Engineering geological disaster intelligent early warning method and system

Abstract

The present application relates to the field of engineering geology disaster monitoring and early warning technology, and more particularly to an engineering geology disaster intelligent early warning method and system. The method comprises: acquiring multi-source monitoring data and static engineering geology data of the same engineering geology unit; extracting a starting event, an acceleration event and a continuous event, constructing a precursor event phase sequence matrix and an event interval vector; constructing a trigger-response hysteresis curve, extracting a hysteresis closure difference vector and an asymmetric vector of the ascending and descending sections; inputting the above features into a machine learning early warning model to output an irreversible instability score and the earliest irreversible time; combining the phase sequence locking condition under the continuous sliding time window to determine the early warning level and generate an early warning instruction. The method solves the technical problem that the prior art is difficult to accurately distinguish between recoverable abnormal fluctuations and irreversible instability precursors.

Description

Technical Field

[0001] This invention relates to the field of engineering geological disaster monitoring and early warning technology, and in particular to an intelligent early warning method and system for engineering geological disasters. Background Technology

[0002] With the continuous advancement of large-scale engineering construction, mineral resource development, and mountain transportation infrastructure construction, the demand for monitoring and early warning of engineering geological hazards such as landslides, collapses, debris flows, and ground subsidence is constantly increasing. Since engineering geological hazards are typically influenced by multiple factors, including topography, soil and rock structure, groundwater activity, rainfall processes, and engineering disturbances, their occurrence and development are characterized by significant complexity, time-varying nature, and uncertainty. Therefore, how to utilize multi-source monitoring data to achieve more accurate and timely disaster early warning has become an important research direction in this field.

[0003] In existing technologies, machine learning has already been incorporated into geological disaster early warning systems. Publication number CN117874499A discloses "A Geological Disaster Early Warning Method and System Based on Dynamic Data Monitoring," which discloses the following steps: acquiring multi-source geological disaster survey data of a target area; screening the locations of potential geological hazards within the target area; generating geological hazard hazard characteristics based on existing deformation features and geological conditions at the hazard locations; further acquiring the influencing factors of the hazard locations; determining geological monitoring indicators and acquiring multi-dimensional monitoring data in real time; constructing a geological disaster early warning model based on this; dynamically judging the development trend of geological disasters; conducting hazard analysis and avoidance decisions based on the development trend; and dynamically correcting the geological disaster early warning model by incorporating interactive information. Therefore, existing technologies can already comprehensively process multi-source geological disaster monitoring data through machine learning and achieve trend judgment and early warning output.

[0004] However, the closest existing technologies mentioned above mainly focus on analysis and early warning based on the comprehensive characteristics and changing trends of multidimensional monitoring data. Their technical focus remains on the fusion modeling of monitoring indicator data and the judgment of disaster development trends. For two types of information that are more significant in the evolution of engineering geological disasters—namely, the relative temporal relationship of precursor events between different monitoring channels, and the hysteresis irreversibility characteristics exhibited when triggering factors and response factors transform from recoverable fluctuations to irreversible instability—existing technologies lack targeted joint characterization and collaborative judgment mechanisms.

[0005] Therefore, the most significant technical problem with existing technologies is that existing machine learning-based engineering geological disaster early warning methods mainly rely on the spatiotemporal characteristics or development trends of multidimensional monitoring data for judgment. They lack joint characterization of the relative temporal relationships of precursor events between different monitoring channels and the irreversibility characteristics in the trigger-response process, making it difficult to accurately distinguish between recoverable abnormal fluctuations and irreversible instability precursors. Summary of the Invention

[0006] To overcome the aforementioned technical deficiencies, the present invention aims to provide an intelligent early warning method and system for engineering geological disasters. The present invention solves the technical problem of existing technologies' inability to accurately distinguish between recoverable abnormal fluctuations and irreversible instability precursors by jointly characterizing the relative temporal relationships and trigger-response irreversibility characteristics of precursor events in multi-source monitoring data, and by combining a machine learning early warning model with phase sequence locking conditions under a continuous sliding time window for joint judgment.

[0007] This invention discloses an intelligent early warning method for engineering geological disasters, comprising:

[0008] S1. Obtain multi-source monitoring data and static engineering geological data of the same engineering geological unit within the target monitoring area. The multi-source monitoring data shall include at least one trigger factor channel, one hydrological response channel, and two deformation response channels at different depths.

[0009] S2. Extract precursor events for each channel using a sliding time window to obtain the start event, acceleration event, and continuous event for each channel. Construct a precursor event phase sequence matrix based on the sequential relationship between each precursor event in different channels, and construct an event interval vector based on the time interval between adjacent precursor events.

[0010] S3. Construct a trigger-response hysteresis curve based on the changes of the trigger factor channel and each response channel within the same sliding time window, and extract the hysteresis closure difference vector and the asymmetric vector of the rising and falling segments;

[0011] S4. Input the precursor event phase sequence matrix, event interval vector, hysteresis closure difference vector, rise and fall segment asymmetry vector, and static engineering geological data into the pre-trained machine learning early warning model to obtain the irreversible instability score and the earliest irreversible moment of the corresponding engineering geological unit.

[0012] S5. When the irreversible instability score is not lower than the preset score threshold, the time interval from the current moment to the earliest irreversible moment is not greater than the preset time threshold, and the precursor event phase sequence matrix meets the phase sequence locking condition within multiple consecutive sliding time windows, the warning level of the corresponding engineering geological unit is determined and a warning instruction is generated.

[0013] Preferably, the extraction of precursor events includes:

[0014] The original monitoring sequences of each channel were denoised and had their baselines stabilized.

[0015] Calculate the local change slope, local curvature, sign holding length, and cumulative deviation within each sliding time window;

[0016] When the slope of the local change exceeds the first preset threshold and the symbol retention length reaches the first preset length, the start event is extracted.

[0017] When the increment of the local change slope exceeds the second preset threshold, the local curvature exceeds the third preset threshold, and the cumulative deviation increases continuously, the acceleration event is extracted.

[0018] When the cumulative deviation exceeds the fourth preset threshold and fails to return to the stable baseline within the preset duration, the continuous event is extracted.

[0019] It retains only the precursor events that are detected in both adjacent sliding time windows and whose event start time deviation is no greater than the preset tolerance.

[0020] Preferably, the two deformation response channels at different depths include a shallow deformation response channel and a deep deformation response channel;

[0021] The matrix elements of the precursor event phase sequence matrix are used to characterize the relative timing state of similar precursor events in different channels. The relative timing state includes leading, synchronous, lagging, and missing.

[0022] The precursor event phase sequence matrix records at least the following relative temporal states:

[0023] The relative temporal state of the initiation event of the triggering factor channel relative to the initiation event of the hydrological response channel;

[0024] The relative temporal state of acceleration events in the hydrological response channel compared to acceleration events in the shallow and deep deformation response channels;

[0025] The relative temporal state of the continuous events of the shallow deformation response channel relative to the continuous events of the deep deformation response channel;

[0026] The state change between the shallow deformation response channel and the deep deformation response channel, from shallow leading to deep leading or vice versa, is denoted as the shallow-deep inversion state.

[0027] Preferably, the event interval vector includes a normalized event interval sub-vector and a phase sequence compression ratio sub-vector;

[0028] The normalized event interval sub-vector is obtained as follows: For the same engineering geological unit, the reference event interval vector is extracted from the historical stable stage of the same engineering geological unit, and the current event interval vector is compared with the reference event interval vector item by item.

[0029] The phase sequence compression ratio subvector is obtained by calculating the ratio of each item in the current event interval vector to the corresponding item in the reference event interval vector.

[0030] The normalized event interval subvector and the phase sequence compression ratio subvector are concatenated to form the event interval vector, which is then input into the machine learning early warning model.

[0031] Preferably, the two deformation response channels at different depths are the shallow deformation response channel and the deep deformation response channel, respectively.

[0032] The trigger-response hysteresis curve includes the hysteresis curves formed by the trigger factor channel and the hydrological response channel, the shallow deformation response channel, and the deep deformation response channel, respectively.

[0033] The hysteresis closure difference vector includes at least the residual opening of each hysteresis curve at the end of a sliding time window;

[0034] The asymmetric vector of the ascending and descending segments includes at least the area difference and slope difference of each hysteresis curve in the ascending and descending segments.

[0035] Preferably, the machine learning early warning model includes:

[0036] An event embedding subnetwork used to encode the phase sequence matrix of precursor events;

[0037] A phase-order attention subnetwork used for weighted modeling of event interval vectors;

[0038] Hysteresis geometric coding subnetwork for geometric coding of hysteresis closed difference vectors and ascending / descending segment asymmetric vectors;

[0039] This is used to receive the output results of the event embedding subnetwork and the output results of the hysteresis geometric coding subnetwork, and to determine whether the phase sequence evolution result is consistent with the hysteresis irreversibility result based on the phase sequence compression ratio subvector in the event interval vector;

[0040] An irreversible fusion subnetwork used to fuse static engineering geological data and the output of a branch consistency constraint subnetwork;

[0041] And a dual-output subnetwork for outputting the irreversible instability score and the earliest irreversible moment, respectively.

[0042] Preferably, the machine learning early warning model is trained in the following way:

[0043] Construct real instability samples using multiple sliding time windows preceding the earliest irreversible moment in historical disaster samples;

[0044] Stable samples are constructed using multiple sliding time windows in historical stable phases where no sustained events have occurred. For each real unstable sample, a matching stable sample is selected from historical stable phases of the same engineering geological unit that are at the same trigger factor level and the same engineering disturbance phase to form a paired sample group of the same unit.

[0045] The precursor event phase sequence matrix, event interval vector, hysteresis closure difference vector, and ascent / depression segment asymmetry vector corresponding to the paired sample groups in the same unit are used as training inputs.

[0046] Preferably, the training of the machine learning early warning model further includes the generation of pseudo-instability samples, which includes:

[0047] Keep the precursor event phase sequence matrix of the real unstable samples unchanged;

[0048] Shorten the event interval in real instability samples according to a preset compression factor;

[0049] Increase the hysteresis closure difference vector and the asymmetric vector of the rising and falling segments in the real instability sample by a preset amplification factor;

[0050] And generate pseudo-instability samples based on the processed samples.

[0051] Preferably, the training of the machine learning early warning model further includes the generation of hard-to-bear samples, which includes at least two of the following:

[0052] Inject transient jump fragments into a single channel of a stable sample;

[0053] Remove intermediate precursor events from stable samples to form a sample of missing events;

[0054] Disrupt the sequential relationship of similar events in different channels within the same sliding time window in stable samples to form phase-order conflict samples.

[0055] Preferably, the training loss function of the machine learning early warning model includes at least a classification loss term and a ranking loss term;

[0056] The ranking loss term constrains the irreversible instability scores of stable samples, hard-to-bear samples, pseudo-instability samples, and real instability samples to satisfy an ordered relationship from low to high.

[0057] Preferably, the phase sequence locking condition satisfies any one of the following:

[0058] The elements of the precursor event phase sequence matrix at the corresponding position remain unchanged within at least three consecutive sliding time windows, and the fluctuation of the event interval vector within at least three consecutive sliding time windows is not greater than a preset fluctuation threshold.

[0059] The elements of the precursor event phase sequence matrix at the corresponding position remain unchanged within at least three consecutive sliding time windows, and the fluctuation of the normalized event interval vector obtained by comparing the event interval vector with the benchmark event interval vector extracted from the historical stable stage of the same engineering geological unit does not exceed the preset fluctuation threshold within at least three consecutive sliding time windows.

[0060] Preferably, it further includes:

[0061] The channel reliability of each channel is calculated based on the packet loss rate, baseline drift, and inconsistency between adjacent channels at the same location.

[0062] When the channel confidence level of the corresponding channel is lower than the preset confidence level threshold, the results of the precursor event extraction for that channel are masked.

[0063] When the shielded channel is either a shallow deformation response channel or a deep deformation response channel, the other unshielded deformation response channel is used to replace the shallow-deep relative timing determination. At the same time, the hysteresis closure difference vector is required to increase monotonically within multiple consecutive sliding time windows before the phase sequence locking condition is satisfied.

[0064] When the blocked channel is a triggering factor channel or a hydrological response channel, the phase sequence locking condition is determined to be met only when the relative temporal state of the same type of precursor event corresponding to the remaining channel remains unchanged and the fluctuation of the hysteresis closure difference vector within multiple consecutive sliding time windows is not greater than the preset hysteresis fluctuation threshold.

[0065] Preferably, the machine learning early warning model outputs the earliest candidate irreversible moment;

[0066] The earliest irreversible moment is the moment obtained by correcting the candidate earliest irreversible moment;

[0067] The earliest irreversible time is obtained by correcting the candidate earliest irreversible time based on the following parameters:

[0068] The number of consecutive sliding time windows;

[0069] The growth rate of the hysteresis closure difference vector over multiple consecutive sliding time windows;

[0070] And the growth rate of the asymmetric vector of the rising and falling segments within multiple consecutive sliding time windows.

[0071] Preferably, the model local calibration step is also included, which includes:

[0072] Collect unlabeled monitoring data of the current engineering geological unit during the stable phase;

[0073] Extracting baseline event interval vectors from unlabeled monitoring data;

[0074] The output calibration parameters of the machine learning early warning model are updated based on the hysteresis closure error distribution formed by unlabeled monitoring data;

[0075] After obtaining the manual verification results or the post-disaster confirmation results, the corresponding samples are written into the incremental sample library. The irreversible fusion subnetwork used to fuse the hysteresis closure difference vector, the asymmetric vector of the rise and fall segment, and the static engineering geological data in the machine learning early warning model, as well as the dual output subnetwork used to output the irreversible instability score and the earliest irreversible moment, are locally retrained.

[0076] In view of this, the present invention also provides an intelligent early warning system for engineering geological disasters, comprising:

[0077] The data acquisition module is used to acquire multi-source monitoring data and static engineering geological data of the same engineering geological unit within the target monitoring area;

[0078] The precursor event extraction module is used to extract the start events, acceleration events, and continuous events of each channel by sliding time windows;

[0079] The phase sequence construction module is used to construct a precursor event phase sequence matrix based on the sequential relationship between each precursor event in different channels, and to construct an event interval vector based on the time interval between adjacent precursor events.

[0080] The hysteresis feature extraction module is used to construct the trigger-response hysteresis curve and extract the hysteresis closure difference vector and the asymmetric vector of the rising and falling segments;

[0081] The machine learning early warning module is used to input the precursor event phase sequence matrix, event interval vector or the normalized event interval vector obtained by comparing the event interval vector with the benchmark event interval vector extracted from the historical stable stage of the same engineering geological unit, hysteresis closure difference vector, rise and fall segment asymmetry vector, and static engineering geological data into a pre-trained machine learning early warning model to obtain the irreversible instability score and the earliest irreversible time. The machine learning early warning model includes an irreversibility fusion subnetwork for fusing the hysteresis closure difference vector, rise and fall segment asymmetry vector, and static engineering geological data, and a dual-output subnetwork for outputting the irreversible instability score and the earliest irreversible time.

[0082] The phase sequence locking determination module is used to determine whether the phase sequence matrix of precursor events meets the phase sequence locking condition; the credibility masking module is used to mask the precursor event extraction results of low credibility channels according to the channel credibility of each channel.

[0083] The local calibration update module is used to update the benchmark event interval vector and the output calibration parameters of the machine learning early warning model, and to perform local retraining on the irreversible fusion subnetwork and the dual-output subnetwork.

[0084] The early warning output module is used to determine the early warning level and generate early warning instructions based on the irreversible instability score, the earliest irreversible moment, and the phase sequence locking condition.

[0085] Compared with existing technologies, the above technical solution has the following advantages:

[0086] 1. Existing technologies typically rely on the temporal trends or comprehensive characteristics of multi-dimensional monitoring data for early warning judgments. They lack more targeted methods to determine whether abnormal fluctuations will further evolve into irreversible instability, thus easily misjudging short-term disturbances, local noise, or recoverable anomalies as true precursors to instability. This invention extracts initial events, acceleration events, and sustained events from multi-source monitoring data to construct a precursor event phase sequence matrix and event interval vector. Furthermore, it constructs a trigger-response hysteresis curve, extracts the hysteresis closure difference vector and the asymmetric vector of the rising and falling segments, and jointly represents the relative temporal relationship with irreversibility characteristics. This is then combined with a machine learning early warning model and phase sequence locking conditions for comprehensive judgment. Therefore, it can more effectively distinguish between recoverable abnormal fluctuations and precursors to irreversible instability, reducing the probability of misjudgment and improving the authenticity and reliability of early warning results.

[0087] 2. In existing technologies, even if an increased risk can be identified, it often only provides a relatively general risk level or development trend, lacking accurate prediction of the leading time for dangerous situations. Furthermore, the warning output is easily affected by short-term fluctuations in a single monitoring value, causing instability. This invention uses a machine learning-based warning model to output an irreversible instability score and the earliest irreversible moment, and combines this with phase sequence locking conditions within multiple consecutive sliding time windows for joint judgment. This ensures that the warning output no longer relies solely on instantaneous anomalies at a single moment, but is based on a continuous evolutionary process. Therefore, it can more accurately determine the warning triggering timing and improve the continuity and stability of the warning output.

[0088] 3. Most existing technologies treat multi-source monitoring data as parallel inputs for comprehensive analysis, but they do not pay enough attention to the evolutionary chain between different channels, such as "which changes first, which responds later, when it accelerates, and when it continues." This invention constructs a precursor event phase sequence matrix to record the relative temporal states between triggering factor channels, hydrological response channels, shallow deformation response channels, and deep deformation response channels. By combining this with event interval vector analysis of the propagation rhythm changes at different stages, it can more clearly reflect the evolutionary path of engineering geological hazards from triggering to response and from shallow to deep, thus improving the ability to identify the hazard formation mechanism and evolution process.

[0089] 4. In existing technologies, an increase in monitoring values ​​does not necessarily mean that the slope or geological body has entered an irreversible unstable state, especially in scenarios such as rainfall, excavation disturbance, or water level fluctuations. Many abnormal changes may still recover after the disturbance disappears. This invention constructs a trigger-response hysteresis curve and extracts the hysteresis closure difference vector and the asymmetric vector of the rise and fall segments to determine the degree of recovery and path difference of the monitored object under the influence of disturbance and after the disturbance subsides. This enables a more accurate identification of whether the system has transitioned from a recoverable state to an irreversible state, enhancing the ability to detect true precursors of instability.

[0090] 5. Existing technologies tend to generate unnecessary warnings when faced with short-term surges, local noise, single-channel glitches, or sporadic disturbances, leading to numerous false alarms and impacting the trust and effectiveness of the warning system among on-site personnel. This invention addresses these issues by setting phase sequence locking conditions, constructing pseudo-instability and difficult-to-bear samples, introducing a sorting loss term, and implementing a channel reliability masking mechanism. This enables the system to learn not only "dangerous samples" but also "samplings that appear abnormal but are not continuous," thus effectively suppressing false alarms caused by short-term abnormal fluctuations and enhancing the system's practical value in actual deployment.

[0091] 6. The engineering site environment is complex, and sensors may experience packet loss, drift, malfunction, or local distortion. Existing technologies, which rely on fixed monitoring channels for judgment, often lead to overall warning failure or unstable judgment after an anomaly in a single critical channel. This invention calculates channel reliability, masks low-reliability channels, and employs a degradation locking mechanism to continue judgment when some channels are abnormal. Therefore, it can maintain the continuity of the warning chain in the event of local measurement anomalies, improving the system's robustness and fault tolerance under complex operating conditions.

[0092] 7. This invention is not limited to rainfall-induced geological disaster scenarios, but can also be applied to monitoring engineering geological disasters such as excavated disturbed slopes, reservoir bank water level fluctuation slopes, and other engineering geological disasters with a "trigger-response" evolution relationship. Because this invention focuses on the relative temporal relationships, event propagation interval compression, and trigger-response irreversibility characteristics in multi-source monitoring data, it has good adaptability to different trigger sources and response channels, thereby improving the invention's applicability under different engineering conditions.

[0093] 8. Different engineering geological units vary in lithology, structural planes, groundwater conditions, support status, and disturbance levels. Using a uniform model for an extended period can easily lead to model drift or insufficient local adaptability. This invention updates the baseline event interval vector using unlabeled monitoring data during the stable phase, corrects the scoring and time prediction results by outputting calibration parameters, and performs local retraining after obtaining manual verification results or post-disaster confirmation results. Therefore, the early warning model can continuously adapt to the actual conditions of the current engineering geological unit, improving the model's stability and adaptability during long-term operation.

[0094] 9. This invention incorporates multiple variables, such as triggering factor channels, hydrological response channels, shallow deformation response channels, and deep deformation response channels, into a unified time synchronization, a unified sliding time window analysis framework, and a unified early warning output mechanism. This enables the formation of multi-variable synchronous records and early warning output records, facilitating subsequent traceability, manual verification, operation and maintenance, and model updates, thereby improving the organization and management efficiency of monitoring information. Attached Figure Description

[0095] Figure 1 This is a schematic diagram of the intelligent early warning system for engineering geological disasters in an embodiment of the present invention.

[0096] Figure 2 This is a flowchart illustrating the intelligent early warning method for engineering geological disasters in an embodiment of the present invention.

[0097] Figure 3 This is a schematic diagram illustrating the construction of the precursor event phase sequence matrix and the extraction of trigger-response hysteresis features in an embodiment of the present invention.

[0098] Figure 4 This is a schematic diagram illustrating the process of model training, phase sequence locking determination, and local calibration update in an embodiment of the present invention.

[0099] Figure 5 This is a schematic diagram of the multi-channel monitoring curve and precursor event markers in an embodiment of the present invention.

[0100] Figure 6 This is a schematic diagram of the trigger-response hysteresis curve and residual opening amount in an embodiment of the present invention.

[0101] Figure 7 This is a schematic diagram of the irreversible instability scoring and phase sequence locking window in an embodiment of the present invention.

[0102] Figure 8 This is a schematic diagram comparing the effects of the embodiments of the present invention with those of comparative embodiments.

[0103] Figure 9 This is a schematic diagram of multivariate synchronous recording and early warning output recording in an embodiment of the present invention.

[0104] Figure 10 This is a schematic diagram illustrating a counterexample in an embodiment of the present invention where abnormal fluctuations can be recovered without triggering an early warning. Detailed Implementation

[0105] The advantages of the present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments.

[0106] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0107] Example 1

[0108] This embodiment uses a high slope on the left side of a mountain highway as an example to describe in detail an intelligent early warning method for engineering geological hazards. The slope is approximately 46m high, with the upper part consisting of strongly weathered argillaceous sandstone, and the middle and lower parts interbedded with thin layers of silty mudstone. A seasonal water catchment zone exists at the toe of the slope, and anti-slide piles and local frame anchor bolts are installed in the middle of the slope. To facilitate continuous monitoring and early warning determination, the slope is first divided into several engineering geological units according to lithological combination, slope structure, and support zones. This embodiment selects one engineering geological unit for description. Rainfall sensors, pore water pressure sensors, shallow displacement gauges, deep inclinometers, crack gauges, and edge acquisition terminals are deployed within this engineering geological unit. The edge acquisition terminal uniformly completes the acquisition, time synchronization, status diagnosis, caching, and uploading of multi-source monitoring data, and writes multiple measured variables into the monitoring record database with a unified time reference, forming a multi-variable synchronous record file and an early warning output record file. Figure 9 The diagram illustrates the form of multivariate synchronous recording and early warning output recording.

[0109] In this embodiment, the multi-source monitoring data includes at least one trigger factor channel, one hydrological response channel, and two deformation response channels at different depths. The trigger factor channel uses rainfall intensity monitoring values, the hydrological response channel uses pore water pressure monitoring values, and the two deformation response channels at different depths are respectively a shallow deformation response channel and a deep deformation response channel. The shallow deformation response channel is jointly reflected by the slope surface GNSS micro-displacement and shallow crack propagation, while the deep deformation response channel is reflected by deep inclinometer displacement. To ensure the repeatability of subsequent calculations, this embodiment sets the sampling period to 10 minutes, and six consecutive sampling periods constitute a sliding time window with a sliding step size of 10 minutes, meaning the sliding time window is updated every 10 minutes.

[0110] For ease of explanation, this embodiment selects a typical time period before the warning as a calculation example. The monitoring values ​​of the four main channels during this time period are shown in Table 1. Figure 5 The corresponding multi-channel monitoring curves and precursor event markers are given.

[0111] Table 1. Examples of multi-channel monitoring data for typical time periods.

[0112]

[0113] This embodiment further provides the original monitoring record segments and event identification results within a certain consecutive 6 sampling periods, as shown in Table 2.

[0114] Table 2 Examples of Original Monitoring Record Fragments

[0115]

[0116] In the data preprocessing stage, the original monitoring sequences are first denoised and the stationary baseline is corrected. This embodiment uses a three-point sliding smoothing and mean regression correction method for the stable phase to eliminate occasional spikes and baseline drift. For any channel, the local change slope, local curvature, sign holding length, and cumulative deviation can be calculated within the corresponding sliding time window. To illustrate the calculation process, the following uses the trigger factor channel as an example to explain the specific calculation method of the local change slope, local curvature, and cumulative deviation; for the hydrological response channel, shallow deformation response channel, and deep deformation response channel, the same calculation logic is used, only the monitoring values ​​and stationary baseline values ​​of the corresponding channels are replaced with the same calculation formula.

[0117] In the At each sampling time, the trigger factor channel monitoring value is recorded as... The hydrological response channel monitoring value is recorded as The shallow deformation response channel monitoring value is recorded as The deep deformation response channel monitoring value is recorded as The slope of the local change in the trigger factor channel is denoted as... The local curvature is denoted as The cumulative deviation is recorded as The sampling time interval is denoted as In this embodiment The slope of the local change in the trigger factor channel is calculated according to equation (1):

[0118]

[0119] The local curvature of the trigger factor channel is calculated according to equation (2):

[0120]

[0121] The cumulative deviation of the trigger factor channel is calculated according to equation (3):

[0122]

[0123] in, To be the stationary baseline value of the trigger factor channel, This represents the number of sampling points participating in the accumulation within the current sliding time window. In this embodiment, the stationary baseline value of the trigger factor channel is 3. For the hydrological response channel, the shallow deformation response channel, and the deep deformation response channel, the values ​​in equations (1) to (3) can be respectively... Replace with , or and will Replace with the stationary baseline value of the corresponding channel to obtain the local slope, local curvature and cumulative deviation of the corresponding channel.

[0124] Taking the trigger factor channel data at 30 min, 40 min, and 50 min as examples, and referring to Table 1, we can see that at 30 min, the trigger factor channel monitoring value decreased from 4 at 20 min. Change to 30 minutes 6 Therefore, the slope of the local change is

[0125]

[0126] At 40 minutes, the trigger factor channel monitoring value changed from 6 at 30 minutes. Change to 12 for 40 minutes Therefore, the slope of the local change is

[0127]

[0128] At 50 minutes, the trigger factor channel monitoring value changed from 12 at 40 minutes. Change to 50 minutes 24 Therefore, the slope of the local change is

[0129]

[0130] Meanwhile, the local curvature corresponding to 40 minutes is

[0131]

[0132] This embodiment further explains the determination method of the relevant thresholds. The first, second, third, and fourth preset thresholds are all determined based on the sample statistics of the historical stable phase of the same engineering geological unit and the results of manual verification. Specifically, the first preset threshold is the 95th quantile of the absolute value of the local change slope in the historical stable phase; the second preset threshold is the mean of the local change slope increment plus 2 standard deviations; the third preset threshold is the 95th quantile of the local curvature; and the fourth preset threshold is the median value between the minimum effective value of the cumulative deviation in the confirmed precursor samples and the maximum value in the stable samples. The preset scoring threshold, preset time threshold, preset fluctuation threshold, preset confidence threshold, and preset hysteresis fluctuation threshold are determined through cross-validation of the training set and the validation set. In this embodiment, the first preset threshold is set to 0.15, the second preset threshold to 0.25, the third preset threshold to 0.03, the fourth preset threshold to 180, the preset scoring threshold to 0.72, the preset time threshold to 20 minutes, the normalized event interval vector fluctuation threshold to 0.05, the channel reliability threshold to 0.70, and the hysteresis fluctuation threshold to 0.08. For different engineering geological units, the same method can be used for calibration separately; uniformity across the entire field is not required.

[0133] To illustrate the process of determining the preset scoring threshold, this embodiment uses a validation set as an example. Assume that the irreversible instability scores of stable samples in the validation set are mainly distributed between 0.10 and 0.58, the irreversible instability scores of recoverable abnormal fluctuation samples are mainly distributed between 0.32 and 0.69, and the irreversible instability scores of true irreversible instability precursor samples are mainly distributed between 0.70 and 0.95. We tried setting the preset scoring threshold to 0.68, 0.70, 0.72, 0.74, and 0.76 respectively, and statistically analyzed the overall identification accuracy, false alarm rate, and false negative rate, obtaining the results shown in Table 3.

[0134] Table 3 Example of Preset Scoring Threshold Selection

[0135]

[0136] As shown in Table 3, when the preset scoring threshold is 0.72, the overall recognition accuracy is the highest, and the false alarm rate and false negative rate are in a relatively good balance. Therefore, this embodiment selects 0.72 as the preset scoring threshold.

[0137] If, in this embodiment, the first preset threshold is 0.15, the second preset threshold is 0.25, the third preset threshold is 0.03, and the fourth preset threshold is 180, then it can be seen that the local change slope is as follows at 30 minutes. The first preset threshold has been exceeded, and the direction of change remains continuous, therefore the starting event of the trigger factor channel can be extracted; during the period from 40 min to 50 min, the local change slope increment increased from 0.20 to 0.60, and then to 1.20, with local curvature... If the cumulative deviation exceeds the third preset threshold, it continues to increase, thus accelerating events in the trigger factor channel can be extracted. After 70 minutes, although the trigger factor channel begins to decline, the cumulative deviations in the hydrological response channel and the deformation response channel continue to increase and have not returned to the stable baseline, thus continuous events can be identified for the response channel. To avoid occasional noise, this embodiment only retains precursor events that are detected in both adjacent sliding time windows and whose event start time deviation is no more than 10 minutes.

[0138] After the above processing, this embodiment identifies the typical event times within this time period as follows: the triggering factor channel's initial event occurs at 30 minutes, and the acceleration event occurs at 50 minutes; the hydrological response channel's initial event occurs at 40 minutes, and the acceleration event occurs at 60 minutes; the shallow deformation response channel's acceleration event occurs at 70 minutes, and the sustained event occurs at 90 minutes; the deep deformation response channel's acceleration event occurs at 80 minutes, and the sustained event occurs at 100 minutes. Based on these event times, a precursor event phase sequence matrix is ​​further constructed. The matrix elements of the precursor event phase sequence matrix are used to characterize the relative temporal state of similar precursor events in different channels. The relative temporal state includes leading, synchronous, lagging, and missing. To facilitate input into the machine learning early warning model, this embodiment optionally denotes leading as... Synchronization is recorded as Delay is recorded as Missing is recorded as ,

[0139] Based on the aforementioned event times, the starting event of the triggering factor channel precedes the starting event of the hydrological response channel by 10 minutes, thus corresponding to a leading relative time sequence. Similarly, the acceleration event of the hydrological response channel precedes the acceleration event of the shallow deformation response channel by 10 minutes, also leading in relative time sequence. Furthermore, the acceleration event of the hydrological response channel precedes the acceleration event of the deep deformation response channel by 20 minutes, resulting in a leading relative time sequence. Finally, the continuation event of the shallow deformation response channel precedes the continuation event of the deep deformation response channel by 10 minutes, also leading in relative time sequence. In this embodiment, there is no instance of the shallow deformation response channel changing from leading to leading, or vice versa, thus eliminating the possibility of a shallow-deep inversion. Figure 3 A schematic diagram of phase sequence matrix construction and trigger-response feature extraction is given.

[0140] While obtaining the precursor event phase sequence matrix, an event interval vector is further constructed. To avoid conceptual confusion, this embodiment first records the event interval before it is compared with the reference event interval vector as the current event interval vector, which is denoted as... The event intervals extracted from the historical stable phases of the same engineering geological unit are denoted as the baseline event interval vector, and the baseline event interval vector is denoted as... Based on this, the normalized event interval sub-vector and the phase order compression ratio sub-vector are calculated separately, and then concatenated to form the event interval vector input to the machine learning early warning model. Therefore, in this embodiment, the "current event interval vector" is the original interval, and the "event interval vector input to the machine learning early warning model" is the final input feature formed by concatenating the two sub-vectors.

[0141] In this embodiment, the following four sets of key event intervals are selected: the interval from the start event of the triggering factor channel to the start event of the hydrological response channel, the interval from the acceleration event of the hydrological response channel to the acceleration event of the shallow deformation response channel, the interval from the acceleration event of the hydrological response channel to the acceleration event of the deep deformation response channel, and the interval from the continuation event of the shallow deformation response channel to the continuation event of the deep deformation response channel. Based on the aforementioned event times, the current event interval vector is:

[0142]

[0143] The baseline event interval vector is obtained from the statistical analysis of the historical stable stages of the same engineering geological unit:

[0144]

[0145] The normalized event interval subvector is denoted as In this embodiment, "item-by-item comparison" is represented by the normalized deviation of the current interval relative to the reference interval, and its calculation formula is:

[0146]

[0147] Therefore, the normalized event interval subvector is used to characterize the degree of deviation of the current event propagation chain from the stable phase.

[0148] The phase sequence compression ratio subvector is denoted as The phase sequence compression ratio sub-vector is calculated by comparing each term in the current event interval vector with the corresponding term in the reference event interval vector; therefore, its calculation formula is:

[0149]

[0150] This shows that the event propagation interval of the current engineering geological unit is significantly compressed during the relatively stable phase within this time period, indicating that the propagation speed of precursor events is accelerating. Subsequently, the normalized event interval subvector and the phase sequence compression ratio subvector are concatenated to form the event interval vector input to the machine learning early warning model.

[0151] In constructing the trigger-response hysteresis characteristics, this embodiment uses the monitoring value of the triggering factor channel on the horizontal axis and the monitoring values ​​of the hydrological response channel, the shallow deformation response channel, and the deep deformation response channel on the vertical axis to form three trigger-response hysteresis curves respectively. Figure 6 A typical hysteresis curve between the triggering factor channel and the hydrological response channel is shown. The single residual opening in the hysteresis closure difference vector is denoted as... In this embodiment, the Euclidean distance between the start and end points of the hysteresis curve is used for characterization, and its calculation formula is as follows:

[0152]

[0153] in, and These are the monitoring values ​​of the triggering factor channel and the hydrological response channel at the starting point of the hysteresis curve, respectively. and These are the monitoring values ​​of the triggering factor channel and the hydrological response channel at the endpoint of the hysteresis curve, respectively. As shown in Table 1, the starting point of the hysteresis curve in this embodiment is approximately... The endpoint is approximately Substituting into equation (6), we get:

[0154]

[0155] This indicates that after the monitoring values ​​of the triggering factor channel basically returned to the initial level, the monitoring values ​​of the hydrological response channel did not return to the initial state synchronously, and still retained a large residual opening, indicating that the system has obvious irreversibility.

[0156] The area difference term in the asymmetric vector of the ascending and descending segments is denoted as... This embodiment interpolates within the common trigger factor channel range for both the ascending and descending branches, and then uses trapezoidal integrals to calculate the areas of the ascending and descending segments. The common range is used as the basis for calculation. For example, the calculated area of ​​the ascending segment is 717.5, and the area of ​​the descending segment is 1096.0. Therefore, the area difference is...

[0157]

[0158] The slope difference term in the asymmetric vector of the ascending and descending segments is denoted as... This embodiment selects the ascending segment. to The slope and the descending segment to The slope is used as a representative, where the ascending segment represents the slope.

[0159]

[0160] The descending segment represents the absolute value of the slope.

[0161]

[0162] The slope difference is

[0163]

[0164] After calculating the phase sequence features and hysteresis features, they are input into the machine learning early warning model. The machine learning early warning model in this embodiment includes an event embedding subnetwork, a phase sequence attention subnetwork, a hysteresis geometric coding subnetwork, a branch consistency constraint subnetwork, an irreversibility fusion subnetwork, and a dual-output subnetwork. The branch consistency constraint subnetwork is used to output consistency constraint values. When the propagation acceleration trend represented by the phase sequence compression ratio subvector is consistent with the irreversibility enhancement trend represented by the hysteresis closure difference vector and the ascending / descending segment asymmetric vector, the consistency constraint value increases. When the phase sequence compression ratio subvector represents accelerated propagation, but the hysteresis closure difference vector or the ascending / descending segment asymmetric vector does not show corresponding enhancement, the consistency constraint value decreases, and the increase in irreversible instability score is suppressed. In this way, premature output of high-risk results based solely on a single branch anomaly can be avoided.

[0165] To illustrate the calculation process, this embodiment provides a specific scoring example. Let the phase sequence compression index be denoted as...

[0166]

[0167] Let the normalized value of the hysteresis closure error be denoted as In this embodiment, 10 is used as the normalization scale, then:

[0168]

[0169] Let the normalized value of asymmetry be denoted as . In this embodiment, the area difference of 400 and the slope difference of 0.5 are used as normalization scales, and the average of the two is taken to obtain:

[0170]

[0171] Let the branch consistency value be denoted as Since the event propagation interval is significantly compressed in this embodiment, and the hysteresis closure error and asymmetry increase simultaneously, the branch consistency value is set to 1. An exemplary calibration output of the machine learning early warning model can be calculated according to equation (7):

[0172]

[0173] Substituting the above values, we get:

[0174]

[0175] Furthermore, the irreversible instability score is denoted as The Sigmoid transform is used for calculation:

[0176]

[0177] Substitution We can obtain:

[0178]

[0179] If the preset scoring threshold in this embodiment is 0.72, then the irreversible instability score corresponding to this time period has exceeded the preset scoring threshold.

[0180] The dual-output subnetwork also outputs candidate earliest irreversible moments. Assuming the current time is 90 minutes, the machine learning warning model outputs a candidate earliest irreversible moment of 108 minutes. This is subsequently corrected based on the number of consecutive sliding time windows, the growth rate of the hysteresis closure difference vector, and the growth rate of the asymmetric vector of the ascending and descending segments. Let the number of consecutive locking windows be denoted as... The growth rate of the hysteresis loop is denoted as The growth rate of the asymmetric vector in the rising and falling segments is denoted as... Then the correction amount It can be calculated by example using formula (8):

[0181]

[0182] If, in this embodiment, the number of consecutive locking windows is 4, the hysteresis closure error growth rate is 1.2, and the ascent / descendance segment asymmetric vector growth rate is 0.3, then the correction amount is:

[0183]

[0184] The earliest irreversible moment after correction is denoted as ,but:

[0185]

[0186] Therefore, at the current time of 90 minutes, the time interval from the current time to the earliest irreversible time is approximately 9 minutes. If the preset time threshold in this embodiment is 20 minutes, this condition is also satisfied. Furthermore, when the growth rate of the hysteresis closure error and the growth rate of the asymmetric vector of the rising and falling segments increase synchronously, it indicates that the irreversibility of the trigger-response process is enhanced; therefore, the earliest irreversible time is corrected towards the current time. When the two do not increase synchronously or the growth trend weakens, the original candidate earliest irreversible time is maintained or the correction magnitude is reduced. In this way, the output of the earliest irreversible time can better reflect the actual instability evolution speed.

[0187] In addition to the output of the machine learning early warning model, this embodiment also uses phase sequence locking conditions for joint determination. Figure 7 The diagram illustrates the changes in irreversible instability score and phase sequence locking window. Assuming that within sliding time windows 9, 10, 11, and 12, the elements of the precursor event phase sequence matrix remain unchanged, and the fluctuations of the normalized event interval vector are 0.02, 0.01, 0.03, and 0.02 respectively, all less than the preset fluctuation threshold of 0.05, then the phase sequence locking condition for multiple consecutive sliding time windows can be determined. At this point, the irreversible instability score exceeds the preset score threshold, the time interval from the current moment to the earliest irreversible moment is not greater than the preset time threshold, and the phase sequence locking condition is met; therefore, a warning level and warning command can be output.

[0188] To further solidify the implementation method of channel reliability, this embodiment provides a complete calculation example. Let the packet loss rate of the shallow deformation response channel within a certain evaluation period be denoted as... The baseline drift is denoted as The degree of inconsistency between adjacent passages at the same location is denoted as The packet loss rate is calculated according to formula (9):

[0189]

[0190] The baseline drift is calculated according to formula (10):

[0191]

[0192] in, This is the average value during the current stable phase. To reference the mean during the stable phase, To prevent small positive numbers with a denominator of zero, this embodiment uses 0.001. The degree of contradiction is calculated according to formula (11):

[0193]

[0194] in, For the evaluated channel in the Smoothed monitoring values ​​at each sampling point These are the corresponding monitoring values ​​of adjacent channels at the same location, after proportional conversion. This represents the number of sampling points.

[0195] Channel reliability is denoted as The calculation is performed using a weighted deduction method:

[0196]

[0197] Assuming the packet loss rate of the shallow deformation response channel is 0.12 during a certain evaluation period, the current stable phase average is 1.90 mm, and the reference stable phase average is 1.00 mm, then the baseline drift is:

[0198]

[0199] Assuming the inconsistency between adjacent channels at the same location is calculated to be 0.15, then the channel reliability is:

[0200]

[0201]

[0202] In this embodiment, the preset confidence threshold is 0.70. Therefore, the channel confidence of the shallow deformation response channel is lower than the preset confidence threshold, and the precursor event extraction results of this channel need to be masked.

[0203] To illustrate the channel reliability and degradation locking mechanism, this embodiment further assumes that the shallow deformation response channel experiences the aforementioned 12% packet loss rate, corresponding baseline drift, and inconsistency with the deep deformation response channel during a certain period. In this case, the calculated channel reliability is lower than a preset reliability threshold, and therefore the precursor event extraction results of the shallow deformation response channel are masked. At this point, instead of relying on complete shallow-deep relative timing determination, the deep deformation response channel replaces part of the shallow-deep relative timing determination. Simultaneously, the hysteresis closure difference vector is required to monotonically increase within multiple consecutive sliding time windows before further determination is allowed to satisfy the phase sequence locking condition. This degradation locking mechanism prevents the failure of a single channel from disrupting the entire early warning chain.

[0204] Furthermore, when multiple channels simultaneously fall below a preset confidence threshold, and the remaining channels are insufficient to form a valid relative temporal state chain, this embodiment does not perform degradation locking. Instead, it directly outputs a "current measurement confidence is insufficient" status flag and maintains the warning level of the previous period unchanged until the effective monitored variable recovers. This boundary handling method can avoid generating erroneous warnings due to insufficient information when multiple channels are simultaneously abnormal.

[0205] To further demonstrate that this embodiment can distinguish between recoverable abnormal fluctuations and irreversible instability precursors, a counterexample scenario is given below. Figure 10 This paper illustrates the multi-channel response changes under a recoverable abnormal fluctuation scenario. In this scenario, the trigger factor channel monitoring value rises rapidly within a short period and then falls back quickly. The hydrological response channel monitoring value and the shallow deformation response channel monitoring value show short-term responses, but return to near the stable baseline within two sliding time windows. Correspondingly, although initiation events and even short-term acceleration events can be detected in individual sliding time windows, the initiation time of these events is not stable in adjacent sliding time windows, and it is impossible to form a precursor event phase sequence matrix element that remains unchanged within multiple consecutive sliding time windows. At the same time, the residual opening of the trigger-response hysteresis curve is small, and the asymmetry of the rise and fall segments is not significant.

[0206] To ensure this counterexample forms a complete numerical closed loop, assume the current event interval vector in this scenario is... The baseline event interval vector is The normalized event interval subvector is then:

[0207]

[0208] The phase sequence compression ratio subvector is:

[0209]

[0210] At this point, it is evident that event propagation has not been significantly compressed. Further assuming the residual opening of the trigger-response hysteresis curve is 1.10, the area difference is 56.0, and the slope difference is 0.08, the hysteresis irreversibility is significantly weaker than in the main implementation scenario. Substituting the same scoring calculation logic, the highest irreversible instability score output by the machine learning early warning model is only 0.61, lower than the preset scoring threshold of 0.72. Simultaneously, the fluctuations of the normalized event interval vector within four consecutive sliding time windows are 0.11, 0.08, 0.12, and 0.09, respectively, all higher than the preset fluctuation threshold of 0.05, thus failing to meet the phase sequence locking condition. Therefore, although the local monitored value shows a significant short-term increase in this scenario, it will not trigger an early warning output. This counterexample illustrates that this embodiment does not directly trigger an alarm whenever an anomaly occurs, but rather effectively distinguishes between recoverable abnormal fluctuations and irreversible instability precursors through a joint determination of the phase sequence stability of the precursor event, the irreversibility of the trigger-response, and the locking relationship of the continuous sliding time windows.

[0211] To more intuitively illustrate the process of multivariate integrated recording and early warning output, Figure 9A schematic diagram of multivariate synchronous recording and early warning output recording is provided. In this recording, the monitoring values ​​of the triggering factor channel, hydrological response channel, shallow deformation response channel, deep deformation response channel, precursor event phase sequence matrix state, event interval vector, hysteresis closure difference vector, irreversible instability score, earliest irreversible time, and early warning output result are recorded simultaneously within the same sliding time window. This unified recording method allows for the traceability of each step of the early warning output and generates a standard record report that can be used for subsequent manual verification and local model calibration. For the instruction manual, this multivariate synchronous recording not only reflects this embodiment but also helps to illustrate that this embodiment is not a simple threshold alarm for a single variable, but rather a comprehensive joint judgment based on multiple monitored variables.

[0212] In terms of model training, this embodiment adopts a paired training method within the same unit. Historical data within the same engineering geological unit are labeled, and multiple sliding time windows before the earliest irreversible moment are selected to form real instability samples. Historical stable phase data under the same trigger factor level and the same engineering disturbance stage are selected to form matched stable samples. Furthermore, pseudo-instability samples and difficult-to-bear samples are generated. Pseudo-instability samples keep the precursor event phase sequence matrix unchanged, while shortening the event interval by a preset compression factor and increasing the hysteresis closure difference vector and the rise-fall segment asymmetry vector by a preset amplification factor, to simulate the state approaching the instability boundary but without actual failure. Difficult-to-bear samples are generated through three methods: instantaneous jump segments, missing events, and phase sequence conflicts, to simulate pseudo-anomaly states that are prone to false alarms. During training, the ranking loss term requires that the irreversible instability scores corresponding to stable samples, difficult-to-bear samples, pseudo-instability samples, and real instability samples satisfy an ordered relationship from low to high, thereby making it easier for the model to learn the boundary of the transition from recoverable anomalous fluctuations to irreversible instability precursors.

[0213] This embodiment further details the engineering training of the machine learning early warning model. Training samples are divided into training, validation, and test sets in chronological order, with the training set accounting for 70%, the validation set for 15%, and the test set for 15%. The training run is 120 rounds, the batch size is 64, and the initial learning rate is 0.001, decaying using cosine annealing. The weight ratio of the classification loss term to the ranking loss term is 3:2. After every 10 rounds of training, the accuracy of identifying irreversible instability precursors, the false alarm rate of recoverable abnormal fluctuations, and the ranking consistency index are calculated on the validation set. If there is no improvement in the validation results for 8 consecutive rounds, early termination is triggered. Output calibration parameters include a scoring offset and a time correction coefficient. The scoring offset is used to correct the overall bias of the irreversible instability score, and the time correction coefficient is used to correct the prediction bias of the earliest irreversible time. The update cycle for output calibration parameters is set to 7 days. The retraining trigger conditions for the local calibration step are: first, the number of newly labeled samples verified by manual verification reaches 30; second, the false alarm rate or false negative rate in the last 14 days increases by more than 20% compared to the historical average; and third, the prediction error of the earliest irreversible moment exceeds the preset error threshold by 10 minutes under multiple consecutive sliding time windows. When any of the above conditions are met, the scoring offset and time correction coefficient are updated, and local retraining is performed on the irreversibility fusion subnetwork and the dual-output subnetwork.

[0214] The test data in this embodiment came from on-site monitoring records of the same slope for eight consecutive months, covering both the high-water and low-water seasons. Manual verification results were obtained from on-site inspection records, crack measurement records, support component inspection records, and post-disaster follow-up records. Sample labeling was jointly completed by monitoring engineers, on-site geological engineers, and data analysis engineers. The earliest irreversible moment was labeled based on factors such as continuous crack propagation, continuous increase in deep deformation, prolonged failure of pore water pressure to decrease, and on-site deformation treatment records.

[0215] To further verify the technical effectiveness of this embodiment, a comparative implementation method is set up. The comparative implementation method adopts the approach of existing technologies in the background art, that is, directly performing machine learning judgment based on the comprehensive characteristics and development trends of multi-dimensional monitoring data, without constructing a precursor event phase sequence matrix, extracting hysteresis closure difference vectors and asymmetric vectors of rising and falling segments, or setting phase sequence locking conditions. The test samples were taken from the same slope, the same engineering geological unit, with the same sampling period and the same proportions for training, validation, and test sets. The distribution of the test samples is shown in Table 4.

[0216] Table 4. Examples of Experimental Sample Distribution

[0217]

[0218] Table 5 compares the effects of this embodiment with those of the comparative embodiment. Figure 8 As shown.

[0219] Table 5 Comparison of the effects of this embodiment and the comparative embodiment.

[0220]

[0221] From Table 5 and Figure 8 It can be seen that this embodiment outperforms the comparative implementation in terms of accuracy in identifying irreversible instability precursors, false alarm rate of recoverable abnormal fluctuations, average early warning lead time, and trigger rate of phase sequence locking across multiple sliding time windows. This indicates that this embodiment, through the combined application of precursor event phase sequence matrix, event interval vector, trigger-response hysteresis characteristics, and phase sequence locking conditions, can more effectively distinguish between recoverable abnormal fluctuations and irreversible instability precursors.

[0222] To further illustrate the role of each key technical step, ablation experiments were conducted. The precursor event phase sequence matrix was removed, retaining only the original trend and hysteresis features; the hysteresis closure difference vector and the asymmetric vector of the rising and falling segments were removed, retaining only the phase sequence features; the phase sequence locking condition was removed, and the warning was directly output by the machine learning warning model. The ablation results are shown in Table 6.

[0223] Table 6 Comparison of Ablation Techniques in Key Technologies

[0224]

[0225] Table 6 shows that the precursor event phase sequence matrix, hysteresis irreversibility feature, and phase sequence locking condition all significantly contribute to the overall effect. Specifically, removing the hysteresis irreversibility feature significantly increases the false alarm rate of recoverable abnormal fluctuations, indicating that the hysteresis closure difference vector and the asymmetric vector of the rising and falling segments have a direct effect on identifying irreversibility. Removing the phase sequence locking condition significantly increases the number of intraday jitters in the warning results, indicating that the phase sequence locking condition has a significant effect on suppressing short-term false triggers.

[0226] Furthermore, this embodiment also compares whether a degradation locking mechanism is used under abnormal channel conditions. Assuming that the shallow deformation response channel has 8% communication packet loss and the hydrological response channel has 5% instantaneous drift, the statistical results are shown in Table 7.

[0227] Table 7. Stability Comparison Under Abnormal Channel Conditions

[0228]

[0229] As shown in Table 7, even when some channels are abnormal or have low reliability, this embodiment can still maintain a high accuracy and a low false alarm rate and false negative rate through the degradation locking mechanism, indicating that this embodiment has good stability under complex field conditions.

[0230] In addition to the scenario where rainfall intensity is used as the triggering factor channel, the method in this embodiment can also be used in excavation disturbance scenarios. For this scenario, the triggering factor channel can use the excavation unloading rate or the monitoring value of blasting vibration intensity, while the hydrological response channel can still use the monitoring value of pore water pressure or the rate of change of groundwater level. The construction methods of the shallow deformation response channel and the deep deformation response channel remain unchanged. In this case, the key relative temporal states in the precursor event phase sequence matrix can be reflected as follows: the excavation unloading initiation event precedes the shallow deformation response channel acceleration event, the shallow deformation response channel acceleration event precedes the deep deformation response channel acceleration event, or the blasting vibration anomaly event precedes the crack widening continuation event. When the phase sequence compression ratio sub-vector caused by excavation disturbance continuously decreases and the hysteresis closure difference vector continuously increases, the method in this embodiment can also be used to output the irreversible instability score and the earliest irreversible moment. Therefore, this embodiment is not limited to the identification of rainfall-induced slope instability precursors, but is also applicable to excavation disturbance-induced slopes or engineering slope scenarios.

[0231] The method in this embodiment can also be applied to scenarios involving reservoir bank water level fluctuations. To provide a simplified but complete example for this scenario, a set of exemplary monitoring data is given below. Assume the trigger factor channel uses the reservoir water level change rate, with units of... The hydrological response channel uses pore water pressure changes in kPa; the shallow deformation response channel and the deep deformation response channel use slope surface displacement increment and deep inclinometer displacement increment, respectively, both in mm. Typical monitoring results for a certain period are shown in Table 8.

[0232] Table 8 Examples of Monitoring Data for Reservoir Bank Water Level Fluctuation Scenarios

[0233]

[0234] Table 8 reveals that: the initial event of the reservoir water level change rate occurs between 6 and 12 hours; the initial event of the pore water pressure occurs at 12 hours; the accelerated event of the shallow deformation response channel occurs between 18 and 24 hours; and the accelerated event of the deep deformation response channel occurs between 24 and 30 hours. Therefore, the current event interval vector is obtained as follows: Assuming the baseline event interval vector extracted from the historical stable phase of the same engineering geological unit is... Then the normalized event interval subvector is:

[0235]

[0236] The phase sequence compression ratio subvector is:

[0237]

[0238] Furthermore, a trigger-response hysteresis curve is constructed with the reservoir water level change rate on the horizontal axis and the pore water pressure change rate on the vertical axis. If the residual opening of this hysteresis curve is calculated to be 5.20, the area difference is 210.0, and the slope difference is 0.19, it indicates that after the rapid change in reservoir water level, the pore water pressure has not synchronously recovered to a stable state, and the system has shown obvious hysteresis irreversibility. After inputting the above phase sequence characteristics and hysteresis characteristics into the machine learning early warning model, assuming the output irreversible instability score is 0.78, and the earliest candidate irreversible time is still 10 hours away from the current time; combined with the fact that the elements of the precursor event phase sequence matrix remain unchanged within three consecutive sliding time windows and the fluctuation of the normalized event interval vector is less than the preset fluctuation threshold, it can be determined that the phase sequence locking condition is met, so an early warning level and early warning command are output. It can be seen that the method of this embodiment does not rely on rainfall as a specific trigger source. As long as the relative temporal relationship of precursor events and the trigger-response irreversibility characteristics in multi-source monitoring data can be formed, the same analysis and judgment logic can be used.

[0239] In summary, this embodiment, through techniques such as precursor event extraction, precursor event phase sequence matrix construction, event interval vector construction, trigger-response hysteresis feature extraction, joint judgment of machine learning early warning models, phase sequence locking condition judgment, channel credibility degradation locking, and local calibration update, can more accurately distinguish between recoverable abnormal fluctuations and irreversible instability precursors in the evolution of engineering geological disasters. It can also maintain early warning capabilities when some multi-source monitoring channels fail. Furthermore, it is applicable to various engineering geological disaster monitoring scenarios such as rainfall-induced, excavation disturbance-induced, and reservoir bank water level fluctuation-induced disasters, thus possessing significant engineering application value.

[0240] Example 2

[0241] This embodiment further illustrates an intelligent early warning system for engineering geological disasters. Since this system is used to implement the steps in the aforementioned method implementations, its monitoring objects, engineering geological unit division methods, multi-source monitoring data sources, precursor event identification rules, phase sequence construction rules, hysteresis feature extraction methods, machine learning early warning models, phase sequence locking conditions, channel reliability calculation methods, and local calibration update methods can all adopt the corresponding settings in the aforementioned method implementations. This embodiment will not repeat the same content in its explanation.

[0242] like Figure 1As shown, the intelligent early warning system for engineering geological disasters includes a data acquisition module, a precursor event extraction module, a phase sequence construction module, a hysteresis feature extraction module, a machine learning early warning module, a phase sequence locking determination module, a credibility shielding module, a local calibration and update module, and an early warning output module. These modules can be integrated into the same industrial computing device, or deployed collaboratively using edge acquisition terminals, on-site industrial control computers, and remote servers. Specifically, the data acquisition module can communicate with rain gauges, pore water pressure sensors, shallow displacement gauges, deep inclinometers, crack gauges, and other on-site monitoring equipment to acquire multi-source monitoring data and static engineering geological data for the same engineering geological unit; the precursor event extraction module processes the monitoring sequences of each channel according to a sliding time window, extracting the initial event, acceleration event, and continuous event; the phase sequence construction module constructs a precursor event phase sequence matrix based on the sequential relationship between different channels and forms an event interval vector based on the time interval between adjacent precursor events; the hysteresis feature extraction module constructs the trigger-response hysteresis curve and extracts the hysteresis closure difference vector and the asymmetric vector of the rising and falling segments; the machine learning early warning module is used for... The precursor event phase sequence matrix, event interval vector, hysteresis closure difference vector, ascent-descent segment asymmetry vector, and static engineering geological data are input into a pre-trained machine learning early warning model to output an irreversible instability score and the earliest irreversible time. The phase sequence locking determination module is used to determine the phase sequence stability within multiple consecutive sliding time windows. The credibility masking module is used to mask the precursor event extraction results of low credibility channels according to the channel credibility of each channel. The local calibration update module is used to update the baseline event interval vector and output calibration parameters, and perform local retraining on the machine learning early warning model. The early warning output module is used to determine the early warning level and output early warning instructions based on the irreversible instability score, the earliest irreversible time, and the phase sequence locking determination results.

[0243] In this embodiment, the data acquisition module is preferably located in the edge acquisition terminal to complete unified time synchronization, data caching, and preliminary quality checks for multiple source channels on site. The precursor event extraction module, phase sequence construction module, and hysteresis feature extraction module can be located in the on-site industrial control computer or edge computing unit to reduce the communication burden caused by directly uploading raw monitoring data. The machine learning early warning module, phase sequence locking determination module, and local calibration update module can be located in the on-site industrial control computer or remote server to complete more complex feature fusion, early warning determination, and model updates. The early warning output module can communicate with audible and visual alarm devices, SMS gateways, monitoring platform display terminals, and duty personnel terminals to achieve multi-form output of early warning results. For slope monitoring scenarios with poor communication conditions, the machine learning early warning module and the phase sequence locking determination module can also be deployed together on the edge side, enabling the system to complete local early warning even in offline or weak network environments.

[0244] When the system is running, the data acquisition module first acquires the monitoring data of the trigger factor channel, hydrological response channel, shallow deformation response channel, and deep deformation response channel of the same engineering geological unit within the target monitoring area, and sends the static engineering geological data into the system as well. Subsequently, the precursor event extraction module extracts the initiation event, acceleration event, and continuous event from the monitoring sequence of each channel; the phase sequence construction module constructs the precursor event phase sequence matrix based on the relative temporal relationship of various events and forms the current event interval vector; the hysteresis feature extraction module forms the trigger-response hysteresis curve based on the changing relationship between the trigger factor channel and each response channel, and further obtains the hysteresis closure difference vector and the asymmetric vector of the rising and falling segments; the machine learning early warning module performs joint analysis on the above features and outputs the irreversible instability score and the earliest irreversible time; the phase sequence locking determination module then determines the phase sequence stability within multiple consecutive sliding time windows; when the irreversible instability score is not lower than the preset score threshold, the time interval from the current time to the earliest irreversible time is not greater than the preset time threshold, and the phase sequence locking condition is met, the early warning output module outputs the corresponding early warning level and early warning instruction. If the credibility masking module determines that the credibility of a certain channel is lower than the preset credibility threshold, it masks the precursor event extraction results of the corresponding channel, and the phase sequence locking determination module performs degradation locking determination based on the remaining effective channels and hysteresis characteristics. If the local calibration update module obtains new manually verified samples or post-disaster confirmed samples, it updates the baseline event interval vector and output calibration parameters, and performs local retraining on the machine learning early warning model to improve the system's adaptability to the current engineering geological unit.

[0245] In a specific deployment example, the data acquisition module is located at the edge acquisition terminal on the slope, while the precursor event extraction module, phase sequence construction module, and hysteresis feature extraction module are located on the on-site industrial control computer. The machine learning early warning module, phase sequence lock determination module, credibility masking module, local calibration update module, and early warning output module are located on the monitoring center server. The edge acquisition terminal sends updated sliding time window data to the on-site industrial control computer every 10 minutes. After completing event extraction, phase sequence construction, and hysteresis feature extraction, the on-site industrial control computer sends the results to the monitoring center server. The monitoring center server calculates the irreversible instability score, predicts the earliest irreversible moment, determines the phase sequence lock, and outputs the early warning level, simultaneously writing the results to the multivariate synchronization record database and the early warning output record database. When the monitoring center server determines that the alert level has been reached, it can simultaneously trigger a platform pop-up window, SMS notification, and on-site audible and visual alarm output.

[0246] Each module in this embodiment can be implemented independently using hardware circuits, by a processor executing program instructions from memory, or by a combination of hardware and software. Data interaction between modules can be achieved through serial ports, Ethernet, industrial buses, wireless communication links, or shared memory areas. Any system capable of performing the aforementioned functions of multi-source monitoring data acquisition, precursor event extraction, precursor event phase sequence matrix construction, event interval vector formation, hysteresis feature extraction, machine learning early warning determination, phase sequence locking determination, reliability masking, local calibration updates, and early warning output can be considered a specific implementation of the intelligent early warning system for engineering geological disasters disclosed in this embodiment.

[0247] It should be noted that the embodiments of the present invention have better implementability and are not intended to limit the present invention in any way. Any person skilled in the art may use the above-disclosed technical content to change or modify it into equivalent effective embodiments. However, any modifications or equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A method for intelligent early warning of engineering geological disasters, characterized in that, include: S1. Obtain multi-source monitoring data and static engineering geological data of the same engineering geological unit within the target monitoring area. The multi-source monitoring data includes at least one trigger factor channel, one hydrological response channel, and two deformation response channels at different depths. S2. Extract precursor events for each channel according to the sliding time window to obtain the start event, acceleration event and continuous event of each channel, and construct the precursor event phase sequence matrix according to the sequential relationship between the precursor events in different channels, and construct the event interval vector according to the time interval between adjacent precursor events. S3. Construct a trigger-response hysteresis curve based on the change process of the trigger factor channel and each of the response channels within the same sliding time window, and extract the hysteresis closure difference vector and the asymmetric vector of the rising and falling segments; S4. Input the precursor event phase sequence matrix, the event interval vector, the hysteresis closure difference vector, the rise and fall segment asymmetry vector, and the static engineering geological data into the pre-trained machine learning early warning model to obtain the irreversible instability score and the earliest irreversible time of the corresponding engineering geological unit. S5. When the irreversible instability score is not lower than the preset score threshold, the time interval from the current time to the earliest irreversible time is not greater than the preset time threshold, and the precursor event phase sequence matrix satisfies the phase sequence locking condition within multiple consecutive sliding time windows, the warning level of the corresponding engineering geological unit is determined and a warning instruction is generated.

2. The intelligent early warning method for engineering geological disasters according to claim 1, characterized in that, The extraction of precursor events includes: The original monitoring sequences of each channel are denoised and have a stable baseline corrected. Within each of the aforementioned sliding time windows, calculate the local change slope, local curvature, sign holding length, and cumulative deviation. When the local change slope exceeds a first preset threshold and the symbol retention length reaches a first preset length, the starting event is extracted; When the increment of the local change slope exceeds the second preset threshold, the local curvature exceeds the third preset threshold, and the cumulative deviation increases continuously, the acceleration event is extracted; When the cumulative deviation exceeds the fourth preset threshold and fails to return to the stable baseline within a preset duration, the continuous event is extracted. It retains only the precursor events that are detected in both adjacent sliding time windows and whose event start time deviation is no greater than the preset tolerance.

3. The intelligent early warning method for engineering geological disasters according to claim 2, characterized in that, The two deformation response channels at different depths include a shallow deformation response channel and a deep deformation response channel; The matrix elements of the precursor event phase sequence matrix are used to characterize the relative timing state of similar precursor events in different channels. The relative timing state includes leading, synchronous, lagging, and missing. The precursor event phase sequence matrix records at least the following relative temporal states: The relative temporal state of the starting event of the triggering factor channel relative to the starting event of the hydrological response channel; The relative temporal state of the acceleration events of the hydrological response channel with respect to the acceleration events of the shallow deformation response channel and the deep deformation response channel; The relative temporal state of the continuous events of the shallow deformation response channel relative to the continuous events of the deep deformation response channel; The state change between the shallow deformation response channel and the deep deformation response channel, from shallow leading to deep leading or from deep leading to shallow leading, is denoted as the shallow-deep inversion state.

4. The intelligent early warning method for engineering geological disasters according to claim 3, characterized in that, The event interval vector includes a normalized event interval subvector and a phase sequence compression ratio subvector; The normalized event interval sub-vector is obtained by extracting a reference event interval vector from the historical stable phase of the same engineering geological unit for the same engineering geological unit, and comparing the current event interval vector with the reference event interval vector item by item. The phase sequence compression ratio sub-vector is obtained by calculating the ratio of each item in the current event interval vector to the corresponding item in the reference event interval vector. The normalized event interval subvector and the phase sequence compression ratio subvector are concatenated to form the event interval vector, which is then input into the machine learning early warning model.

5. The intelligent early warning method for engineering geological disasters according to claim 1, characterized in that, The two deformation response channels at different depths are the shallow deformation response channel and the deep deformation response channel, respectively. The trigger-response hysteresis curve includes the hysteresis curve formed by the trigger factor channel and the hydrological response channel, the shallow deformation response channel and the deep deformation response channel, respectively; The hysteresis closure difference vector includes at least the residual opening of each hysteresis curve at the end of a sliding time window; The asymmetric vector of the ascending and descending segments includes at least the area difference and slope difference of each of the hysteresis curves in the ascending and descending segments.

6. The intelligent early warning method for engineering geological disasters according to claim 5, characterized in that, The machine learning early warning model includes: An event embedding subnetwork used to encode the precursor event phase sequence matrix; A phase-order attention subnetwork used for weighted modeling of the event interval vector; Hysteresis geometric coding subnetwork for geometrically coding the hysteresis closure difference vector and the ascending / descending segment asymmetric vector; A branch consistency constraint subnetwork used to receive the output results of the event embedding subnetwork and the output results of the hysteresis geometric coding subnetwork, and to determine whether the phase sequence evolution result is consistent with the hysteresis irreversibility result based on the phase sequence compression ratio subvector in the event interval vector; An irreversible fusion subnetwork used to fuse the static engineering geological data and the output of the branch consistency constraint subnetwork; And a dual-output subnetwork for outputting the irreversible instability score and the earliest irreversible moment, respectively.

7. The intelligent early warning method for engineering geological disasters according to claim 6, characterized in that, The machine learning early warning model was trained in the following way: Construct real instability samples using multiple sliding time windows preceding the earliest irreversible moment in historical disaster samples; Stable samples are constructed using multiple sliding time windows in which the aforementioned persistent events have not occurred during historical stable phases. For each real unstable sample, a matching stable sample is selected from historical stable phases of the same engineering geological unit that are at the same trigger factor level and the same engineering disturbance phase to form a paired sample group within the same unit. The precursor event phase sequence matrix, the event interval vector, the hysteresis closure difference vector, and the ascent-depression segment asymmetry vector corresponding to the paired sample group in the same unit are used as training inputs.

8. The intelligent early warning method for engineering geological disasters according to claim 7, characterized in that, The training of the machine learning early warning model also includes the generation of pseudo-instability samples, which includes: The precursor event phase sequence matrix of the real instability sample remains unchanged; The event interval in the real instability sample is shortened according to a preset compression factor; Increase the hysteresis closure difference vector and the rise / fall segment asymmetry vector in the real instability sample by a preset amplification factor; And generate pseudo-instability samples based on the processed samples.

9. The intelligent early warning method for engineering geological disasters according to claim 8, characterized in that, The training of the machine learning early warning model also includes the generation of hard-to-bear samples, which includes at least two of the following: Inject transient jump fragments into a single channel of the stable sample; The intermediate precursor events in the stable samples are deleted to form a missing event sample; The sequential relationship of similar events in each channel within the same sliding time window in the stable sample is disrupted to form phase-order conflict samples.

10. An intelligent early warning system for engineering geological disasters, characterized in that, include: The data acquisition module is used to acquire multi-source monitoring data and static engineering geological data of the same engineering geological unit within the target monitoring area; The precursor event extraction module is used to extract the start events, acceleration events, and continuous events of each channel by sliding time windows; The phase sequence construction module is used to construct a precursor event phase sequence matrix based on the sequential relationship between each precursor event in different channels, and to construct an event interval vector based on the time interval between adjacent precursor events. The hysteresis feature extraction module is used to construct the trigger-response hysteresis curve and extract the hysteresis closure difference vector and the asymmetric vector of the rising and falling segments; The machine learning early warning module is used to input the precursor event phase sequence matrix, the event interval vector or the normalized event interval vector obtained by comparing the event interval vector with the benchmark event interval vector extracted from the historical stable stage of the same engineering geological unit, the hysteresis closure difference vector, the rise and fall segment asymmetry vector, and the static engineering geological data into a pre-trained machine learning early warning model to obtain an irreversible instability score and the earliest irreversible moment. The machine learning early warning model includes an irreversibility fusion subnetwork for fusing the hysteresis closure difference vector, the rise and fall segment asymmetry vector, and the static engineering geological data, and a dual-output subnetwork for outputting the irreversible instability score and the earliest irreversible moment. The phase sequence locking determination module is used to determine whether the precursor event phase sequence matrix satisfies the phase sequence locking condition; the credibility masking module is used to mask the precursor event extraction results of low credibility channels according to the channel credibility of each channel. The local calibration update module is used to update the baseline event interval vector and the output calibration parameters of the machine learning early warning model, and to perform local retraining on the irreversible fusion subnetwork and the dual-output subnetwork. The early warning output module is used to determine the early warning level and generate an early warning command based on the irreversible instability score, the earliest irreversible moment, and the phase sequence locking condition.

Images