Multi-scale acoustic emission and deep learning integrated rock mass instability early warning method and system

Abstract

The present application relates to the technical field of rock mass instability early warning, and discloses a rock mass instability early warning method and system combining multi-scale acoustic emission and deep learning, comprising: collecting rock mass acoustic emission waveform signals, carrying out denoising, filtering and normalization preprocessing on the signals, and carrying out event-level segmentation on the signals based on an acoustic emission event triggering mechanism; constructing time sequence characteristics reflecting the rock mass damage evolution process according to the occurrence time, energy change and event density of a plurality of continuous acoustic emission events; inputting the time sequence characteristics into a state recognition model to determine whether the rock mass is in a stable stage, a damage evolution stage or a instability critical stage, determine the rock mass risk level, and output corresponding early warning information; and updating the state recognition model when a new collected sample reaches a preset condition during monitoring. The present application reduces the dependence of artificial experience parameters on rock mass instability determination, and improves the consistency and reliability of early warning results.

Description

Technical Field

[0001] This invention relates to the field of rock mass instability early warning technology, specifically to a rock mass instability early warning method and system that integrates multi-scale acoustic emission and deep learning. Background Technology

[0002] Rock mass instability early warning is an important means of dynamic monitoring and risk assessment of the stability of rock mass structures in geological bodies, slopes, mines, or underground engineering. Traditional rock mass monitoring mostly relies on acoustic emission technology, which uses sensors deployed on or inside the rock mass to collect signals released by microcracks generated during the stress and deformation process of the rock mass. The internal damage evolution state of the rock mass is reflected by analyzing characteristic parameters such as the count rate, energy, and amplitude of acoustic emission events. Acoustic emission-based monitoring methods have the advantages of high response sensitivity and applicability to complex environments, and are an important research direction in the field of rock mass safety monitoring. At present, acoustic emission early warning methods generally use the method of manually selecting characteristic parameters and setting fixed thresholds for state determination.

[0003] However, in current technology, the judgment is made by manually selecting feature parameters and setting fixed thresholds, which is easily affected by noise, environment and rock mass type. The thresholds need to be adjusted frequently according to the working conditions, making it difficult to maintain the consistency of the judgment. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a rock mass instability early warning method and system based on acoustic emission signal processing and state recognition. By performing structured processing and state determination on acoustic emission signals, it solves the problems that acoustic emission early warning is easily affected by noise, environment and rock mass type, and that the threshold needs to be frequently adjusted according to working conditions, making it difficult to maintain consistent judgments.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a rock mass instability early warning method integrating multi-scale acoustic emission and deep learning, comprising the following steps:

[0006] S1. Acoustic emission waveform signals are collected by acoustic emission sensors deployed on or inside the rock mass;

[0007] S2. Perform noise reduction, filtering and normalization preprocessing on the acoustic emission waveform signal, and perform event-level segmentation on the signal based on the acoustic emission event triggering mechanism to obtain acoustic emission event waveform data;

[0008] S3. For each acoustic emission event, extract parameters including the occurrence time, energy value, amplitude, duration, and ringing count. Based on multiple consecutive acoustic emission events, calculate the event time interval sequence within at least three time windows of different time scales, and count the number of events within each unit time window to obtain the event density.

[0009] S4. Within each time window, calculate at least one statistical feature among the event time interval sequence, event energy value, and event density. The statistical feature is selected from a group consisting of mean, standard deviation, maximum value, minimum value, and rate of change. Concatenate the statistical features of each time window to construct a multi-scale time series feature vector.

[0010] S5. Input the multi-scale time series feature vector into a deep learning state recognition model that can automatically extract the time series features of acoustic emission signals and perform multi-classification. The model is trained by supervised learning based on indoor rock sample test data and field monitoring data to obtain the judgment probability that the rock mass is in a stable stage, a damage evolution stage and an instability critical stage, and to determine the rock mass state of the current time window according to the maximum probability criterion.

[0011] S6, Record Continuous The rock mass condition within a time window, among which The value is a preset positive integer; the number of consecutive integers is counted. Within each time window, the frequencies at which the rock mass was identified as being in the stable stage, the damage evolution stage, and the critical instability stage were respectively: , , ,and ; Calculate the state evolution index , ;

[0012] S7. Based on the current rock mass state and the aforementioned state evolution index Determine the risk level of the rock mass and output early warning information; the risk level includes at least low risk, medium risk, and high risk. respectively with the first threshold With the second threshold The risk level is determined by comparison, wherein ,when The risk level is low when The risk level is medium risk. The risk level is high.

[0013] S8. During the monitoring process, when the number of newly collected samples reaches the preset conditions, the state recognition model is incrementally updated based on the actual damage of the rock mass on site or the results of manual annotation.

[0014] Furthermore, the at least three time windows with different time scales include a short-term window, a medium-term window, and a long-term window; wherein, the short-term window is 5 seconds, used to capture instantaneous sudden events and rapid crack propagation; the medium-term window is 30 seconds, used to reflect the short-term damage accumulation process; and the long-term window is 120 seconds, used to reveal long-term evolution trends.

[0015] Furthermore, the state recognition model is a deep learning model capable of automatically extracting the temporal features of acoustic emission signals and performing multi-classification. The deep learning model is a convolutional neural network, a recurrent neural network, or a hybrid structure thereof.

[0016] Furthermore, it also includes step S6a: calculating the state evolution index. Rate of change within adjacent calculation periods , ,in This is the state evolution index for the current calculation period. The state evolution index of the previous calculation period; when Greater than the third threshold At that time, a trend acceleration warning is triggered, indicating that the rock mass condition is rapidly deteriorating.

[0017] Furthermore, in step S5, when training the deep learning state recognition model, the labeling rules for training samples are as follows: for indoor rock sample loading test data, stages are divided according to the stress-strain curve, and the stress ratio reflects the stress stage of the rock mass. When the stress ratio... The time is marked as "stable phase" when The time is marked as "damage evolution stage", when The time was marked as "critical stage of instability";

[0018] For on-site monitoring data, retrospective labeling is performed based on the actual time of damage. The first 0-10 minutes before damage is labeled as the "critical instability stage", the first 10-60 minutes before damage is labeled as the "damage evolution stage", and the remaining time period is labeled as the "stable stage".

[0019] A multi-scale acoustic emission and deep learning-integrated rock mass instability early warning system is disclosed for implementing the aforementioned multi-scale acoustic emission and deep learning-integrated rock mass instability early warning method. The system includes: an acoustic emission signal acquisition module, a signal preprocessing module, a time series construction module, a state identification module, a state evolution analysis module, a risk determination module, and an adaptive update module. The time series construction module is used to construct multi-scale time series feature vectors within at least three time windows of different scales. The state evolution analysis module is used to record the rock mass state over K consecutive time windows and calculate the state evolution index. With the rate of change of the evolutionary index The risk assessment module is used to determine the risk level based on... threshold and The trigger conditions output risk warnings and trend acceleration warnings; the adaptive update module is used to perform incremental training and update model parameters when the number of new samples reaches a preset threshold.

[0020] Furthermore, the time series construction module further includes an event identification unit, a parameter extraction unit, and a multi-scale feature calculation unit. The event identification unit is used to identify the start and end times of events based on short-time energy detection or STA / LTA algorithms. The parameter extraction unit extracts parameters including the event occurrence time, energy, amplitude, duration, and ringing count.

[0021] Furthermore, the state recognition module includes: a feature extraction unit, a state classification unit, and a model storage unit. The feature extraction unit performs deep feature extraction on the input multi-scale time series vector, the state classification unit outputs a three-class probability vector, and the model storage unit stores the trained model parameters and configuration.

[0022] Furthermore, the adaptive update module includes: a feedback data collection unit, a sample relabeling unit, and an incremental learning unit, wherein the incremental update can be executed on an edge computing device or a server.

[0023] Furthermore, the data storage module is used to store historical acoustic emission signals, time series characteristics, identification results, and actual damage conditions; the visualization display module is used to display the rock mass stability state, risk level, and state evolution curve in real time; and the remote communication module is used to output early warning information to the monitoring center or mobile terminal.

[0024] This invention provides a method and system for early warning of rock mass instability based on multi-scale acoustic emission and deep learning. It has the following beneficial effects:

[0025] 1. This invention utilizes a deep learning model to fuse and analyze multi-scale acoustic emission data. It can automatically extract nonlinear, high-dimensional instability precursor features that are difficult to identify using traditional methods from the complex signals of crack initiation, propagation, and penetration within rock masses. Compared to existing early warning methods based on single physical quantities or simple threshold judgments, this invention significantly improves the accuracy and reliability of early warning results, substantially reduces false alarms and missed alarms, and makes early warning decisions more reliable.

[0026] 2. Addressing the shortcomings of existing early warning systems, such as short warning windows and low risk differentiation, this invention utilizes deep learning to deeply model the rock mass instability evolution process, enabling the capture of weaker and earlier precursory instability information. By combining a multi-stage judgment mechanism encompassing stability, damage evolution, instability threshold, and instability acceleration, along with a quantified evolution index E, this invention provides ample advance warning time and allows for detailed classification of rock mass instability risks. This provides engineering managers with valuable emergency response time and supports tiered and differentiated safety management measures.

[0027] 3. Existing early warning methods are often limited by specific geological conditions and are susceptible to performance degradation due to environmental noise interference. In this invention, the deep learning model, trained with large datasets, possesses powerful feature generalization and noise suppression capabilities, enabling it to adapt to different rock types, geological structures, and complex external disturbances such as blasting and rainfall. Simultaneously, the redundant information from multi-scale acoustic emission signals helps identify and reduce invalid noise, ensuring the system maintains stable and efficient early warning performance even in complex and variable working environments, effectively reducing the frequent parameter calibration and maintenance required by traditional systems.

[0028] 4. This invention provides an integrated rock mass instability early warning system, realizing a complete process from automatic acquisition and processing of multi-scale acoustic emission data, deep feature extraction, model inference, multi-stage judgment, evolution index calculation to automatic release of graded early warning information. This automated and intelligent decision-making process significantly reduces reliance on traditional manual experience and subjectivity, effectively avoids risks caused by human error or delays, and ensures the timeliness, objectivity, and consistency of early warning response, thereby improving the operational efficiency and reliability of the entire early warning system.

[0029] 5. The evolution index E and clear graded early warning results output by this invention provide engineering managers with an intuitive and quantifiable risk assessment tool. Based on these precise quantitative indicators and early warning levels, managers can formulate more targeted monitoring enhancement plans, resource allocation plans, or activate emergency response plans, elevating rock mass instability risk from qualitative judgment to more refined quantitative management. This greatly improves the scientific nature, predictability, and efficiency of decision-making, effectively ensuring the long-term stable operation of engineering facilities and the safety of personnel and property. Attached Figure Description

[0030] Figure 1 This is a flowchart of the rock mass instability early warning method integrating multi-scale acoustic emission and deep learning according to the present invention.

[0031] Figure 2 This is an architecture diagram of the rock mass instability early warning system that integrates multi-scale acoustic emission and deep learning according to the present invention. Detailed Implementation

[0032] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Please see the appendix Figure 1 This invention provides a rock mass instability early warning method that integrates multi-scale acoustic emission and deep learning, comprising:

[0034] S1. Using high-sensitivity acoustic emission sensors deployed on or inside the rock mass, the original acoustic emission waveform signals generated by the rock mass under stress are collected in real time.

[0035] The acoustic emission signal acquisition includes: firstly, deploying multiple acoustic emission sensors on the rock surface and at different depths; then, using synchronous acquisition to achieve multi-channel parallel sampling and clock calibration; and finally, performing pre-amplification and anti-aliasing filtering on the acquired signals to obtain the original waveform signals for subsequent preprocessing.

[0036] Specifically, acoustic emission signal acquisition is achieved through high-sensitivity broadband acoustic emission sensors deployed on the rock surface and at different depths. These sensors are fixed to the monitoring points within the rock mass using anchor bolts or adhesives. Their number and distribution can be determined based on the rock mass structure characteristics, the size of the monitoring area, and geological conditions. For example, in tunnel surrounding rock monitoring, several acquisition points can be set along the arch, side walls, and floor to form a three-dimensional monitoring array; in slopes or open-pit mines, sensors can be arranged in layers to acquire fracture signals at different depths.

[0037] Each sensor is connected to the central acquisition terminal via a multi-channel data acquisition unit, and a unified clock source is used for sampling synchronization to ensure time consistency of data acquired from different channels. The sampling frequency is generally selected between 500 kHz and 2 MHz to cover the main frequency bands generated during rock fracturing. Before entering the data acquisition equipment, the acquired signal is processed by a preamplifier and an anti-aliasing filter circuit. The preamplifier is used to enhance the amplitude of the weak acoustic emission signal, and the anti-aliasing filter is used to suppress high-frequency noise components to ensure the fidelity of the signal waveform.

[0038] Taking laboratory rock sample loading tests as an example, four acoustic emission sensors can be deployed on the surface of a cylindrical rock sample, evenly distributed around its circumference. During loading, as stress increases, the sensors collect acoustic emission events at different time points. The waveforms are pre-amplified and sent to a data acquisition card, with trigger time and amplitude information recorded simultaneously. In field monitoring, using the same principle, multi-point acquisition can be connected to a portable acquisition host or cloud acquisition terminal to achieve long-term continuous monitoring. This allows for the acquisition of raw acoustic emission waveform data reflecting the propagation of microcracks, shear slip, and energy release processes within the rock mass, providing a reliable data foundation for subsequent signal preprocessing and deep learning analysis.

[0039] S2. Preprocess the original acoustic emission signal, including signal denoising, filtering, normalization and event trigger detection, to obtain event-level waveform data with a uniform sampling rate and amplitude range.

[0040] The signal preprocessing includes: first, denoising the original waveform signal using wavelet thresholding or bandpass filtering; then, normalizing the amplitude of the denoised signal; and finally, performing event-level segmentation based on the acoustic emission event triggering mechanism to obtain waveform data corresponding to each acoustic emission event.

[0041] After the acoustic emission signal is acquired, preprocessing of the original waveform signal is necessary to ensure the consistency and analyzability of the data input. First, wavelet threshold denoising or bandpass filtering is used to suppress noise in the original waveform. Wavelet threshold denoising employs a Daubechies wavelet basis for four-level multi-scale decomposition, dividing the signal into different frequency bands, and setting a soft threshold for high-frequency noise components.

[0042] ;

[0043] in Where is the noise standard deviation and N is the signal length.

[0044] The bandpass filter uses a Butterworth IIR filter with upper and lower cutoff frequencies set to 100 kHz (lower cutoff) and 500 kHz (upper cutoff) respectively, to effectively filter out low-frequency background vibrations (<100 kHz) and high-frequency interference components (>500 kHz), while retaining the effective frequency band that reflects the activity of microcracks inside the rock mass, thereby reducing environmental noise while ensuring signal fidelity.

[0045] After denoising and filtering, the resulting signal undergoes amplitude normalization. The normalization method uses a standardization approach:

[0046] ;

[0047] Map the signal amplitude to the [0,1] interval.

[0048] Subsequently, event-level segmentation is performed based on short-time energy (STA) detection or the STA / LTA ratio algorithm. The event triggering threshold is set as follows:

[0049] ;

[0050] in and These are the mean and standard deviation of the background noise energy, calculated from a data segment that did not produce significant acoustic emission activity in the preceding 60 seconds. When the real-time energy exceeds this adaptive threshold, the system automatically identifies it as the start point of an acoustic emission event; when the energy falls back below the threshold, it is identified as the end point of the event, thus segmenting the continuous waveform signal into a series of discrete acoustic emission events.

[0051] S3. Based on the event-level waveform data obtained in step S2, construct multi-scale time series features reflecting the rock mass damage evolution process, specifically including:

[0052] For each identified acoustic emission event, the following parameters are extracted:

[0053] When it happened The event start time is determined by the event trigger detection algorithm.

[0054] Energy value The energy of the event waveform is obtained by integrating the square of the event waveform signal. The calculation formula is:

[0055] ;

[0056] in The amplitude of the event waveform signal. The starting point of the event. This is the time when the event terminates. In numerical calculations, for a sampling frequency of... For discrete signals, the above integral can be discretized as:

[0057] ;

[0058] in The number of sampling points included in the event. For the first The amplitude of each sampling point.

[0059] amplitude , which is the maximum amplitude value of the event waveform, reflects the instantaneous intensity of crack release.

[0060] Duration The duration of the event is calculated using the following formula:

[0061] ;

[0062] in The starting point of the event. This is the moment the event ends.

[0063] Ringing count The number of times the waveform exceeds the threshold represents the complexity of the event.

[0064] During the stress process of rock mass, acoustic emission events in the stable stage have lower energy and longer time intervals; after entering the damage evolution stage, the event energy gradually increases and the frequency of occurrence increases significantly; near the instability stage, high-energy events occur densely and the time intervals shorten sharply. By extracting the above parameters, the spatiotemporal evolution law of microfracture activity inside the rock mass can be quantitatively characterized.

[0065] Within the preset time window Within (e.g., 30 seconds), for the identified For each acoustic emission event, calculate the following timing characteristics:

[0066] Event time interval sequence: for continuously identified A sequence of acoustic emission events, whose time intervals are defined as follows: , of which The formula for calculating each time interval is:

[0067] ;

[0068] in For the first The moment the event occurred, For the first The timing of each event. This parameter reflects the temporal frequency of acoustic emission events and is a key indicator for judging the rate of rock mass damage accumulation. Actual data examples show that in the stable phase, the average time interval is approximately 2-5 seconds, with events occurring sparsely and randomly; in the damage evolution phase, the average time interval decreases to 0.5-1 seconds, and the event frequency increases significantly; in the critical instability phase, the average time interval shortens to below 0.1 seconds, exhibiting continuous burst characteristics.

[0069] Event energy sequence: Energy reflects the scale of microfractures within the rock mass and the magnitude of released strain energy, serving as a direct physical quantity for assessing the degree of rock mass damage. Measured data examples show that, in the stable phase, the energy is mostly [missing information]. Low-energy events of aJ (Ajoule); during the damage evolution stage, [the following occurs]. The proportion of mid-energy events in aJ is gradually increasing; in the critical stage of instability, A dense surge of high-energy events exceeding aJ occurred, with individual events reaching energies of up to [aJ]. aJ.

[0070] Event density sequence: In sliding window analysis, the first... Event density per time window Defined as the number of events per unit of time, the calculation formula is:

[0071] ;

[0072] in For the first The total number of acoustic emission events detected within a sliding window. This represents the time span of the window (in seconds). Event density characterizes the frequency of acoustic emission activity per unit time, reflecting the activity level of the rock mass fracturing process. Example of measured data shows that during the steady phase... Approximately 0.2-0.5 events / second; during the damage evolution stage, Approximately 1-3 events / second; during the critical stage of instability, Greater than 5 events / second, even reaching 10-20 events / second.

[0073] To comprehensively capture the characteristics of rock mass damage evolution at different time scales, three time windows were set: short-time window... = 5 seconds, medium time window = 30 seconds, long window = 120 seconds.

[0074] According to the multi-scale theory of fracture mechanics, the short-time window corresponds to the stress concentration zone near the crack tip. At this scale, the acoustic emission event interval... With crack propagation rate The relationship reflects the rapid crack propagation process:

[0075] ;

[0076] in It is the spatial distance between adjacent acoustic emission sources (typically 1-5 mm). Within 5 seconds, it can capture 1-2 events to establish a baseline during the steady phase, 5-10 events to reflect the acceleration process during the damage phase, and at least 50 events to trigger an early warning during the critical phase, providing extremely strong distinguishability.

[0077] The intermediate time window corresponds to the formation process of the crack network. The convergence time of multiple microcracks is typically 10-60 seconds. Within 30 seconds, the gradual accumulation process of damage can be observed completely, while avoiding the influence of individual sudden events. The time interval with the 5-second window (6 times the interval) is large enough to ensure the independence of the two scales.

[0078] The long-term window corresponds to the formation of the overall fracture surface. According to the energy release theory, the transformation of rock mass from microscopic damage to macroscopic fracture requires an energy accumulation process of 80-120 seconds. Within this window, the system can track the acceleration of the overall energy release.

[0079] For each time window, statistical characteristics of the event time interval, energy, and event density are calculated, including the mean (mean) to reflect the average level of the parameters within the window, the standard deviation (std) to reflect the degree of parameter fluctuation, the maximum value (max) to capture extreme events, the minimum value (min) to identify the baseline level of acoustic emission activity, and the rate of change (rate) to represent the rate of change of parameters between adjacent windows.

[0080] The above statistical features are combined into a multi-scale time series feature vector, and the specific construction method is as follows:

[0081] ;

[0082] The eigenvectors at each scale are defined as follows:

[0083] ;

[0084] ;

[0085] ;

[0086] in This represents the mean. Indicates standard deviation, Indicates the maximum value. This represents the minimum value. This represents the rate of change between adjacent windows. This represents the average energy of events within that time window, reflecting the average energy release rate of the rock mass. The standard deviation of energy is used to characterize the degree of fluctuation in energy release. During the stable phase of a rock mass, energy fluctuations are relatively small. Typically below 10³; as damage evolves, energy fluctuations increase. It rises to 104-105; in the critical stage of instability, energy release is violent. It can reach over 106, and its growth pattern directly reflects the accelerated process of damage accumulation inside the rock mass.

[0087] It represents the maximum amplitude of the event within the time window, reflecting the instantaneous intensity of the strongest pulse within the window, and is a key indicator for quickly identifying sudden rupture events; The amplitude represents the average fracture intensity of events within the window. The combination of amplitude and energy can more accurately distinguish different rock mass states: high amplitude + high energy indicates strong fracture (critical instability state), while low amplitude + low energy indicates weak acoustic emission (stable state). The synergistic change of these two parameters enhances the model's ability to discriminate rock mass states.

[0088] This represents the average duration of an event, used to distinguish event types. Pulsating short events (release from a single microcrack) typically last less than 10 milliseconds, while complex fracturing events (caused by the cascading convergence of multiple adjacent cracks) can last 50-200 milliseconds. `std(d_i)` represents the standard deviation of the duration, reflecting the difference in event complexity within the window. In the steady-state phase, event types are relatively simple, and the duration distribution is more concentrated. The value is relatively small; in the critical stage of instability, events are complex and diverse, with a dispersed distribution of duration. The value increased significantly, and this characteristic change effectively distinguishes different stages of rock mass evolution.

[0089] Ring counting characteristic parameters: This represents the average ring count of events within the given time window, used to quantify the complexity of the event. The ring count for each acoustic emission event reflects the number of times the waveform exceeds a threshold, and is an important indicator for distinguishing between pulse-type events and complex rupture events. Low ring counts (1-2 times) indicate a single pulse-type event, corresponding to the rapid release of simple microcracks; high ring counts (5-10 times or more) indicate complex events that exceed the threshold multiple times, reflecting the hierarchical convergence and interaction of multiple cracks.

[0090] At different stages of rock mass instability evolution, the average ring count shows a clear increasing trend: in the stable stage, Typically below 2, the event type is simple and straightforward; during the damage evolution stage, An increase to the 2-5 range reflects the gradual formation of a microcrack network, making individual events more complex; in the critical stage of instability, A value exceeding 5, or even reaching 10 or higher, indicates an extremely complex rupture mode, marking a critical state of macroscopic destruction. Through analysis of... Monitoring can effectively distinguish between the characteristic changes caused by the "increased number" of events and the "increased complexity of individual events" in the changes of rock mass state, and enhance the model's ability to describe the detailed evolution process of fracture modes.

[0091] By fusing multi-scale features, it can respond quickly to sudden instability (based on a short time window) and track gradual damage over a long period of time (based on a long time window), while using a medium time window to balance real-time performance and stability.

[0092] For example, in a uniaxial compression test of a granite sample, at the initial stage of loading, when the stress ratio... Short-time window event density With an event rate of 0.3 events / second and a long-term window mean change rate of less than 5%, the system is identified as being in a stable phase. During the crack propagation phase, when the stress ratio is 0.7, the short-term window... When the energy level rises to 1.5 events / second, the average energy over the long-term window increases by 30%, indicating a damage evolution stage in the system. Before peak failure, when the stress ratio is 0.95, the short-term window... The frequency surged to 8.2 events / second, the event interval decreased to 0.08 seconds, and the maximum energy over a long window reached [a certain value]. aJ, the system has triggered a critical instability warning.

[0093] The time-series feature vectors constructed using the above method are correlated and labeled with the rock mass stability state. For indoor rock sample test data, stages are defined based on the stress-strain curves. The labeling rules are based on the physical stages of rock fracture:

[0094] Stable phase (stress ratio) The first stage corresponds to the compaction and elastic stages. Primary cracks within the rock mass close, elastic deformation is dominant, acoustic emission is minimal, and energy release is extremely low. This stage is designated as the "stable stage." 0.5 is the upper limit of elasticity for the safe service of the rock mass. Beyond this point, primary cracks begin to propagate on a large scale, marking the transition from "controllable elasticity" to "microscopic fracturing."

[0095] Damage evolution stages ( This corresponds to the crack propagation stage, where new cracks begin to initiate, propagate, and connect with each other within the rock mass. The frequency and energy of acoustic emission activity increase significantly, and the rock mass's bearing capacity begins to decline but has not yet become unstable. This stage is labeled the "damage evolution stage." A stress ratio of 0.9 is the critical point at which the rock mass transitions from recoverable damage to irreversible instability.

[0096] Instability Critical Stage ( This corresponds to the post-peak failure stage, where macroscopic cracks have formed inside the rock mass, the bearing capacity decreases rapidly, and macroscopic instability and failure may occur at any time. Acoustic emission activity is intense and the energy is high. This stage is marked as the "critical instability stage".

[0097] For on-site monitoring data, retrospective labeling is performed based on the actual time of failure. The period from 0 to 10 minutes before failure is labeled the "critical instability stage." During this extremely short period before macroscopic instability of the rock mass, acoustic emission activity reaches its peak, with extremely high event density and energy. The system is already in an irreversible unstable state, and this time window provides a final early warning for emergency rescue operations. The period from 10 to 60 minutes before failure is labeled the "damage evolution stage." Before macroscopic instability, internal damage to the rock mass has accumulated to a certain extent, crack networks gradually form, and acoustic emission signals show an accelerating trend, but have not yet reached the final instability peak. The remaining time periods are labeled the "stable stage." Data from periods far from the failure occurrence usually reflects the normal working state or slow creep of the rock mass, with stable acoustic emission activity.

[0098] By using the above annotation rules, a mapping relationship between time series feature vectors and the physical state of rock mass is established, forming training and validation sample sets for subsequent training and validation of the state recognition model. The innovation of this annotation system lies in corresponding acoustic emission signals with rock mass mechanical behavior, forming a physically meaningful structured sample set, ensuring the correspondence between the features learned by the model and the actual stable state of the rock mass.

[0099] S4. After preprocessing and sample labeling, the acoustic emission signal is input into a deep neural network model for training. The deep neural network model can adopt different structural forms according to task requirements and data characteristics, including one-dimensional convolutional neural network (1D-CNN), long short-term memory network (LSTM), convolutional and recurrent network combined model (CNN-LSTM), or sequence model based on Transformer.

[0100] The theoretical basis for model selection: Acoustic emission time series exhibits both local spatial characteristics and long-term dependencies. Local spatial characteristics refer to the strong correlation between features at adjacent time points, which CNNs excel at extracting. Long-term dependencies refer to the causal relationship between the current rock mass state and its state 30-120 seconds ago, which LSTMs excel at capturing. Using CNNs alone is prone to misjudgments in the initial stage, while using LSTMs alone cannot fully extract local features. Therefore, the CNN-LSTM hybrid model can fully utilize the advantages of both networks. The convolutional part is responsible for extracting local features (energy mutations, envelope morphology, etc.), while the recurrent structure captures the temporal dependencies of the waveform, enabling the model to identify the evolution trend of precursory signals of rock mass instability.

[0101] A one-dimensional convolutional neural network is used to process the normalized acoustic emission time series signal. This model automatically extracts local waveform features of the signal, such as energy abrupt changes, envelope morphology, and event duration, through sliding computation across multiple convolutional kernels in the time dimension. Pooling layers and fully connected layers follow the convolutional layers to achieve multi-scale feature aggregation and classification output, thus completing the feature learning of the acoustic emission signal. Through the training process of this model, the characteristic distribution patterns of acoustic emission signals under different rock mass stability states can be automatically learned, achieving a continuous mapping from acoustic emission waveform features to rock mass stability determination. Compared with traditional methods that rely on manual feature extraction, this reduces the need for manual parameter setting, making the feature learning process more complete and adaptable to the characteristics of acoustic emission signals under different geological conditions and monitoring environments, providing a reliable model foundation for subsequent real-time identification and early warning.

[0102] Pre-training phase: Before being deployed to the on-site monitoring system, the pre-trained state recognition model needs to undergo an offline training phase to obtain initial model parameters. The data sources used for pre-training include three categories:

[0103] Indoor rock sample loading test data: Standard samples of various lithologies, including granite, sandstone, and marble, were subjected to uniaxial compression, triaxial compression, or Brazilian splitting tests. Acoustic emission signals were simultaneously acquired during the loading process, and the signals at each stage were divided into stable stage, damage evolution stage, and instability critical stage based on the stress-strain curve. Taking the uniaxial compression test of granite as an example, the compaction and elastic stages ( The corresponding acoustic emission events have low density and low energy, and are labeled as the "stable stage"; the crack propagation stage ( The corresponding acoustic emission event density increases significantly and energy continues to grow, marked as the "damage evolution stage"; the post-peak destruction stage ( The corresponding acoustic emission event density surges and energy is released in a concentrated manner, which is marked as the "instability critical stage".

[0104] Existing engineering monitoring data: Collect acoustic emission data from long-term monitoring records of mine roadways, tunnel surrounding rock, slope engineering, etc., and retrospectively label them in conjunction with actual rock mass failure events. That is, using the known failure time point as a benchmark, the acoustic emission data in the period before failure is labeled as the instability critical stage, and the earlier signals are labeled as the stable stage or damage evolution stage according to the event density and energy change trends.

[0105] Data augmentation samples: When the actual amount of collected data is insufficient, data augmentation can be performed by adding Gaussian noise to existing samples, performing time stretching or compression, applying amplitude scaling, etc., in order to expand the number of training samples and improve the robustness of the model to changes in signal characteristics under different acquisition conditions.

[0106] Pre-training process:

[0107] Step 1: Data Preparation. Preprocess the acoustic emission data from the three sources mentioned above according to step S2, construct multi-scale time series vectors according to step S3, and pair them with the corresponding rock mass stability state labels to form a structured training dataset. Randomly divide the dataset into training and validation sets at an 8:2 ratio.

[0108] Step 2: Model Initialization. Based on the preset network structure (such as a one-dimensional convolutional neural network, LSTM, CNN-LSTM, or Transformer), set the model parameters using random initialization or preset weight initialization methods.

[0109] Step 3: Iterative Training. Input the multi-scale time series vectors from the training set into the model batch by batch. The model outputs a three-class probability vector (probability of stable phase, probability of damage evolution phase, and probability of instability critical phase). Calculate the difference between the predicted output and the true label using the cross-entropy loss function or Focal Loss function, and update the network weights using the backpropagation algorithm.

[0110] Step 4: Validation and Tuning. After each training round, the model's classification accuracy and loss function value are evaluated using the validation set. During training, a learning rate decay strategy (e.g., the initial learning rate is set to 0.001, and it decays to 0.5 times the original rate every 20 epochs) and an early stopping mechanism (training is stopped if the validation set loss does not improve for 10 consecutive epochs) are employed to prevent the model from overfitting.

[0111] Step 5: Model Consolidation. After training is complete, save the model parameters and network configuration file corresponding to the optimal validation accuracy to generate a deployable pre-trained state recognition model for subsequent rock mass state recognition in step S5.

[0112] Specific parameter configuration: In one specific implementation, the pre-training parameter configuration is as follows: A CNN-LSTM hybrid network structure is adopted, wherein there are 3 convolutional layers with a kernel size of 5, and the number of filters in each layer is 32, 64, and 128 respectively; there are 2 LSTM layers with 128 hidden units; the number of neurons in the fully connected output layer is 3, corresponding to the three rock mass stability states; the activation function is ReLU in the intermediate layers and Softmax in the output layer. The number of convolutional layers and filters is based on the progressive feature extraction theory. The first layer (32 filters) extracts low-level features (energy mutation, basic waveform), the second layer (64 filters) extracts mid-level features (feature combination, event pattern), and the third layer (128 filters) extracts high-level features (stage features, damage pattern). More than 3 layers will lead to overparameterization and overfitting under limited data; the multi-scale feature vector dimension design is as follows: short-time window , medium time window Long window Each feature contains 13 dimensions (statistical features of time interval, energy, density, amplitude, duration, and ring count), forming a 39-dimensional basic multi-scale time series vector. Through further feature fusion and nonlinear transformation, it is extended to a high-dimensional feature space of 90-99 dimensions in the convolutional network.

[0113] The 128 hidden units in the LSTM are set based on a multiple of the dimensionality of the convolutional output features, which can fully accommodate the expanded feature information and establish "shallow time dependencies" (between two computation cycles) and "deep time dependencies" (states from 30-60 seconds ago). The 128 hidden units correspond to 128 memory cells in the LSTM unit, sufficient to transmit long-term dependencies spanning multiple time steps (30-120 seconds corresponds to 30-40 time steps).

[0114] Training parameters: Batch size is set to 64, initial learning rate is set to 0.001, the optimizer is Adam, the learning rate decay strategy is StepLR (decay by 50% every 20 epochs), the maximum number of training epochs is set to 200 epochs, and the early stopping patience parameter is set to 10.

[0115] The selection criteria for the learning rate parameter are as follows: The initial learning rate of 0.001 is based on the analysis of 50 training cycles. This learning rate results in the fastest convergence (10-15 epochs), while 0.0005 results in slower convergence (20-25 epochs) and is prone to oscillation. The learning rate decay strategy (decreasing by 50% every 20 epochs) is based on the fact that a larger learning rate is needed for the system to quickly learn the main features in the first 20 epochs. After that, in the refinement stage, the learning rate needs to be reduced to obtain more refined parameter adjustments. A decay coefficient of 0.5 is the most reasonable compared to 0.3 (too fast) or 0.7 (too slow).

[0116] Loss function: adopted The weights for the critical instability stage are set relatively high to alleviate the sample imbalance problem.

[0117] Model training method: Supervised learning is adopted, that is, the multi-scale time series feature vector constructed in step S3 and its corresponding state labels are provided at the input end, and the stable state of the rock mass (stable stage, damage evolution stage, and critical instability stage) is used as a supervision signal at the output end to guide the iterative update of the model parameters. The model training uses the backpropagation algorithm to adjust the parameters, so that the difference between the predicted output and the true label is gradually reduced.

[0118] To measure the difference between the model output and the actual labels, the cross-entropy loss function is used. For optimization purposes, the loss function for classification tasks is defined as follows:

[0119] ;

[0120] in The total number of samples, The number of classification categories in this invention =3, corresponding to the stable stage, the damage evolution stage, and the critical instability stage, respectively. For the sample In category The true label on the sample (when the sample Category hour ,otherwise ), Samples predicted by the model Category The probability satisfies .

[0121] During training, optimization algorithms (such as Adam or SGD) minimize The model parameters are updated with values ​​to ensure the predicted results are as consistent as possible with the label distribution. During model training, cross-validation and early stopping conditions can be set to prevent overfitting. Cross-validation involves dividing the data into several folds and repeatedly training and validating under different combinations to evaluate the model's generalization performance. Early stopping terminates training when the validation set loss function does not decrease in several consecutive iterations, thus preventing overfitting on the training set. For example, in a typical training task, approximately 80% of the samples can be used as the training set, and 20% as the validation set. During training, if the validation set loss function does not improve in 10 consecutive iterations, the early stopping mechanism is automatically triggered. Through this supervised learning and optimization strategy, the model can converge quickly under limited data conditions, obtaining a parameter set that can stably distinguish different rock mass states, providing a reliable identification model for subsequent real-time identification and risk assessment steps.

[0122] S5. Based on the judgment probability output in step S4, determine the current stable state of the rock mass, and analyze the evolution trend of the rock mass state by combining the judgment results of multiple consecutive time windows. During the rock mass monitoring process, the multi-scale time series vectors constructed in real time are input into the trained state recognition model to obtain real-time recognition results corresponding to different rock mass stable states.

[0123] The state evolution analysis and real-time inference steps involve inputting the multi-scale time series vectors constructed in real time into a trained state recognition model to achieve intelligent determination of the rock mass's stability state and analysis of its evolution trend. First, following steps S2-S3, the system preprocesses the real-time acquired acoustic emission signals, performs event identification, parameter extraction, and multi-scale feature construction to generate a time series vector. Then, this vector is input into the state recognition model for forward computation to extract deep features and output the probability of determining whether the rock mass is in a stable stage, a damage evolution stage, or a critical instability stage.

[0124] In one specific implementation, the decision probability vector output by the model is defined as:

[0125] ;

[0126] in , , These represent the probabilities of the rock mass being in the stable stage, the damage evolution stage, and the critical instability stage, respectively, satisfying the following conditions: .

[0127] Current rock mass condition category Determined by the maximum probability criterion:

[0128] ;

[0129] To quantitatively analyze the evolution trend of rock mass conditions, the system records the judgment results for K consecutive time windows and counts the frequency of each condition. , , And define the state evolution index. The calculation formula is:

[0130] ;

[0131] in The number of time windows for continuous observation. , , They are respectively The frequency of occurrence of the stable phase, damage evolution phase, and instability critical phase within each window satisfies , The value range is [1, 3]. The larger the value, the more unstable the rock mass is.

[0132] The above The value is used to quantitatively assess the acceleration of rock mass damage evolution, and its value is positively correlated with the degree of rock mass failure risk. The weighting coefficient reflects the difference in the degree of instability risk corresponding to each stage.

[0133] The steady-state phase has a weight of 1, which is defined as the baseline state, corresponding to an extremely low energy release rate (<1 mJ / s) and the lowest probability of instability.

[0134] The damage evolution stage weight is 2, corresponding to the crack network formation process, where the energy release rate increases to 1-100 mJ / s, which is 10-100 times that of the baseline stage. However, since the crack network has not yet been fully connected and there is still an energy dissipation mechanism, the actual risk weight is about twice that of the baseline.

[0135] The weight of the instability critical stage is 3, which corresponds to the macroscopic crack propagation and concentrated energy release (>100 mJ / s), which is more than 100 times that of the baseline stage. However, since this is the final "controllable" stage, the weight is set to 3 (a multiple of the relatively stable stage).

[0136] The criterion for determining evolutionary trends is: when The risk level is low when The risk level is medium risk. The risk level is high.

[0137] Specifically, when When < 1.3, it is judged as "stable evolution", indicating that the rock mass state remains basically stable; when When this is determined to be "evolution towards damage," it indicates that the rock mass is transitioning from stability to damage; when When this occurs, it is determined to be "accelerated instability," indicating that the rock mass is rapidly evolving towards a critical state of instability.

[0138] The threshold selection is based on statistical analysis of a large amount of rock mass test data. Analysis of 100 sets of uniaxial compression test data for granite, sandstone, and marble shows that at a stress ratio of 0.5, the system output... The average value is Therefore, set =1.3 is the dividing line; when the stress ratio is 0.9, the average E value of the system output is Therefore, E=2.0 is set as the dividing line. These thresholds ensure that the system can identify critical state transition points in a timely manner.

[0139] For example, during a continuous monitoring process, the system outputs a judgment probability of 1 / 2 at a certain moment. = 0.15、 =0.65、 = 0.20, then according to the maximum probability criterion, the system determines that the current rock mass is in the "damage evolution stage"; if the damage evolution stage occurs 6 times and the instability critical stage occurs 4 times in the subsequent 10 time windows, then the evolution index... The system detected "instability and acceleration" and immediately prompted the need to strengthen monitoring and prepare for emergency response.

[0140] If the probability of determining the "critical stage of instability" continues to increase over several consecutive time windows, the system will trigger an upper-level early warning mechanism, thereby enabling continuous intelligent identification and evolution trend tracking of the rock mass state. This will keep the model output synchronized with the acoustic emission time series characteristics, establish a continuous risk identification chain, and provide real-time, data-driven decision-making basis for subsequent risk level determination.

[0141] S6. Determine the risk level of the rock mass based on the identification results and output the corresponding early warning information. (Based on evolution index) Define a quantitative criterion for determining whether a "trend of transitioning to stage XX" exists:

[0142] Undirected damage evolution stage shift trend: when When the value is less than 1.3, it is classified as "stable evolution," indicating that the rock mass has remained in or is close to a stable stage for multiple consecutive time windows, with no obvious signs of accelerated damage. At this point, the current state is determined to be in a stable stage, and the risk level is low.

[0143] There is a tendency to shift towards the damage evolution stage: when When this occurs, it is classified as "accelerated damage," indicating that although the rock mass may currently be in a stable phase or have entered a damage evolution phase, significant damage evolution signals have appeared within a continuous time window, and the frequency of damage occurrences has increased significantly. At this point, the risk level is medium risk. For example, in... =In 10 consecutive time windows, if 4 windows are identified as the damage evolution stage and 6 windows as the stable stage, then ,satisfy It was determined to be of medium risk.

[0144] There is a tendency to shift towards the critical stage of instability: when When this occurs, it is classified as a "critical warning," indicating that the rock mass has frequently reached the critical stage of instability or the stage of high-proportion damage evolution within a continuous time window, and the rock mass is rapidly approaching instability. At this time, the risk level is high, and the system triggers an emergency warning. For example, in... If, within a continuous time window, 3 windows are identified as the critical stage of instability, 5 as the damage evolution stage, and 2 as the stable stage, then... ,satisfy The risk level was determined to be high, triggering an emergency warning.

[0145] In addition, the system also monitors evolutionary indices. rate of change:

[0146] ;

[0147] This threshold reflects the rate of change of the state evolution index over two computational cycles. In the analysis of historical data from 100 rock mass fracturing tests, during the stable evolution period... The range of variation is During the damage evolution period During the period of unstable acceleration, it can reach The warning threshold of 0.3 was chosen because it corresponds to the transition zone where damage evolution accelerates towards instability: when When the value first exceeds 0.3, damage usually occurs within 10-30 minutes (corresponding to 2-6 30-second calculation cycles), allowing sufficient time for monitoring and emergency response.

[0148] when Even at present Even if the value has not yet reached the threshold of the next level, the system will still trigger a "trend acceleration" warning, prompting monitoring personnel to pay attention to the rapid deterioration of the rock mass condition.

[0149] In a verification example of a granite uniaxial compression test, during the initial loading stage ( evolution index The system is stable around version 1.0 and judged as low risk; when loaded to hour, The value rose to 1.6, the system determined it to be medium risk and issued a warning; when loaded to When = 0.95, The value jumped to 2.5, and = 0.9, the system judged it as high risk and triggered an emergency warning. The actual rock sample was in The occurrence of macroscopic damage verifies the effectiveness of the aforementioned evolutionary trend judgment criteria.

[0150] S7. When the number of new samples collected during the monitoring process reaches the preset condition, the new samples will be used for incremental training of the model.

[0151] The characteristics of detected acoustic emission signals may shift due to changes in rock mass properties, stress environment adjustments, or differences in sensor conditions. Therefore, when the accumulated number of new samples reaches a preset threshold, an incremental training process for the model is triggered. This involves automatically recording validated acoustic emission waveform data and their corresponding rock mass condition results during daily operation and storing them in an updated sample library. When the number of new samples in the library reaches a set threshold, the original model structure is invoked, and incremental training is performed using the new sample data to update some model parameters. This can be executed on edge computing devices or servers. Edge devices are suitable for real-time updates in the field, while servers are suitable for centralized computing or collaborative updates across multiple monitoring points. For example, in mining slope monitoring applications, after continuously collecting and confirming more than one thousand new acoustic emission samples, an incremental training task can be automatically initiated. After the update is completed, the new model parameters are replaced in the operating system to continue real-time identification and early warning, allowing the model to continuously adapt to signal characteristics under different time periods and geological conditions, thereby improving the continuous reliability and versatility of the rock mass instability early warning system.

[0152] Please see the appendix Figure 2 A multi-scale acoustic emission and deep learning-integrated rock mass instability early warning system is proposed to implement the aforementioned early warning method. The system includes an acoustic emission signal acquisition module, a signal preprocessing module, a time series construction module, a state identification module, a state evolution analysis module, a risk determination module, and an adaptive update module.

[0153] The time series construction module is used to construct multi-scale time series feature vectors within at least three different time windows. The state evolution analysis module is used to record continuous... The rock mass state was measured for each time window, and the state evolution index E and the rate of change of the evolution index were calculated. The risk assessment module is used for E-based thresholds and... The system triggers risk warnings and trend acceleration alerts. The adaptive update module performs incremental training and updates model parameters when the number of new samples reaches a preset threshold.

[0154] The acoustic emission signal acquisition module is used to acquire acoustic emission waveform signals generated during the stress deformation process of the rock mass through high-sensitivity acoustic emission sensors deployed on or inside the rock mass. The signal preprocessing module is used to filter, denoise, and perform event trigger detection on the acoustic emission waveform signals to obtain event-level waveform data.

[0155] The time series construction module further includes an event identification unit, a parameter extraction unit, and a multi-scale feature calculation unit. The event identification unit is used to identify the start and end times of events based on short-time energy detection or the STA / LTA algorithm. The parameter extraction unit extracts parameters such as the event occurrence time, energy, amplitude, duration, and ring count. The multi-scale feature calculation unit is used to calculate the statistical features of the event time interval, energy, and density under different time windows and combine them into a multi-scale time series vector.

[0156] The state recognition module includes a feature extraction unit, a state classification unit, and a model storage unit. The feature extraction unit performs deep feature extraction on the input multi-scale time series vector. The state classification unit outputs three-class probability vectors, corresponding to the determination probabilities of the stable stage, the damage evolution stage, and the instability critical stage, respectively. The model storage unit stores the trained model parameters and configuration.

[0157] The state evolution analysis module is used to determine the current stable state of the rock mass based on the judgment probability, and analyze the evolution trend of the rock mass state by combining the judgment results of multiple consecutive time windows. The risk judgment module is used to determine the risk level of the rock mass based on the rock mass's stable state and evolution trend, and trigger corresponding early warning signals.

[0158] The adaptive update module includes a feedback data collection unit, a sample relabeling unit, and an incremental learning unit. The feedback data collection unit is used to collect the actual damage status of the rock mass on site and the results of manual annotation. The sample relabeling unit is used to annotate the status of historical data according to the actual damage status. The incremental learning unit is used to fine-tune the parameters of the status recognition model based on the newly annotated samples. The incremental update can be executed on an edge computing device or a server.

[0159] The system also includes a data storage module for storing historical acoustic emission signals, time series characteristics, identification results and actual damage conditions; a visualization display module for displaying the rock mass stability state, risk level and state evolution curve in real time; and a remote communication module for outputting early warning information to the monitoring center or mobile terminal.

[0160] The specific workflow is as follows: First, a high-sensitivity acoustic emission sensor is deployed on the rock surface and at different depths using an acoustic emission signal acquisition module to collect the raw acoustic emission waveform signals generated by the rock mass under stress in real time. The acquired data is processed by a signal preprocessing module, which sequentially performs filtering and noise reduction, and event trigger detection to obtain event-level waveform data. Subsequently, a time series construction module analyzes the event-level waveform data, identifies multiple consecutive acoustic emission events, extracts event parameters (time interval, energy value, amplitude, duration, ring count, etc.), and constructs time series feature vectors at multiple time scales. The state recognition module loads a pre-trained state recognition model, performs feature extraction and state classification on the input multi-scale time series vectors, and outputs the probability of determining whether the rock mass is in a stable stage, a damage evolution stage, or a critical instability stage. The state evolution analysis module determines the current stable state based on the determination probability and analyzes the evolution trend by combining the determination results of multiple consecutive time windows. The risk determination module receives the identification results and evolution trend information, determines the risk level of the rock mass, and generates a corresponding early warning signal. During system monitoring, the adaptive update module automatically determines whether incremental training is needed based on the number of newly collected samples. When preset conditions are met, the model parameters are updated to achieve dynamic model adaptation.

[0161] In a preferred embodiment, the system may further include a data storage module, a visualization display module, and a remote communication module. The data storage module stores historical data, the visualization display module visually displays the rock mass risk level and monitoring status on a terminal interface, and the remote communication module transmits early warning information to a monitoring center or mobile terminal.

[0162] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A rock mass instability early warning method integrating multi-scale acoustic emission and deep learning, characterized in that, Includes the following steps: S1. Acoustic emission waveform signals are collected by acoustic emission sensors deployed on or inside the rock mass; S2. Perform noise reduction, filtering and normalization preprocessing on the acoustic emission waveform signal, and perform event-level segmentation on the signal based on the acoustic emission event triggering mechanism to obtain acoustic emission event waveform data; S3. For each acoustic emission event, extract parameters including the time of occurrence, energy value, amplitude, duration, and ring count. Based on multiple consecutive acoustic emission events, the event time interval sequence is calculated within at least three time windows of different time scales, and the number of events within each unit time window is counted to obtain the event density; S4. Within each time window, calculate at least one statistical feature among the event time interval sequence, event energy value, and event density. The statistical feature is selected from a group consisting of mean, standard deviation, maximum value, minimum value, and rate of change. Concatenate the statistical features of each time window to construct a multi-scale time series feature vector. S5. Input the multi-scale time series feature vector into a deep learning state recognition model that can automatically extract the time series features of acoustic emission signals and perform multi-classification. The model is trained by supervised learning based on indoor rock sample test data and field monitoring data to obtain the judgment probability that the rock mass is in a stable stage, a damage evolution stage and an instability critical stage, and to determine the rock mass state of the current time window according to the maximum probability criterion. S6, Record Continuous The rock mass condition within a time window, among which The value is a preset positive integer; the number of consecutive integers is counted. Within each time window, the frequencies at which the rock mass was identified as being in the stable stage, the damage evolution stage, and the critical instability stage were respectively: , , ,and ; Calculate the state evolution index , ; S7. Based on the current rock mass state and the aforementioned state evolution index Determine the risk level of the rock mass and output early warning information; the risk level includes at least low risk, medium risk, and high risk. respectively with the first threshold With the second threshold The risk level is determined by comparison, wherein ,when The risk level is low when The risk level is medium risk. The risk level is high. S8. During the monitoring process, when the number of newly collected samples reaches the preset conditions, the state recognition model is incrementally updated based on the actual damage of the rock mass on site or the results of manual annotation.

2. The method according to claim 1, characterized in that: The at least three time windows with different time scales include a short-term window, a medium-term window, and a long-term window; wherein, the short-term window is 5 seconds and is used to capture instantaneous sudden events and rapid crack propagation; the medium-term window is 30 seconds and is used to reflect the short-term damage accumulation process; and the long-term window is 120 seconds and is used to reveal long-term evolution trends.

3. The method according to claim 1, characterized in that: The state recognition model is a deep learning model that can automatically extract the temporal features of acoustic emission signals and perform multi-classification. The deep learning model is a convolutional neural network, a recurrent neural network, or a hybrid structure thereof.

4. The method according to claim 1, characterized in that: It also includes step S6a: calculating the state evolution index. Rate of change within adjacent calculation periods , ,in This is the state evolution index for the current calculation period. The state evolution index of the previous calculation period; when Greater than the third threshold At that time, a trend acceleration warning is triggered, indicating that the rock mass condition is rapidly deteriorating.

5. The method according to claim 1, characterized in that: In step S5, when training the deep learning state recognition model, the annotation rules for training samples are as follows: For indoor rock sample loading test data, stages are divided according to the stress-strain curve, and the stress ratio reflects the stress stage of the rock mass. When the stress ratio... The time is marked as "stable phase" when The time is marked as "damage evolution stage", when The time was marked as "critical stage of instability"; For on-site monitoring data, retrospective labeling is performed based on the actual time of damage. The first 0-10 minutes before damage is labeled as the "critical instability stage", the first 10-60 minutes before damage is labeled as the "damage evolution stage", and the remaining time period is labeled as the "stable stage".

6. A rock mass instability early warning system integrating multi-scale acoustic emission and deep learning, characterized in that, The system for implementing the method of any one of claims 1-5 comprises: an acoustic emission signal acquisition module, a signal preprocessing module, a time series construction module, a state identification module, a state evolution analysis module, a risk determination module, and an adaptive update module; wherein the time series construction module is used to construct a multi-scale time series feature vector within at least three time windows of different scales; the state evolution analysis module is used to record the rock mass state for K consecutive time windows and calculate the state evolution index. With the rate of change of the evolutionary index The risk assessment module is used to determine the risk level based on... threshold and The trigger conditions output risk warnings and trend acceleration warnings; the adaptive update module is used to perform incremental training and update model parameters when the number of new samples reaches a preset threshold.

7. The system according to claim 6, characterized in that: The time series construction module further includes an event identification unit, a parameter extraction unit, and a multi-scale feature calculation unit. The event identification unit is used to identify the start and end times of events based on short-time energy detection or STA / LTA algorithms. The parameter extraction unit extracts parameters including the event occurrence time, energy, amplitude, duration, and ringing count.

8. The system according to claim 6, characterized in that: The state recognition module includes a feature extraction unit, a state classification unit, and a model storage unit. The feature extraction unit performs deep feature extraction on the input multi-scale time series vector. The state classification unit outputs a three-class probability vector. The model storage unit stores the trained model parameters and configuration.

9. The system according to claim 6, characterized in that: The adaptive update module includes: a feedback data collection unit, a sample relabeling unit, and an incremental learning unit. The incremental update can be executed on an edge computing device or a server.

10. The system according to claim 6, characterized in that: The data storage module stores historical acoustic emission signals, time-series characteristics, identification results, and actual damage conditions; the visualization module displays the rock mass stability, risk level, and state evolution curve in real time; and the remote communication module outputs early warning information to the monitoring center or mobile terminal.

Images