A slope acoustic emission dynamic baseline multi-scale self-adaptive early warning method and system

Abstract

The application discloses a kind of slope acoustic emission dynamic baseline multiscale self-adaptive early warning method and system, belong to slope safety monitoring and geological disaster early warning field.This method collects stable period no-label acoustic emission signal, uses multi-task self-supervised encoder to establish multiple prototype dynamic baseline and calculates abnormal deviation degree;Extract short-time transient precursor and long-time progressive precursor, after fusion by causal time sequence coding, combined with phase transition constraint, identify five stages from stable to impending slide;Dynamic baseline is only controlled update under the control of multiple conditions;According to the integrated risk value, dynamically adjust the sampling frequency, trigger threshold and waveform preservation time, output the damage stage, warning level and control instruction.It is suitable for long-term online monitoring and instability early warning of soil, rock and rock-soil mixed slope.

Description

Technical Field

[0001] This invention relates to the fields of slope engineering safety monitoring, acoustic emission signal processing, and intelligent early warning of geological disasters, specifically to a multi-scale adaptive early warning method and system for dynamic baseline acoustic emission of slopes. This method collects time-series acoustic emission signals generated by particle slippage, crack initiation and propagation, and structural surface displacement within the slope. It establishes a baseline of normal acoustic state using unlabeled data from the stable monitoring phase and identifies slope failure stages by combining short-term anomaly responses and long-term damage accumulation characteristics. Furthermore, it adaptively adjusts the early warning threshold, signal sampling frequency, original waveform retention time, and data transmission cycle. This method is suitable for long-term online monitoring and instability early warning of soil slopes, rock slopes, and mixed soil-rock slopes. Background Technology

[0002] Slopes are widely used in highways, railways, mines, water conservancy and hydropower projects, and urban construction projects. Their stability is influenced by a variety of factors, including the structural characteristics of the soil and rock mass, the degree of joint and fissure development, rainfall infiltration, changes in groundwater level, excavation unloading, vehicle vibration, and long-term weathering. Slope instability usually does not occur instantaneously, but rather goes through stages such as local particle slippage, micro-crack initiation, crack propagation, structural surface displacement, continuous damage accumulation, and accelerated macroscopic deformation. If the precursory information appearing during the transition from a stable state to a critical instability state is not identified in time, it is easy to miss the opportunity for early warning and emergency response.

[0003] Current slope safety monitoring mainly employs technologies such as surface displacement, deep displacement, tilt, stress, pore water pressure, groundwater level, global navigation satellite systems, and ground-based radar. While these methods can reflect macroscopic slope deformation and hydrological changes, some monitoring indicators typically only show a significant response after damage within the soil and rock mass has accumulated to a certain extent. They are insufficient to fully characterize the entire process of slope evolution from a stable to a critical state, particularly regarding early-stage failure activities such as crack initiation, localized particle slippage, and minor structural surface displacement.

[0004] When soil and rock masses experience particle friction, crack initiation, crack propagation, and structural surface slippage, they release transient elastic waves. Acoustic emission monitoring technology collects these elastic wave signals through sensors and analyzes their amplitude, energy, duration, rise time, ring count, frequency, event rate, and event interval. This technology can be used to characterize subtle internal damage activities in slopes, as changes can occur before a significant increase in macroscopic slope displacement. Therefore, it has the potential to conduct early warning identification and continuous online early warning.

[0005] Existing slope acoustic emission early warning methods mostly rely on parameters such as event rate, ring count rate, cumulative energy, or amplitude to set fixed thresholds. When the monitored value exceeds the preset threshold, an early warning of the corresponding level is issued. This type of method is simple in structure and easy to implement. However, different slopes have significant differences in terms of soil and rock types, initial stress, water content, sensor coupling conditions, and environmental noise levels. The same acoustic emission parameters may correspond to different steady states in different projects, which can easily lead to problems such as insufficient threshold shift capability, poor field adaptability, and the coexistence of false alarms and missed alarms.

[0006] Furthermore, slope conditions are typically affected by factors such as rainfall impact, vehicle traffic, mechanical construction, blasting vibrations, electromagnetic interference, and changes in sensor coupling. These external disturbances may cause a sudden increase in acoustic emission event rate or signal energy in a short period of time, but this does not necessarily correspond to the continuous development of internal slope damage. If judgment is made solely based on abnormal peak values ​​or instantaneous thresholds within a single time window, short-term disturbances can easily be misidentified as precursors to instability. Conversely, some gradual damage processes may manifest as a slow increase in event activity, a gradual shortening of event intervals, continuous spectral migration, and long-term accumulation of abnormal states. Monitoring methods with a single short timescale are insufficient to capture such gradual changes in a timely manner.

[0007] Existing intelligent early warning methods typically employ supervised classification models to identify acoustic emission signals or artificial features, requiring a large number of labeled stable, damaged, and unstable samples during training. However, the probability of actual overall slope instability events is low, and the number of field samples that fully record the entire process from stability to instability is limited. Labels for different failure stages often rely on manual judgment or other monitoring methods for post-event determination, resulting in problems such as sample imbalance, ambiguous stage boundaries, and high labeling costs. Models trained based on a small number of instability samples are prone to overfitting and are difficult to directly transfer to different slopes.

[0008] During long-term monitoring, the normal acoustic state of the slope is not always constant. Slow changes in seasonal temperature, moisture content, background noise, sensor coupling status, and construction environment can all cause drift in normal acoustic emission characteristics. If the normal baseline is fixed for a long time, the system may continuously identify normal environmental changes as abnormal; if the normal baseline is updated without restriction based on newly acquired data, damage initiation or precursory landslides may be incorrectly written into the normal state model, leading to a decrease in anomaly sensitivity.

[0009] On the other hand, existing acoustic emission monitoring systems typically employ fixed sampling frequencies, fixed trigger thresholds, and fixed data upload cycles. Continuously storing and transmitting high-sampling-rate raw waveforms during the long-term stable phase of a slope results in significant storage, communication, and energy consumption; while using lower sampling rates or longer upload cycles during the rapid damage development phase may miss crucial precursor waveforms. Furthermore, existing early warning algorithms and the operating parameters of the acquisition equipment are usually independent of each other, making it difficult to adjust the sampling frequency, waveform storage duration, and data transmission cycle in real time according to changes in risk. Summary of the Invention

[0010] To address the problems that existing slope acoustic emission early warning methods rely on fixed empirical thresholds and a large number of unstable label samples, making it difficult to adapt to different soil and rock types, sensor coupling states, and environmental background changes, and that a single time window index cannot simultaneously identify short-term abnormal outbreaks and long-term damage accumulation, this invention provides a slope acoustic emission dynamic baseline multi-scale adaptive early warning method and system.

[0011] This invention utilizes unlabeled acoustic emission data generated during the target slope stability monitoring phase for multi-view self-supervised representation learning to establish a dynamic baseline of normal state reflecting the individual acoustic characteristics of the target slope. Through controlled update gating, abnormal sample isolation, and backtracking verification mechanisms, the normal baseline is allowed to be updated to a limited extent with seasonality, water content, and slow equipment drift, while preventing damage precursors from being mistakenly absorbed as normal states. Furthermore, by combining short-timescale sudden precursors, long-timescale cumulative precursors, failure stage constraints, and uncertainty estimation, a closed loop of risk assessment, graded early warning, and acquisition parameter control is formed, thereby reducing the dependence on instability labels and fixed thresholds and improving the stability of early damage identification, continuous anomaly discrimination, and critical stage early warning in complex environments.

[0012] To achieve the above objectives, the present invention adopts the following technical solution.

[0013] A multi-scale adaptive early warning method for dynamic baseline acoustic emission of slopes includes the following steps: S1, multiple acoustic emission monitoring nodes are set up on the slope to be monitored, and multi-channel acoustic emission waveforms are continuously collected according to a unified time reference, while environmental parameters, equipment status, and channel health information are acquired simultaneously; the first... The continuous waveforms of each monitoring channel are arranged according to the window length. and sliding step size Divided into signal segments Robust normalization, event detection, time-frequency transformation, and basic feature extraction are performed on each signal segment. The channel health is determined based on the channel signal-to-noise ratio, the correlation with the reference waveform compared to the standard waveform, gain deviation, clock jitter, and the proportion of missing data. ; S2, select stable candidate windows from low-risk monitoring intervals that are either in the early stages of deployment or verified on-site, and implement at least two of the following enhancements for each stable candidate window: time masking, frequency band masking, amplitude scaling, noise injection, channel deactivation, and sub-window pruning, to construct multi-view positive sample pairs and negative sample candidates that maintain semantic consistency in the destruction. S3. Construct a self-supervised encoder comprising a temporal coding branch, a frequency coding branch, and a cross-branch fusion layer. Jointly train the self-supervised encoder using contrastive consistency loss, mask reconstruction loss, future segment prediction loss, and temporal order discrimination loss to obtain the latent representation of the current monitoring window. And the reconstruction residual of the current window; S4. Establish a dynamic baseline in the latent representation space containing multiple normal state prototypes. Calculate the anomaly deviation based on the distance between the current window representation and each normal state prototype, the reconstruction residual obtained in step S3, and the prototype attribution uncertainty. The abnormal deviation is standardized based on the statistical distribution of the controlled normal samples. S5. Extract corresponding acoustic emission features at at least two short timescales, at least one medium timescale, and at least one long timescale, wherein the short timescales are on the order of seconds or minutes, the medium timescales are on the order of hours, and the long timescales are on the order of days or longer; wherein the short timescale features include at least event rate spikes, shortened event intervals, high-energy event proportions, spectral abrupt changes, and bursts of pulses; the medium timescale features include at least event rate trends and frequency band energy evolution; and the long timescale features include at least cumulative energy growth, anomaly persistence proportions, frequency band shifts, trend slopes, and trend accelerations; calculate a short-term precursor score based on the extracted short-term, medium-term, and long-term timescale features. Long-term cumulative damage and persistent outlier statistics The system utilizes Bayesian online change point detection and a mechanism-competitive posterior among the normal fluctuation hypothesis, environmental disturbance hypothesis, equipment anomaly hypothesis, and true damage hypothesis to identify the current anomaly source. The mechanism-competitive posterior includes the recursive posterior probabilities of each of the normal fluctuation hypothesis, the environmental disturbance hypothesis, the equipment anomaly hypothesis, and the true damage hypothesis. S6, employing a causal temporal encoder that uses only the current window and historical windows, the short-timescale features, the medium-timescale features, and the long-timescale features are fused. A stage recognition network, combined with stage transition constraints and stage duration constraints, outputs the stage probabilities for stable stages, locally activated stages, persistent damage stages, critical acceleration stages, and near-slip stages. These stage probabilities include the stable stage probability. Further, the stage prediction set and overall uncertainty are determined through evidence output, repeated reasoning, and distribution-independent confidence calibration. ; S7, based on the abnormal deviation degree The probability of the stable phase Abnormal persistence The aforementioned comprehensive uncertainty and channel health Generate a baseline update gate, wherein the anomaly duration For the most recent The percentage of abnormal windows in each window. The duration of the anomaly is a positive integer. The persistent anomaly statistic is used to determine whether the current window is eligible to update the normal baseline. Used to determine whether an anomaly has a continuous accumulation characteristic; the baseline update gating Represented as: ; in, This is an indicator function that takes the value 1 when the condition is met and 0 when the condition is not met. , , , and These are the abnormal deviation threshold, the stable phase probability threshold, the abnormal persistence threshold, the comprehensive uncertainty threshold, and the channel health threshold; only windows that meet the update conditions are written to the normal cache and the matched normal state prototype is updated; windows that do not meet the update conditions are written to the observation cache or the isolation cache respectively, and the normal baseline version, the updated sample index, and the rollback checkpoint are saved. S8, the abnormal deviation The aforementioned short-term precursor score The cumulative amount of long-term damage The aforementioned persistent abnormal statistics The stage probability, the recursive posterior probability of each mechanism hypothesis, and the overall uncertainty. Integrate into a comprehensive risk value The recursive posterior probabilities of each mechanism hypothesis include the recursive posterior probabilities of the normal fluctuation hypothesis, the environmental disturbance hypothesis, the equipment anomaly hypothesis, and the actual damage hypothesis. The fusion process uses a weighted summation method to convert the recursive posterior probabilities of each mechanism hypothesis into a mechanism risk contribution value. The mechanism risk contribution value is then used in the calculation of the overall risk value. The value is a scalar value obtained by weighted summation of the recursive posterior probabilities of the hypotheses of each mechanism; the graded early warning threshold is dynamically determined based on the risk distribution in the controlled normal baseline and the environmental context, and a continuous confirmation and hysteresis fallback mechanism is used to output blue, yellow, orange, or red warnings; the blue warning indicates that the slope is in a stable state, the yellow warning indicates that the slope is in a state requiring attention, the orange warning indicates that the slope is in a state requiring vigilance, and the red warning indicates that the slope is in a dangerous state; the continuous confirmation is the latest requirement for warning escalation. At least one window Each window satisfies the corresponding risk level's amplitude, duration, stage set, mechanism posterior, and data quality conditions, among which... and are positive integers and The hysteresis fallback is defined as a warning downgrade requirement whereby the comprehensive risk value continuously falls below the current level threshold minus the hysteresis bandwidth for a set period of time; simultaneously, based on the comprehensive risk value... The expected information gain, as well as energy constraints, communication constraints, and storage constraints, are used to adjust the sampling frequency, trigger threshold, original waveform retention time, number of participating channels, and data upload cycle. The warning results are then linked to the baseline version, model version, and parameter control records for output.

[0014] Furthermore, in step S1, the robust normalization uses the median and median absolute deviation of the current window to eliminate the effects of DC drift, gain difference, and occasional spikes; the basic characteristics include at least five of the following: event rate, peak amplitude, energy, duration, rise time, ring count, center frequency, band energy ratio, event interval, and cross-channel synchronization ratio; monitoring channels with channel health below the set value are downweighted, isolated, or generate maintenance prompts, and are not allowed to trigger orange or red alerts independently.

[0015] Furthermore, in step S3, the joint loss of the self-supervised encoder Represented as: ; in, For the aforementioned consistency loss, The mask reconstruction loss is... Predict the loss for the future segment. To determine the loss in the time sequence, , , and The weights are non-negative and satisfy the following conditions: The contrast consistency loss is used to maintain the consistency of the representation of the same stable window under noise, amplitude and channel perturbation. The mask reconstruction loss, the future segment prediction loss and the time order discrimination loss are used to retain information on local mutations, frequency band migration and event organization changes.

[0016] Further, in step S4, the dynamic baseline includes The first normal state prototype, the... A normal state prototype includes at least a mixed weight. mean Covariance ; Current window and the first Square Mahalanobis distance between normal state prototypes Represented as: ; The abnormal deviation Based on the squared Mahalanobis distance corresponding to each normal state prototype, the reconstructed residual and the probability of belonging to the maximum prototype calculate: ; in, This is the minimum squared Mahalanobis distance of the current window relative to all normal state prototypes. , and The weights are non-negative and satisfy the following conditions: When the number of normal samples is insufficient, the covariance is ill-conditioned, or the prototype has not been accessed for a long time, at least one of the following methods can be used to maintain dynamic baseline stability: covariance shrinkage, minimum sample limit, prototype merging, or prototype splitting.

[0017] Furthermore, in step S5, the burstiness of the current window acoustic emission event interval sequence... Represented as: ; in, and These are the mean and standard deviation of the event intervals, respectively. Positive number; cumulative damage over long period of time and persistent outlier statistics They are represented as follows: Initial time Set to 0; Initial time Set to 0; in, The memory coefficient varies with time intervals. For long-term feature mapping function, This is a long-term scale feature vector. The reference drift amount is determined based on the normal dynamic baseline; when the short-term precursor score is... Compared to the increase in the adjacent historical window, the cumulative amount of long-term damage... and the aforementioned persistent anomaly statistics When there is no sustained increase, the current anomaly is prioritized as a transient disturbance or a recoverable local activity; when the accumulated long-term damage... and the aforementioned persistent anomaly statistics When there is continuous growth accompanied by enhanced consistency across multiple channels, the current anomaly is considered a candidate for persistent damage.

[0018] Furthermore, in step S7, only when the baseline update gate is... When the value is equal to 1, the dynamic baseline can be updated at the following rate: Represented as: ; The mean of the matched normal state prototype is updated according to the following formula: ; in, For maximum update rate, This is the abnormal suppression coefficient. For the current window to the first The probability of a normal state prototype is assigned; the covariance and mixed weights are updated recursively according to the same gating principle, and the maximum drift per unit time, the upper limit of prototype center movement, the observation cache, the isolation cache, and the baseline rollback checkpoint are set.

[0019] Furthermore, in step S6, the fusion weights of the short-timescale features, the medium-timescale features, and the long-timescale features are... and fusion characterization They are represented as follows: ; ; in, For the first Attention score at each time scale within the current window, , and The results are causal coding at short, medium, and long time scales, respectively; the stage decoding of the stage recognition network uses an allowed transition matrix. and phase duration constraints, where , =0,1,2,3,4, corresponding to the stable phase, the local activation phase, the continuous damage phase, the critical acceleration phase, and the near-slip phase, respectively. Indicates by the first The first phase transitions to the second. The permissible costs for each stage prioritize maintaining the current stage or transitioning to an adjacent high-risk stage, and allow skipping stages to the critical acceleration stage when short-term bursts, long-term accumulations, and multi-channel consistency are simultaneously satisfied; the evidence parameters and stage probabilities output by the stage identification network are expressed as follows: ; in, For the first The original output of each stage, As non-negative evidence, For the evidence distribution parameters, The total parameters for all stages, The stage probability is determined by using verified samples to determine the distribution-independent calibration quantile. When the stage prediction set contains multiple adjacent stages, the stage with higher risk is used for control. When the stage prediction set is empty or contains multiple stages with a large span, the sampling intensity is increased and manual review is initiated.

[0020] Furthermore, in step S8, the comprehensive risk value Represented as: ; in, It is a monotonic mapping function. For bias terms, , , and These are, respectively, the abnormal deviation degree, the short-term precursor score, the long-term cumulative damage, and the persistent abnormality statistic. For the stage The mean probability, The risk factors are for continuous damage, critical acceleration, and the skid-end phase. , , , , and For the corresponding weights; the first Level dynamic threshold is expressed as: ; in, For the corresponding quantiles, This represents the combined risk value for each historical window within the controlled normal baseline. Update the gating status for the corresponding window. This indicates that the window belongs to a controlled normal baseline. For the first Environmental correction parameters for the threshold level, For the environment context vector, For the first The significance level corresponding to the level warning threshold. Choose 1, 2, and 3, which correspond to the yellow, orange, and red warning thresholds, respectively; the warning escalation requirement is the most recent At least one window Each window meets the corresponding risk level's risk magnitude, duration, stage set, mechanism posterior, and data quality conditions, and the warning downgrade requires the comprehensive risk value mentioned above. The value is continuously below the current level threshold minus the hysteresis bandwidth and continues to be below the set time.

[0021] Furthermore, in step S8, the sampling frequency, pre-gain, trigger threshold, pre-trigger storage duration, post-trigger storage duration, number of participating channels, feature calculation cycle, and data upload cycle are combined to form candidate monitoring actions. The utility of candidate monitoring actions is calculated according to the following formula. : ; in, The expected information gain of candidate monitoring actions for the next window stage judgment or damage characterization. , and These are energy consumption, communication volume, and storage volume, respectively. , and The system selects candidate monitoring actions that meet the constraints of minimum sampling frequency, minimum number of healthy channels, longest upload period, and minimum original waveform retention ratio, and have the greatest utility. When an orange or red alert is reached, key channels are not allowed to be closed or the sampling frequency is not allowed to be reduced below a safe value. Each normal baseline update saves the prototype parameters, sample index, environmental context range, and verification status. When the baseline drift exceeds the limit, the abnormal score of historical stable verification samples increases, the false alarm rate continues to increase after the update, or the conclusions of the old and new baselines are reversed, the online update is frozen and rolled back to the most recently verified baseline version.

[0022] This invention also provides a multi-scale adaptive early warning system for dynamic baselines of slope acoustic emission, used to execute the aforementioned multi-scale adaptive early warning method for dynamic baselines of slope acoustic emission. The self-supervised dynamic baseline established in the method, the calculated multi-timescale precursor scores, and the stage identification results constitute the same specific technical features of the system and the method. The system includes: The acoustic emission acquisition and synchronization module is used to continuously acquire acoustic emission waveforms from multiple monitoring channels and simultaneously obtain environmental parameters, equipment status, and channel health information. The robust preprocessing module is used for sliding window partitioning, robust normalization, event detection, time-frequency transformation, and basic feature extraction. The stable sample and multi-view building module is used to filter stable candidate windows and generate self-supervised positive sample pairs and negative sample candidates; The self-supervised representation learning module is used to obtain the monitoring window representation through time-domain coding branch, frequency-domain coding branch and joint self-supervised task; The dynamic baseline and controlled update module is used to establish a multi-prototype normal state model, calculate the abnormal deviation, and manage the normal cache, observation cache, and isolation cache according to the update gating. A multi-timescale precursor module is used to extract short-term sudden precursors, medium-timescale evolution trends, and long-term damage accumulation features; The variable point and mechanism competition module is used to identify normal fluctuations, environmental disturbances, equipment anomalies and real damage using Bayesian online variable point detection; The causal fusion and stage identification module is used to fuse features from multiple time scales through causal temporal coding and output the probability of the destruction stage, the stage prediction set, and the comprehensive uncertainty. The sequential risk and adaptive early warning module is used to generate comprehensive risk values, dynamic thresholds, and blue, yellow, orange, and red early warnings. The active monitoring and control module is used to adjust the sampling frequency, trigger threshold, original waveform retention time, number of participating channels, and data upload cycle. It also includes a version governance and communication module, which is used to perform baseline version saving, rollback, and verification record management, and to send early warning information and monitoring and control instructions to the field server, remote monitoring platform or mobile terminal.

[0023] Compared with the prior art, the present invention has at least the following beneficial effects: (1) This invention uses a large amount of unlabeled acoustic emission data of the target slope during the stable operation phase as the main learning object. It forms a normal state characterization by comparing consistency, mask reconstruction, future prediction and time sequence discrimination, thereby reducing the dependence on complete instability process and artificial stage labels.

[0024] (2) The present invention adopts a multi-prototype normal state dynamic baseline, which can express the multi-modal normal acoustic state caused by seasonal, water content, temperature and sensor coupling changes, and has stronger individual slope adaptability compared with a unified empirical threshold.

[0025] (3) The present invention sets a baseline controlled update gating, observation cache, isolation cache and rollback checkpoint, and only writes samples with low anomaly, high stability probability, low uncertainty and healthy channel into the normal baseline, which can avoid continuous damage and the precursor of slippage being mistakenly absorbed as normal state.

[0026] (4) The present invention combines the sudden precursors at the second or minute level with the damage accumulation features at the hour or day level, and can simultaneously identify rapid anomalies such as sudden increase in event rate and clustering of high-energy events, as well as gradual anomalies such as accelerated cumulative energy, continuous frequency band migration and long-term rise of anomalies.

[0027] (5) This invention utilizes the orthogonal and antagonistic decoupling of damage characterization and environmental disturbance characterization, combined with rainfall, vehicle, construction and equipment status information, to reduce false alarms caused by short-term environmental impact and periodic mechanical noise.

[0028] (6) The present invention adopts causal temporal coding and duration-aware semi-Markov stage constraints, which can reduce stage jumps caused by a single abnormal window and output interpretable stage results according to the evolution process of stability, local activation, continuous damage, critical acceleration and slippage.

[0029] (7) The present invention outputs risk value, stage posterior probability and comprehensive uncertainty simultaneously, and distinguishes between states with insufficient data quality, large model divergence or low channel health, so as to avoid low confidence anomalies directly triggering high-level warnings.

[0030] (8) The present invention dynamically determines the yellow, orange and red thresholds based on the controlled normal baseline risk distribution and environmental context, and adopts the m-of-n continuous confirmation and hysteresis fallback mechanism, which can take into account both early warning sensitivity and level stability.

[0031] (9) The present invention feeds back the risk identification results to the acquisition end. As the risk increases, the sampling frequency is increased, the original waveform storage time is extended and the upload cycle is shortened. As the risk decreases, the amount of data is reduced. It can take into account both the long-term monitoring resource consumption and the information integrity of the critical stage.

[0032] (10) This invention retains the input window, baseline version, model version, threshold, environmental status and control instructions corresponding to each warning, and performs closed-loop verification through on-site inspection or subsequent monitoring, which facilitates the review of warning results, model auditing and cross-stage traceability.

[0033] (11) This invention can be implemented without relying on potential slip surface spatial reconstruction or high-precision sound source localization. It is suitable for soil slopes, rock slopes and soil-rock mixed slopes with limited number of sensors, insufficient localization conditions or greater concern for long-term state evolution.

[0034] (12) This invention uses Bayesian online variable point detection to compete with multiple mechanisms such as normal fluctuations, environmental disturbances, equipment abnormalities and real damage for posterior arithmetic. When the abnormal amplitudes are similar, the main source can be identified based on persistence, environmental synchronicity and channel health differences.

[0035] (13) The present invention uses evidence deep learning, repeated reasoning uncertainty and distribution-independent confidence calibration to jointly constrain the stage output, which can output a conservative stage set when the environment is unknown or the data quality is degraded, rather than giving an overly certain single result.

[0036] (14) The present invention adopts the expected information gain and energy, communication and storage costs to jointly optimize the monitoring action, and sets hard security constraints related to the warning level, so as to prioritize the collection of the most valuable data for risk judgment under limited resource conditions.

[0037] (15) The present invention sets up multiple versions of baselines, candidate sub-baselines, drift monitoring, verification cost and automatic rollback mechanism, which can reduce the risk of irreversible drift of normal models during long-term online learning. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 This is a flowchart of the multi-scale adaptive early warning method for dynamic baseline of slope acoustic emission according to the present invention.

[0039] Figure 2 This is a performance comparison diagram of different early warning methods for open-pit rock slopes in Embodiment 1 of the present invention.

[0040] Figure 3 This is a performance comparison diagram of different early warning methods for reservoir bank slopes in Embodiment 2 of the present invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0042] like Figure 1 As shown, a multi-scale adaptive early warning method for dynamic baseline acoustic emission of slopes includes the following steps.

[0043] S1 involves deploying multiple acoustic emission monitoring nodes at the top, middle, and toe of the slope to be monitored, or near key structural surfaces. Each monitoring node includes an acoustic emission sensor, a preamplifier unit, an analog-to-digital converter unit, a synchronization clock unit, an environmental parameter interface, and an edge computing unit. Each node continuously acquires acoustic emission waveforms according to a unified time reference and simultaneously records one or more auxiliary information, including rainfall intensity, temperature, humidity, equipment power supply status, channel gain, and sensor coupling status.

[0044] The first The continuous waveforms of each monitoring channel are arranged according to the window length. and sliding step size Divided into signal segments To reduce the impact of amplitude drift, occasional spikes, and differences in gain between different channels, robust normalization is performed on each window using the median and median absolute deviation. ; in, For the first The first channel in the The first window The original amplitude of each sampling point; For median operators; This represents the absolute deviation of the median. To prevent positive numbers with a denominator of zero, after normalization, the envelope, spectrum, time-frequency plot, and pulse sequence are extracted from the waveform level, and the event rate, amplitude, energy, duration, rise time, ring count, center frequency, band energy ratio, event interval, and cross-channel synchronization ratio are extracted from the event level, forming a multi-channel observation set for the current window. This normalization method can effectively eliminate the effects of DC drift, gain differences, and occasional spikes, making the acoustic emission characteristics of the same slope comparable under different sensor gain conditions.

[0045] At the same time, a health status is established for each monitoring channel. Health status is determined by the signal-to-noise ratio, the correlation with the reference waveform compared to the standard waveform, gain deviation, clock jitter, and the proportion of missing data. It is used for subsequent feature fusion, anomaly scoring, and warning confidence correction. Channels with health status below the set value are downweighted, isolated, or have maintenance prompts generated, but high-level warnings are not directly triggered to avoid false alarms caused by a single faulty channel.

[0046] S2, a stable baseline learning period is set during the initial system deployment. Stable candidate windows are selected based on a combination of factors, including no sustained increase in acoustic emission activity, meeting channel health requirements, no significant deformation detected during manual inspections, and optional parameters such as displacement rate, pore water pressure change rate, or crack width below corresponding control values. For slopes lacking auxiliary monitoring data, an initial stable candidate set is determined using continuous low-anomaly intervals, inter-window characteristic stability, and manual confirmation results. This multi-condition selection ensures that the samples used to establish a normal baseline accurately reflect the normal acoustic state of the slope.

[0047] For each stable candidate window, at least two enhanced views are generated that maintain semantic consistency in preventing disruption. The enhancement methods include two or more of the following: local temporal masking, band masking, small-amplitude noise injection, amplitude scaling, time stretching, random channel deactivation, sub-window pruning, and adjacent window splicing. The enhancement magnitude is constrained by the signal peak retention rate, the dominant frequency offset, and the rate of change of the number of events, to avoid erasing genuine sudden events or artificially generating instability precursors.

[0048] Different enhanced views formed from the same original window are defined as positive sample pairs, while windows from different time intervals, different channel combinations, or different operating states are defined as negative sample candidates. For windows that are temporally adjacent and whose features change slowly, soft positive sample weights are set to avoid mistakenly pushing adjacent segments in the same steady state apart. Through multi-view enhancement and the reasonable construction of positive and negative sample pairs, the self-supervised encoder can enhance its sensitivity to precursor features such as local mutations and frequency band shifts while maintaining the consistency of the representation of the same steady state.

[0049] S3 constructs a self-supervised encoder comprising a temporal coding branch, a frequency coding branch, and a cross-branch fusion layer. The temporal coding branch uses one-dimensional convolution, dilated causal convolution, or a temporal Transformer to extract pulse morphology, event intervals, and local mutation features; the frequency coding branch uses two-dimensional convolution, a spectral Transformer, or a learnable filter bank to extract dominant frequency shift, bandwidth energy redistribution, and high-frequency attenuation features; the cross-branch fusion layer outputs a window representing the vector. .

[0050] The encoder employs a joint self-supervised objective function consisting of contrast consistency, mask reconstruction, future segment prediction, and temporal order discrimination. ; in, To compare the loss of consistency; Loss due to mask reconstruction; Predict loss for future potential representations; To determine the loss in the time sequence of the segments; , , and The weights are non-negative. The consistency loss is expressed as: ; in, This refers to the number of samples in a small batch. and Representation of two enhanced views within the same window; Cosine similarity; Let be the temperature parameter. The mask reconstruction loss is expressed as: ; in, For the first A set of sampling locations or time-frequency blocks where a window is obscured; For decoder reconstructed values, The signal segment is defined as robustly normalized; n is the sampling point index. Through joint training, the representation vector maintains the consistency of the same steady state under noise, amplitude, and channel perturbations, while retaining sensitivity to local mutations, frequency band shifts, and changes in event organization. This self-supervised learning framework fully utilizes the normal state information contained in the large amount of unlabeled data during the slope stability monitoring period, avoiding the dependence of supervised models on unstable labeled samples.

[0051] The joint loss weights are determined jointly based on dimensional normalization and validation set performance. (Previous training period...) rounds ( (A positive integer, used to determine the statistical starting stage for dimension normalization) Statistically analyze separately , , and The median or exponential moving average of the loss is used to divide each loss by its corresponding statistic, preventing a single task from dominating the gradient due to differences in numerical magnitude; all weights are non-negative and satisfy the following conditions: .

[0052] After completing dimensional normalization, the enhanced view of the stable validation set is used. Figure 1 Consistency, mask reconstruction error, future segment prediction error, order discrimination accuracy, potential characterization non-collapse degree, and environmental disturbance false alarm rate constitute a comprehensive evaluation function. Weights are determined using grid search, Bayesian optimization, or an adaptive balancing method based on gradient norm. The preferred range is... , , , When independent validation samples are lacking, the initial values ​​can be taken sequentially. , , and .

[0053] When the gradient norm corresponding to any loss exceeds three times the median of the gradient norms of other losses for multiple consecutive training epochs, reduce the weight of that term; when the accuracy of future prediction or order discrimination remains close to random levels for a long period, moderately increase the weight. or Weight adjustments are only performed during offline training or verified model update phases, and remain fixed during online alert inference phases.

[0054] S4, considering that the normal acoustic emission state of the slope may exhibit a multimodal distribution due to changes in season, water content, temperature, and equipment coupling, a model is established in the potential characterization space based on... A hybrid dynamic baseline composed of normal state prototypes . No. Each prototype includes weights Mean vector Covariance Matrix . No. The window belongs to the first The soft attribution probability of a normal prototype is expressed as: ; Where r is the summation index variable, Indicated by It is the mean vector. Let be the probability density function of the multidimensional normal distribution of the covariance matrix; The original anomaly deviation is calculated based on the Mahalanobis distance between the window representation and the normal prototype, the mask reconstruction residual, and the prototype attribution uncertainty: ; in, This indicates that the minimum value is taken over all normal state prototypes. This indicates taking the maximum value over all normal state prototypes. for The squared Mahalanobis distance between the m-th normal prototype; Reconstruct the residual for the window; , and Non-negative fusion weights are used. To improve the comparability of anomaly scores across different slopes and monitoring stages, the current controlled baseline is used for caching. Robust standardization of the median and median absolute deviation: ; Standardized outlier deviation A value greater than zero indicates that the current window has deviated from the normal baseline, and the larger the value, the greater the degree of deviation. For cases with insufficient baseline samples, ill-conditioned covariance matrices, or where a prototype has not been accessed for a long time, covariance contraction, minimum sample constraints, and prototype merging or splitting mechanisms are used to maintain baseline stability. The multi-prototype structure can express the multimodal distribution characteristics of the normal acoustic state of a slope.

[0055] Number of prototypes in normal state It is not arbitrarily fixed by humans. During the initial modeling... Within the candidate range, multiple prototype models are fitted, and the Bayesian information criterion, silhouette coefficient, minimum prototype occupancy, covariance condition number, and false alarm rate on the stable validation set are calculated. The minimum prototype occupancy rate is determined when the Bayesian information criterion no longer significantly decreases and the silhouette coefficient and false alarm rate meet the requirements. Stop adding prototypes at this point. The sample size of each prototype should preferably be no less than 3%, and the sample size should preferably be no less than 5 times the representation dimension. Preferred satisfaction .

[0056] During online operation, prototype splitting is performed when the internal dispersion of a prototype continuously exceeds the high quantile value of the stable cache and the number of samples is sufficient; prototype merging is performed when the Mahalanobis distance between the centers of two prototypes is lower than the merging threshold, the environmental contexts overlap, and the false positive rate does not increase after merging. It can be controlled and adjusted according to the changes in normal state modes.

[0057] The distance term, reconstruction residual term, and attribution uncertainty term in the anomaly deviation are first standardized using the median and median absolute deviation of the controlled stable cache, respectively, before being determined. , and and satisfy The parameter selection aims for low false alarms within a stable window, high detection within a verified damage window, and monotonicity of anomaly scores; the optimal range is... , , The initial value can be taken as follows: , and .

[0058] S5. Extract corresponding acoustic emission features at at least two short timescales, at least one medium timescale, and at least one long timescale. The short timescales are on the order of seconds or minutes, the medium timescales are on the order of hours, and the long timescales are on the order of days or longer. The short timescale features include at least event rate spikes, shortened event intervals, a higher proportion of high-energy events, spectral abrupt changes, and bursts of pulses; the medium timescale features include at least event rate trends and frequency band energy evolution; and the long timescale features include at least cumulative energy growth, anomaly persistence ratio, frequency band shift, trend slope, and trend acceleration. This multi-timescale feature extraction framework can simultaneously capture precursor information for two different types of temporal features: transient anomaly bursts and progressive damage accumulation.

[0059] Short-term precursor scores are calculated based on the extracted short-term, medium-term, and long-term features. Long-term cumulative damage and persistent outlier statistics .

[0060] Burstness of the current window acoustic emission event interval sequence Represented as: ; in, and These are the mean and standard deviation of the event intervals, respectively. A value close to 1 indicates that events are clearly clustered, close to 0 indicates approximately randomness, and close to -1 indicates that event intervals are relatively regular. The features at each short time scale are then combined into a vector. Short-term aura scores were obtained using scalar attention: ; in, For short-term precursor mapping networks; For the r-th short-time scale, the adaptive weights are used. and Here, S is a set of short timescales, and j is a summation index variable. The short-term precursor score is used to identify sudden increases in event rate, concentrations of high-energy pulses, rapid shortening of event intervals, and instantaneous changes in the spectrum, but it is not used as the sole basis for high-level early warning. The adaptive scale attention mechanism enables the short-term precursor score to automatically select the most sensitive timescale based on the current signal characteristics.

[0061] The long-term damage accumulation and the aforementioned persistent anomaly statistics They are represented as follows: Initial time Set to 0; Initial time Set to 0; in, The memory coefficient varies with time intervals. For long-term feature mapping function, This is a long-term scale feature vector. This is the reference drift amount determined based on the normal dynamic baseline. The long-term damage accumulation is achieved through a forgetting factor mechanism to weighted memorize historical damage information, and the persistent anomalous statistic is detected by a cumulative sum method to detect the continuous upward trend of anomalous deviation.

[0062] This study utilizes Bayesian online change point detection and a competing posterior mechanism among the normal fluctuation hypothesis, environmental disturbance hypothesis, equipment anomaly hypothesis, and actual damage hypothesis to identify the current anomaly source. The anomaly deviation, short-term precursor score, long-term damage accumulation, channel health, and environmental context are combined to form an observation vector. Bayesian online variable point detection is used to recursively calculate the current running length. Posterior probability: ; in, Indicates the length of operation since the most recent mechanism change; Determined by the hazard rate function; This represents the predicted probability of the current observation within the existing operating segment.

[0063] Establish the normal fluctuation hypothesis Environmental disturbance hypothesis Equipment malfunction assumption And the true damage hypothesis Based on the current multi-source observation sequence Post-hoc recursive mechanism: ; in, The competing posterior probability of the mechanisms includes the recursive posterior probabilities of the normal fluctuation hypothesis, the environmental disturbance hypothesis, the equipment anomaly hypothesis, and the true damage hypothesis. The true damage hypothesis requires at least two of the following to be true: anomaly persistence, damage mitigation characterization, multi-channel consistency, and change point evidence. The environmental disturbance hypothesis is determined based on the synchronicity of the anomaly with rainfall, vehicles, construction, or blasting conditions. The equipment anomaly hypothesis is determined based on single-channel anomalies, decreased channel health, and increased distortion or missing data in the reference waveform. When the equipment anomaly hypothesis prevails, channel isolation, activation of backup channels, or maintenance prompts are implemented, without directly raising the slope disaster warning level. The multi-mechanism competing discrimination framework can effectively distinguish anomalies from different sources.

[0064] When the short-term precursor score Compared to the increase in the adjacent historical window, the cumulative amount of long-term damage... and the aforementioned persistent anomaly statistics When there is no sustained increase, the current anomaly is prioritized as a transient disturbance or a recoverable local activity; when the accumulated long-term damage... and the aforementioned persistent anomaly statistics When the anomaly continues to grow and is accompanied by enhanced consistency across multiple channels, the current anomaly is considered a candidate for persistent damage. This criterion can effectively distinguish between transient environmental disturbances and persistent real damage.

[0065] S6 employs a causal temporal encoder that uses only the current window and historical windows to fuse the short-term, medium-term, and long-term features. Causal encoding utilizes only the current moment and historical windows, avoiding the use of future data that could lead to warning time leaks. This is combined with an environmental context vector composed of rainfall, temperature, time period, construction status, and channel health. Calculate the weights for each scale and form a fusion representation: ; ; in, For the first Attention score at each time scale within the current window, , and The results represent causal encoding at short, medium, and long time scales, respectively. This mechanism ensures that the fusion of features across multiple time scales respects temporal causality while adaptively adjusting the contribution weights of each scale according to the environmental context.

[0066] The system uses a stage recognition network combined with stage transition constraints and stage duration constraints to output stage probabilities for stable stages, locally activated stages, persistent damage stages, critical acceleration stages, and near-slip stages. These stage probabilities include the probability of stable stages. Further, the stage prediction set and overall uncertainty are determined through evidence output, repeated inference, and distribution-independent confidence calibration. .

[0067] Based on the fusion representation, the output includes five probability categories: stable phase, local activation phase, continuous damage phase, critical acceleration phase, and slip phase. To avoid irregular jumps in the destruction phase caused by a single anomalous window, a duration-aware semi-Markov constraint decoding is employed: ; in, For stage transition constraints; For the stage continued The duration constraint term for each window. The stage decoding of the stage recognition network uses an allowed transition matrix. and phase duration constraints, where , =0,1,2,3,4, corresponding to the stable phase, the local activation phase, the continuous damage phase, the critical acceleration phase, and the near-slip phase, respectively. Indicates by the first The first phase transitions to the second. The permissible costs for each stage are defined. The stage transition matrix preferentially allows maintaining the current stage or transitioning to an adjacent high-risk stage; when short-term high-energy events, long-term cumulative amounts, risk posteriors, and channel consistency simultaneously satisfy strong triggering conditions, skipping stages to the critical acceleration stage is permitted. For cases where short-term indicators recover after local activation but long-term damage does not increase, returning to the stable stage is allowed. Duration-aware semi-Markov constraint decoding ensures temporal consistency in stage identification results.

[0068] The phase transition matrix defines five states in order of increasing risk of destruction: For stability, For local activation, For continuous damage, For critical acceleration, For sliding. Matrix elements Indicates the state Transfer The cost: as well as Adjacent transfers are subject to limited costs; rapid regression in high-risk phases without recovery evidence and escalation across phases without strong triggering evidence are subject to prohibition costs; and strong triggering conditions are subject to... or The transitions impose conditions and penalties. The design of this transition matrix strictly follows the physical evolution laws of slope instability.

[0069] The duration probability of each stage can be obtained by statistically analyzing the number of duration windows of the verified samples. When the sample size is insufficient, a truncated geometric distribution, Poisson distribution, or log-normal distribution is used, and a minimum and maximum duration window are set. During decoding, the duration of each candidate stage and the previous allowed stage are accumulated. The stage sequence with the minimum cumulative cost is obtained by backtracking through dynamic programming.

[0070] The evidence parameters and stage probabilities output by the stage identification network are represented as follows: , , ; in, . For the first The original output of each stage, As non-negative evidence, For the evidence distribution parameters, The total parameters for all stages, The stage probability is determined using a distribution-independent calibration quantile based on verified samples. When the stage prediction set contains multiple adjacent stages, the stage with the higher risk is used for control. When the stage prediction set is empty or contains multiple stages with large spans, the sampling intensity is increased and manual review is initiated. Evidence-based deep learning and confidence calibration mechanisms enable the system to output a conservative stage prediction set when data quality is insufficient or model discrepancies are significant.

[0071] S7, based on the abnormal deviation degree The probability of the stable phase Abnormal persistence The aforementioned comprehensive uncertainty and channel health Generate a baseline update gate, wherein the anomaly duration For the most recent The percentage of abnormal windows in each window. The duration of the anomaly is a positive integer. The persistent anomaly statistic is used to determine whether the current window is eligible to update the normal baseline. Used to determine whether an anomaly has a continuous accumulation characteristic; the baseline update gating Represented as: ; in, This is an indicator function that takes the value 1 when the condition is met and 0 when the condition is not met. , , , and These five thresholds are: anomaly deviation threshold, stable phase probability threshold, anomaly persistence threshold, comprehensive uncertainty threshold, and channel health threshold. This five-fold gating ensures that only windows simultaneously meeting the criteria of low anomaly, high stability probability, low anomaly persistence, low uncertainty, and high channel health are allowed to enter the normal baseline, thus preventing the possibility of damage precursors contaminating the normal model at the source. Only windows meeting the update criteria are written to the normal cache and their matched normal state prototypes are updated; windows not meeting the update criteria are written to the observation cache or isolation cache, and the normal baseline version, updated sample index, and rollback checkpoint are saved.

[0072] Only when When the value is equal to 1, the dynamic baseline can be updated at the following rate: Represented as: ; The mean of the matched normal state prototype is updated according to the following formula: ; in, For maximum update rate, This is the abnormal suppression coefficient. For the current window to the first The probability of a normal prototype being assigned is determined. The update rate decays exponentially with the degree of anomaly; windows with higher anomaly levels have less impact on the baseline, further enhancing the anomaly contamination blocking effect. Covariance and mixed weights are updated recursively according to the same gating principle, and the maximum drift per unit time, the upper limit of prototype center movement, the observation cache, the isolation cache, and the baseline rollback checkpoint are set.

[0073] When subsequent manual inspections, displacement monitoring, or continuous low-risk operation prove that the window in the observation cache represents a normal environmental change, the window can be added to the normal baseline with a reduced learning rate after verification. When subsequent evidence indicates that the window corresponds to damage development, it is transferred to the precursor sample library. This forms a data management structure in which the normal cache, observation cache, isolation cache, and verification sample library are independent of each other.

[0074] The update threshold is determined using a method of "initial selection based on stable statistics—verification of outlier samples—environmental stratification correction." First, it is determined using the 90%–97.5% quantile of the standardized outlier deviation in the controlled stable cache. Determined by the low quantile value of the posterior in the stable phase. The high quantile values ​​of the abnormal persistence ratio and the overall uncertainty were determined respectively. and And determined based on the historical distribution of health channels The preferred range is... , , , , .

[0075] Subsequently, playback was performed using verified normal drift, environmental disturbance, and damage precursor windows. Under the conditions that the normal baseline write error rate does not exceed a preset upper limit, the damage window does not enter the normal buffer, and the healthy channel coverage requirements are met, a threshold combination that maximizes both normal drift acceptance and damage rejection rates was selected. For operating conditions such as rainfall, blasting, diurnal temperature variations, or sensor recoupling, separate environmental context threshold groups were saved, and a hysteresis mechanism was used during switching to avoid frequent jumps.

[0076] S8, the abnormal deviation The aforementioned short-term precursor score The cumulative amount of long-term damage The aforementioned persistent anomaly statistics The stage probability, the recursive posterior probability of each mechanism hypothesis, and the overall uncertainty. Integrate into a comprehensive risk value The recursive posterior probabilities of each mechanism hypothesis include the recursive posterior probabilities of the normal fluctuation hypothesis, the environmental disturbance hypothesis, the equipment anomaly hypothesis, and the actual damage hypothesis. The fusion process uses a weighted summation method to convert the recursive posterior probabilities of each mechanism hypothesis into a mechanism risk contribution value. This is then used in the calculation of the overall risk value. The mechanism's risk contribution value. The scalar value is the weighted summation of the recursive posterior probabilities of the hypotheses of each mechanism.

[0077] The comprehensive risk value Represented as: ; in, It is a monotonic mapping function. For bias terms, , , and These are, respectively, the abnormal deviation degree, the short-term precursor score, the long-term cumulative damage, and the persistent abnormality statistic. For the stage The mean probability, The risk factors are for the continuous damage, critical acceleration, and skidding phases. , , , , and The corresponding weights are used. This multi-source information fusion framework integrates anomaly deviation, multi-scale precursors, stage probability, mechanism posterior, and uncertainty into a unified risk quantification system.

[0078] Using stage transition matrix Sequential Bayesian smoothing is applied to the stage probabilities of multiple windows to obtain the posterior probability of the current stage: ; in, For the first The window is in a phase. The posterior probability, For the stage The mean probability, For the stage Transfer to stage Allowable transition matrix elements, and This is a stage index variable, with a value range of 0 to 4.

[0079] The warning threshold is not a fixed value across projects, but is dynamically determined based on the risk value quantiles in the controlled normal baseline and environmental context correction terms. The ℓth level dynamic threshold is expressed as: ; in, For the corresponding quantiles, This represents the combined risk value for each historical window within the controlled normal baseline. Update the gating status for the corresponding window. This indicates that the window belongs to a controlled normal baseline. For the first Environmental correction parameters for the threshold level, For the environment context vector, For the first The significance level corresponding to the level warning threshold. Choose 1, 2, and 3, which correspond to the yellow, orange, and red warning thresholds, respectively; the warning escalation requirement is the most recent At least one window Each window meets the corresponding risk level's risk magnitude, duration, stage set, mechanism posterior, and data quality conditions, and the warning downgrade requirement is based on the comprehensive risk value. The value is continuously below the current level threshold minus the hysteresis bandwidth and continues to be below the set time.

[0080] The warning levels are as follows: 1. Blue stable state: the overall risk value is lower than the yellow threshold, and the posterior a posteriori are mainly stable; 2. Yellow concern state: the overall risk value reaches the yellow threshold, or the local activation phase continues to occur; 3. Orange alert state: the overall risk value reaches the orange threshold, and the posterior a posteriori of continuous damage or critical acceleration is established within a set number of windows; 4. Red danger state: the overall risk value reaches the red threshold, the posterior a posteriori of critical acceleration or near-slip phase exceeds the set value, and there are at least two of the following: short-term outbreak, long-term accumulation, or multi-channel consistency evidence.

[0081] Based on the risk distribution and environmental context within a controlled normal baseline, tiered early warning thresholds are dynamically determined. A continuous confirmation and delayed retracement mechanism is used to output blue, yellow, orange, or red warnings. A blue warning indicates the slope is in a stable state; a yellow warning indicates the slope requires attention; an orange warning indicates the slope requires alert; and a red warning indicates the slope is in a dangerous state. Continuous confirmation is the requirement for the most recent warning escalation. At least one window Each window satisfies the corresponding risk level's amplitude, duration, stage set, mechanism posterior, and data quality conditions, among which... and are positive integers and The hysteresis fallback mechanism requires the comprehensive risk value to remain continuously below the current level threshold minus the hysteresis bandwidth for a set period of time before the warning level is downgraded. Continuous confirmation and the hysteresis fallback mechanism ensure temporal consistency in the raising and lowering of the warning level.

[0082] Using m-of- Continuous confirmation and delayed retracement mechanisms. Early warning escalation requirements are most recent. At least one window Each window must meet the corresponding level conditions; the warning downgrade requirement is that the risk value is continuously lower than the current level threshold minus the hysteresis bandwidth, and this continues for a set time. For high-energy, multi-channel synchronized, and well-healthy strong events, a hard trigger can be set for the channel to directly enter the orange or red manual review state.

[0083] At the same time, based on the aforementioned comprehensive risk value The expected information gain, along with energy, communication, and storage constraints, are used to adjust the sampling frequency, trigger threshold, original waveform retention time, number of participating channels, and data upload cycle. The sampling frequency, pre-gain, trigger threshold, pre-trigger retention time, post-trigger retention time, number of participating channels, feature calculation cycle, and data upload cycle are combined to form candidate monitoring action 'a'. The utility of the candidate monitoring action is calculated using the following formula. : ; in, The expected information gain of candidate monitoring actions for the next window stage judgment or damage characterization. , and These are energy consumption, communication volume, and storage volume, respectively. , and The system selects candidate monitoring actions that meet the constraints of lowest sampling frequency, fewest healthy channels, longest upload period, and lowest original waveform retention ratio, while maximizing utility. This proactive monitoring strategy enables the system to prioritize the collection of data most valuable for risk assessment under limited resource constraints.

[0084] When an orange or red alert is reached, critical channels must not be shut down or the sampling frequency reduced below a safe value. Each normal baseline update saves prototype parameters, sample indexes, environmental context ranges, and verification status. When baseline drift exceeds limits, the abnormal score of historically stable verification samples increases, the false alarm rate continues to increase after the update, or the conclusions of the old and new baselines are reversed, the online update is frozen and rolled back to the most recently verified baseline version.

[0085] Calculate the normalized risk control quantity based on the comprehensive risk value and the dynamic early warning threshold. : ; in This indicates the degree to which the current overall risk value exceeds the yellow warning threshold; This indicates the risk range between the red alert threshold and the yellow alert threshold; Adaptive adjustment of acoustic emission sampling frequency based on normalized risk control quantity and data upload cycle : ; in, and These are the lowest and highest sampling frequencies, respectively. and These represent the longest and shortest upload periods, respectively. The trigger threshold and the original waveform storage duration are adjusted according to the following formula: ; in, and These are the highest and lowest trigger thresholds, respectively; and These are the shortest and longest waveform storage times, respectively. and These represent the sensitivity indices for adjusting the sampling frequency and upload cycle, respectively. As the risk increases, the system simultaneously lowers the waveform trigger threshold, extends the retention time of the original waveforms before and after triggering, increases the number of channels participating in fusion, and shortens the feature calculation cycle at the edge. In the long-term stable phase, the system reduces the proportion of original waveforms retained, only retaining statistical features, representative waveforms, and periodically sampled waveforms. This risk-driven parameter control mechanism enables the system to acquire richer information for early warning decisions in high-risk phases, while effectively saving energy and storage resources in low-risk phases.

[0086] The warning results are linked to the baseline version, model version, and parameter adjustment records for output.

[0087] This invention also provides a multi-scale adaptive early warning system for slope acoustic emission dynamic baseline, used to execute the above-described method. The self-supervised dynamic baseline established in the method, the calculated multi-timescale precursor scores, and the stage identification results constitute the same specific technical features of the system and the method. The system includes: The acoustic emission acquisition and synchronization module is used to continuously acquire acoustic emission waveforms from multiple monitoring channels and simultaneously obtain environmental parameters, equipment status, and channel health information. The robust preprocessing module is used for sliding window partitioning, robust normalization, event detection, time-frequency transformation, and basic feature extraction. The stable sample and multi-view building module is used to filter stable candidate windows and generate self-supervised positive sample pairs and negative sample candidates; The self-supervised representation learning module is used to obtain the monitoring window representation through time-domain coding branch, frequency-domain coding branch and joint self-supervised task; The dynamic baseline and controlled update module is used to establish a multi-prototype normal state model, calculate the abnormal deviation, and manage the normal cache, observation cache, and isolation cache according to the update gating. A multi-timescale precursor module is used to extract short-term sudden precursors, medium-timescale evolution trends, and long-term damage accumulation features; The variable point and mechanism competition module is used to identify normal fluctuations, environmental disturbances, equipment anomalies and real damage using Bayesian online variable point detection; The causal fusion and stage identification module is used to fuse features from multiple time scales through causal temporal coding and output the probability of the destruction stage, the stage prediction set, and the comprehensive uncertainty. The sequential risk and adaptive early warning module is used to generate comprehensive risk values, dynamic thresholds, and blue, yellow, orange, and red early warnings. The active monitoring and control module is used to adjust the sampling frequency, trigger threshold, original waveform retention time, number of participating channels, and data upload cycle. It also includes a version governance and communication module, which is used to perform baseline version saving, rollback, and verification record management, and to send early warning information and monitoring and control instructions to the field server, remote monitoring platform or mobile terminal.

[0088] The above modules can be deployed entirely on the field server, or they can be deployed in a layered manner with the edge device completing signal preprocessing, event detection and preliminary anomaly scoring, the field server completing multi-timescale fusion and hierarchical early warning, and the cloud completing model version management, verified sample training and multi-slope operation management.

[0089] The present invention will be further described below with reference to specific embodiments. The following monitoring parameters and test data are used to illustrate the implementation process of the present invention and can be adjusted according to the actual slope type, monitoring scale and site environment.

[0090] Example 1: Comparative Examples of Early Warning Systems under Blasting Disturbance Conditions on Rock Slopes in Open-Pit Mines like Figure 1 As shown, the slope acoustic emission adaptive early warning method of the present invention is executed in the order of steps S1 to S8, forming a complete technical closed loop from data acquisition, self-supervised representation learning, dynamic baseline establishment, multi-timescale precursor extraction, stage identification to risk early warning and parameter control.

[0091] A rock slope in an open-pit mine was selected as the monitoring target. The slope has a total height of approximately 68 m, with well-developed joints and fissures, and is subject to periodic production blasting disturbances. Twelve waveguide-coupled acoustic emission monitoring nodes were deployed along the slope top, middle, and toe, simultaneously collecting data on blasting times, rainfall, channel health status, and manual inspection records. After 20 days of continuous and stable system operation, 18,400 stable windows were selected from periods with no significant displacement growth, no new cracks, and a monitoring channel health score higher than 0.85. These windows were used for self-supervised characterization learning and normal dynamic baseline construction.

[0092] In this embodiment, the four self-supervised losses are divided by their medians on stable samples in the early training phase to ensure they are of similar magnitude. Grid selection is then performed using representation consistency, reconstruction error, and false alarm rate of blast perturbations on the stable validation set as objectives. Finally, the grid selection is... , , , .exist Within the candidate range, the Bayesian information criterion is... When the value reaches a small value and the profile coefficient no longer increases significantly, four normal state prototypes are selected.

[0093] After robust standardization, the Mahalanobis distance term, the reconstructed residual term, and the prototype attribution uncertainty term of the anomaly deviation are respectively taken as follows: , , The update threshold is determined using a combination of stable sample statistics and manual verification results. In this embodiment, the threshold is... , , , , Parameter change At that time, the false alarm rate and damage detection rate of the stable validation set did not show any significant changes. Therefore, the above values ​​were used as the initial values ​​in the field, and only minor adjustments were made during the online operation within the update cycle of the verified baseline version.

[0094] To verify the performance of the method of this invention, it was compared with the fixed threshold method and the conventional adaptive threshold method. The test samples included four categories: stable state, normal blasting disturbance state, post-blasting continuous damage state, and critical acceleration state, totaling 480 labeled windows. Evaluation metrics included accuracy, precision, recall, F1 score, warning lead time, and number of false alarms.

[0095] Test results are as follows Figure 2As shown: the fixed threshold method has an accuracy of 78.6%, precision of 71.4%, recall of 76.9%, F1 score of 74.0%, warning lead time of 0.8 hours, and 11 false alarms; the conventional adaptive threshold method has an accuracy of 85.2%, precision of 82.1%, recall of 84.6%, F1 score of 83.3%, warning lead time of 1.6 hours, and 5 false alarms; the method of this invention has an accuracy of 93.8%, precision of 91.7%, recall of 95.4%, F1 score of 93.5%, warning lead time of 2.5 hours, and only 1 false alarm.

[0096] The results show that the method of the present invention can better distinguish between short-term high-energy disturbances caused by normal blasting and real damage that continues to develop after blasting. It not only has higher identification accuracy but also significantly reduces false alarms and can issue effective early warnings earlier, indicating that it has good engineering application value in open-pit mine rock slope monitoring scenarios.

[0097] Example 2: Comparative Examples of Early Warning Systems Based on Reservoir Bank Slope Water Level Fluctuations A reservoir bank slope was selected as the monitoring target. The slope is approximately 35m high, with the upper part mainly consisting of silty clay and gravelly soil, and the lower part of strongly weathered sandstone. It is significantly affected by the periodic rise and fall of the reservoir water level. Ten acoustic emission monitoring nodes were deployed along the slope, simultaneously collecting data on reservoir water level, rainfall, pore water pressure, and surface displacement. The system operated continuously for 60 days, with the first 30 days used to establish a normal dynamic baseline, and the remaining time used for online identification and performance evaluation.

[0098] In this embodiment, the self-supervised joint loss is first normalized in terms of dimensions. , , , .exist Within the candidate range, screening was performed using the Bayesian information criterion, silhouette coefficient, and false alarm rate of normal window cross-validation; since the reservoir water level rise and fall correspond to multiple repeatable normal operating conditions, the final selection was... The weight of abnormal deviation is taken as follows: , , .

[0099] The update threshold was calibrated separately for controlled stable samples in different reservoir water level ranges, among which... Using the 95th percentile of the standardized anomaly deviation under the corresponding environmental stratification, this embodiment takes... , , , , The phase identification network adopts... Evidence mapping and Dirichlet parameterization, semi-Markov decoding only allow hold, adjacent stage transitions, and cross-stage entry into the critical acceleration stage after strong trigger confirmation.

[0100] To verify the applicability of this invention in environments with fluctuating water levels, the method of this invention was compared with the fixed threshold method and conventional time series models. The test samples included four categories: stable state, normal water storage state, yellow warning state for rapid water recession, and orange warning state for rapid water recession, totaling 520 labeled windows. The evaluation metrics also included accuracy, precision, recall, F1 score, warning lead time, and number of false alarms.

[0101] Test results are as follows Figure 3 As shown: the fixed threshold method has an accuracy of 76.9%, precision of 69.8%, recall of 74.2%, F1 score of 72.0%, warning lead time of 1.3 hours, and 9 false alarms; the conventional time series model has an accuracy of 84.7%, precision of 81.5%, recall of 83.9%, F1 score of 82.7%, warning lead time of 2.4 hours, and 4 false alarms; the method of this invention has an accuracy of 92.6%, precision of 90.8%, recall of 94.1%, F1 score of 92.4%, warning lead time of 4.0 hours, and only 1 false alarm.

[0102] The results show that, under the background fluctuation caused by normal rise and fall of reservoir water level, the method of the present invention is less likely to misjudge the normal water storage process as a dangerous state; while in the stage where rapid water receding leads to the gradual development of internal slope damage, the method of the present invention can maintain high identification accuracy and issue early warnings, indicating that it has better adaptability to environmental changes.

[0103] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-scale adaptive early warning method for dynamic baseline acoustic emission of slopes, characterized in that, Includes the following steps: S1, acquire multi-channel acoustic emission waveforms of the slope to be monitored, simultaneously obtain environmental parameters and channel status, preprocess the acoustic emission waveforms to obtain the acoustic emission characteristics and channel health of the monitoring window; S2, filter stable candidate windows, perform at least two enhancements on stable candidate windows to maintain semantic consistency, and construct multi-view positive sample pairs and negative sample candidates; S3 uses contrast consistency loss, mask reconstruction loss, future segment prediction loss and temporal order discrimination loss to train a self-supervised encoder to obtain the latent representation and reconstruction residual of the monitoring window; S4. Establish a dynamic baseline containing multiple normal state prototypes based on latent representations, and calculate the anomaly deviation based on latent representations, reconstruction residuals, and prototype attribution uncertainties. S5 extracts acoustic emission precursor features at short, medium, and long time scales, calculates short-term precursor scores, long-term damage accumulation, and persistent anomaly statistics, and uses the competition mechanism among normal fluctuations, environmental disturbances, equipment anomalies, and actual damage to identify the current anomaly source. S6 employs a causal temporal encoder that uses only the current window and historical windows to fuse precursor features at various time scales. Combined with stage transition constraints and stage duration constraints, it outputs the stage probabilities and comprehensive uncertainties of stable stage, local activation stage, continuous damage stage, critical acceleration stage, and near-slip stage. S7. Generate a baseline update gate based on the abnormal deviation, the stage probability, the continuous abnormal statistics, the comprehensive uncertainty and the channel health. Only write windows that meet the update conditions into the normal cache and update the dynamic baseline. Write windows that do not meet the update conditions into the observation cache or the isolation cache. S8, integrate the abnormal deviation, the short-term precursor score, the long-term damage accumulation, the continuous abnormality statistics, the stage probability, and the comprehensive uncertainty to obtain a comprehensive risk value, output a warning level based on the comprehensive risk value and the dynamic warning threshold, and adjust the sampling frequency, trigger threshold, original waveform storage duration, and data upload cycle.

2. The method according to claim 1, characterized in that, In step S1, the preprocessing includes robust normalization and basic feature extraction. The robust normalization uses the median and median absolute deviation of the current window to eliminate the effects of DC drift, gain difference, and occasional spikes. The basic features include at least five of the following: event rate, peak amplitude, energy, duration, rise time, ring count, center frequency, band energy ratio, event interval, and cross-channel synchronization ratio. Monitoring channels with channel health below a set value are downweighted, isolated, or generate maintenance prompts, and are not allowed to trigger orange or red alerts independently.

3. The method according to claim 1, characterized in that, In step S3, the joint loss of the self-supervised encoder Represented as: ; in, For the aforementioned consistency loss, The mask reconstruction loss is... Predict the loss for the future segment. To determine the loss in the time sequence, , , and The weights are non-negative and satisfy the following conditions: The contrast consistency loss is used to maintain the consistency of the representation of the same stable window under noise, amplitude and channel perturbation. The mask reconstruction loss, the future segment prediction loss and the time order discrimination loss are used to retain information on local mutations, frequency band migration and event organization changes.

4. The method according to claim 1, characterized in that, In step S4, the dynamic baseline includes The first normal state prototype, the... A normal state prototype includes at least a mixed weight. mean Covariance ; Current window and the first Square Mahalanobis distance between normal state prototypes Represented as: ; The abnormal deviation Based on the squared Mahalanobis distance corresponding to each normal state prototype, the reconstructed residual and the probability of belonging to the maximum prototype calculate: ; in, This is the minimum squared Mahalanobis distance of the current window relative to all normal state prototypes. , and The weights are non-negative and satisfy the following conditions: When the number of normal samples is insufficient, the covariance is ill-conditioned, or the prototype has not been accessed for a long time, at least one of the following methods can be used to maintain dynamic baseline stability: covariance shrinkage, minimum sample limit, prototype merging, or prototype splitting.

5. The method according to claim 1, characterized in that, In step S5, the burstiness of the current window acoustic emission event interval sequence Represented as: ; in, and These are the mean and standard deviation of the event intervals, respectively. Positive number; cumulative damage over long period of time and persistent outlier statistics They are represented as follows: Initial time Set to 0; Initial time Set to 0; in, The memory coefficient varies with time intervals. For long-term feature mapping function, This is a long-term scale feature vector. The reference drift amount is determined based on the normal dynamic baseline; when the short-term precursor score is... Compared to the increase in the adjacent historical window, the cumulative amount of long-term damage... and the aforementioned persistent anomaly statistics When there is no sustained increase, the current anomaly is prioritized as a transient disturbance or a recoverable local activity; when the accumulated long-term damage... and the aforementioned persistent anomaly statistics When there is continuous growth accompanied by enhanced consistency across multiple channels, the current anomaly is considered a candidate for persistent damage.

6. The method according to claim 1, characterized in that, In step S7, only when the baseline update gate is... When the value is 1, dynamic baseline updates are allowed, and the update rate is... Represented as: ; The mean of the matched normal state prototype is updated according to the following formula: ; in, For maximum update rate, This is the abnormal suppression coefficient. For the current window to the first The probability of a normal state prototype is assigned; the covariance and mixed weights are updated recursively according to the same gating principle, and the maximum drift per unit time, the upper limit of prototype center movement, the observation cache, the isolation cache, and the baseline rollback checkpoint are set.

7. The method according to claim 1, characterized in that, In step S6, the precursor features at each time scale include short-time scale features, medium-time scale features, and long-time scale features extracted at the short-time scale, medium-time scale, and long-time scale, respectively, and the fusion weights of the short-time scale features, medium-time scale features, and long-time scale features. and fusion characterization They are represented as follows: ; ; in, For the first Attention score at each time scale within the current window. , and The causal coding results are given for short, medium, and long time scales, respectively. The fused representation is input into a stage identification network, where stage decoding employs an allowed transition matrix. and phase duration constraints, where , =0,1,2,3,4, corresponding to the stable phase, the local activation phase, the continuous damage phase, the critical acceleration phase, and the near-slip phase, respectively. Indicates by the first The first phase transitions to the second. The permissible costs for each stage prioritize maintaining the current stage or transitioning to an adjacent high-risk stage, and allow skipping stages to the critical acceleration stage when short-term bursts, long-term accumulations, and multi-channel consistency are simultaneously satisfied; the evidence parameters and stage probabilities output by the stage identification network are expressed as follows: ; in, For the first The original output of each stage, As non-negative evidence, For the evidence distribution parameters, The total parameters for all stages, The stage probability is determined using verified samples to establish a distribution-independent calibration quantile. A stage prediction set is then determined based on the stage probability and the distribution-independent calibration quantile. When the stage prediction set contains multiple adjacent stages, the stage with the higher risk is used for control. When the stage prediction set is empty or contains multiple stages with a large span, the sampling intensity is increased and manual review is initiated.

8. The method according to claim 1, characterized in that, In step S8, the comprehensive risk value Represented as: ; in, It is a monotonic mapping function. For bias terms, , , and These are, respectively, the abnormal deviation degree, the short-term precursor score, the long-term cumulative damage, and the persistent abnormality statistic. For the stage The mean probability, The risk factors are for continuous damage, critical acceleration, and the skid-end phase. , , , , and For the corresponding weights; the first Level dynamic threshold is expressed as: ; in, For the corresponding quantiles, This represents the combined risk value for each historical window within the controlled normal baseline. Update the gating status for the corresponding window. This indicates that the window belongs to a controlled normal baseline. For the first Environmental correction parameters for the threshold level, For the environment context vector, For the first The significance level corresponding to the level warning threshold. Choose 1, 2, and 3, which correspond to the yellow, orange, and red warning thresholds, respectively; the warning escalation requirement is the most recent At least one window Each window meets the corresponding risk level's risk magnitude, duration, stage set, mechanism posterior, and data quality conditions, and the warning downgrade requirement is based on the comprehensive risk value. The value is continuously below the current level threshold minus the hysteresis bandwidth and continues to be below the set time.

9. The method according to claim 1, characterized in that, In step S8, the sampling frequency, pre-gain, trigger threshold, pre-trigger storage duration, post-trigger storage duration, number of participating channels, feature calculation cycle, and data upload cycle are combined to form candidate monitoring actions. The utility of candidate monitoring actions is calculated according to the following formula. : ; in, The expected information gain of candidate monitoring actions for the next window stage judgment or damage characterization. , and These are energy consumption, communication volume, and storage volume, respectively. , and The system selects candidate monitoring actions that meet the constraints of minimum sampling frequency, minimum number of healthy channels, longest upload period, and minimum original waveform retention ratio, and have the greatest utility. When an orange or red alert is reached, key channels are not allowed to be closed or the sampling frequency is not allowed to be reduced below a safe value. Each normal baseline update saves the prototype parameters, sample index, environmental context range, and verification status. When the baseline drift exceeds the limit, the abnormal score of historical stable verification samples increases, the false alarm rate continues to increase after the update, or the conclusions of the old and new baselines are reversed, the online update is frozen and rolled back to the most recently verified baseline version.

10. A slope acoustic emission dynamic baseline multi-scale adaptive early warning system, used to execute the slope acoustic emission dynamic baseline multi-scale adaptive early warning method according to any one of claims 1 to 9, wherein the self-supervised dynamic baseline established in the method, the calculated multi-timescale precursor scores, and the stage identification results constitute the same specific technical features of the system and the method, and the system includes: The acoustic emission acquisition and synchronization module is used to continuously acquire acoustic emission waveforms from multiple monitoring channels and simultaneously obtain environmental parameters, equipment status, and channel health information. The robust preprocessing module is used for sliding window partitioning, robust normalization, event detection, time-frequency transformation, and basic feature extraction. The stable sample and multi-view building module is used to filter stable candidate windows and generate self-supervised positive sample pairs and negative sample candidates; The self-supervised representation learning module is used to obtain the monitoring window representation through time-domain coding branch, frequency-domain coding branch and joint self-supervised task; The dynamic baseline and controlled update module is used to establish a multi-prototype normal state model, calculate the abnormal deviation, and manage the normal cache, observation cache, and isolation cache according to the update gating. A multi-timescale precursor module is used to extract short-term sudden precursors, medium-timescale evolution trends, and long-term damage accumulation features; The variable point and mechanism competition module is used to identify normal fluctuations, environmental disturbances, equipment anomalies and real damage using Bayesian online variable point detection; The causal fusion and stage identification module is used to fuse features from multiple time scales through causal temporal coding and output the probability of the destruction stage, the stage prediction set, and the comprehensive uncertainty. The sequential risk and adaptive early warning module is used to generate comprehensive risk values, dynamic thresholds, and blue, yellow, orange, and red early warnings. The active monitoring and control module is used to adjust the sampling frequency, trigger threshold, original waveform retention time, number of participating channels, and data upload cycle. It also includes a version governance and communication module, which is used to perform baseline version saving, rollback, and verification record management, and to send early warning information and monitoring and control instructions to the field server, remote monitoring platform or mobile terminal.

Images