A high and steep slope partition dynamic early warning method for open-pit mine

Abstract

The application discloses a kind of open-pit mine high and steep slope partition dynamic early warning method, belong to open-pit mine slope safety monitoring and disaster early warning technical field, to solve the problem that unified threshold of whole slope is difficult to adapt to partition difference, background drift and local anomaly cross-zone expansion, the application is by constructing version early warning partition and physical directed partition graph, event isolation and stable background update are implemented to monitoring data, generate partition dynamic threshold under the constraint of whole mine false alarm budget, and form propagation evidence with direction, time lag, same direction residual and different kind index mutual evidence, combined with delay rule output partition linkage early warning, realize the technical effect of reducing false alarm caused by event disturbance and identifying local anomaly expansion.

Description

Technical Field

[0001] This invention relates to the field of open-pit mine slope safety monitoring and disaster early warning technology, and in particular to a method for dynamic early warning of steep open-pit mine slopes by zone. Background Technology

[0002] Steep slopes in open-pit mines typically consist of multiple steps and sections, with each spatial unit differing in lithology, structural plane combinations, potential slip surfaces, excavation disturbance, water catchment conditions, and support status. Existing slope monitoring systems usually collect data on displacement, displacement rate, displacement acceleration, pore water pressure or water level, rainfall, vibration, cracks, rockfalls, and support stress, and issue early warnings based on pre-set thresholds.

[0003] When using uniform displacement, displacement rate, or acceleration thresholds for the entire slope, it is difficult to represent the background differences between different steps and slope sections. As the mining area advances, seasonal and water-bearing conditions change, and monitoring baselines drift, the deviation between the fixed thresholds and the actual background will continue to accumulate. If the event responses caused by blasting, rainfall, and disposal operations are directly written into the background sample, it will raise or distort subsequent thresholds, while equipment anomalies and short-term fluctuations may cause frequent false alarms.

[0004] Furthermore, when local anomalies extend to adjacent zones, simply comparing whether each zone independently crosses the threshold makes it difficult to distinguish between propagation processes with structural, hydraulic, unloading, or support transmission basis and anomalies that coincidentally overlap in time, resulting in insufficient identification of the linkage between local damage and the overall instability development stage.

[0005] Therefore, a dynamic early warning method for high and steep slopes in open-pit mines is needed to address the shortcomings of existing technologies. Summary of the Invention

[0006] One objective of this invention is to propose a dynamic early warning method for steep slopes in open-pit mines. Addressing the shortcomings of existing technologies that use uniform or fixed thresholds to fail to reflect differences in zoning backgrounds, where event data contaminates stable backgrounds, and where the propagation of local anomalies lacks physical constraints, this invention proposes a technical solution combining versioned early warning zoning with directed zoning maps, isolation between real-time risk channels and stable learning channels, stable sample background updates, hierarchical dynamic thresholds under mine-wide false alarm budget constraints, and cross-regional propagation gating. This invention sequentially converts monitored quantities into event labels, background versions, real-time residuals, inconsistencies in indicators, hierarchical dynamic thresholds, and evidence of cross-regional propagation. It maintains output stability through threshold freezing, amplitude limiting, rollback, and hysteresis-based grading, achieving the technical effects of adapting to zoning differences and background changes, suppressing false alarms caused by event disturbances, and identifying the expansion of local anomalies.

[0007] This invention provides a method for dynamic early warning of steep slopes in open-pit mines, comprising:

[0008] S1. Obtain engineering geological data, hydrological data and mining status data of open-pit mine slopes, divide early warning zones, and construct a directed partition map with edge direction and allowable time delay based on the physical influence relationship between the zones.

[0009] S2. Obtain monitoring data and event information for each early warning zone, identify events and their recovery periods, and isolate the monitoring data into real-time risk data and stable learning candidate data;

[0010] S3. Filter the stable learning candidate data to update the partitioned stable sample pool and partitioned background quantity, calculate the real-time residual quantity of real-time risk data relative to the partitioned background quantity, and generate stable sample index inconsistency quantity and real-time index inconsistency quantity based on the deviation of stable samples relative to the partitioned background quantity and the real-time residual quantity, respectively.

[0011] S4. Based on the distribution of inconsistencies in the stable sample indicators, generate and update the hierarchical dynamic thresholds for each early warning zone under the constraint of the whole mine's false alarm budget.

[0012] S5. Compare the real-time inconsistency of the indicators with the hierarchical dynamic threshold to determine the local early warning level; when the source early warning partition is abnormally persistent, the propagation direction is consistent with the edge direction of the directed partition graph, and the target early warning partition shows the same direction deformation real-time residual within the allowable time delay and is mutually verified by monitoring indicators of different physical types, cross-regional propagation evidence is formed and a partition linkage early warning is generated.

[0013] Optionally, S1 includes: forming initial spatial units with step boundaries and slope boundaries, and merging or splitting the initial spatial units according to the combination of lithology and structural surfaces, potential slip surfaces, excavation sequence, water catchment relationship and support continuity to obtain early warning zones;

[0014] When the excavation outline, support status, or water catchment relationship changes, a new partition version identifier is generated, and the spatial correspondence between the old and new warning partitions is saved.

[0015] The edge types are structural continuity, hydraulic connection, unloading transfer relationship and weak support continuity, respectively. The edge direction is determined according to the direction of action transfer, and the allowable time delay is determined according to the historical response time difference from the source early warning zone to the target early warning zone for the corresponding action.

[0016] Optionally, S2 includes: collecting displacement, displacement rate, displacement acceleration, pore water pressure or water level, rainfall, vibration, crack change or rockfall count, support stress, and work log.

[0017] Transform all data to a unified time base and perform health checks on missing measurements, over-range measurements, timestamp jumps, persistent zero values, and sensor drift.

[0018] The event start point is identified based on the blasting time, the construction period, and the rainfall event threshold determined by the rainfall duration and rainfall amount in the work log. The recovery period end point is determined by the moment when the post-event monitoring data returns to the preset quantile interval of the pre-event background distribution and remains continuously for a preset duration, and the event tag is generated.

[0019] Furthermore, the real-time risk data includes all data that has currently passed health verification;

[0020] The stable learning candidate data is selected from data that has passed health verification during non-event periods and non-recovery periods, so that health data during non-event periods can be used for both current anomaly judgment and stable learning screening.

[0021] Data that passes health verification during the event occurrence and recovery period is used only as real-time risk data and is used to determine the current real-time residual based on the difference between the real-time risk data and the corresponding partition background amount, and is prohibited from being written into the stable sample pool;

[0022] For monitoring equipment that fails the health verification, stop background updates for the corresponding indicators and retain the equipment abnormality flag.

[0023] Optionally, S3 includes: establishing a sample pool index for combinations of early warning zones, seasons, excavation stages, and water-bearing conditions;

[0024] Data that do not have event tags, whose corresponding monitoring equipment has passed health verification, whose trend changes of each deformation index within the stability window do not exceed the deformation stability boundary determined by the quantiles of historical stable samples, and whose changes of at least one hydrological index or support force index have passed health verification and are located within the stability boundary of heterogeneous index determined by their respective historical stable sample quantiles, are identified as samples that meet the stability conditions and are mutually verified by heterogeneous indicators.

[0025] The partition background quantity is updated based on the time-weighted statistics of the sample, and the update time, sample pool index, sample range and verification results are saved to form a background version.

[0026] Furthermore, the real-time residual is the difference between the real-time risk data and the corresponding partition background data;

[0027] The deviations of real-time residuals and stable samples relative to the partition background are scaled and normalized according to the historical stable sample distribution of each indicator. The indicator weights are determined according to the health status of each indicator and the mutual verification results of heterogeneous indicators. The normalized real-time residuals and stable sample deviations are weighted and aggregated using the same indicator weights and aggregation rules to obtain the real-time indicator inconsistency and stable sample indicator inconsistency.

[0028] When the number of false alarms in the historical stable sample replay check after the background version is updated exceeds the false alarm benchmark determined by the historical replay results of the uncontaminated background version, or when the proportion of data in the sample that is subsequently detected as event labels exceeds the contamination criterion determined by the historical uncontaminated background version, the background version is marked as a contaminated version and rolled back to the most recent historical background version that was not marked as a contaminated version.

[0029] Optionally, S4 includes: determining the budget allocation weight based on the static risk score of each early warning zone, the importance of the protected object, and the number of cross-regional propagation evidences formed by the previous threshold statistical period, propagating along the edge pointing to the early warning zone in the directed partitioning graph and still valid within the current allowable time delay, and the corresponding edge type; and allocating the entire mine's false alarm budget to the false alarm budget share of each early warning zone according to the budget allocation weight.

[0030] Using the inconsistency of the stable sample index as calibration data, the hierarchical dynamic thresholds for each early warning zone are generated based on the empirical quantiles corresponding to each false alarm budget share.

[0031] The online coverage error is determined based on the difference between the actual false alarm ratio and the false alarm budget share within the online statistical period, and the empirical quantile position of the next threshold statistical period is corrected based on the online coverage error.

[0032] Furthermore, when the number of valid monitoring indicators in any early warning zone is less than the minimum number of indicators configured for that early warning zone, or when the abnormality determined by the monitoring equipment status continues to exist until the preset equipment verification time, the hierarchical dynamic threshold of the most recent verification of that early warning zone is frozen.

[0033] When the change of the corrected hierarchical dynamic threshold relative to the previous threshold exceeds the single-cycle change upper limit determined by the historical stable threshold change distribution, the threshold change is limited according to the single-cycle change upper limit.

[0034] When the online coverage error increases for two consecutive statistical periods, it is rolled back to the threshold version where the online coverage error did not increase.

[0035] Optionally, S5 includes: when the real-time inconsistency of the indicators exceeds the corresponding level of the dynamic threshold and is continuously maintained at the confirmation duration determined by the quantile of the response duration of the historical events in the warning zone, and when the deformation indicator and at least one of the hydrological indicators, support stress indicators, crack change or rockfall count have a common direction of change pointing to the development of instability, the corresponding local warning level is determined.

[0036] The propagation judgment is made only along the edge from the source warning partition to the target warning partition in the directed partitioning graph. The moment when the source warning partition crosses the hierarchical dynamic threshold is taken as the starting point of the allowable time delay. The potential slip direction of the target warning partition is determined according to the potential slip surface geometry information in the spatial structure data. The cross-regional propagation evidence and its propagation confidence score are determined by the mutual verification result of the displacement residual component along the potential slip direction in the real-time residual and at least one of the heterogeneous indicators.

[0037] Furthermore, a hysteresis control is implemented on the partitioned linkage early warning level using preset linkage upgrade level, propagation upgrade threshold, local release threshold and propagation release threshold. The local release threshold is obtained by subtracting the local hysteresis amplitude from the hierarchical dynamic threshold corresponding to the current local early warning level, and the propagation release threshold is obtained by subtracting the propagation hysteresis amplitude from the propagation upgrade threshold.

[0038] When the local warning level reaches the preset linkage upgrade level and the propagation confidence score reaches the propagation upgrade threshold, the regional linkage warning level is upgraded. When the real-time indicator inconsistency does not reach the local release threshold, the propagation confidence score does not reach the propagation release threshold, and the cross-regional propagation evidence is continuously missing for the preset release time after the corresponding allowable time delay ends, the regional linkage warning level is downgraded or released.

[0039] The partitioned linkage early warning record includes the starting early warning partition, partition version identifier, propagation path, edge type, trigger indicator, propagation confidence score, early warning start and end time, and handling scope;

[0040] The cross-regional propagation evidence is used to form a threshold sensitivity correction amount, and the budget allocation weight of the corresponding early warning partition is adjusted in the next threshold statistical period. The cross-regional propagation evidence is prohibited from being written into the stable sample pool.

[0041] The beneficial effects of this invention are:

[0042] 1. By keeping the data of the event occurrence period and recovery period in the real-time risk channel and excluding it from the stable learning channel, and only using samples that meet the stability conditions and are mutually verified by heterogeneous indicators to update the versioned background, the contamination of the background baseline by blasting, rainfall, disposal construction and equipment anomalies can be reduced.

[0043] 2. By using the inconsistency of indicators from stable samples as calibration data, under the constraint of the mine-wide false alarm budget, a hierarchical dynamic threshold is generated according to the static risk, importance, and effective evidence of edge propagation in each zone. Combined with coverage error correction, equipment anomaly freezing, threshold change limit, and version rollback, the threshold can be updated according to the zone's operating conditions and background status.

[0044] 3. By verifying the continuous anomalies of the source partition, the side direction and allowable time delay, the same-direction deformation residuals of the target partition and the heterogeneous indicators on four types of directed edges (structure, hydraulic, unloading and support), and by implementing hysteresis judgment using upgrade threshold and release threshold, it is possible to suppress accidental synchronous anomalies and level fluctuations, and at the same time form a traceable propagation path and handling scope. Attached Figure Description

[0045] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0046] Figure 1 This is a flowchart of a dynamic early warning method for high and steep slopes in open-pit mines according to the present invention.

[0047] Figure 2 This is a flowchart of the present invention, S4, which generates and updates the dynamic thresholds of each early warning zone under the constraint of the whole mine false alarm budget. Detailed Implementation

[0048] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0049] refer to Figures 1-2 A dynamic early warning method for high and steep slopes in open-pit mines, comprising:

[0050] S1. Obtain engineering geological data, hydrological data, and mining status data of open-pit mine slopes, divide the early warning zones, and construct a directed partition map with edge direction and allowable time delay based on the physical influence relationship between the zones.

[0051] S2. Obtain monitoring data and event information for each early warning zone, identify events and their recovery periods, and isolate the monitoring data into real-time risk data and stable learning candidate data;

[0052] S3. Filter the stable learning candidate data to update the partitioned stable sample pool and partitioned background quantity, calculate the real-time residual quantity of real-time risk data relative to the partitioned background quantity, and generate stable sample index inconsistency quantity and real-time index inconsistency quantity based on the deviation of stable samples relative to the partitioned background quantity and the real-time residual quantity, respectively.

[0053] S4. Based on the distribution of inconsistencies in the stable sample indicators, generate and update the hierarchical dynamic thresholds for each early warning zone under the constraint of the whole mine's false alarm budget.

[0054] S5. Compare the real-time inconsistency of the indicators with the hierarchical dynamic threshold to determine the local early warning level; when the source early warning partition is abnormally persistent, the propagation direction is consistent with the edge direction of the directed partition graph, and the target early warning partition shows the same direction deformation real-time residual within the allowable time delay and is mutually verified by monitoring indicators of different physical types, cross-regional propagation evidence is formed and a partition linkage early warning is generated.

[0055] In this specific embodiment, S1 includes:

[0056] The slope early warning server reads the three-dimensional design model of the mine, monthly geological logging, borehole exposure, groundwater observation and mining plan, and converts them into the engineering coordinate system of the mining area. The spatial unit records are composed of partition version identifier, spatial unit identifier, step number, slope segment number, lithology code, structural surface group, potential slip surface number, water catchment unit number, support segment number and effective time. In this specific embodiment, the closed polygon obtained by the intersection of the step elevation boundary and the slope segment plane boundary is used as the initial spatial unit. The boundary coincidence error limit is set to 0.20 m. Small units with an area of ​​less than 25 m² are merged into the adjacent unit with the longest shared boundary and the same lithology.

[0057] The data preprocessor removes duplicate records by design model version number, logging batch and mining plan version number, linearly fills in missing elevation points by two adjacent valid points on the same step, splits self-intersecting polygons into two closed polygons along the first intersection point, and writes records with unclosed boundaries, unmatched lithology codes or coordinates exceeding the mining rights boundary into the queue to be reviewed. Only data that passes the coordinate range, topological closure and version consistency verification will enter the initial spatial unit construction.

[0058] The zoning module performs merging or splitting based on lithology and structural plane combination, potential slip surface connectivity, excavation sequence, water catchment relationship, and support continuity. Adjacent initial spatial units are merged when they simultaneously meet the following conditions: consistent lithology coding, angle difference between the main control structural plane groups not exceeding 10°, belonging to the same potential slip surface projection range, planned excavation time difference not exceeding 7 days, and continuous support sections. Any unit is split along the corresponding boundary when it crosses two potential slip surfaces, two independent water catchment outlets, or abrupt change in support stiffness boundary. The resulting early warning zoning record saves the set of member spatial units, three-dimensional boundary, area, slope aspect, slope angle, main control slip direction, and current mining stage.

[0059] When the excavation profile changes by more than 2.0 m, the status of anchor cables or retaining walls changes from valid to invalid, or the connectivity of the water catchment outlet changes, the version manager closes the validity period of the current partition version and generates an incrementing new partition version identifier, calculating the spatial inheritance rate for the old and new warning partitions. , Indicates the old partition To the new partition The spatial inheritance rate is a dimensionless number within [0,1]. This indicates that the old version was numbered as The warning zone polygon, This indicates that the number in the new version is The warning zone polygon, Indicates the old partition index. Indicates a new partition index. This represents the spatial intersection operation between the old and new partition polygons. The result is the set of polygons in the overlapping area, and the units are not applicable. This represents an area operator that calculates the horizontal projected area of ​​a polygon based on the engineering coordinates of the mining area, with the output unit being m².

[0060] When the spatial inheritance rate is not less than 0.60, the version manager writes the old partition identifier, new partition identifier, inheritance rate, overlapping area, reason for change and effective time into the partition correspondence table to save the spatial correspondence between the old and new warning partitions. The historical stable samples of the old partition are used only as candidate sources for the background initialization of the new partition. When there is no new partition with an inheritance rate not less than 0.60, the old partition is marked as the termination mapping. When a new partition corresponds to multiple old partitions, the mapping order is recorded in descending order of overlapping area. When the areas are the same, the order is determined in ascending order of the old partition identifier.

[0061] The graph construction module uses the current version's early warning zones as nodes and uses structural continuity, hydraulic connection, unloading transfer relationship, and weak support continuity as edge types. Structural continuity edges are triggered when the same main control structural surface or potential sliding surface passes through two zones without a complete rock mass in between. Hydraulic connection edges point from the upstream outlet of the same water catchment path to the downstream water-receiving zone. Unloading transfer edges point from the excavation zone to its normal unloading influence zone. Weak support continuity edges point from the section with reduced support stiffness to the adjacent affected zone along the continuous support strip. The four types of edges respectively store the source zone, target zone, edge type, direction of action, source of evidence, start and end effective time, and allowable time delay.

[0062] For each edge type, the graph construction module replays the effective hierarchical dynamic thresholds from the verified event database of the most recent 24 months. The moment when the source partition first crosses the corresponding warning level threshold is taken as the starting point of the historical source time delay. The moment when the same level and same direction deformation response of the target partition first appears and is mutually verified by heterogeneous indicators is taken as the starting point of the target response. The time difference between the two is extracted, and records with less than 0 days or more than 30 days are removed. The allowable time delay is determined by grouping according to edge type, mining stage, and warning level. , Indicates the source partition To target partition The allowable time delay under the corresponding warning level, in hours. Indicates the source warning partition index. This indicates the target warning partition index. This represents the operator that takes the 0.90 empirical quantile for the set of effective historical response time differences. Indicates the first In each verification event, the time difference between the target response start point and the moment when the source partition first crosses the corresponding level threshold is denoted as h. Indicates the event sample index; if the historical record has no replayable threshold version or the source partition has not exceeded the corresponding level threshold, it will not enter the time delay sample set of that level, so that the output of this S1 and S5 use the same time delay starting point benchmark.

[0063] When there are fewer than 8 valid historical response time differences in the same group, the allowable time delay is taken as the 0.90 empirical quantile of the whole mine sample of the same edge type and marked as the mine-level backtracking source. When there are still fewer than 8 samples of the same edge type in the whole mine, the physical transfer upper limit confirmed by the geological engineer is taken and the confirmer, the supporting document and the due date for review are saved in the edge record. When 5 new historical events are added or the partition version is updated, the allowable time delay is recalculated. The calculation result is limited to 0.5 h to 720 h.

[0064] The graph construction module removes self-loops that are identical in the source and target partitions, merges duplicate edges that are identical in the source, target, and edge types, retains edges with opposite directions but valid evidence, and writes the node table, edge table, partition correspondence table, and allowed time-delay version into the directed partition graph object for S2 to associate monitoring records, S4 to count incoming edge propagation evidence, and S5 to perform directed propagation judgment.

[0065] In this specific embodiment, S2 includes:

[0066] The monitoring access module uses the current partition version formed by S1 as the association benchmark, and binds the measuring points to the early warning partition according to the spatial coordinates of the equipment and the three-dimensional boundary of the partition. It collects GNSS data and inclination displacement (mm), displacement rate (mm / d), displacement acceleration (mm / d²), pore water pressure (kPa) or water level (m), rainfall (mm), vibration velocity (mm / s), crack change (mm) or rockfall count, anchor cable axial force (kN), and operation logs. The monitoring records include equipment identifier, partition identifier, partition version identifier, index type, original value, unit, sampling time, receiving time, quality code, and calibration version.

[0067] The time conversion module reads the device time synchronization table and... Transform all data to a unified time base. This indicates the unified timekeeping system for the mine, with the unit being seconds (s). Indicates the device's original time in seconds. Indicates equipment The clock correction amount in the current time interval, in seconds. The device index is represented by the time synchronization table, which stores the correction amount, network time synchronization deviation, and verification time using the device identifier and time synchronization interval as keys. When the corrected time is earlier than the previous record time, the record is marked as a timestamp and backtracked.

[0068] The time aligner distributes the converted records to a 10-minute time grid. Displacement, water pressure, water level, and support stress are taken as the median within the grid. Rainfall is taken as the cumulative value, vibration as the peak value, crack change as the final value minus the initial value, and rockfall count as the cumulative increment. The operation log retains the original start and end times. When there are no valid values ​​in the grid, a missing measurement code is written. No interpolation is performed for gaps spanning 60 minutes. The aligned records are written to the partitioned time-series cache.

[0069] For the rate derived from the displacement sequence, calculate , Indicates partition At any moment The displacement rate is expressed in mm / d. This represents the median displacement of the current 10-minute grid, in mm. This represents the median displacement of the previous grid cell, in mm. The time difference between adjacent grids is represented and converted to d. If the previous valid grid is missing or the time difference exceeds 60 min, the rate is not calculated and the missing test code is written.

[0070] Data completeness according to calculate, Indicates partition At any moment The completeness of the index is a dimensionless number within [0,1]. This indicates the number of validly formatted indicators received by the current grid, expressed in terms of items. This indicates the number of indicators that should be included in the equipment configuration list and must be a positive integer. When the indicator completeness is less than 0.60, the grid is marked as having incomplete partition data and the reason for the missing test is transmitted to the health verification module.

[0071] The health verification module reads the device range table and the most recent calibration record, performs verification for missing measurements, exceeding the range, timestamp retrogression or jump exceeding 30 minutes, unchanged zero values ​​for 12 consecutive sampling points, and sensor drift, and calculates the drift amount. , Indicates equipment At any moment The drift amount and the unit are consistent with the specifications of the equipment. Indicates the end time The median of the most recent 24-hour health data and its correlation with Rolling updates Indicates equipment The median of a fixed calibration baseline is formed over a stable 30-day period prior to the most recent field calibration completion time. Its source is recorded as the calibration version, the start and end dates of the 30 days, and the effective date. It is not updated until a new version is generated for the next field calibration. If both units are consistent with the equipment specifications, and the drift exceeds the upper limit of the same unit listed in the calibration certificate, it is judged as an abnormal drift.

[0072] A record in the equipment range table uses the equipment identifier and index type as a joint key to store the lower limit, upper limit, unit, number of zero value persistence points, drift upper limit, calibration version, and validity period. When the value is equal to the range boundary, it is treated as a valid record. When the value exceeds any boundary, an out-of-range anomaly is written. When the range table lacks a matching key, the index is set to pending verification and prohibited from entering the two purpose data sets. After the on-site calibration is completed, the health check of the most recent 24 hours is re-executed with the new version record.

[0073] For records that pass the health check, write health flag 1; for records that fail, write health flag 0 and abnormal codes for missing measurement, over-range, clock, zero value or drift. Equipment indicators that fail the check are retained in the audit database but are not included in real-time risk data and stable learning candidate data. At the same time, send an equipment abnormality flag to S3 to stop background updates of the indicator. The abnormality flag is cleared only after the equipment completes on-site verification and passes the check for 12 consecutive sampling points.

[0074] The event identifier reads the start and end times of blasting and the duration of subsequent operations from the work log, and calculates the cumulative rainfall over 6 hours. , Indicates partition At any moment The cumulative rainfall in the first 6 hours, in mm. Indicates the forward first Rainfall in 10-minute grids, in mm. The grid backtracking index indicates that when the cumulative rainfall is not less than 25 mm or the 24-hour cumulative rainfall calculated using the same method is not less than 50 mm, the moment when the threshold is first exceeded is taken as the starting point of the rainfall event. After the starting point of the rainfall event, when the rainfall in each 10-minute grid is less than 0.1 mm for 6 consecutive hours and the 6-hour cumulative rainfall at the end of the continuous window is less than 25 mm and the 24-hour cumulative rainfall is less than 50 mm, the end of the window is determined as the end of the rainfall event, and the recovery period determination is performed from the next 10-minute grid. If the rainfall in a single grid is not less than 0.1 mm or any cumulative rainfall exceeds the limit again before the end of the event is formed, the original rainfall event is extended and the continuous rainless timer is restarted, and no new duplicate events are created.

[0075] Rainfall thresholds are stored in the event rule table and are jointly calibrated by the rainfall sequence of the most recent 5 years and the verified slope response. The rule table stores the event type, window length, rainfall threshold, unit, version and effective time. When two rainfall conditions are crossed at the same time, they are merged into the same event and the first crossing time is taken. When the starting point of blasting, treatment construction and rainfall events coincide, the main event type is recorded in the order of blasting, treatment construction and rainfall and all sub-event references are retained.

[0076] Event recovery determination is performed by reading non-event data from the 30 days prior to the event that has passed health checks, categorized by partition and indicator type. Each indicator must have at least 144 valid 10-minute grids and retain at least one healthy deformable indicator. Background quantile intervals are then formed using empirical quantiles of 0.10 and 0.90, and the recovery coverage is calculated. , Indicates partition In continuous window The proportion of health indicators falling within the background quantile interval and being a dimensionless number within [0,1]. This indicates the number of times a health indicator falls within a given range. This represents the total number of health indicators to be assessed, and is a positive integer. This represents a continuous 6-hour window starting from the candidate recovery time.

[0077] When there are insufficient valid samples before the event, the recovery decision-maker sequentially reads the old partitions with a spatial inheritance rate of not less than 0.60 in the corresponding table of S1 partitions, the most recently uncontaminated background of the old partitions, the same mining stage and water-bearing conditions and verified mine-grade partitions of the same type, and the artificial baselines signed and confirmed by the geological engineer and marked with the validity period; if the above baselines are still unavailable or there are no healthy deformation indicators at present, the recovery period remains open, writing records to the stable learning candidate data set is prohibited, and an artificial review task containing missing baselines, missing indicators, partition versions and review deadlines is generated until a computable baseline that meets the minimum number of samples and the conditions of valid deformation indicators is formed;

[0078] When the recovery coverage rate remains above 0.90 for 6 consecutive hours and all deformation indicators fall within the background quantile range, the start point of this continuous window is determined as the end point of the recovery period. If the conditions are not met within 72 hours, the recovery period remains open and a duty review task is generated. The review task includes the out-of-bounds indicator, duration, event label, and the most recent valid background version. Blasting and disposal construction events are subject to the same recovery judgment starting from the end time of the log. Rainfall events are subject to recovery judgment starting from the next 10-minute grid after the end point of the aforementioned rainfall event in this S2. When rainfall, blasting, and disposal construction events overlap, they are merged into a composite event. The latest time among the end points of each event is taken as the unified recovery start point. If any event is extended or exceeds the limit again before this time, the composite event is extended simultaneously and the recovery coverage rate timer is prohibited from starting.

[0079] Event tag records save event identifier, event type, partition identifier, partition version identifier, event start point, recovery period end point, trigger log or rain window, rule version and review status. Overlapping events are merged into composite events, and the recovery period is the latest time of the recovery period end point of each sub-event. When the log lacks an end time, the corresponding event is marked as unclosed and its data is prohibited from entering the stable learning candidate data.

[0080] The isolator writes all monitoring records that have passed the health check to the real-time risk data set. During periods when there are no event tags and the data is not in the recovery period, the same health record is synchronously written to the stable learning candidate data set. This allows health data from non-event periods to be used for both current anomaly judgment and stable learning screening. Records that have passed the health check during the event period and the recovery period are only written to the real-time risk dataset and a flag is set to prohibit stable writing. Indicators that have not passed the health check stop background updates and retain the device anomaly flag.

[0081] The stable learning candidate data filter selects records in sequence according to event status, recovery period status, and health indicators. When there are parallel healthy records for the same idempotent key, the one with the latest calibration version is retained first. When the calibration versions are the same, the one with the earliest reception time is retained. When the filter result is empty, an empty set and missing reason code are generated instead of using the previous grid record, thereby avoiding cross-time data duplication for stable learning purposes.

[0082] When a record satisfies multiple set conditions simultaneously, the isolator writes it into the purpose bitmap according to the rule of prioritizing real-time risk use and then stable learning use. It also removes duplicate writes by forming an idempotent key composed of device identifier, sampling time, indicator type, and partition version identifier. Both sets carry event tag references, health check version, unified time base time, and original record pointers. Event and recovery period data are set with stable pool rejection constraints in the storage layer. The final set is used by S3 to calculate real-time residuals and update the stable sample pool, respectively.

[0083] In this specific embodiment, S3 includes:

[0084] The background calculation module takes the stable learning candidate data output by S2 as input and establishes a sample pool index according to the warning zone identifier, zone version identifier, season, excavation stage and water-bearing conditions. The season is divided into rainy season and dry season according to local meteorological data. The excavation stage is taken from the stage code in the mining plan version. The water-bearing conditions are divided into three levels: low, medium and high according to the 30-day quantile status of water level or pore water pressure in the zone. The index records save the sample time range, indicator list, equipment health version and event tag audit results.

[0085] The stable sample filter calculates the trend change for each deformation index within a 24-hour stability window. , Indicators In partition and time The trend change is expressed in units of this indicator per hour. Indicates time Health verification monitoring values ​​and units using the same indicators physical units, The same indicator monitoring value representing the starting point of the stable window. This indicates the stable window length, which is 24 h in this specific embodiment. Indicates the index of indicators. This indicates the warning partition index. Indicates the sampling time under a unified time reference;

[0086] The filter reads the 0.90 empirical quantile of the absolute value of the trend change from the historical stable samples corresponding to the current index as the deformation stability boundary, and reads the 0.90 empirical quantile of the absolute value of the change from the historical stable samples of hydrological indicators or support stress indicators as the stability boundary of heterogeneous indicators. Candidate records must have no event labels, the equipment health flag must be 1, the trend change of each deformation indicator must not exceed the deformation stability boundary, and the change of at least one healthy hydrological indicator or support stress indicator must be within the stability boundary of heterogeneous indicators before they are written into the stable sample pool.

[0087] When the current index has fewer than 200 historical stable samples, the stable boundary inherits the corresponding quantiles of samples from adjacent seasons and excavation stages within the same partition and marks the inheritance source. If the inherited samples are still fewer than 200, the stable boundary of the most recent uncontaminated background version is used. When a new partition is put into operation for the first time or has no historical stable samples or uncontaminated background versions, the old partition background with a spatial inheritance rate of not less than 0.60 is first introduced according to the S1 partition correspondence table. If it still does not exist, the same type of partition background of the same lithology, mining stage, and water-bearing conditions is read and reviewed by a geological engineer to generate a validity period of no more than 30 years. The initial background version of d records the source partition, reviewer, validity period, and low confidence flag. If the initial background is still unavailable, the output data is insufficient and the monitoring-only status is disabled. Automatic background updates and automatic warning cancellation are prohibited, and the on-duty personnel execute manual warnings based on the original monitoring data. During the monitoring-only period, the geological engineer reviews the continuous observation window daily according to the S2 event label. There are no events and all equipment health checks have passed. After signing to confirm that there is no continuous unidirectional deformation or adverse changes in abnormal indicators, the samples in the window are written into the cold start candidate pool isolated from the formal stable sample pool. The signed record saves the original curve, event exclusion list, equipment health version, reviewer, and 7. d. Validity period; When the number of candidate samples reaches 200, the 0.90 empirical quantile of the absolute value of the change in deformation trend and the change in anomaly index in the candidate pool is used as a temporary stability boundary. All candidate samples are replayed using this temporary boundary. Only when no less than 90% of the candidate samples pass the stability screening conditions mentioned above in S3 will a formal background version be generated and the monitoring-only and low confidence flags be withdrawn. Otherwise, manual warnings are maintained and the verification is repeated. If the geological engineer's signature confirmation cannot be obtained, a candidate stable sample is not formed, and the monitoring-only status continues.

[0088] Update the partition background value of the selected samples according to the time decay weight from the current time. , Indicates background version Medium indicators and partitions Time-weighted background quantity and unit and indicator Consistent Indicates the background version index. This represents the number of valid samples in the current sample pool and is a positive integer. Indicates the stable sample index. Indicates the first The time decay weights for each sample are dimensionless numbers within the range (0,1). Indicates the first The index value of a stable sample and its unit is the same as the index. Consistent;

[0089] The time decay weight decays to 0.5 of the previous value every 30 days according to the sample age and is normalized before calculation. The background version simultaneously saves the update time, sample pool index, sample start and end time, number of valid samples, stability boundary, background quantity, device health check result and parent version identifier. The stable sample pool and background version are stored in the append-only version library. Data marked by S2 as prohibited from stable writing is directly rejected at the write transaction entry.

[0090] The scaling module obtains a robust scale by multiplying the median absolute deviation of historical stable samples relative to the background by 1.4826. , Indicators In partition robust scale and units and indicators Consistently, when the scale is less than 0.001 times the range of the same index, 0.001 times the range is taken as the zero protection quantity of the same unit, and then the real-time residual is calculated. , Represents real-time residual quantity and its unit is the same as the index. Consistent This indicates the time in the S2 real-time risk data. The indicator value and the unit of the indicator Consistent and We will continue to use the time-weighted background quantity formula and its complete explanation mentioned above in S3;

[0091] The scaling module calculates the real-time normalized residuals separately. and stable sample normalization bias , This represents the real-time normalized residual, which is a non-negative dimensionless number. This represents the normalization bias of stable samples and is a non-negative dimensionless number. This indicates the zero protection quantity for the same indicator and the same partition, and the unit is the same as the indicator. Consistent Following the real-time residual formula and its complete explanation mentioned above in S3, Using the robustness criterion definition from the previous section, , and Using the time-weighted background quantity formula and its complete explanation mentioned above in S3, the absolute value operator retains the input unit, and the denominator and numerator have the same unit, thus making the two results dimensionless.

[0092] Before calculating the weights, the weighting module generates real-time heterogeneous index cross-verification results: it reads the real-time residuals, equipment health indicators, and unfavorable direction mapping tables of the current 10-minute grid and the previous effective grid in the same partition. When the residual of the health deformation index increases along the potential instability direction, and at least one of the different physical types of indicators, such as health water level or pore water pressure, support stress, crack change, or rockfall count, changes in the unfavorable direction relative to the previous effective grid in the table, the real-time heterogeneous cross-verification result of that deformation index is set to 1; otherwise, it is set to 0. When the previous effective grid is missing, all heterogeneous indicators are missing, or the change direction conflicts, 0 is output and the reason code is recorded. The real-time heterogeneous cross-verification results are only used for the current grid weights. The heterogeneous cross-verification status of stable samples follows the stable cross-verification conclusion of the same sampling grid when screening stable samples as described in S3. The two are saved separately and cannot be mixed. The weighting module first resets the weights of indicators with a health indicator of 0 to zero, then assigns a 1.2-fold cross-verification coefficient to deformation indicators with a real-time heterogeneous cross-verification result of 1, assigns a 1.0-fold coefficient to the remaining health indicators, and normalizes them within the current effective index set. , Indicators In partition and time The real-time normalized index weights are located in [0,1] and satisfy the weight sum of 1; for the th Each stable sample is generated using the same rules based on its device health indicators, stable mutual verification status, and set of valid indicators at the time of sampling. , Indicators In partition No. Each stable sample corresponds to a time-weighted sample located in the range [0,1] and the sum of the weights is 1. The stable calibration dataset simultaneously saves the time-weighted sample version, the set of effective indicators, and the generation time, ensuring that historical samples do not change with the current time. Weight changes;

[0093] The aggregation module calculates the inconsistency of real-time metrics. and the inconsistency of stable sample indicators , Indicates partition At any moment The real-time indicators are inconsistent and are non-negative dimensionless numbers. Indicates partition No. The inconsistency of the indicators for a stable sample is a non-negative dimensionless number, with superscript... To indicate the use of stable samples, in the two summations Each time point or sample's current set of valid indicators is traversed. Following the aforementioned real-time weights and their complete explanation in S3, Following the real-time normalized residual formula and its complete explanation mentioned above in S3, Using the aforementioned sample weights and their immediately following complete interpretations in S3, Using the stable sample normalization bias formula and its complete explanation mentioned above in S3, the larger the value, the stronger the overall deviation from the relatively stable background. The two results are written into the real-time inconsistency buffer and the stable calibration dataset, respectively, and then passed to S4 and S5.

[0094] The background parameter library stores stability windows, quantile positions, time decay half-life, zero protection ratios, and mutual verification coefficients according to partition type and indicator type. Fields include parameter name, value, unit, effective range, sample source, version, effective time, and update trigger condition. Among them, the stability window recorded for displacement indicators is 24 hours, the robustness scale zero protection ratio is 0.001 of the range, and the time decay half-life is 30 days. The parameters are verified by the replay false alarm results of the most recent uncontaminated background version. When the partition version changes or the number of replay false alarms exceeds the benchmark, automatic reuse stops and a review begins. After the review is passed, a new parameter version is generated while the old version audit chain is retained. A one-to-one reference relationship is established between the parameter version and the background version, and both are written with the audit timestamp, verification personnel, and replay data range.

[0095] After each background version update, the audit module replays the new background using historical stable samples locked before the update. If the number of false alarms exceeding the previous effective threshold exceeds 1.25 times the median of the number of false alarms replayed in the three most recent uncontaminated versions, or if the proportion of data with post-detection event labels written to the current sample exceeds the 0.95 empirical quantile of the contamination proportion in the three most recent uncontaminated versions, then the current version is marked as a contaminated version, its active pointer is revoked, and it is rolled back to the most recent background version that was not marked as contaminated. The rollback record saves the trigger criteria, sample range, version chain, and execution time.

[0096] In this specific embodiment, S4 includes:

[0097] At the beginning of each 7-day threshold statistical cycle, the threshold management module reads the stable calibration dataset generated by S3, the current background version, and the directed partitioning map of S1, and reads the partition risk configuration table. The configuration table uses the partition identifier and the partition version identifier as a joint key to store the static risk score, the importance of the protected object, the minimum number of effective indicators, the equipment verification time, and the configuration version, which are assessed by geological survey and normalized to [0,1]. When the partition version changes or the scope of the protected object changes, a review is triggered.

[0098] For each partition, cross-regional propagation evidence that arrives along the incoming edge within the previous threshold period, remains valid within the corresponding allowable time delay, and has been confirmed by S5 is statistically analyzed. The product of the edge type coefficient and the propagation confidence score is summed and divided by the maximum value of the number of incoming edges in the partition and 1. The result is then truncated to [0,1] to form the propagation evidence strength. The coefficients for the four types of edges—structural continuity, hydraulic connection, unloading transfer, and weak support—are set to 1.0, 0.9, 0.8, and 0.7, respectively. The coefficients are obtained by normalizing the hit rate of verified propagation events in the most recent 24 months and are updated every 20 new verified events.

[0099] The budget module calculates the budget base for each zone. , Indicates partition The budget allocation base is a dimensionless number within [0,1]. This indicates a static risk score located in [0,1]. This indicates the importance of the protected object and is located in the range [0,1]. Indicates the strength of the evidence of dissemination and is located in [0,1]. , and These represent three non-negative weights that sum to 1. In this specific implementation, they are set to 0.4, 0.35, and 0.25 respectively. The larger the budget allocation base, the more stringent the false alarm control resources the partition needs.

[0100] The budget module reads the number of deduplicated and verified early warning events for each partition in the last 12 completed statistical periods, and rounds up the median while keeping at least 1 to obtain the expected exposure for the next period. Then calculate the exposure percentage. , Indicates partition The expected number of verification events, expressed in terms of items. Indicates index The expected number of verification events for the next cycle in the corresponding partition, expressed in units of items, is statistically sourced from [source missing]. Same and at least 1, Indicates partition The expected verification exposure percentage is located in (0,1). Indicates the partition index in the current partition version; budget allocation cardinality. and The budget base formula and its complete explanation described above in S4 are used as is.

[0101] Budget module by Calculate the budget weight for the number of false alarms in each partition, and then... Calculate the allowed false alarm tail probability for the partition. Indicates partition The false alarm count is budgeted and weighted, and is located in (0,1) and the sum of the weights is 1. Following the aforementioned exposure percentage formula and its complete explanation in S4, Indicates index The expected verification exposure percentage for the corresponding partition, located in (0,1], is calculated using the exposure percentage formula and its subsequent complete explanation described in S4 above. This represents a dimensionless protection value to prevent the highest-risk partition from having a weight of zero, and is set to 0.05. Indicates partition The false alarm tail probability is allowed in the next statistical period and is located at (0,1). This indicates the overall mine's false alarm rate, set at 0.02 by the safety management system; thus satisfying... The summation index Iterate through all partitions of the current partition version, when When the value approaches 1, the corresponding partition receives a smaller allowed tail probability;

[0102] The threshold generator reads inconsistencies from at least 500 stable samples over the past 180 days for each partition, setting level coefficients of 1, 0.5, and 0.25 for low, medium, and high alert levels, respectively; calculations are performed for the first statistical period. And calculate , Indicates partition and warning level It is located at the empirical quantile position in the first statistical period and lies between the corresponding upper and lower limits. This represents the hierarchical dynamic threshold for the first statistical period and is a non-negative dimensionless number. This represents an index of warning levels from low to high. Indicates level The false alarm budget coefficient is a dimensionless number within (0,1]. and Representing levels Lower and upper limits of quantile positions Indicates partition The set of stable sample inconsistencies. Indicates the position of the empirical quantile over the set. Operators corresponding to numerical values, This represents an operator that truncates the input to a given closed interval. Using the formula for the allowed false alarm tail probability of the aforementioned S4 and its complete explanation, the threshold increases as the warning level increases;

[0103] When there are fewer than 500 valid stable calibration samples in a partition, the threshold generator first reads the stable calibration data of the most recent uncontaminated background version in the same partition. If this is still insufficient, samples from the old partition are introduced according to the S1 partition correspondence table with spatial inheritance rate weighting and a cumulative weight of not less than 300. If this is still not met, the most recently verified graded dynamic threshold is frozen and a sample shortage flag is written. When the system is put into operation for the first time and there is no freezeable threshold, the mine-grade threshold template of the same partition type that has been approved by replay and the offline initial threshold approved by the safety management person and marked with a 30-day validity period are used in sequence. If neither of these exists, the output is insufficient data and only monitoring status is displayed. Automatic local classification and automatic release are prohibited. The on-duty personnel perform manual warnings based on the original monitoring volume. After obtaining 500 valid stable calibration samples, the system switches to the quantile threshold and retains the switch audit. Samples from any event period, recovery period, or contaminated version are not included in the quantile calculation.

[0104] The online statistics module uses partitions and statistical periods as statistical keys, and retains the warning level as an event detail field; only warning events that are deduplicated by event identifier within that period and have a final conclusion before the verification deadline are counted. When there are no fewer than 10 items, The denominator is the number of events with the final conclusion being a false alarm. Use the numerator to calculate the actual false alarm rate. ; Indicates partition In the statistical period The actual false alarm rate is located in [0,1]. This indicates the number of false alarms, expressed in terms of individual items. This indicates the number of verified events, which must be a positive integer. Unverified events are not included in the numerator or denominator of this period but are carried over to the next week's expected verification list. Cross-period events are assigned according to the starting point of the first warning. Repeated status commands are not counted repeatedly. Events that are re-verified after being canceled are still identified by the same event identifier, and only the final conclusion before the deadline is taken. If there are fewer than 10 records, the statistical status is insufficient. The actual false alarm ratio and online coverage error of this partition are not generated for this period. The current empirical quantile position and activity threshold remain unchanged. The verified samples of this period are incorporated into the cumulative statistical list of the next period until 10 records are reached.

[0105] The online statistics module simultaneously calculates the actual false alarm rate for the entire mine. And calculate , Indicates the statistical period The actual false alarm rate for the entire mine is located in [0,1]. The summation of the two equations... Iterate through the partitions that have reached the minimum number of 10 verification events in this period. and Following the formula for the actual false alarm ratio and its complete explanation mentioned above in S4, Indicates partition In the statistical period The online coverage error is a dimensionless number within the range [-1, 1]. Following the formula for the actual false alarm ratio and its complete explanation mentioned above in S4, Using the formula for the allowed tail probability of false alarms in the aforementioned S4 and its complete explanation, a positive online coverage error indicates that the actual false alarm is higher than the allowed tail probability; the actual false alarm ratio of the whole mine is not calculated when no zone reaches the minimum number of verification events, and all activity thresholds remain unchanged; the numerator, denominator, verification cutoff time, sample list and the aggregated result of the whole mine are all written into the threshold audit record;

[0106] Threshold generator with Adjust the position of the empirical quantile for the next period, and use Generate candidate dynamic thresholds for the next cycle; Indicates the next period partition grade The empirical quantile position is located between the corresponding upper and lower limits. Indicates the current period's empirical quantile position. Indicates the next period partition grade The candidate hierarchical dynamic threshold is a non-negative dimensionless number. This indicates a dimensionless correction step size, set to 0.20. , , and Following the first statistical period threshold formula and its complete explanation mentioned above in S4, This represents an operator that truncates the input to a given closed interval. Following the online coverage error formula and its complete explanation mentioned above in S4, the coverage error is characterized by an increase in the positive quantile position and a tightening of the threshold; during rollback or locking, the corresponding active threshold version is read and saved. During the lockout period and the period of insufficient statistics, the position of this quantile and the activity threshold remain unchanged. After the lockout is lifted and the cumulative number of verified events reaches 10, the candidate threshold is generated according to this formula.

[0107] The threshold configuration table saves the false alarm budget coefficient, quantile position upper and lower limits, correction step size, and single-cycle change upper limit calculation window according to the warning level. A medium-level warning record includes a level coefficient of 0.5, a quantile position lower limit of 0.95, an upper limit of 0.999, a correction step size of 0.20, and a change upper limit sample window of 180 days. The configuration parameters are derived from the verified local warning records of the most recent 24 months, and a new configuration version is generated when 30 new verification records are added or the safety management system is adjusted to adjust the false alarm budget of the whole mine.

[0108] After generating candidate thresholds, the threshold management module uses stable samples from the previous statistical period that were not calibrated to replay the candidate version and the previous active version using the same holdout set. It checks the empirical false alarm ratio of each partition, the strict increasing relationship of the three-level thresholds, and the expected exposure weighted false alarm ratio level by level. If the empirical false alarm ratio of any candidate in any partition exceeds 1.20 times the corresponding allowable tail probability, the expected exposure weighted false alarm ratio of the candidate exceeds the target false alarm ratio for the entire mine, or the threshold order is reversed, the candidate version is rejected, and the previous active version is retained. If the actual false alarm ratio for the entire mine in the previous period exceeds the target false alarm ratio for the entire mine, it is only used as a trigger condition for tightening online coverage error and key verification, not as an independent rejection condition. In this case, the candidate version is also required to... If the total number of false alarms on the same holdout set decreases by at least one compared to the previous activity version, or the weighted false alarm ratio of the candidate expected exposure decreases by at least 0.001 compared to the previous activity version, the activity threshold pointer is switched when the aforementioned candidate access control and the improvement conditions are met. If the improvement conditions are not met, the empirical quantile position is increased and replay is performed within the single-cycle change limit. If there is still no improvement after reaching the limit, the previous activity version is maintained and a threshold manual review task is generated. The holdout replay record saves the number of times the limit is exceeded at each level for the candidate and the old version, the total number of samples, the empirical false alarm ratio, the allowed tail probability, the total aggregate value, the improvement range, and the verification conclusion, so that the threshold, background version, and budget configuration read by S5 are kept in the same statistical period.

[0109] When the number of effective monitoring indicators is less than the minimum number of effective indicators (3) in the risk configuration table, or when equipment malfunctions occur continuously for 6 hours to reach the verification time, the threshold management module freezes the threshold version that was most recently verified for that partition. When the absolute change of the new threshold relative to the previous threshold exceeds the single-cycle change limit determined by the empirical quantile of the historical stable threshold change amount of 0.95, the limit is applied to the upper limit, and the lower threshold is kept less than the middle threshold and the middle threshold is less than the higher threshold.

[0110] When the online coverage error increases for two consecutive statistical cycles, the threshold management module rolls back to the latest threshold version in which the online coverage error did not increase, and writes the partition identifier, background version, expected verification exposure, allowable false alarm tail probability, quantile position, three-level threshold, sample range, propagation evidence summary, freezing or clipping state, and parent version identifier into the threshold version library. After rollback, it locks for one complete statistical cycle: during the locking period, only verification results are accumulated without updating the quantile, and after the period expires, online correction is restarted with the false alarm ratio of the locking period. If the number of valid verification events in the locking period is less than 10, the locking is continued for one more cycle while maintaining the original threshold; after the number reaches 10, the actual false alarm ratio is calculated after deduplication by event identifier. Upon restart, the locking reason, verification sample list, partition and full-mine false alarm statistics, and unlocking moment are appended to the threshold version audit record. The audit record adopts a unique key composed of the partition identifier, statistical cycle and threshold version, and retains the creator and generation time. The active threshold version is transmitted to S5 to execute local and linked early warning.

[0111] In this specific embodiment, S5 includes:

[0112] The early warning judgment module reads the S3 real-time inconsistency cache, S4 active threshold version, S1 directed partition graph and S2 event tags at a 10-minute cycle, compares the real-time indicator inconsistency with the low-level, intermediate-level and high-level graded dynamic thresholds for each partition, and takes the 0.75 empirical quantile from the set of response durations of at least 10 verified historical events of the same partition and level as the confirmation duration. When there are fewer than 10 samples, the 0.75 empirical quantile of samples of adjacent levels in the same partition, the same level in old partitions with an S1 spatial inheritance rate not lower than 0.60, and the same level at the mine level of the same type of partition is adopted in sequence. If there are still no valid samples, the default values of 60 minutes for low level, 40 minutes for intermediate level and 20 minutes for high level confirmed by geological engineers are adopted. The confirmation duration is uniformly limited to 20 minutes to 6 hours. The confirmation rule table records the sample source, sample number, fallback level, and validity period, and recalculates after the number of samples of the same partition and same level reaches 10. When the output data of S3 or S4 is insufficient or in the monitoring-only state, automatic grade judgment and automatic release are stopped, and a manual early warning task containing the original monitoring quantity is generated;

[0113] When the real-time indicator inconsistency exceeds a certain level threshold and continuously reaches the confirmation duration of that level, the direction discriminator checks whether the residual of the deformation indicator increases along the potential instability direction, and simultaneously checks at least one heterogeneous indicator among rising water level or pore water pressure, support force changing in an unfavorable direction, crack propagation or falling rock count increase. The local early warning level is determined only when both the deformation indicator and the heterogeneous indicator point to the development of instability; if the continuous duration is not reached or heterogeneous mutual verification is missing, the state is kept as observation;

[0114] For each entry from the source partition pointing to the target partition For valid edges, the propagation judge records the moment when the source partition first crosses the corresponding hierarchical dynamic threshold as the time delay start point. It only accepts the mutual verification results of the target partition's same-direction residual and heterogeneous indicators when the difference between the first occurrence time and the time delay start point is not greater than the allowed time delay of the edge, and writes them together with the first occurrence time, device health version, edge version and allowed time delay into the candidate propagation evidence cache. When the source partition subsequently reaches the confirmation duration consecutively, it is allowed to read the cache to complete the verification judgment, and it is no longer required that the verification time is still within the allowed time delay. However, if the source partition falls back below the upgrade threshold before confirmation, the device health corresponding to the cached evidence fails, or the edge version changes, the cache is immediately invalidated. Propagation along reverse edges or failed edges is prohibited. The edge type, partition version and allowed time delay must be consistent with the current directed partition graph record of S1.

[0115] The propagation determiner reads the unit slip direction vector from the potential slip surface geometry record of the target partition and calculates... , Indicates the target partition At any moment The displacement residual components along the potential slip direction, in mm. This represents the real-time residual vector of the three-dimensional displacement of the target partition, with units in mm. The unit vector representing the potential slip direction of the target partition is dimensionless, and a positive dot product indicates that the displacement residual is in the same direction as the potential slip direction.

[0116] The source persistence is obtained by dividing the source partition threshold overrun duration by the corresponding confirmation duration and truncating to [0,1]. The denominator of the target response is the empirical quantile of the absolute deviation of displacement in the same direction of at least 200 stable samples in the current official background version of the target partition, which is 0.95. If there are insufficient samples, the stable samples in the same direction of the old partition with an S1 spatial inheritance rate of not less than 0.60 and the stable samples of the same type of partition of the same lithology, mining stage and water-bearing conditions are read in sequence, and the inheritance level and background version are recorded. When the obtained quantile is less than 0.001 times the range of the target displacement equipment, 0.001 of that range is taken. Using the zero-protection denominator as the unit, the residual components of the same-direction displacement of the target partition are divided by this denominator and truncated to [0,1] to obtain the target responsivity; when there is only the initial migration background and no more than 200 effective stable sample sources, or when all backtracking sources are still unavailable, the propagation data insufficiency status is output, the source partition is kept under local warning, linkage upgrade is prohibited, and manual review tasks are assigned. Automatic propagation judgment is restored after the denominator can be calculated; the number of outlier indicators that pass the health check and jointly point to unstable development is divided by the maximum value of the current number of healthy outlier indicators and 1 and truncated to [0,1] to obtain the mutual verification degree;

[0117] Propagation confidence score by calculate, Representing an edge to At any moment The propagation confidence score is located in [0,1]. This indicates that the source partition has exceeded the persistence limit and is located in [0,1]. This indicates the target partition's responsiveness and it lies within the range [0,1]. This indicates the cross-verification degree of heterogeneous indicators in the target partition and is located in [0,1]. , and The values ​​represent non-negative weights with a sum of 1. In this specific implementation, they are set to 0.35, 0.40, and 0.25 respectively. The larger the score, the stronger the evidence of cross-regional transmission.

[0118] When the source partition anomaly persists for the confirmation time, the target partition displacement residual component in the candidate propagation evidence cache is positive, at least one heterogeneous indicator completes mutual verification and the propagation confidence score reaches 0.70, the propagation judge forms cross-regional propagation evidence corresponding to the valid edge and writes it into the edge evidence set. Then, the path filter uses the edge evidence set to generate candidate propagation paths. The main path is not determined in advance in this section. The main path is determined only by the unified path-level sorting rules described later.

[0119] The path filter generates candidate paths with no more than 3 edges from the directed partition graph of the current partition version. Paths containing invalid edges, reverse edges, duplicate nodes, cumulative elapsed time exceeding the sum of the allowable time delays of each edge, or the target partition lacking a displacement residual in the same direction are deleted. The minimum value of the propagation confidence score of each edge on the path is used as the filtering score. Paths with a filtering score of 0.70 or higher are added to the sorting set, and paths with a score lower than 0.70 are written to the reason field for not being used.

[0120] The sorting module determines the main path using the filtering score as the first descending key, the total allowed time delay of the path as the second ascending key, the path length as the third ascending key, and the partition identifier sequence as the final ascending key. Only one main path is retained for the same source partition and target partition in a judgment cycle. If the sorting set is empty, no linkage upgrade command is generated and the local warning of the source partition is retained. If the main path is different from the previous cycle, the path switching reason and the filtering scores of the two paths are saved.

[0121] The propagation confidence score is rounded to a precision of 0.001 before comparison. When the score is not lower than 0.70, it is prioritized for cross-regional propagation verification and linkage upgrade branches. Records with scores not lower than 0.60 but lower than 0.70 are marked as propagation pending confirmation by the early warning service and continue to be collected for a confirmation period. During this period, when new same-direction deformation residuals appear in the target partition and the heterogeneous mutual verification degree increases, the score is recalculated. If the upgrade threshold is not reached by the end of the confirmation period, the pending confirmation status is closed. When the target displacement vector is missing due to equipment abnormality, the local early warning of the source partition is maintained and the measurement point verification task is dispatched. It is forbidden to use spatial proximity relationship to replace the missing orientation residual.

[0122] The hysteresis controller sets the intermediate local warning level to the linkage escalation level, sets 0.70 as the propagation escalation threshold, and calculates... and , Indicates partition grade The local release threshold is a non-negative dimensionless number. This represents the upgrade threshold for the corresponding level and is equal to the S4 level dynamic threshold. This represents the magnitude of local hysteresis and is taken as 10% of the corresponding upgrade threshold. This indicates the propagation termination threshold and is located in [0,1]. This indicates a propagation escalation threshold of 0.70. This indicates the propagation lag magnitude, which is 0.10.

[0123] When a local alert reaches the linkage escalation level and the propagation confidence score reaches the propagation escalation threshold, the regional linkage alert level is upgraded. When there is evidence of cross-regional propagation of an activity or a regional linkage alert exists, the regional linkage alert is downgraded or lifted only when all three conditions are met: the real-time indicator inconsistency is below the local lifting threshold, the propagation confidence score is below the propagation lifting threshold, and the cross-regional propagation evidence has been missing for 2 consecutive hours after the corresponding allowable time lag ends. If any condition is not met, the current linkage level is maintained. When no evidence of cross-regional propagation of an activity has been formed and only a local alert exists, if the real-time indicator inconsistency is below the current level's local lifting threshold and remains below it for 2 consecutive hours, and the number of effective monitoring indicators is not less than the minimum number of effective indicators in the risk configuration table, the local alert is downgraded by one level and the 2-hour local lifting timer is restarted. The lower-level local alert is lifted when the condition is met again. When heterogeneous indicators disappear but the real-time indicator inconsistency has not yet fallen below the local lifting threshold, observation is maintained and the alert is not lifted directly. During periods of equipment abnormality, missing key indicators, insufficient S3 or S4 data, and manual locking, the current level is maintained and marked for verification, and automatic lifting is prohibited. The joint lifting rule for the activity linkage status takes precedence over the local alert lifting rule only.

[0124] The hysteresis state machine stores the current level, upgrade candidate level, continuous over-limit timer, continuous release timer, the most recent occurrence time of propagation evidence, and manual lock flag for each partition. Upgrade commands only allow the level to increase step by step, and release commands only allow the level to decrease step by step. After the on-duty personnel set a manual lock, the automatic command only updates the criterion record without changing the level. When the manual lock is released, the confirmation timer is restarted from the current real-time inconsistency. All state transitions record the previous state, triggering condition, subsequent state, execution time, operation source, and associated propagation path identifier. Repeated state commands are rejected for re-execution by using the partition identifier, level, and judgment period as an idempotent key.

[0125] The early warning service generates the scope of action based on the action rules table: when there is an effective linkage main path, the scope is based on the partition boundaries of all nodes of the main path and the protected objects within the target partition. Then, candidate path nodes with a screening score of not less than 0.70 are included in the inspection scope according to the linkage level. The action rules table uses the early warning level and the importance of the protected object as keys, mapping low, medium and high levels to encrypted inspection, personnel and equipment control and personnel evacuation scopes, respectively. When there is no effective linkage path, only the local scope of the source partition is used. When the main path is switched, the linkage level changes or the protected object version is updated, the scope of action is recalculated and the previous and subsequent versions, triggering reasons and effective times are saved.

[0126] The early warning service writes the starting early warning zone, zone version identifier, main propagation path, alternative paths, edge type, trigger indicator, propagation confidence score, local and linked early warning levels, early warning start and end time, handling scope and its rule version, threshold version and background version into the zone linked early warning record. The sensitivity correction input for the next threshold cycle is formed by multiplying the edge type coefficient and the propagation confidence score. The propagation evidence strength is calculated by S4. Cross-regional propagation evidence and its derived records are only entered into the early warning audit database and are prohibited from being written into the S3 stable sample pool.

[0127] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0128] This invention uses versioned early warning partitions as spatial units for background calculation and threshold allocation, and uses event tags to isolate real-time judgment data from stable learning data, so that the inconsistency between partition background quantity and index quantity can correspond to specific seasons, excavation stages and water-bearing conditions; the whole mine false alarm budget is used to determine the empirical quantile position after partition weight allocation, so as to establish a traceable link between the hierarchical dynamic threshold and the stable sample distribution and monitoring equipment status.

[0129] This invention changes the linkage structure between local early warnings: cross-regional propagation is not directly triggered by the time overlap of early warning levels, but is constrained by the edge type, edge direction and allowable time delay of the directed partition graph, and propagation evidence is formed by mutual verification of the target partition's same-direction deformation residual and heterogeneous indicators; the propagation evidence only corrects the subsequent threshold sensitivity and is not written into the stable sample pool, thereby maintaining the isolation between the use of background learning and risk feedback.

Claims

1. A method for dynamic early warning of steep slopes in open-pit mines, characterized in that, include: S1. Obtain engineering geological data, hydrological data and mining status data of open-pit mine slopes, divide early warning zones, and construct a directed partition map with edge direction and allowable time delay based on the physical influence relationship between the zones. S2. Obtain monitoring data and event information for each early warning zone, identify events and their recovery periods, and isolate the monitoring data into real-time risk data and stable learning candidate data; S3. Filter stable learning candidate data to update the partitioned stable sample pool and partitioned background quantity, calculate the real-time residual quantity of real-time risk data relative to the partitioned background quantity, and generate stable sample inconsistency quantity and real-time indicator inconsistency quantity based on the deviation of stable samples relative to the partitioned background quantity and the real-time residual quantity. S4. Based on the distribution of inconsistent indicators of stable samples, generate and update the hierarchical dynamic thresholds of each early warning zone under the budget constraint of the whole mine false alarm. S5. Compare the real-time inconsistency of indicators with the hierarchical dynamic threshold to determine the local early warning level; When the source warning zone continues to be abnormal, the propagation direction is consistent with the edge direction of the directed partition graph, and the target warning zone shows real-time residuals of deformation in the same direction within the allowable time lag and is mutually verified by monitoring indicators of different physical types, cross-regional propagation evidence is formed and a zone linkage warning is generated.

2. The method for dynamic early warning of steep slopes in open-pit mines according to claim 1, characterized in that, S1 includes: Initial spatial units are formed using step boundaries and slope boundaries. These initial spatial units are then merged or split based on lithology and structural surface combinations, potential slip surfaces, excavation sequence, water catchment relationships, and support continuity to obtain early warning zones. When the excavation profile, support status, or water catchment relationship changes, a new zone version identifier is generated, and the spatial correspondence between the old and new early warning zones is preserved. Structural continuity, hydraulic connection, unloading transfer relationship, and weak support continuity are used as edge types, respectively. The edge direction is determined based on the direction of action transfer, and the allowable time delay is determined based on the historical response time difference from the source early warning zone to the target early warning zone for the corresponding action.

3. The method for dynamic early warning of steep slopes in open-pit mines according to claim 1, characterized in that, S2 includes: The system collects displacement, displacement rate, displacement acceleration, pore water pressure or water level, rainfall, vibration, crack changes or rockfall count, support stress, and operation logs. All data are transformed to a unified time reference, and health checks are performed on missing measurements, exceeding measurement ranges, timestamp jumps, persistent zero values, and sensor drift. The system identifies the event starting point based on the blasting time, treatment construction period, and rainfall event threshold determined by rainfall duration and amount in the operation log. The recovery period endpoint is determined by the moment when the post-event monitoring data returns to a preset quantile interval of the pre-event background distribution and remains continuously for a preset duration, thus generating the event tag.

4. The method for dynamic early warning of steep slopes in open-pit mines according to claim 3, characterized in that, The real-time risk data includes all data that has currently passed health verification; The stable learning candidate data is selected from data that has passed health verification during non-event periods and non-recovery periods, so that health data during non-event periods can be used for both current anomaly judgment and stable learning screening. Data that passes health verification during the event occurrence and recovery period is used only as real-time risk data and is used to determine the current real-time residual based on the difference between the real-time risk data and the corresponding partition background amount, and is prohibited from being written into the stable sample pool; For monitoring equipment that fails the health verification, stop background updates for the corresponding indicators and retain the equipment abnormality flag.

5. The method for dynamic early warning of steep slopes in open-pit mines according to claim 1, characterized in that, S3 includes: establishing a sample pool index for combinations of early warning zones, seasons, excavation stages, and water-bearing conditions; identifying data that do not have event tags, whose corresponding monitoring equipment passes health verification, whose trend changes in each deformation index within the stability window do not exceed the deformation stability boundary determined by the historical stable sample quantiles, and whose changes in at least one hydrological index or support stress index pass health verification and are located within the heterogeneous index stability boundary determined by their respective historical stable sample quantiles, as samples that meet the stability conditions and are mutually verified by heterogeneous indicators; updating the zone background quantity according to the time-weighted statistics of the samples, and saving the update time, sample pool index, sample range, and verification results to form a background version.

6. The method for dynamic early warning of steep slopes in open-pit mines according to claim 5, characterized in that, The real-time residual is the difference between the real-time risk data and the corresponding partition background data. The deviations of the real-time residual and the stable samples relative to the partition background data are scaled and normalized according to the historical stable sample distribution of each indicator. Indicator weights are determined based on the health status of each indicator and the cross-verification results of dissimilar indicators. The normalized real-time residual and stable sample deviations are then weighted and aggregated using the same indicator weights and aggregation rules to obtain the real-time indicator inconsistency and the stable sample indicator inconsistency. When the number of false alarms in the historical stable sample replay check after the background version is updated exceeds the false alarm benchmark determined by the historical replay results of the uncontaminated background version, or when the proportion of data with post-detection event labels written into the sample exceeds the contamination criterion determined by the historical uncontaminated background version, the background version is marked as a contaminated version and rolled back to the most recent historical background version that was not marked as contaminated.

7. The method for dynamic early warning of steep slopes in open-pit mines according to claim 6, characterized in that, S4 includes: The budget allocation weight is determined based on the static risk score of each early warning zone, the importance of the protected object, the number of cross-regional propagation evidences formed in the previous threshold statistical period and propagated along the edge pointing to the early warning zone in the directed partitioning graph and still valid within the current allowable time delay, and the corresponding edge type. The total mine false alarm budget is allocated to the false alarm budget share of each early warning zone according to the budget allocation weight. Using the inconsistency of the stable sample index as calibration data, the hierarchical dynamic thresholds for each early warning zone are generated based on the empirical quantiles corresponding to each false alarm budget share; the online coverage error is determined based on the difference between the actual false alarm ratio and the false alarm budget share within the online statistical period, and the empirical quantile position for the next threshold statistical period is corrected based on the online coverage error.

8. The method for dynamic early warning of steep slopes in open-pit mines according to claim 7, characterized in that, When the number of valid monitoring indicators in any early warning zone is less than the minimum number of indicators configured for that early warning zone, or when an anomaly determined by the status of the monitoring equipment persists for a preset equipment verification period, the most recently verified tiered dynamic threshold of that early warning zone is frozen; when the change in the corrected tiered dynamic threshold relative to the previous threshold exceeds the single-cycle change upper limit determined by the historical stable threshold change distribution, the threshold change is limited according to the single-cycle change upper limit; when the online coverage error increases for two consecutive statistical periods, it is rolled back to the threshold version where the online coverage error has not increased.

9. The method for dynamic early warning of steep slopes in open-pit mines according to claim 7, characterized in that, S5 includes: when the real-time inconsistency of the indicators exceeds the corresponding level of the dynamic threshold and is continuously maintained at the confirmation duration determined by the quantile of the historical event response duration of the warning zone, and when the deformation indicator and at least one of the hydrological indicators, support stress indicators, crack change, or rockfall count have a common direction of change pointing to the development of instability, the corresponding local warning level is determined; propagation judgment is only performed along the edge from the source warning zone to the target warning zone in the directed zoning diagram, the moment when the source warning zone exceeds the dynamic threshold is taken as the starting point of the allowable time delay, the potential slip direction of the target warning zone is determined according to the potential slip surface geometry information in the spatial structure data, and the cross-regional propagation evidence and its propagation confidence score are determined by the mutual verification result of the displacement residual component along the potential slip direction in the real-time residual and at least one of the heterogeneous indicators.

10. The method for dynamic early warning of steep slopes in open-pit mines according to claim 9, characterized in that, The system employs preset linkage upgrade levels, propagation upgrade thresholds, local release thresholds, and propagation release thresholds to implement hysteresis control on the regional linkage early warning level. The local release threshold is obtained by subtracting the local hysteresis amplitude from the hierarchical dynamic threshold corresponding to the current local early warning level, and the propagation release threshold is obtained by subtracting the propagation hysteresis amplitude from the propagation upgrade threshold. When the local early warning level reaches the preset linkage upgrade level and the propagation confidence score reaches the propagation upgrade threshold, the regional linkage early warning level is upgraded. When the real-time indicator inconsistency does not reach the local release threshold, the propagation confidence score does not reach the propagation release threshold, and cross-regional propagation evidence is continuously missing for the preset release duration after the corresponding allowable time delay ends, the regional linkage early warning level is downgraded or released. The partitioned linkage early warning record includes the starting early warning partition, partition version identifier, propagation path, edge type, trigger indicator, propagation confidence score, early warning start and end time, and handling scope; cross-regional propagation evidence is used to form a threshold sensitivity correction amount and the budget allocation weight of the corresponding early warning partition is corrected in the next threshold statistical cycle, and the cross-regional propagation evidence is prohibited from being written into the stable sample pool.

Images