groundwater level depth in the instantaneous interaction matrix is ​​1.2, which indicates that the hydrological factors on that day are mutually reinforcing. If the... The decoupling characteristic value corresponding to the peak velocity of the daily blasting vibration is relatively high, but the decoupling characteristic value corresponding to the daily displacement rate is not high. Therefore, the interaction value between the blasting disturbance and the deformation response is low, which helps the subsequent model to identify that the blasting event did not cause significant and continuous deformation.

[0063] 2. Based on the physical mechanism of the slope, three physical correlation kernel matrices are constructed. These kernel matrices are used to assign different weights to different interaction locations in the instantaneous interaction matrix. The number of physical correlation kernel matrices is denoted as... , This represents the number of physical correlation kernel matrices, with a preferred value of 3. This represents the index of the physical correlation kernel matrix, with values ​​ranging from 1 to... ; Indicates the first There are physical correlation kernel matrices, with a size of . ; Indicates the first In the physical correlation kernel matrix, the th Physical quantities of slope type and the first Weights of the interaction relationships of physical quantities in slope-like structures.

[0064] In practical implementation, the first physical correlation kernel matrix can be set as a fully interactive retention kernel matrix to preserve the basic interaction relationships between all slope physical quantities. Specifically, the diagonal weight is set to 1, and the off-diagonal weight is set to 0.5, so that the strength of a single slope physical quantity and its pairwise interaction relationships are both preserved.

[0065] Furthermore, the second physical correlation kernel matrix can be set as a hydrological cumulative kernel matrix to emphasize the slow-varying coupling relationship between daily rainfall, groundwater level depth, daily displacement rate, and daily cumulative displacement. Specifically, the weight between daily rainfall and groundwater level depth is set to 1.0, the weight between daily rainfall and daily displacement rate, and the weight between groundwater level depth and daily displacement rate are set to 0.8, the weight between daily rainfall and daily cumulative displacement, and the weight between groundwater level depth and daily cumulative displacement are set to 0.6, and the weight involving daily peak blasting vibration velocity is set to 0.1 to 0.3.

[0066] Furthermore, the third physical correlation kernel matrix can be set as a blasting disturbance isolation kernel matrix to emphasize the short-term interaction between the daily peak blasting velocity and the daily displacement rate, while reducing the direct mixing of blasting disturbances and persistent hydrological factors. Specifically, the weight between the daily peak blasting velocity and the daily displacement rate is set to 1.0, the weight between the daily peak blasting velocity and the daily cumulative displacement is set to 0.5, the weight between the daily peak blasting velocity and the daily rainfall and groundwater level depth is set to 0.1, and the weights for other non-blasting interaction locations are set to 0.3 to 0.5.

[0067] In one embodiment, for example, a hydrological cumulative kernel matrix is ​​used. For example, some of its weights are set as follows: (The ratio between daily rainfall and groundwater level depth) is taken as 1.0; The relationship between daily rainfall and daily displacement rate is taken as 0.8. (The relationship between groundwater level depth and daily displacement rate) is taken as 0.8; (The ratio of daily rainfall to daily cumulative displacement) is taken as 0.6; (The relationship between groundwater level depth and daily cumulative displacement) is taken as 0.6; (The ratio of daily peak blast vibration velocity to daily displacement rate) is taken as 0.2; (The ratio between the daily peak blasting vibration velocity and the daily rainfall) is 0.1; for other locations not specified, it is 0.3.

[0068] 3. Calculate the statistical correlation strength among the five types of slope physical quantities within the training set, and perform a light correction on the three physical correlation kernel matrices. The statistical correlation strength is denoted as... , Indicates the first Physical quantities of slope type and the first The absolute value of the Pearson correlation coefficient of the physical quantities of slope-like structures in the decoupling feature matrix of the training set, ranging from 0 to 1; when season .

[0069] In practical implementation, all training samples and the first date from all dates will be included. The decoupled eigenvalues ​​of slope-like physical quantities are concatenated into a long sequence, and the first... The decoupled eigenvalues ​​of slope-like physical quantities are spliced ​​into another long sequence, and then the absolute values ​​of the Pearson correlation coefficients of the two long sequences are calculated.

[0070] Furthermore, the weights in each physical correlation kernel matrix are lightly adjusted element-wise to simultaneously integrate the fundamental relationships defined by the physical mechanism and the statistical correlation strengths in the training data. This approach assigns higher weights to interactions of slope physical quantities that have historically been more stable, while preserving the fundamental relationships derived from physical mechanisms. This modification does not alter the main structure of the physical correlation kernel matrix; it is only used to adapt to local geological and monitoring differences in different open-pit mine slopes. This represents the number of elements after correction for statistical correlation strength based on the training set. In the physical correlation kernel matrix, the th Physical quantities of slope type and the first The weights of the interaction relationships between physical quantities similar to slope, with dimensions of .

[0071] In one embodiment, as an example, suppose The original value was 0.8, representing the absolute value of the Pearson correlation coefficient between the decoupled eigenvalues ​​of daily rainfall and daily displacement rate in the training set. Then the corrected weights If a certain physical quantity has a weak correlation with the others, The correction factor is The original weights were appropriately reduced. Based on this, the basic structure given by the physical mechanism was preserved through modification, while minor adjustments were made according to the statistical correlation strength of the actual monitoring data.

[0072] 4> The first Instantaneous Interaction Matrix of the Sky Multiply each element-wise with each physical correlation kernel matrix (after correction) to obtain the . Heavenly Interactive feature map under each core channel ,in, Indicates the first In the training samples, the th Heaven is in The interaction feature map obtained under the action of a physical correlation kernel matrix has a size of [missing information]. .

[0073] In practical implementation, element-wise multiplication refers to the multiplication of the first element in the instantaneous interaction matrix. Line number The column elements and the physical correlation kernel matrix of the first Line number Multiply the column weights to obtain the weighted interaction value for the same position.

[0074] 5> Stack all interaction feature maps from the past 60 days in the order of date, physical quantity row, physical quantity column, and kernel channel to obtain the physical correlation input tensor. , Indicates the first The physical association input tensor corresponding to each training sample has a size of . The first dimension represents the 60-day time dimension, the second and third dimensions represent the interaction matrix positions between the five types of slope physical quantities, and the fourth dimension represents the physical correlation core channel.

[0075] It should be noted that the physical correlation input tensor is not simply an increase in input dimension, but rather incorporates the slope's physical mechanisms into the input structure. The fully interactive kernel matrix ensures that the model does not lose unknown interactions; the hydrological cumulative kernel matrix highlights the slow-varying influence of continuous rainfall and groundwater activity on displacement development; and the blasting disturbance isolation kernel matrix allows blasting impacts to be expressed in a dedicated channel, reducing the risk of premature mixing of blasting pulses and continuous deformation signals.

[0076] S3. Accelerated Deformation Recognition Model Training Based on Grouped Spatiotemporal Convolution and Physical Activity Gated State Network Input tensors with physical association As input, window-level labels As a supervisory signal, the training of the accelerated deformation recognition model is carried out. This invention first extracts the multi-physics interaction spatial pattern of each day through a grouped spatiotemporal convolutional encoder, and then updates the hidden state in the order of dates through a physical activity gating state network. The physical activity intensity corresponding to the daily cumulative displacement, daily displacement rate, daily rainfall, daily peak velocity of blasting vibration and groundwater level depth is input into the gating calculation, so that the model can adopt a differentiated memory strategy for different physical factors.

[0077] The accelerated deformation recognition model consists of a grouped spatiotemporal convolutional encoder and a physical activity-gated state network connected in series. The physical association input tensor first enters the grouped spatiotemporal convolutional encoder, where cumulative hydrological features and short-term disturbance features are extracted by grouping according to the physical association kernel channels. After cross-group fusion, a daily spatial encoding vector is generated. The spatial encoding sequence then enters the physical activity-gated state network, which incorporates the daily physical activity intensity vector into the gating calculation, updates the hidden state in date order, and outputs a window-level accelerated deformation prediction probability. The specific steps are as follows: S301, Grouped Spatiotemporal Convolutional Encoder Grouped spatiotemporal convolutional encoders are used to convert each day's physical correlation into a tensor. The interactive feature map is compressed into a spatial encoding vector. The encoder first groups the channels according to the meaning of the physical association kernel channels, extracts hydrological cumulative features and blasting perturbation features respectively, and then performs cross-group fusion to avoid premature mixing of different physical mechanisms in the shallow convolution stage. The specific steps are as follows: 1> From the first The physical association input tensor of each training sample Take out the first one A slice of the day was obtained. Tian Interactive Feature Map ,in, Indicates the first In the training samples, the th The multi-channel interactive feature map of the day, with a size of .

[0078] 2> According to the function of the physically associated nuclear channel, the first Tian Interactive Feature Map Divided into 2 channel groups, of which: The first channel group is the cumulative hydrological channel group, which includes channels corresponding to the fully interactive kernel matrix and channels corresponding to the hydrological cumulative kernel matrix. It is used to extract the slow-varying interaction patterns between continuous rainfall, groundwater level depth and displacement changes. The second channel group is the short-time disturbance channel group, which includes the channel corresponding to the blasting disturbance isolation core matrix. It is used to extract the short-time response pattern between the daily blasting vibration peak velocity and the daily displacement rate.

[0079] In the specific implementation, grouping is performed according to the function of the physically associated kernel channels, and the specific division method is as follows: Cumulative hydrological channel group: indexed by physical association kernel channel satisfy (Corresponding to the fully interactive kernel-preserving matrix) and The channel composition (corresponding to the hydrological cumulative kernel matrix) consists of 2 kernel channels; Short-term perturbation channel group: indexed by physically associated kernel channel satisfy The core consists of one channel (corresponding to the explosion disturbance isolation core matrix).

[0080] 3> Spatial convolution is performed on the cumulative hydrological channel group and the short-term disturbance channel group respectively. The spatial convolution adopts... The convolution kernel has a stride of 1 and edge padding of 1, ensuring that the spatial dimensions of the convolution output remain the same. Each channel group first undergoes intra-group convolution to extract local interaction structures, and then... The number of channels is adjusted during convolution, followed by batch normalization and ReLU activation. If each channel group outputs 32 channels, the size of the feature map output by the cumulative hydrological channel group and the feature map output by the short-term perturbation channel group is [size missing]. .

[0081] Furthermore, the output feature maps of the two channel groups are concatenated along the channel dimension to obtain the intermediate interactive feature map. ,in, Indicates the first In the training samples, the th The intermediate interaction feature map of the day, with a size of Intermediate interaction feature map It includes both cumulative hydrological effects and short-term blasting disturbance effects, but the two effects have been extracted separately in the shallow layer, which can reduce feature confusion.

[0082] 4> Intermediate interaction feature map Perform cross-group fusion.

[0083] In the specific implementation, one is used. The convolutional layer remixes the 64 channels and then uses global average pooling to... Spatial location is compressed into channel vectors to obtain the first... The spatial encoding vector of the day ,in, Indicates the first In the training samples, the th The spatial encoding vector of the sky, with dimension denoted as . , This represents the spatial encoding dimension, for example, 128. If the number of channels after global average pooling is less than 128, another fully connected layer can be added to map to 128 dimensions, and the hyperbolic tangent activation function can be used to limit the value to 128. Within the range.

[0084] 5> Regarding the first The above processing is performed on the 60-day interaction feature maps of each training sample day by day to obtain the spatial coding sequence. ,in, Indicates the first The spatial encoding sequence corresponding to each training sample is given by Arranged in chronological order by date, with a size of [missing information]. .

[0085] It should be noted that the cumulative impact of hydrology has a continuous duration over multiple days, while the blasting disturbance has a short-term impact. The two are different in terms of physical mechanism. This step adopts the method of first extracting in groups and then fusing across groups, which can enable the accelerated deformation identification model to retain the independent expression of different disturbance mechanisms before learning the high-level state, thereby improving the ability to distinguish between the two types of situations: "no continuous deformation after blasting" and "continuous accelerated deformation after rainfall".

[0086] S302, Physical Activity Gated State Network Physical activity-gated state networks are used to process spatially encoded sequences in chronological order. The network outputs the probability that the 60-day observation window belongs to the accelerated deformation state. Based on a gated cyclic structure, the network incorporates a physical activity intensity vector into the update and reset gates, enabling the gating process to perceive which type of slope physical quantity dominates the current date. The specific steps are as follows: 1> Input tensors based on physical association Calculate the first The physical activity intensity vector of the day ,in, Indicates the first In the training samples, the th The five-dimensional physical activity intensity vector of the day has five components that correspond to the daily cumulative displacement, daily displacement rate, daily rainfall, daily peak velocity of blasting vibration, and groundwater level depth, respectively.

[0087] In the specific implementation, the physical association input tensor Middle and the first The first type of slope physical quantity related The absolute values ​​of the row interaction values ​​are averaged to obtain the first row. The daily interaction activity values ​​of slope-like physical quantities are then normalized using the 95th percentile of these activity values ​​from the training set, and the results are truncated to the range of 0 to 1 to obtain the physical activity intensity vector. The Each component. Regarding the groundwater level depth, since step S201 has already performed directional consistency processing, the greater the normalized activity intensity, the stronger the risk of groundwater activity.

[0088] In one embodiment, as an example, let the physical association input tensor of the nth training sample on day t be... The five interaction values ​​in the fourth row (corresponding to the peak daily blast vibration velocity) are as follows: Take the average of the absolute values ​​to get If the 95th quantile of the daily peak velocity of blasting vibrations in the training set is 2.0, then the normalized activity intensity is: This value is the physical activity intensity vector. The fourth component.

[0089] 2> The first The spatial encoding vector of the day Hidden status the previous day and physical activity intensity vector Common input gating calculation yields the updated gate. and reset door .in, Indicates the first In the training samples, the th The hidden state of the sky, with dimensions of , This represents the dimension of the hidden state, for example, it can be 128; when hour, Initialize as a vector of all zeros.

[0090] It should be noted that the door was updated. Used to control the proportion of new information written to the hidden state, and to reset the gate. Used to control the degree to which candidate hidden states depend on historical hidden states.

[0091] In one implementation, both the update and reset gates employ a sigmoid activation function to output a gating value between 0 and 1, and a trainable mapping term corresponding to the physical activity intensity vector is added to the gating linear transformation. This trainable mapping term can learn the influence of different slope physical quantities on the gating behavior. For example, when the daily peak velocity of blasting vibration is high but the daily displacement rate is low, the network can learn to reduce the proportion of this short-term impact written into the long-term hidden state; when the daily rainfall and groundwater level depth are continuously high and the daily displacement rate gradually increases, the network can learn to enhance the retention of historical states and the expression of risk accumulation.

[0092] In the specific implementation, update the gate. and reset door The physical activity intensity vector is introduced for calculation, and it is expressed as: ; ; in, This represents the sigmoid activation function, with an output range of... ; , This represents the trainable weight matrix corresponding to the spatial encoding vector, with size . ; , This represents the trainable weight matrix corresponding to the previous hidden state, with size . ; , This represents the trainable weight matrix corresponding to the physical activity intensity vector, with size . ; , This represents the trainable bias vector for the corresponding gate, with dimension . .

[0093] In one embodiment, for example, let the hidden state dimension be... , The dimension is 128. The dimension is 128. The dimension is 5. For size The trainable matrix. When the daily peak velocity of blasting vibration is 0.9 and the daily displacement rate is 0.2, The output is a 128-dimensional vector, in which the values ​​are significantly affected by the explosion-related dimensions. After training, this mapping can make the corresponding update gate components approach 0, thereby reducing the proportion of explosion impact information written into the long-term hidden state.

[0094] 3> According to the reset door Generate candidate hidden states , Indicates the first In the training samples, the th The candidate hidden states of the day, with dimensions of .

[0095] In the actual implementation, the previous day's state will be hidden. With Reset Door Element-wise multiplication, then with the spatial encoding vector The candidate hidden states are obtained by combining a linear layer of candidate states with a hyperbolic tangent activation function.

[0096] 4> According to the update gate merge the hidden state from the previous day and candidate hidden state , obtained the Hidden state ,in, Indicates the first In the training samples, the th The slope status accumulated over the days up to the current date is represented by a dimension of [missing information]. The larger the update gate, the higher the proportion of the current candidate hidden state being written; the smaller the update gate, the higher the proportion of the previous day's hidden state being retained.

[0097] In the specific implementation, the hidden state update is represented as follows: ; in, This represents element-wise multiplication, where both vectors involved in the operation have dimensions of 1. The result dimension is still ; Indicates and A vector of all 1s with the same dimension.

[0098] 5> Perform hidden state updates sequentially from day 1 to day 60 to obtain the hidden state on day 60. Hidden status on day 60 Indicates the first The training sample represents the overall slope condition over the entire 60-day observation window, including multiphysics interaction spatial patterns, date evolution information, and physical activity gating information.

[0099] Furthermore, the 60th day will be in a hidden state. Input the binary classification output layer to obtain the accelerated deformation prediction probability. ,in, Indicates the first The predicted probability of a training sample belonging to the accelerated deformation state, with a value ranging from 0 to 1.

[0100] In practical implementation, the binary classification output layer can consist of a fully connected layer and a sigmoid activation function; when When the value exceeds the preset discrimination threshold, the corresponding 60-day observation window is determined to be in an accelerated deformation state. The preset discrimination threshold can be 0.5, or it can be adjusted to 0.6 or 0.7 according to the validation set recall requirements.

[0101] 6> In hidden state every day Upper structure auxiliary daily risk score ,in, Indicates the first In the training samples, the th The auxiliary daily risk score ranges from 0 to 1. This score is obtained by passing the hidden state through a fully connected layer and a sigmoid activation function. It is not used as the final alarm result, but rather to constrain the risk evolution process within the model.

[0102] In one implementation, daily risk scoring is used as an aid. From hidden state It is obtained through a fully connected layer and a sigmoid activation function, and the calculation method is expressed as follows: ; in, The trainable weight matrix representing the auxiliary risk score has a size of . ; The trainable bias of the auxiliary risk score is represented by a scalar.

[0103] In one embodiment, for example, let the hidden state dimension be... One day ,but This indicates that the auxiliary risk score for that day was approximately 0.77.

[0104] Furthermore, the difference between the daily risk scores for adjacent dates is averaged squared to obtain a trend smoothing regularization term. The trend smoothing regularization term is used to suppress drastic jumps in risk scores caused by single-day blasting pulses, sensor noise, or occasional data anomalies, but it does not prevent the risk score from gradually increasing during continuous rainfall, shallower groundwater levels, and increased daily displacement rates.

[0105] In the specific implementation, for the nth training sample, the trend smoothing regularization term... Then for a batch training samples, batch trend smoothing regularization term ,in, This indicates the number of training samples in a training batch, i.e., the batch size.

[0106] S303, Loss Function Calculation and Model Parameter Training The total loss function of the accelerated deformation recognition model is defined as a weighted sum of weighted binary cross-entropy loss and trend smoothing regularization term. The former drives the model to output accurate window-level accelerated deformation prediction probability, while the latter constrains the temporal smoothness of the daily risk scores within the model. The specific steps are as follows: 1> Based on the probability prediction of accelerated deformation and real category labels Calculate the weighted binary cross-entropy loss Specifically, it is implemented using the standard weighted binary cross-entropy loss, which mitigates the impact of class imbalance on model training by assigning different weight factors to training samples of different classes.

[0107] In practical implementation, since the historical monitoring data of open-pit mine slopes typically contains more stable state samples than accelerated deformation state samples, class weights can be set based on the number of stable state samples and accelerated deformation state samples in the training set, allowing accelerated deformation state samples to receive a higher loss weight during training. If the number of stable state samples is... The number of accelerated deformation state samples is Then the weight of the accelerated deformation state can be taken as The steady-state weight can be 1.

[0108] 2> Calculate the trend smoothing regularization term based on the auxiliary daily risk score.

[0109] In practice, the difference between the auxiliary daily risk scores of two adjacent days within each training sample is squared, and then the average is calculated from day 2 to day 60 to obtain the trend smoothing regularization term for that training sample; then the average is calculated for all training samples within a training batch to obtain the batch trend smoothing regularization term.

[0110] 3> The weighted binary cross-entropy loss and the trend smoothing regularization term are weighted and summed, i.e. The total loss function is obtained. Trend smoothing regularization term weights The value can be between 0.05 and 0.2, with 0.1 being the preferred value. If the weight of the trend smoothing regularization term is too small, it will be difficult to suppress the risk jumps caused by daily noise; if the weight of the trend smoothing regularization term is too large, it may excessively restrict the rapid rise of the real acceleration process. Therefore, it can be selected through the validation set.

[0111] 4. A labeled historical open-pit mine slope monitoring sample is used to train an accelerated deformation recognition model. The adaptive fractional-order decoupling in step S201 and the physical association input tensor construction in step S202 are performed as fixed preprocessing steps; the convolution parameters of the grouped spatiotemporal convolutional encoder, the gating parameters of the physical activity gating state network, the physical activity intensity mapping parameters, the binary classification output layer parameters, and the auxiliary daily risk scoring parameters are updated through backpropagation. The optimizer can be the Adam optimizer, with an initial learning rate of 0.001, a batch size of 16 or 32, and 100 to 200 training epochs, and early stopping is performed based on the validation set loss or the validation set F1 score.

[0112] After each training epoch, the accelerated deformation recognition model is evaluated using a validation set, and accuracy, recall, F1 score, and area under the receiver operating characteristic (AUC) curve are calculated. Due to the high risk of missed detections in accelerated deformation, the model parameters that achieve the highest recall rate for accelerated deformation states and the highest F1 score are prioritized as the final model parameters.

[0113] S4. Online monitoring and accelerated deformation early warning of open-pit mine slopes After training the accelerated deformation recognition model, the trained model is deployed to the online monitoring system for open-pit mine slopes. During online application, the system reads the latest monitoring records daily, maintaining the same data organization, preprocessing, and judgment process as during the training phase. The specific steps are as follows: The online monitoring system reads daily cumulative displacement, daily displacement rate, daily rainfall, daily peak blasting velocity, and groundwater level depth in the open-pit mine slope monitoring area, maintaining a sliding observation window of 60 days. Whenever new day's monitoring data arrives, the earliest day's data is deleted and the latest day's data is added, forming an online raw monitoring matrix. ,in, This represents the 60-day online open-pit mine slope monitoring data to be identified, with dimensions of [size missing]. The order of the five columns of slope physical quantities is completely consistent with that of the training phase.

[0114] Furthermore, the online original monitoring matrix Perform missing value imputation, outlier truncation, adaptive fractional-order decoupling, and robust scaling uniformity consistent with the training phase to obtain the online decoupled feature matrix. ,in, This represents the decoupled feature matrix corresponding to the current observation window to be identified, with a size of [missing information]. Online robust scaling must use the median and interquartile range saved during the training phase, without recalculating the scaling parameters using the current 60-day window, to ensure that the online input and training input are on the same numerical scale.

[0115] Furthermore, based on the three physical correlation kernel matrices saved during the training phase, the online decoupling feature matrix is... Convert to online physical correlation input tensor ,in, This represents the physical association input tensor corresponding to the current observation window to be identified, with a size of [size missing]. The physical correlation kernel matrix remains fixed during the online phase and does not change with a single online window, thus avoiding unstable fluctuations in the early warning results due to short-term noise.

[0116] Furthermore, the online physical correlation input tensor Input the trained accelerated deformation recognition model. A grouped spatiotemporal convolutional encoder extracts multiphysics interaction spatial patterns daily; a physical activity gating state network updates the hidden states in date order; and a binary classification output layer outputs the online accelerated deformation prediction probability. ,in, This represents the predicted probability that the current 60-day online observation window belongs to the accelerated deformation state, with a value ranging from 0 to 1.

[0117] Furthermore, based on the online accelerated deformation prediction probability Output the warning result. When When the value is less than 0.5, the output is in a stable state; when... When the value is greater than or equal to 0.5 and less than 0.7, an alert status is displayed, prompting technical personnel to strengthen the review; when... When the value is greater than or equal to 0.7, an accelerated deformation warning is issued. To reduce false alarms caused by single-day anomalies, a continuous confirmation rule can be set, such as triggering a formal alarm when two consecutive online observation windows reach the accelerated deformation warning threshold.

[0118] While outputting the early warning results, the average value of the physical activity intensity vector over the past 7 days is calculated to obtain the dominant physical factor score. If the scores corresponding to daily rainfall and groundwater level depth are high, it indicates that the current risk is mainly related to hydrological accumulation; if the daily peak velocity score of blasting vibration is high but the daily displacement rate score is not high, it indicates that the current risk may be a short-term blasting disturbance, which needs to be verified in conjunction with subsequent displacement development; if the scores of daily displacement rate and daily cumulative displacement continue to rise, it indicates that the current risk has manifested as enhanced deformation response, and the corresponding slope area needs to be inspected first.

Claims

1. A method for identifying the displacement trend of open-pit mine slopes, characterized in that, Includes the following steps: Data on open-pit mine slope monitoring was collected, including daily cumulative displacement, daily displacement rate, daily rainfall, daily peak velocity of blasting vibration, and groundwater level depth. The collected monitoring data were labeled, and an original monitoring matrix containing time series features was constructed. The original monitoring matrix was then preprocessed. An adaptive fractional-order decoupling process is performed on each original monitoring matrix to obtain a decoupling feature matrix. A physical correlation input tensor is constructed based on the physical coupling relationship between slope physical quantities. An accelerated deformation recognition model is constructed based on a grouped spatiotemporal convolutional encoder and a physical activity gating state network. The physical correlation input tensor is input into the accelerated deformation recognition model, and the accelerated deformation prediction probability is output to obtain the early warning result. Specifically, an adaptive fractional-order decoupling process is performed on each original monitoring matrix to obtain the decoupling feature matrix, as follows: For the first The original monitoring matrix of each training sample was read column by column to obtain 5 one-dimensional time series; For the first The first-order backward difference is used to calculate the physical quantities of slope-like structures to obtain local variations. The mutation ratio is calculated based on the local change amount on the current date and the background change level in the local neighborhood; According to the Calculate the trend continuity coefficient of the changes in physical quantities of slope-like structures over the past 3 days. An adaptive fractional order is generated based on the mutation ratio and the trend continuity coefficient. The fractional-order memory weights are generated based on the adaptive fractional-order order, and the original monitoring values ​​of the current date and historical dates are back-weighted to obtain the fractional-order memory results. Calculate the first Short-term trend terms of physical quantities similar to slopes; By fusing the fractional memory results and the short-term trend term, we obtain the eigenvalues ​​before decoupling. Robust scaling was applied to the undecoupled eigenvalues ​​of the five types of slope physical quantities to obtain the undecoupled eigenvalues. The first The decoupling feature values ​​of 60 days and 5 types of slope physical quantities in the training samples are arranged in order of date and physical quantity to obtain the decoupling feature matrix; The physical correlation input tensor is constructed based on the physical coupling relationship between the physical quantities of the slope, specifically as follows: For the first The training sample at the th ... The five decoupling eigenvalues ​​of a day are arranged in a fixed order to form a decoupling eigenvector; Based on the physical mechanism of slope, three physical correlation kernel matrices are constructed. The physical correlation kernel matrices are used to assign different weights to different interaction positions in the instantaneous interaction matrix. Calculate the statistical correlation strength among five types of slope physical quantities within the training set, and perform a light correction on the three physical correlation kernel matrices; The first The instantaneous interaction matrix of day is multiplied element-wise by each physical correlation kernel matrix to obtain the first... Heavenly Interactive feature maps under each core channel; Stack all interactive feature maps from the past 60 days in the order of date, physical quantity row, physical quantity column, and kernel channel to obtain the physical correlation input tensor; The method for accelerating the training of deformation recognition models is as follows: The grouped spatiotemporal convolutional encoder groups channels according to the meaning of the physical association kernel channels, extracts hydrological cumulative features and blasting disturbance features respectively, and then performs cross-group fusion. The physical activity gating state network is based on a gating loop structure. The physical activity intensity vector is added to the update gate and reset gate, so that the gating process can perceive which type of slope physical quantity dominates the current date. The total loss function is composed of a weighted sum of the weighted binary cross-entropy loss and the trend smoothing regularization term. A deformation recognition model was trained using labeled historical open-pit mine slope monitoring samples.

2. The method for identifying the displacement trend of an open-pit mine slope according to claim 1, characterized in that, Using a continuous 60-day observation window, training samples are extracted from historical open-pit mine slope monitoring data. The training samples were organized into the original monitoring matrix. , Indicates the first Each training sample corresponds to 60 days of open-pit mine slope monitoring data, with a size of [missing information]. .

3. The method for identifying the displacement trend of an open-pit mine slope according to claim 1, characterized in that, The grouped spatiotemporal convolutional encoder groups channels according to the meaning of the physically associated kernel channels, extracts hydrological cumulative features and blasting disturbance features separately, and then performs cross-group fusion, specifically as follows: From the Extract the th training sample's physical association input tensor. A slice of the day was obtained. Sky interaction feature map; According to the function of the physically associated nuclear channels, the first The interactive feature map is divided into two channel groups: the first channel group is the cumulative hydrological channel group, and the second channel group is the short-term disturbance channel group. Spatial convolution is performed on the cumulative hydrological channel group and the short-term disturbance channel group respectively. The output feature maps of the two channel groups are concatenated along the channel dimension to obtain the intermediate interactive feature map. Perform cross-group fusion on intermediate interaction feature maps; For the first The above processing is performed on the 60-day interaction feature maps of each training sample day by day to obtain the spatial coding sequence.

4. The method for identifying the displacement trend of an open-pit mine slope according to claim 1, characterized in that, The physical activity gating state network is based on a gating loop structure. It incorporates physical activity intensity vectors into the update and reset gates, enabling the gating process to perceive which type of slope physical quantity dominates the current date. Specifically: Calculate the first based on the physically associated input tensor. The physical activity intensity vector of the day; The first The spatial encoding vector of the day, the hidden state of the previous day, and the physical activity intensity vector are input into the gating calculation to obtain the update gate and the reset gate. The previous day's hidden state is multiplied by the reset gate element, and then fed together with the spatial encoding vector into the candidate state linear layer and the hyperbolic tangent activation function to obtain the candidate hidden state; Based on the hidden state and candidate hidden state of the day before the update gate fusion, we obtain the... Hidden state; The hidden state is updated sequentially from day 1 to day 60 to obtain the hidden state on day 60. An auxiliary daily risk score is constructed on the hidden state of each day. The auxiliary daily risk score is obtained by passing the hidden state through a fully connected layer and a sigmoid activation function. It is not used as the final alarm result, but is used to constrain the risk evolution process inside the model.

5. The method for identifying the displacement trend of an open-pit mine slope according to claim 1, characterized in that, The total loss function is composed of a weighted sum of the weighted binary cross-entropy loss and the trend smoothing regularization term, specifically: Calculate the weighted binary cross-entropy loss based on the accelerated deformation prediction probability and the true category label; The difference between the auxiliary daily risk scores of two consecutive days within each training sample is squared, and then averaged over the period from day 2 to day 60 to obtain the trend smoothing regularization term for that training sample. Then, the average of all training samples within a training batch is calculated to obtain the batch trend smoothing regularization term; The total loss function is obtained by weighting and summing the weighted binary cross-entropy loss with the trend smoothing regularization term.

6. The method for identifying the displacement trend of an open-pit mine slope according to claim 1, characterized in that, The warning results are as follows: when the predicted probability is less than 0.5, a stable state is output; when the predicted probability is greater than or equal to 0.5 and less than 0.7, a state of concern is output, prompting technical personnel to strengthen the review; when the predicted probability is greater than or equal to 0.7, an accelerated deformation warning is output.

7. The method for identifying the displacement trend of an open-pit mine slope according to claim 6, characterized in that, While outputting the early warning results, the average value of the physical activity intensity vector over the past 7 days is calculated to obtain the dominant physical factor score: if the scores corresponding to daily rainfall and groundwater level depth are high, it indicates that the current risk is mainly related to hydrological accumulation; if the daily peak velocity score of blasting vibration is high but the daily displacement rate score is not high, it indicates that the current risk may be a short-term blasting disturbance, which needs to be reviewed in conjunction with subsequent displacement development. If the scores for daily displacement rate and daily cumulative displacement continue to rise, it indicates that the current risk has manifested as enhanced deformation response, and the corresponding slope area needs to be inspected first.

Images