Rockfall precursor identification method based on acoustic vibration signal and self-supervised representation learning

Abstract

The application relates to the technical field of geological disaster monitoring and early warning, and discloses a dangerous rock collapse precursor identification method based on acoustic vibration signals and self-supervised feature learning. First, distributed acoustic vibration monitoring data of a target dangerous rock body within a preset time length is acquired, filtering processing is performed, and segmentation is performed in a preset fixed time length sliding window to generate multiple signal segments. Each signal segment is subjected to standardization processing to generate a standardized feature vector, which is input into a pre-trained neural network model to output probabilities that each signal segment respectively belongs to a normal state, a collapse precursor state and a collapse state. A preset number of signal segments are used as an early warning window, the average value of the precursor state probabilities corresponding to each signal segment in the early warning window is calculated, and the proportion of the signal segments in the early warning window whose precursor state probabilities exceed a preset basic threshold value is calculated. According to the average value and the proportion, comparison with a preset early warning triggering threshold value is respectively performed to generate an early warning result.

Description

Technical Field

[0001] This application relates to the field of geological disaster monitoring and early warning technology, and in particular to a method for identifying precursors of rockfall based on acoustic and vibration signals and self-supervised representation learning. Background Technology

[0002] Rockfalls are a common type of geological hazard in mountainous areas, mining areas, and along transportation routes. They are characterized by their suddenness, wide destructive range, and extremely rapid destructive speed, seriously threatening the lives and property of the people and the operational safety of major infrastructure. With the continuous advancement of infrastructure construction, the number of potential rockfall hazards along mountain roads, in mining areas, on the banks of reservoirs, and in the cliff areas of scenic spots is increasing, making the need for monitoring and early warning of rockfall disasters increasingly urgent.

[0003] Currently, the identification of precursors to rockfalls mainly relies on the following technical solutions. The first category is manual inspection and simple sensor monitoring. This involves periodically checking the surface of the rock mass for apparent changes such as crack development and displacement deformation, or deploying isolated simple sensors like surface displacement and crack gauges for fixed-point observation. However, this method relies on the rock mass already undergoing deformation and often only issues an alarm when a collapse is imminent, resulting in extremely limited warning time. The second category is single-mode acoustic emission or vibration signal monitoring. This involves deploying acoustic emission sensors or accelerometers at key locations on the rock mass to collect single-mode signals and make warning judgments based on set empirical thresholds. Acoustic emission technology can reflect the micro-fracture process inside the rock and has a certain application foundation in the field of rockfall monitoring. However, single-mode signals are easily interfered with by environmental noise, have limited feature dimensions, and are difficult to comprehensively characterize the complex evolution process of a rock mass from stability to instability. The third category is identification methods based on supervised machine learning. This involves collecting acoustic and vibration signals, manually labeling state samples, and training a classification model for state identification.

[0004] However, the aforementioned existing technical solutions all have significant limitations. Manual inspections cannot achieve continuous, 24 / 7 monitoring, making it difficult to capture early, subtle precursor signals of rockfalls. Single-modal monitoring cannot effectively distinguish between environmental interference such as wind noise and vehicle vibration and actual crack propagation signals, resulting in high false alarm and missed alarm rates. More importantly, rockfalls are low-frequency disasters, making it extremely difficult to obtain a large amount of labeled data containing precursor stages. This leads to a severe shortage of training samples for supervised machine learning models, making them prone to overfitting and exhibiting poor generalization ability. Therefore, there is an urgent need for an automatic rockfall precursor identification method that can overcome the bottleneck of scarce labeled samples and integrate multi-source acoustic and vibration signals. Summary of the Invention

[0005] (a) Technical problems to be solved In view of the above-mentioned shortcomings and deficiencies of the prior art, this application provides a method for identifying early signs of rockfall based on acoustic and vibration signals and self-supervised representation learning, in order to solve the following problems in the prior art: (1) it is difficult for manual inspection and single-point sensor monitoring to capture early micro-amplitude early signs of rockfall; (2) rockfall is a low-frequency disaster, and early signs of rockfall are extremely scarce, making it difficult to train supervised models with manual annotation and having weak generalization ability; (3) single-mode acoustic and vibration monitoring has limited feature dimensions, which easily leads to misjudging environmental interference such as wind noise and mechanical vibration as early signs of rockfall, resulting in high false alarm and false alarm rates.

[0006] (II) Technical Solution To achieve the above objectives, the main technical solutions adopted in this application include: This application provides a method for identifying precursors of rockfalls based on acoustic vibration signals and self-supervised representation learning. The specific steps include: S1. Obtain distributed acoustic and vibration monitoring data of the target unstable rock mass within a preset time period, filter the distributed acoustic and vibration monitoring data, and divide it into multiple signal segments using a preset fixed-time sliding window. S2, extract time-domain features, frequency-domain features, and time-frequency-domain features for each signal segment, and perform splicing and standardization to generate a standardized feature vector; S3, input the standardized feature vectors corresponding to multiple signal segments into a pre-trained neural network model, and output the probability that each signal segment belongs to the normal state, the pre-collapse state, and the collapse state, respectively; wherein, the neural network model is obtained in advance through a self-supervised representation learning training method; S4. Using a preset number of signal segments as the early warning window, calculate the average probability of the collapse precursor state corresponding to each signal segment within the early warning window, and the proportion of signal segments within the early warning window whose precursor state probability exceeds a preset basic threshold; compare the average value and the proportion with the preset early warning trigger threshold respectively to generate an early warning result.

[0007] Optionally, in some embodiments of this application, obtaining distributed acoustic and vibration monitoring data of the target unstable rock mass within a preset time period specifically includes: Based on the pre-obtained geological survey report, the surface fracture zone, stress concentration zone, and internal borehole location of the target unstable rock mass are identified as key monitoring areas. Acquire multi-channel signals from key monitoring areas, including triaxial vibration signals, acoustic emission signals, and ambient temperature and humidity signals; The triaxial vibration signal, acoustic emission signal, and ambient temperature and humidity signal are extracted and retained separately. The retained triaxial vibration signal, acoustic emission signal, and ambient temperature and humidity signal are then aggregated into distributed acoustic vibration monitoring data after time alignment.

[0008] Optionally, in some embodiments of this application, the triaxial vibration signal, acoustic emission signal, and ambient temperature and humidity signal are respectively extracted and retained, including: Calculate the ratio of the short-time average to the long-time average of the acoustic emission signal. When the ratio exceeds a preset event trigger threshold, take the moment when the ratio exceeds the threshold as the trigger point, extract data from the range of a first preset time before the trigger point to a second preset time after the trigger point, and record this time range as the first time window. Simultaneously capture and retain the triaxial vibration signal and ambient temperature and humidity signal within the first time window.

[0009] Optionally, in some embodiments of this application, the neural network model is pre-obtained through a training method of self-supervised representation learning, specifically including: S01: Obtain a large number of unlabeled acoustic and vibration signal fragments of the target unstable rock mass during historical monitoring periods to construct a training dataset; S02, based on the acoustic and vibration signal segments in the training dataset, generate positive sample pairs and negative sample pairs, where the positive sample pairs consist of the first enhancement signal and the second enhancement signal; S03, Construct a pre-trained network architecture that includes an encoder and a projector head; S04, input each signal segment in the positive sample pair and the signal segment in the negative sample pair into the encoder respectively to extract the corresponding original features; perform nonlinear mapping on each of the original features through the projection head to obtain the corresponding feature vector; S05, construct a pre-training task based on the pre-constructed contrastive loss function, with the goal of maximizing the similarity between the feature vectors corresponding to the two signal segments in the positive sample pair and minimizing the similarity between the positive sample pair and the negative sample pair, update the parameters of the encoder and the projection head until convergence, and obtain the pre-trained encoder; The contrast loss function includes a temporal consistency constraint term, which is used to constrain the representation distance of signal segments collected by the same monitoring node in adjacent time windows in the embedding space and is positively correlated with the time interval between the adjacent time windows. S06. Connect the classification head to the output of the pre-trained encoder, and use a small number of labeled samples with state labels to perform supervised fine-tuning of the pre-trained encoder and the classification head to obtain a trained neural network model.

[0010] Optionally, in some embodiments of this application, positive sample pairs are generated based on acoustic signal segments in the training dataset, specifically including: Based on two enhanced signals generated by applying different data enhancement operations to the same acoustic and vibration signal segment, construct a first-class positive sample pair; Based on acoustic and vibration signal segments collected at the same monitoring node in adjacent time windows, a second type of positive sample pair is constructed. The second type of positive sample pair is used to constrain the model to learn the continuous evolution characteristics of the rupture signal in the time dimension. Based on acoustic and vibration signal fragments collected by different monitoring nodes in the same dangerous rock fracture zone at the same time period, a third type of positive sample pair is constructed. The third type of positive sample pair is used to constrain the model to learn the consistent representation of the same rupture event in the spatial dimension. Positive sample pairs are composed of positive sample pairs of type I, type II, and type III.

[0011] Optionally, in some embodiments of this application, a first type of positive sample pair is constructed based on two enhanced signals generated by applying different data enhancement operations to the same acoustic signal segment, specifically including: Obtain an acoustic signal segment from the training dataset as the source signal segment; A combination of enhancement operations is randomly applied to the source signal segment to obtain the first enhanced signal; the combination of enhancement operations includes one or more of the following: noise addition based on the actual environmental noise spectrum of the unstable rock mass, time clipping based on the duration of the rock fracture signal, and frequency domain masking based on the main frequency band of the rock fracture. A second enhanced signal is obtained by randomly applying a combination of enhancement operations different from the first enhanced signal to the source signal segment. The first and second enhanced signals are used to construct a first-class positive sample pair.

[0012] Optionally, in some embodiments of this application, step S03 involves constructing a pre-trained network architecture that includes an encoder and a projection head; the encoder is a multi-scale temporal convolutional network, which includes multiple temporal convolutional kernels of different sizes arranged in parallel, including a first-size convolutional kernel for extracting high-frequency transient features of acoustic emission signals and a second-size convolutional kernel for extracting low-frequency continuous features of vibration signals; The encoder also includes an attention module, which is used to perform weighted fusion of features at different time scales.

[0013] Optionally, in some embodiments of this application, the average value and the proportion are compared with preset warning trigger thresholds to generate warning results, specifically including: When the average value exceeds the first threshold and the proportion exceeds the preset proportion threshold, a first-level warning is triggered. When the average value exceeds the second threshold and the proportion exceeds the preset proportion threshold, a second-level warning is triggered; where the second threshold is greater than the first threshold.

[0014] Optionally, in some embodiments of this application, after step S4, an incremental learning and updating step of the neural network model is further included: The inference results of the deployed neural network model on real-time data are correlated with the feedback information on whether a collapse has occurred, forming new labeled samples; When the number of newly labeled samples reaches a preset threshold or a fixed update interval is reached, the model parameters are updated using an elastic weight consolidation strategy.

[0015] Optionally, in some embodiments of this application, an elastic weight consolidation strategy is used to update the model parameters, specifically including: Get newly labeled samples; Obtain the labeled samples used in the fine-tuning stage in step S06. Assign higher sampling weights to the samples labeled as precursor states in the labeled samples used in the fine-tuning stage than to the samples in normal states. Calculate the Fisher information matrix based on the weighted samples. The Fisher information matrix is ​​used to measure the importance of each parameter of the neural network model to the task of identifying the state of dangerous rock mass. Based on the contribution of each network layer in the encoder to the identification results of the precursor state of unstable rock mass, different interlayer amplification factors are set for different network layers, with the amplification factor of deep convolutional layers being higher than that of shallow convolutional layers. Model training is performed based on newly labeled samples. The product of Fisher information matrix and inter-layer amplification coefficient is used as the regularization constraint strength. Penalty terms are applied to parameters whose importance is higher than a preset threshold to constrain their deviation from historical values ​​during training and complete the model parameter update.

[0016] (III) Beneficial Effects This application presents a method for identifying precursors of rockfall collapses based on acoustic vibration signals and self-supervised representation learning. By filtering and segmenting distributed acoustic vibration monitoring data using a sliding window, it effectively denoises and standardizes multi-source signals from the rock mass, providing a high-quality, time-aligned data foundation for subsequent analysis. By extracting multi-dimensional features in the time, frequency, and time-frequency domains and then splicing and standardizing them, it fully explores the multi-level information related to changes in the structural state of the rock mass within the acoustic vibration signals, enhancing the model's ability to distinguish different rockfall states. The standardized feature vectors are then input into a self-supervised representation learning pre-trained neural network. The model enables the learning of general representations from massive amounts of unlabeled acoustic and vibration data, significantly reducing the reliance on expensive manually labeled samples. It overcomes the training bottleneck of scarce precursor samples in low-frequency disaster scenarios and improves the model's generalization ability. By constructing an early warning window with a preset number of signal segments, calculating the average probability of precursor states within the window and the proportion of segments exceeding the threshold, and comparing them with the early warning trigger threshold, a hierarchical early warning mechanism based on comprehensive judgment over continuous time periods is realized. This avoids false alarms caused by the occasional high probability of a single segment, improves the accuracy and robustness of the early warning, and can provide timely and reliable technical support for the prevention and control of rockfall disasters. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method for identifying precursors of rockfalls based on acoustic and vibration signals and self-supervised representation learning, as proposed in this application. Figure 2 This is a flowchart of the neural network model training process in one embodiment of this application. Detailed Implementation

[0018] To better explain and facilitate understanding of this application, the following detailed description of the application is provided in conjunction with the accompanying drawings and specific embodiments.

[0019] In existing technologies, methods for identifying precursors to rockfalls can be mainly categorized into the following three types: The first category is manual inspection and simple sensor monitoring methods. This involves periodically checking for surface changes such as crack development and displacement deformation on the dangerous rock mass, or deploying simple sensors such as surface displacement gauges and crack gauges for fixed-point observation. This type of method relies on manual experience and judgment, making it impossible to achieve continuous monitoring around the clock and difficult to capture early, subtle precursor signals of dangerous rock masses. Furthermore, the simple sensors are deployed in isolation, resulting in severely fragmented data that cannot comprehensively reflect the coordinated evolution of stress on the surface and inside the dangerous rock mass, and thus failing to capture all precursor features.

[0020] The second category is single-mode acoustic emission or vibration signal monitoring methods. These methods deploy acoustic emission sensors or accelerometers at key locations on unstable rock formations to collect single-mode signals and use fixed empirical thresholds for early warning. However, these methods have limited feature dimensions and cannot effectively distinguish between environmental disturbances such as wind noise, vehicle vibration, and rainfall, and actual crack propagation signals, resulting in extremely high false alarm and false alarm rates. Furthermore, the fixed thresholds cannot adapt to the differences in geological structure and seasonal environmental changes of different unstable rock formations, making it difficult to guarantee the accuracy of early warnings.

[0021] The third category is conventional data-driven supervised deep learning methods. After collecting acoustic and vibration signals, these methods rely on manually labeled samples of normal, precursor, and collapse states to train a classification model for state recognition. This type of method heavily depends on a large amount of high-quality labeled data. However, rockfalls are low-frequency disaster events, and precursor samples are extremely scarce. This makes model training difficult, prone to overfitting, and has weak generalization ability across geological scenarios. At the same time, once deployed, the model becomes fixed and cannot be continuously optimized and iterated using newly collected data, making it difficult to adapt to feature drift during the long-term evolution of rock masses.

[0022] To address this, the rockfall precursor identification method based on acoustic vibration signals and self-supervised representation learning provided in this application acquires multi-channel signals from the surface fissure zone, stress concentration zone, and internal borehole locations of the target rock mass through distributed acoustic vibration signal acquisition and STA / LTA event-triggered interception, and aggregates them into distributed acoustic vibration monitoring data. Multi-domain feature extraction and splicing standardization generate model input feature vectors. A pre-trained neural network model outputs the probability that each signal segment belongs to the normal state, the precursor state, and the collapse state, respectively. This neural network model is pre-trained using self-supervised contrastive learning and fine-tuned with a small number of labeled samples. A warning window is constructed using a preset number of signal segments, and the average probability of the precursor state within the window and the proportion of segments exceeding the threshold are calculated and compared with a preset warning trigger threshold to generate a graded warning result. The model parameters are continuously updated using an elastic weight consolidation strategy through incremental learning updates. This achieves all-weather synchronous monitoring, low label dependency, high anti-interference capability, graded and accurate warning, and sustainable iterative optimization for rockfall precursor identification, overcoming the shortcomings of the aforementioned prior art.

[0023] To better explain and facilitate understanding of this application, a detailed description of its embodiments is provided below in conjunction with the accompanying drawings. While exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a clearer and more thorough understanding of this application and to fully convey the scope of this application to those skilled in the art.

[0024] Example 1: Figure 1 This is a flowchart illustrating a method for identifying precursors to rockfalls based on acoustic vibration signals and self-supervised representation learning, according to this application. This method can be applied to monitoring and early warning scenarios of precursors to rockfalls at potential rockfall sites, such as along mountain roads, slopes in mining areas, banks of reservoirs, and cliff-side areas in scenic spots. It aims to solve the technical problems in existing technologies, such as incomplete monitoring coverage, scarce labeled samples leading to poor model generalization, inaccurate warnings with fixed thresholds, and the inability of models to continuously iterate. Figure 1 As shown, this method for identifying precursors to rockfalls based on acoustic and vibration signals and self-supervised representation learning includes: Step S1: Obtain distributed acoustic and vibration monitoring data of the target unstable rock mass within a preset time period, filter the distributed acoustic and vibration monitoring data, and divide it into multiple signal segments using a preset fixed-time sliding window. Step S2: Extract time-domain features, frequency-domain features, and time-frequency-domain features from each signal segment, and perform splicing and standardization processing to generate a standardized feature vector; Step S3: Input the standardized feature vectors corresponding to the multiple signal segments into the pre-trained neural network model, and output the probability that each signal segment belongs to the normal state, the pre-collapse state, and the collapse state, respectively; the neural network model is obtained in advance through a self-supervised representation learning training method. Step S4: Using a preset number of signal segments as an early warning window, calculate the average probability of the collapse precursor state corresponding to each signal segment within the early warning window, and the proportion of signal segments within the early warning window whose precursor state probability exceeds a preset basic threshold; compare the average value and the proportion with a preset early warning trigger threshold respectively to generate an early warning result.

[0025] This application provides a method for identifying precursors of rockfall collapses based on acoustic vibration signals and self-supervised representation learning. By acquiring distributed acoustic vibration monitoring data of the target rock mass within a preset time period, and performing filtering and sliding window segmentation, it achieves effective denoising and standardized segmentation of multi-source acoustic vibration signals, providing a high-quality, time-aligned data foundation for subsequent analysis and supporting continuous monitoring around the clock. The method extracts multi-dimensional features in the time domain, frequency domain, and time-frequency domain for each signal segment, and performs splicing and standardization processing, fully exploring the multi-level information related to changes in the structural state of the rock mass within the acoustic vibration signals. Compared with single-dimensional features, this effectively improves the model's ability to distinguish different states of the rock mass and the stability of the input representation. By inputting the standardized feature vector into the self-supervised representation learning method… Supervised representation learning pre-trained neural network models utilize massive amounts of unlabeled acoustic and vibration data to learn general representations. Only a small number of labeled samples are needed to complete classification task adaptation, significantly reducing reliance on expensive manual labeling samples. This overcomes the training bottleneck of scarce precursor samples in low-frequency disaster scenarios, improving the model's generalization ability and recognition accuracy. By constructing an early warning window with a preset number of signal segments, calculating the average probability of precursor states within the window and the proportion of segments exceeding the threshold, and comparing this with a preset early warning trigger threshold, an early warning mechanism based on comprehensive judgment over continuous time periods is realized. This avoids false alarms caused by the occasional high probability of a single segment, improving the overall accuracy and robustness of the early warning results, and providing timely and reliable technical support for the prevention and control of rockfall disasters.

[0026] Example 2: The method described in this application is applicable to the identification of pre-collapse precursors of different types of unstable rock masses. This system is deployed in key monitoring areas of potential rockfall hazards, such as along mountain roads, slopes in mining areas, banks of reservoirs, and cliff-side areas in scenic spots. Once deployed, the monitoring nodes enter a continuous monitoring state, continuously collecting acoustic and vibration signals within a preset fixed-duration window. The system continuously filters, segments, extracts features, and performs model inference on the signals. When the average probability of the precursor state accumulated within the warning window and the proportion of segments exceeding the threshold meet the preset warning triggering conditions, the system automatically triggers the generation of a warning result of the corresponding level.

[0027] This embodiment uses steep slopes and dangerous rock masses along mountain highways as an example for detailed explanation.

[0028] Step S1: Obtain distributed acoustic and vibration monitoring data of the target unstable rock mass within a preset time period, filter the distributed acoustic and vibration monitoring data, and divide it into multiple signal segments using a preset sliding window of fixed duration.

[0029] Specifically, based on the pre-obtained geological survey report, the surface fracture zone, stress concentration zone, and internal borehole location corresponding to the target unstable rock mass are identified and designated as key monitoring areas.

[0030] IoT monitoring nodes, which integrate triaxial accelerometers, acoustic emission sensors, and temperature and humidity sensors, are deployed in key monitoring areas. These IoT monitoring nodes collect multi-channel signals from the key monitoring areas in real time. The multi-channel signals specifically include triaxial vibration signals, acoustic emission signals, and ambient temperature and humidity signals.

[0031] Among them, acoustic emission signals are used to capture high-frequency transient signals generated by micro-fractures inside the rock, which is the core indicator for judging internal damage to the rock mass; triaxial vibration signals are used to capture the dynamic response of the target rock mass to the external environment under natural conditions, including vibration responses in the X, Y, and Z directions. For example, passing vehicles, wind, and rain will cause rock mass vibration, and the expansion of micro-fractures inside the rock will also generate vibration signals that are captured by the sensor. By analyzing the time-domain and frequency-domain characteristics of these natural vibrations, the stability state of the dangerous rock mass can be judged; environmental temperature and humidity signals include recorded environmental temperature and humidity, which are used to eliminate signal drift interference caused by temperature and humidity changes.

[0032] In a preferred embodiment, considering that unstable rock masses are typically located along remote mountain roads, and that monitoring nodes rely on battery power and have limited wireless communication bandwidth, uploading all data collected by sensors in real time would deplete the batteries within days, failing to meet the needs of long-term continuous monitoring. Therefore, this embodiment employs an event-triggered interception strategy, intercepting and retaining key signal data only when a suspected rock micro-fracture event is detected. This minimizes transmission load and power consumption while ensuring that precursor signals are not lost, making it suitable for unattended monitoring scenarios in the field.

[0033] In this embodiment, acoustic emission signals are selected as the primary trigger source for event detection because the rock micro-fracture process is essentially a stress wave release process. When micro-fractures occur inside the rock, elastic strain energy is released in the form of stress waves. Acoustic emission signals originate directly from the fracture source, and their high-frequency components (typically tens to hundreds of kHz) are significantly distinguishable in the spectrum from natural environmental noise (wind noise, water flow, etc., usually below 1 kHz) and mechanical vibration (typically below hundreds of Hz). Therefore, acoustic emission signals are the most direct and sensitive physical signals reflecting internal rock damage. In contrast, triaxial vibration signals represent the overall rock mass response triggered by rock fracture, are more affected by rock mass structural surface attenuation and filtering effects, and are easily interfered with by external environmental vibrations. Their sensitivity and specificity are inferior to acoustic emission signals. Therefore, this embodiment uses acoustic emission STA / LTA triggering as the trigger condition for event interception, ensuring that any potential rock micro-fracture event is first captured by the acoustic emission sensor, guaranteeing the sensitivity of early precursor signal capture. Triaxial vibration signals are recorded synchronously as auxiliary reference information for subsequent analysis to assist in judging event intensity, but do not participate in the interception decision.

[0034] Specifically, the triaxial vibration signal, acoustic emission signal, and ambient temperature and humidity signal obtained above are extracted and retained, including: The acoustic emission sensor continuously acquires acoustic emission signals at a preset fixed sampling rate (e.g., 1 MHz). The system calculates the ratio R(t) of the short-term average (STA) to the long-term average (LTA) of the acoustic emission signal in real time. The short-term average STA(t) reflects the instantaneous change in signal amplitude near the current moment, while the long-term average LTA(t) reflects the average level of background noise over a longer period. When a rock micro-fracture event occurs, the acoustic emission signal amplitude rises rapidly within a short time, and the short-term average STA(t) changes faster than the long-term average LTA(t), resulting in a significant peak in the STA / LTA ratio.

[0035] The durations of the short and long time windows can be preset based on the geological conditions of the target unstable rock mass and the level of environmental noise. As an example, the short window duration can be set to 0.5–2 seconds (e.g., 1 second), and the long window duration can be set to 10–60 seconds (e.g., 30 seconds). The starting points of the short and long windows can coincide or use different alignment methods. The short window is used to capture instantaneous signal changes in micro-fracture events, while the long window is used to establish a baseline level of the current environmental noise.

[0036] The system calculates the ratio R(t) of the short-term average to the long-term average in real time. When R(t) exceeds the preset event trigger threshold (which can be calibrated according to the environmental noise level of the target dangerous rock mass monitoring site during the silent period, with a typical value range of 2.0~5.0, and a value of 3.0 in this embodiment), it is determined to be a valid acoustic emission event, and this moment is recorded as the trigger point. Data within the range from the first preset time before the trigger point (e.g., 0.1~0.5 seconds) to the second preset time after the trigger point (e.g., 0.3~1.0 seconds) is extracted and recorded as the first time window to preserve the complete decay process of the acoustic emission event.

[0037] In a preferred embodiment, the system can dynamically adjust the trigger threshold according to time periods to adapt to the diurnal difference in environmental noise. During daytime periods (e.g., 6:00 to 20:00), environmental noise levels are high due to human activities such as vehicle traffic and machinery operations, and the system automatically adopts a higher trigger threshold to avoid false triggering caused by noise. During nighttime periods (e.g., 20:00 to 6:00 the next day), the environment is relatively quiet, and the system automatically switches to a lower trigger threshold to maintain sensitivity in capturing weak and minor fracturing events. The division of the above-mentioned daytime and nighttime periods can be flexibly set according to the actual sunshine hours in the area where the target rock mass is located. Through the above-mentioned diurnal adaptive threshold adjustment, the system can maintain reasonable event trigger sensitivity at different times, effectively balancing the dual requirements of false trigger suppression and precursor signal capture.

[0038] The system synchronously captures and retains triaxial vibration signals and ambient temperature and humidity signals within a first time window. Specifically, the triaxial accelerometer continuously acquires triaxial vibration signals, including X-axis, Y-axis, and Z-axis vibration signals, at a preset sampling rate (e.g., 2kHz). After the acoustic emission signal triggers and determines the first time window, the system synchronously captures and retains the triaxial vibration signals within the first time window.

[0039] The ambient temperature and humidity sensor samples and records ambient temperature and humidity signals at a preset cycle (e.g., every 5-10 minutes). When an acoustic emission event is triggered, the system synchronously extracts the ambient temperature and humidity data within the time period corresponding to the first time window and associates and records them with the acoustic emission signal and triaxial vibration signal.

[0040] Rock microfracture events are detected in real time by using the STA / LTA ratio of acoustic emission signals. Only when an event is triggered is a valid data window captured and retained, which greatly reduces the wireless transmission load and the amount of data for subsequent model processing. At the same time, it ensures that key precursor signals are not lost, enabling the system to achieve long-term continuous monitoring under low power conditions in the field.

[0041] The triaxial vibration signal, acoustic emission signal, and ambient temperature and humidity signal that were extracted and retained were time-aligned according to a unified time base and then aggregated into distributed acoustic and vibration monitoring data. This data was then uploaded to the server through a gateway. The uploaded data carried a time tag for the first time window, which included the trigger time and the start and end times of the window. This time tag was used in subsequent steps to correlate and analyze signals collected by different sensors within the same time period.

[0042] In a preferred embodiment, to further improve the accuracy of acoustic emission event interception and avoid frequent false triggers caused by environmental noise fluctuations, the system can also introduce a threshold cancellation mechanism. Specifically, during the real-time calculation of the STA / LTA ratio R(t) of the acoustic emission signal, in addition to the preset event triggering threshold T... on In addition, a cancellation threshold T is preset. off And T off <T on When R(t) exceeds T on When the acoustic emission event begins, the trigger time t is recorded. trigger After the event begins, the system continuously monitors the change in R(t). When R(t) falls back to T... off When the following condition is met, the acoustic emission event is considered to have ended, and the end time t is recorded. end .

[0043] Based on the above determination, the system intercepts and retains the data from the trigger time t. trigger First preset duration Δt pre At the end time t end The second preset duration Δt post The acoustic emission data within the range, i.e., the intercepted interval, is: [t trigger -Δt pre , t end +Δt post ]; Where, Δt pre Δt is used to preserve background signals prior to the event for comparative analysis. post This is used to preserve the complete decay process after the event ends, ensuring that the event waveform is captured completely. The resulting first time window is Δt from the trigger moment. pre After the end time Δt post Continuous intervals.

[0044] This threshold cancellation mechanism can dynamically adjust the length of the interception window according to the actual duration of the acoustic emission event, avoiding the problems of long events being truncated or short events containing too much invalid noise under the fixed window method. It is especially suitable for monitoring scenarios where the environmental noise level fluctuates greatly or the duration of micro-fracture events varies greatly.

[0045] This embodiment uses the first time window method as an example for illustration. This method is simple to implement, has low computational load, and is suitable for monitoring scenarios where environmental noise is relatively stable. The threshold cancellation mechanism, as another preferred implementation, can be used in scenarios with complex environmental noise or significant variations in event duration. Both methods can be flexibly selected according to actual monitoring conditions, and neither departs from the scope of protection of this application.

[0046] In a preferred embodiment, to further improve the accuracy of event determination, the system adds a vibration signal as an auxiliary criterion based on STA / LTA acoustic emission triggering.

[0047] Specifically, after the acoustic emission signal STA / LTA ratio exceeds the event trigger threshold and triggers interception, the system synchronously acquires the triaxial vibration vector amplitude A at the same moment. vm (t). The vibration amplitude is not a necessary condition for interception. Regardless of whether the vibration amplitude exceeds the silence threshold, the system will execute the data interception and retention process normally. The system synchronously records the vibration vector amplitude corresponding to the triggering event and includes this vibration amplitude information as an auxiliary annotation in the uploaded data.

[0048] When acoustic emission is triggered and the vibration amplitude significantly exceeds the quiescent threshold at the same time, it indicates that the event is accompanied by a significant rock mass vibration response and a large energy release intensity. It can be used as a key reference marker in subsequent data processing and early warning analysis for reference by models or early warning logic (such as appropriately increasing the early warning weight of the data in this period).

[0049] When acoustic emission is triggered but the vibration amplitude does not exceed the silence threshold at the same time, it indicates that the energy release of the event is weak or that it is a small-scale micro-fracture near the sensor. In this case, the data is still captured and retained normally, but it is marked as "weak vibration response" in the auxiliary information for reference in subsequent analysis. The data is not abandoned because the vibration signal is weak.

[0050] In this way, acoustic emission triggering is always the only triggering condition for event interception, while vibration signal is used as auxiliary reference information rather than a necessary condition. This ensures the complete capture of micro-fracture events and provides additional physical dimension reference for subsequent data analysis without increasing the risk of missed reports, but only enhancing the physical understanding of triggered events.

[0051] Then, the distributed acoustic and vibration monitoring data is filtered using a zero-phase finite impulse response filter. The group delay is eliminated by first forward filtering and then reverse filtering, ensuring that the filtered signal is strictly aligned with the unfiltered signal in time.

[0052] After filtering, the filtered continuous signal is segmented into multiple signal segments of equal length using a preset sliding window of fixed duration. The window duration can be set according to the typical duration of the micro-fracture signal of the target unstable rock mass, preferably to cover a complete micro-fracture event waveform. As an optional implementation, the window duration can be set to 1 to 10 seconds, preferably 2 to 5 seconds. In this embodiment, 2 seconds is used as an example. The sliding step size can be set to 1 second (i.e., 50% overlap) to ensure that potential micro-fracture transient events can be captured by at least one complete window, while increasing the number of training samples.

[0053] Step S2: Extract time-domain features, frequency-domain features, and time-frequency-domain features from each signal segment, and perform splicing and standardization processing to generate a standardized feature vector.

[0054] Specifically, the following three types of features are extracted for each signal segment: (1) Time-domain characteristics, including mean, standard deviation, peak-to-peak value, skewness and kurtosis. Among them, kurtosis is extremely sensitive to transient impact components in the signal. When micro-fractures occur inside the rock, the acoustic emission signal will show a sudden change in amplitude, and the kurtosis value will increase significantly. It is a key indicator for identifying micro-fracture events.

[0055] (2) Frequency domain characteristics, including the spectral centroid and spectral variance. The spectral centroid reflects the concentration of signal energy in the frequency domain. When the rock fractures, the signal energy will shift to the lower frequency direction, and the spectral centroid will shift downward. The spectral variance reflects the degree of dispersion of the spectrum. The spectral distribution of the fracture signal is usually more dispersed than that of the environmental noise.

[0056] (3) Time-frequency domain features, including Mel frequency cepstral coefficients (MFCC). MFCC can effectively compress spectral information, focusing on frequency bands related to the acoustic characteristics of rock fracture, and enhancing the model's ability to capture the essential characteristics of acoustic emission signals.

[0057] The three types of features extracted above are concatenated into a high-dimensional feature vector in the order of time-domain features, frequency-domain features, and finally time-frequency-domain features. All feature vectors are then Z-score standardized to eliminate the dimensional differences between different features and generate standardized feature vectors.

[0058] In a preferred embodiment, after feature extraction, the features can be standardized based on lithological constraints according to the lithological information of the target unstable rock mass. The acoustic emission and vibration signals of unstable rock masses of different lithologies (such as limestone, sandstone, and shale) differ significantly in amplitude range and spectral distribution. Specifically, the training samples are first grouped according to lithology, and the mean and standard deviation of each feature dimension within each group are calculated. Then, Z-score standardization is performed on the sample features within the same lithology group.

[0059] By standardizing based on lithological constraints, data from different lithologies are made comparable in the feature space, preventing the model from misinterpreting differences in lithological characteristics as changes in the state of dangerous rock masses. When the lithological information of the target dangerous rock mass is unknown, global Z-score standardization can still be used, and this application does not impose any restrictions on this.

[0060] Furthermore, considering that the environmental noise level at the monitoring site of the unstable rock mass will slowly drift with seasonal changes and weather variations, if the feature standardization parameters remain fixed for a long period, signals under the same physical state may be mapped to different feature vectors at different times, affecting the model's recognition stability. Therefore, in a preferred embodiment, the system can periodically (e.g., hourly) update the environmental noise baseline and accordingly update the mean and standard deviation parameters used in Z-score standardization. The environmental noise baseline can be obtained by statistically calculating signal segments during periods without event triggering at the monitoring site. Through this periodic update, the standardization process always uses the signal statistical distribution under the current environmental conditions as a reference benchmark, avoiding feature drift and maintaining the consistency and stability of the model's input features.

[0061] Step S3: Input the standardized feature vectors corresponding to multiple signal segments into the pre-trained neural network model, and output the probability that each signal segment belongs to the normal state, the pre-collapse state, and the collapse state, respectively.

[0062] The neural network model includes an encoder and a classifier. The encoder maps the input feature vector to a low-dimensional embedding vector, and the classifier outputs three probability values ​​based on the embedding vector, indicating whether each signal segment belongs to the normal state, the pre-collapse state, or the collapse state.

[0063] For example, the model output is as follows: Signal Segment 1 → Model → [Normal: 95%, Precursors: 3%, Collapse: 2%]; Signal Segment 2 → Model → [Normal: 92%, Precursors: 5%, Collapse: 3%]; Signal segment 3 → Model → [Normal: 90%, Precursors: 7%, Collapse: 3%]; ...... Because the evolution cycle of unstable rock masses from stability to collapse is long, the time window for the appearance of collapse precursor signals is extremely short. In continuous monitoring data, "normal" samples constitute the vast majority, while "collapse precursor" and "collapse" samples are extremely rare, making it difficult to support the training of conventional supervised models. Furthermore, the determination of precursor states requires experienced geological experts to make comprehensive judgments based on multi-source information; manual annotation is costly and difficult to standardize. In addition, models trained on a very small number of labeled samples are prone to overfitting and are difficult to generalize to unstable rock masses with different geological conditions. Therefore, in this embodiment, the neural network model is pre-trained using a self-supervised representation learning method. Figure 2 This is a flowchart of the neural network model training process in this embodiment. The neural network model training process specifically includes: Step S01: Obtain a large number of unlabeled acoustic and vibration signal fragments of the target unstable rock mass during historical monitoring periods and construct a training dataset.

[0064] Specifically, a large number of unlabeled acoustic and vibration signal segments are extracted from the distributed acoustic and vibration monitoring data collected and stored during the historical monitoring period of the target unstable rock mass as training samples. "Unlabeled" refers to these signal segments that have not yet been labeled. These unlabeled acoustic and vibration signal segments originate from data collected at different times, monitoring nodes, and under different environmental conditions, covering various states of the unstable rock mass, including stable periods, periods of environmental disturbance, and potential precursory evolution periods. Constructing the training dataset only requires organizing and storing the aforementioned signal segments according to a preset format, without any manual labeling, thus allowing for the collection of a large number of samples. As an optional implementation method, continuous monitoring data of the target unstable rock mass over the past 6 to 12 months can be selected, divided into fixed-time windows, and used to construct a training dataset containing tens of thousands to hundreds of thousands of unlabeled signal segments to support the data volume requirements for subsequent self-supervised comparative learning pre-training.

[0065] Step S02: Generate positive sample pairs and negative sample pairs based on the acoustic and vibration signal segments in the training dataset.

[0066] Among them, positive sample pairs include three categories, and the process of obtaining the three categories of positive sample pairs specifically includes: (1) Two enhanced signals generated by applying different data enhancement operations to the same acoustic vibration signal segment are used to construct the first type of positive sample pair.

[0067] (2) For a single acoustic signal segment in the training dataset, another acoustic signal segment collected by the same monitoring node within an adjacent time window is obtained, and the two are constructed as a second type of positive sample pair. Rock fracturing is a gradual evolutionary physical process, and the signals within adjacent time windows are continuous in time, and their representations should remain similar in the embedding space. The second type of positive sample pair is used to constrain the model to learn the continuous evolutionary characteristics of the fracturing signal in the time dimension.

[0068] (3) For a single acoustic-vibration signal segment in the training dataset, acquire acoustic-vibration signal segments collected by different monitoring nodes within the same unstable rock fracture zone during the same time period, and construct the two as a third type of positive sample pair. The same rock mass fracture event will be captured simultaneously by sensors at multiple spatial locations. Although the signals collected by different nodes are slightly different due to differences in propagation paths, they reflect the same physical event, and their representation should be consistent in the embedding space. The third type of positive sample pair is used to constrain the model to learn the consistent representation of the same fracture event in the spatial dimension.

[0069] The positive sample pairs are composed of the first, second, and third types of positive sample pairs mentioned above. Thus, the first type of positive sample pairs enables the model to learn invariance to enhancement transformations, the second type of positive sample pairs enables the model to perceive the temporal evolution of rupture signals, and the third type of positive sample pairs enables the model to establish a multi-node spatially consistent representation. The three types of positive sample pairs jointly constrain the model to comprehensively learn the physical characteristics related to the rupture of unstable rock masses from three dimensions: temporal enhancement, temporal evolution, and spatial synchronization.

[0070] Furthermore, using an enhanced signal of the current acoustic vibration signal segment as an anchor point, the enhanced signals of other acoustic vibration signal segments in the same training batch are used as negative samples to form a negative sample pair.

[0071] In a preferred embodiment, in addition to random negative samples within a batch, the following two additional negative sample types can be added to enrich the diversity of negative samples and improve the model's discriminative ability: (1) Signal segments collected at the same monitoring node in distant time windows (such as 24 hours or a week ago) are used as cross-time negative samples. These negative samples and anchor point signals come from the same spatial location but are collected at different times. Their characteristic differences mainly reflect long-term changes in environmental conditions or rock mass conditions. By introducing cross-time negative samples, the model is forced to learn to distinguish the signal differences of "the same location at different times", avoiding the model from misjudging time drift caused by seasonal changes, slow environmental changes, etc. as changes in the state of dangerous rock mass, thereby improving the stability of the model in long-term monitoring scenarios.

[0072] (2) Collect typical environmental noise samples (such as wind noise, vehicle vibration, rainfall, etc.) from the monitoring site in advance and add them to the training as negative environmental interference samples. These negative samples have certain similarities to micro-fracture signals in terms of spectral characteristics, but their physical sources are completely different. By directly allowing the model to learn which signals are not precursors during training, the model's ability to reject common environmental interferences can be effectively improved, and the model's anti-interference performance in actual field environments can be enhanced.

[0073] The two additional negative samples mentioned above, together with the random negative samples within the batch, constitute a complete set of negative samples. They are uniformly used in the training of the contrastive loss function, enabling the model to distinguish different signals while being robust to time drift and environmental interference.

[0074] In a preferred embodiment, a first type of positive sample pair is constructed based on two enhanced signals generated by applying different data enhancement operations to the same acoustic signal segment, specifically including: Obtain an acoustic signal segment from the training dataset as the source signal segment, and randomly apply a combination of enhancement operations to the source signal segment to obtain the first enhanced signal.

[0075] The enhancement operation combination includes one or more of the following: noise addition based on the actual environmental noise spectrum of the unstable rock mass, time clipping based on the duration of the rock fracture signal, and frequency domain masking based on the main frequency band of the rock fracture.

[0076] A second enhanced signal is obtained by reapplying a combination of enhancement operations, different from the first enhanced signal, to the source signal segment. The first enhanced signal and the second enhanced signal are used to construct the first type of positive sample pair.

[0077] Step S03: Construct a pre-trained network architecture including an encoder and a projection head. The encoder is used to extract features from the acoustic and vibration signal segments in the positive and negative sample pairs constructed in step S02.

[0078] The encoder is a multi-scale temporal convolutional network, which includes at least two temporal convolutional kernels of different sizes arranged in parallel, used to extract features of the acoustic vibration signal at different time scales.

[0079] The first-size convolutional kernel is smaller and is used to capture high-frequency transient impact features in acoustic emission signals; the second-size convolutional kernel is larger and is used to capture low-frequency continuous evolution features in vibration signals.

[0080] Additionally, the encoder includes an attention module for weighted fusion of features at different time scales.

[0081] The projection head is used to nonlinearly map the embedding vectors output by the encoder to the contrastive learning space to calculate the similarity between the feature vectors corresponding to two signal segments in a positive sample pair. After pre-training, the projection head is discarded, and only the encoder is retained for subsequent fine-tuning and inference.

[0082] In a preferred embodiment, the attention module employs a physical prior-guided attention mechanism to replace the purely data-driven attention weight calculation method.

[0083] Conventional attention modules learn their weights entirely through data-driven learning, lacking physical constraints. This can lead to focusing on frequency bands unrelated to rock fracture under conditions of insufficient training data or noise interference. Therefore, this implementation utilizes the physical frequency band characteristics of acoustic emission and vibration signals as prior knowledge to guide the computational process of the attention module.

[0084] Specifically, based on the typical dominant frequency band range of rock fracture acoustic emission signals (e.g., 20–100 kHz) and the dominant frequency band range of triaxial vibration signals (e.g., 1–500 Hz), the attention module presets initial weight biases for the frequency bands corresponding to features at different time scales. For high-frequency acoustic emission features extracted by small-sized convolutional kernels, the attention module assigns higher initial attention weights during calculation; for low-frequency vibration features extracted by large-sized convolutional kernels, corresponding initial attention weights are also assigned based on their physical frequency bands. On this basis, the attention module fine-tunes these initial weights using training data, enabling the model, guided by physical priors, to prioritize frequency band information related to the acoustic features of rock fractures and suppress the attention response to frequency band features corresponding to environmental noise such as wind noise and mechanical vibration.

[0085] By combining the domain knowledge of rock fracture with data-driven learning through the aforementioned physical prior-guided attention mechanism, the model can have a reasonable frequency band attention tendency in the early stage of training, reduce the dependence on a large amount of labeled data, and at the same time improve the model's ability to extract features from micro-fracture signals and its robustness to environmental noise.

[0086] In a preferred embodiment, to further enhance the interpretability of the model decision-making process and facilitate field engineers' understanding of the model's judgment basis, a lightweight interpretation branch is added to the encoder output in this embodiment.

[0087] Specifically, the interpretation branch runs parallel to the classification head, consisting of a global average pooling layer and a fully connected layer. Its input is the embedding vector output by the encoder, and its output is the frequency band label and time position label that contributes most to the discrimination result of the current input signal segment. During training, the interpretation branch shares encoder parameters with the main classification network. Through Gradient Weighted Class Activation Mapping (Grad-CAM) or a similar mechanism, it learns to map the encoder's feature responses onto the spectral-temporal dimension of the input signal, generating a contribution distribution in the form of a heatmap. During inference, the interpretation branch outputs a two-dimensional contribution heatmap, where the horizontal axis represents time position, the vertical axis represents frequency range, and the color intensity indicates the degree of contribution of that time-frequency region to the precursor judgment result.

[0088] To maintain the overall lightweight nature of the model, the explanation branch contains only a small number of parameters (such as 1-2 fully connected layers), and the additional computational cost during inference is less than 5% of that of the main network. The output of this explanation branch does not participate in the early warning decision; it is only used as auxiliary information and pushed to the monitoring platform along with the early warning result for reference by on-site engineers.

[0089] By adding interpretable output branches, the system can provide visual judgment criteria while outputting early warning results, which helps field engineers understand the model's decision logic, assists in troubleshooting when false alarms or missed alarms occur, and enhances the credibility of the monitoring system in actual engineering applications.

[0090] In step S03, a pre-trained network architecture including an encoder and a projection head is constructed; the encoder is a multi-scale temporal convolutional network, which includes multiple temporal convolutional kernels of different sizes arranged in parallel. The multiple temporal convolutional kernels of different sizes include a first-size convolutional kernel for extracting high-frequency transient features of acoustic emission signals and a second-size convolutional kernel for extracting low-frequency continuous features of vibration signals. The encoder also includes an attention module, which is used to perform weighted fusion of features at different time scales.

[0091] Step S04: Input each signal segment in the positive sample pair and the signal segment in the negative sample pair into the encoder to extract the corresponding original features; perform nonlinear mapping on each of the original features through the projection head to obtain the corresponding feature vector.

[0092] Step S05: Construct a pre-training task based on the pre-built contrastive loss function, with the goal of maximizing the similarity between the feature vectors corresponding to the two signal segments in the positive sample pair and minimizing the similarity between the positive sample pair and the negative sample pair. Update the parameters of the encoder and the projection head until convergence, and obtain the pre-trained encoder.

[0093] The contrastive loss function includes a temporal consistency constraint term, which constrains the representation distance of signal segments collected by the same monitoring node in adjacent time windows in the embedding space, and is positively correlated with the time interval between the adjacent time windows.

[0094] It should be noted that the contrastive loss function does not distinguish between the source types of positive sample pairs. For all three types of positive sample pairs, the training objective is uniformly adopted: "maximizing the similarity between the feature vectors corresponding to the two signal segments in the same positive sample pair, and minimizing the similarity between the positive and negative sample pairs." Each of the three types of positive sample pairs is calculated independently in the same training batch, and the final loss is the sum (or average) of the losses of the three types of positive sample pairs. These losses collectively update the parameters of the encoder and projector head without interfering with each other.

[0095] In a preferred embodiment, the pre-training task further includes a temporal consistency prediction task and a frequency domain energy distribution prediction task.

[0096] The temporal consistency prediction task is used to constrain the consistency of representation across the time dimension of different augmented versions of the same signal. Specifically, for two augmented signals generated from the same source signal segment through different data augmentation operations, they are required to have similar representations along the time axis in the embedding space of the encoder output. This task enhances the model's ability to perceive the gradual process of rock mass fracturing by forcing the model to pay attention to the temporal evolution of the signal.

[0097] The frequency domain energy distribution prediction task is used to constrain the model's accurate perception of the energy in the main frequency band of the signal. Specifically, the spectrogram of the input signal is randomly masked in the frequency domain, and the model is required to predict the energy distribution of the masked frequency band based on the masked spectrogram. Since the energy of rock fracture signals is concentrated in a specific main frequency band (such as the high-frequency components of acoustic emission signals), this task forces the encoder to focus on frequency domain features related to rock fracture, enhancing the model's sensitivity to micro-fracture signals.

[0098] The two auxiliary tasks mentioned above, together with the original contrastive loss function, constitute the multi-task pre-training objective. They share the same encoder for joint optimization, and the total loss function is a weighted sum of contrastive loss, temporal consistency loss, and frequency domain energy distribution prediction loss.

[0099] By introducing two auxiliary tasks—temporal consistency prediction and frequency domain energy distribution prediction—into the pre-training task, the encoder, building upon its existing comparative learning to distinguish different signals, further enhances its perception of the physical characteristics of acoustic and vibration signals from unstable rock masses from both temporal and frequency domain dimensions. Temporal consistency constraints enable the model to fully capture the continuous and gradual changes in signals during the progressive evolution of the rock mass from stability to precursors and then to collapse, avoiding misjudging isolated noise as a trend precursor. Frequency domain energy distribution constraints focus the model on the dominant frequency band where rock fracture signals are concentrated, enhancing its sensitivity to the acoustic emission characteristics of micro-fractures and effectively suppressing frequency domain interference from environmental noise such as wind noise and mechanical vibration. After joint pre-training of these three tasks, the embedding vector output by the encoder possesses instance discriminative ability, temporal evolution perception capability, and frequency domain physical sensitivity.

[0100] Step S06: Connect the classification head to the output of the pre-trained encoder. After the pre-trained encoder converges, use a small number of labeled samples with state labels to perform joint supervised fine-tuning of the pre-trained encoder and the classification head to obtain the trained neural network model.

[0101] Specifically, a classification head is constructed, which is a fully connected layer. Its input dimension is consistent with the output dimension of the pre-trained encoder, and its output dimension is consistent with the number of states of the unstable rock mass (i.e., three states: normal state, pre-collapse state, and collapse state). The classification head is connected to the output of the pre-trained encoder, so that the embedding vector output by the pre-trained encoder is used as the input of the classification head.

[0102] A small number of acoustic and vibration signal samples annotated by experts are obtained. Each signal segment in the annotated samples is labeled with a corresponding state label, which includes three categories: normal state, pre-collapse state, and collapse state. The annotated samples are input into the pre-trained encoder to extract embedding vectors, and then the state probability distribution is output through the classification head. The pre-trained encoder and the classification head are jointly supervisedly fine-tuned with the cross-entropy loss function as the target.

[0103] As an optional implementation, a phased training strategy can be adopted during fine-tuning: in the first phase, all parameters of the pre-trained encoder are frozen, and only the classification head is trained; in the second phase, the top 1-2 layers of the pre-trained encoder are unfrozen and fine-tuned together with the classification head to alleviate the overfitting problem in small sample scenarios.

[0104] After fine-tuning, the pre-trained encoder and the classification head together form a trained neural network model, which is used to perform real-time state recognition of input signal segments.

[0105] Step S4: Using a preset number of signal segments as an early warning window, calculate the average probability of the collapse precursor state corresponding to each signal segment within the early warning window, and the proportion of signal segments within the early warning window whose precursor state probability exceeds a preset basic threshold; compare the average value and the proportion with a preset early warning trigger threshold respectively to generate an early warning result.

[0106] In a preferred embodiment, the early warning window slides and updates as new signal segments arrive. That is, whenever a new signal segment completes state probability inference, the system adds it to the current early warning window and removes the oldest signal segment from the window, always maintaining a preset number of the latest signal segments within the window. The system only calculates the average and proportion of the signal segments within the current window after the current window is updated, and outputs a single early warning result for that moment, rather than performing batch backtesting of historical windows. Through this method, the system achieves real-time online monitoring and continuous early warning output of the unstable rock mass's condition.

[0107] When the average value exceeds a first threshold and the proportion exceeds a preset proportion threshold, a first-level warning is triggered; when the average value exceeds a second threshold and the proportion exceeds the preset proportion threshold, a second-level warning is triggered; wherein the second threshold is greater than the first threshold.

[0108] Specifically, suppose the warning window contains N consecutive signal segments (the value of N can be set according to the warning sensitivity requirements of the monitoring scenario; as an example, N can be 10, that is, the warning window covers the time span corresponding to the most recent 10 signal segments), and obtain the precursor state probability value p corresponding to each of these N segments. i (i=1,2,...,N), calculate their average value; at the same time, count the probability of the precursor exceeding the preset base threshold T in these N segments. base The number of segments is calculated, and its proportion to the total number of segments in the window is determined. The calculated average value is compared with the preset first threshold T1 and the second threshold T2 (T2>T1), and the calculated proportion is compared with the preset proportion threshold P.

[0109] When the average value exceeds T1 and the proportion exceeds P, a Level 1 warning (attention level) is triggered; when the average value exceeds T2 and the proportion exceeds P, a Level 2 warning (alarm level) is triggered. The warning information, along with the original data fragments, identification probability, and timestamp, is packaged and pushed to the monitoring center or the terminal of on-site management personnel to realize real-time assessment of the dangerous rock mass status and graded warning push.

[0110] Furthermore, to reduce false alarms caused by occasional noise in a single window, the system can be set up with a continuous compliance confirmation mechanism: the system will only officially output the warning result of that level when a preset number (e.g., 2) of adjacent warning windows meet the triggering conditions of the same warning level; if the window compliance is an isolated event, it will only be recorded as a short-term anomaly and will not trigger a formal warning.

[0111] In this embodiment, the early warning judgment is primarily based on the probability of a pre-collapse state output by the model. Simultaneously, the probabilities of a normal state and a collapse state output by the model serve as supplementary reference information: when the probability of a normal state is significantly higher than the other two probabilities, it indicates that the current period is in a stable state; when the probability of a collapse state increases significantly, it indicates that the unstable rock mass may be approaching or is in the process of becoming unstable. Even if the pre-collapse probability has not yet triggered an early warning, the system must generate an emergency alert for management personnel to make a comprehensive judgment based on the on-site situation.

[0112] Furthermore, as a preferred implementation, this embodiment also provides two supplementary criteria, which can be used alone or in combination to help enhance the reliability and lead time of the early warning: (1) The system calculates the confidence level of the current warning judgment based on the distribution of the three probability values ​​output by the model. Specifically, the three probability values ​​are arranged in descending order, and the difference between the highest probability and the second highest probability is taken as the confidence level index. The larger the difference, the clearer the model's distinction between different states, and the higher the confidence level; the smaller the difference, the more ambiguous the model's judgment, and the lower the confidence level. When the confidence level is lower than the preset threshold, the system adds a low confidence level mark to the warning information for monitoring personnel to refer to, so as to avoid blindly relying on the model output.

[0113] (2) The system monitors the temporal change trend of the precursor state probability within the warning window and calculates the first-order rate of change (velocity) and the second-order rate of change (acceleration) of the precursor probability as the window slides. When the first-order rate of change is continuously positive and the second-order rate of change exceeds the preset threshold, it indicates that the precursor probability is not only continuously rising, but the rate of increase itself is accelerating (i.e., positive acceleration). At this time, even if the average precursor probability within the current window has not exceeded the first-level warning threshold, the system still issues an early warning notice. This supplementary criterion is designed for the physical characteristics of accelerated instability of rockfalls. Precursors to collapses often show an accelerated evolution trend before the disaster, rather than a uniform increase. By capturing the acceleration change, potential dangerous trends can be identified in advance before the traditional threshold is triggered, giving on-site personnel more time to prepare for response.

[0114] The two supplementary criteria mentioned above do not replace the original early warning logic; they are only intended as supplementary information for on-site personnel to refer to. The triggering rules for the original Level 1 and Level 2 early warnings remain unchanged to ensure the stability and reliability of the tiered early warning mechanism.

[0115] Following step S4, after the model has conducted real-time inference for a period of time and accumulated sufficient new data, an incremental learning and update step for the neural network model is also included: Step S5 involves associating the inference results of the deployed neural network model on real-time data with the feedback information on whether a collapse ultimately occurs, thus forming new labeled samples.

[0116] Specifically, during system operation, the neural network model continuously infers from the real-time acquired signal segments, outputting the probability of precursor states corresponding to each signal segment. When the target unstable rock mass eventually collapses or is confirmed to be in a stable state on-site, the system associates the inferred signal segments from the aforementioned period with this final feedback information. If a collapse occurs, the signal segments from the period preceding the collapse are labeled as "precursor to collapse" or "collapse"; if no collapse occurs, they are confirmed as "normal." In this way, the system associates the original inference results with the actual feedback, forming newly labeled samples without the need for manual segment-by-segment labeling.

[0117] In a preferred embodiment, when the number of newly labeled samples reaches a preset threshold or a fixed update time interval, an elastic weight consolidation strategy is used to update the model parameters. The specific steps are as follows: First, obtain the newly labeled samples mentioned above; Obtain the labeled samples used in the fine-tuning stage in step S06, and assign a higher sampling weight to the samples labeled as precursor states than to the samples in normal states. Calculate the Fisher information matrix based on the weighted samples. The Fisher information matrix is ​​used to measure the importance of each parameter of the neural network model to the task of identifying dangerous rock mass states. The larger the value of its diagonal element, the more important the parameter is to the learned knowledge of dangerous rock mass state identification, and it should be more strongly protected in subsequent updates.

[0118] Then, based on the contribution of each network layer in the encoder to the identification results of precursor states of unstable rock masses, different inter-layer amplification coefficients were set for different network layers. The shallow convolutional layers of the encoder mainly extract the general time-frequency features of acoustic and vibration signals, and their contribution to the specificity of precursor state identification is relatively low; the deep convolutional layers extract high-level semantic features related to changes in the structural state of unstable rock masses, and their contribution to precursor identification is higher. Therefore, the amplification coefficient of the deep convolutional layers is higher than that of the shallow convolutional layers.

[0119] Finally, model training is performed based on the newly labeled samples. The product of the Fisher information matrix and the interlayer amplification coefficient is used as the regularization constraint strength. Parameters with importance exceeding a preset threshold are penalized to constrain their deviation from historical values ​​during training. This allows the system to absorb new knowledge while retaining the memory of old knowledge, thus completing the model parameter update. Through these incremental learning and update steps, the system can continuously optimize model performance using the newly labeled samples, while avoiding catastrophic forgetting and adapting to feature drift during the long-term evolution of unstable rock masses.

[0120] This embodiment provides an all-weather, low-power, highly reliable, and continuously optimizable intelligent monitoring and early warning solution for the prevention and control of rockfall disasters in mountainous highways, mine slopes, water conservancy reservoirs, and scenic cliffside areas. It achieves low-power, high-efficiency field data acquisition through a multi-sensor fusion deployment strategy guided by geological surveys and an event-triggered interception strategy driven by acoustic emission events. Through multi-domain feature extraction and lithological constraint standardization, it effectively mines multi-level information related to changes in the structural state of unstable rock masses from acoustic vibration signals. During model training, it overcomes the challenge of training in rockfall scenarios where labeled samples are scarce by combining self-supervised comparative learning with a small amount of annotation fine-tuning. The bottleneck is addressed by employing mechanisms such as three types of physical constraint positive sample pairs, scenario-based data augmentation, multi-scale temporal convolutional encoders, and multi-task joint pre-training. This enables the model to possess high sensitivity to micro-fracture signals, generalization ability across rock types, and environmental anti-interference capabilities. In early warning decision-making, a hierarchical early warning mechanism based on precursor probability windows, combined with auxiliary criteria such as confidence assessment and trend acceleration early warning, effectively reduces the false alarm rate and false negative rate, providing sufficient response time for disaster prevention and control. During iterative evolution, incremental learning loops continuously absorb new data to optimize model performance, enabling the system to adapt to feature drift during the long-term evolution of unstable rock masses.

[0121] Example 3: Finally, this application also proposes a computing device, which includes a processor and a memory. The memory stores computer programs, and the processor executes the instructions stored in the memory so that the computer device performs the rockfall precursor identification method based on acoustic vibration signals and self-supervised representation learning described in the above embodiments.

[0122] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for identifying precursors of rockfall based on acoustic vibration signals and self-supervised representation learning, characterized in that, include: S1, acquire distributed acoustic and vibration monitoring data of the target unstable rock mass within a preset time period, filter the distributed acoustic and vibration monitoring data, and divide it into multiple signal segments using a preset fixed-time sliding window; S2, extract time-domain features, frequency-domain features, and time-frequency-domain features for each signal segment, and perform splicing and standardization processing to generate a standardized feature vector; S3, input the standardized feature vectors corresponding to the multiple signal segments into a pre-trained neural network model, and output the probability that each signal segment belongs to the normal state, the pre-collapse state, and the collapse state, respectively; the neural network model is obtained in advance through a self-supervised representation learning training method; S4, using a preset number of signal segments as an early warning window, calculate the average probability of the collapse precursor state corresponding to each signal segment within the early warning window, and the proportion of signal segments within the early warning window whose precursor state probability exceeds a preset basic threshold; compare the average value and the proportion with a preset early warning trigger threshold respectively to generate an early warning result.

2. The method for identifying precursors of rockfalls based on acoustic and vibration signals and self-supervised representation learning according to claim 1, characterized in that, The acquisition of distributed acoustic and vibration monitoring data of the target unstable rock mass within a preset time period specifically includes: Based on the pre-obtained geological survey report, the surface fracture zone, stress concentration zone, and internal borehole location of the target unstable rock mass are identified as key monitoring areas. Acquire multi-channel signals from the key monitoring area, including triaxial vibration signals, acoustic emission signals, and ambient temperature and humidity signals; The triaxial vibration signal, acoustic emission signal, and ambient temperature and humidity signal are extracted and retained respectively. The retained triaxial vibration signal, acoustic emission signal, and ambient temperature and humidity signal are then aggregated into distributed acoustic vibration monitoring data after time alignment.

3. The method for identifying precursors of rockfalls based on acoustic vibration signals and self-supervised representation learning according to claim 2, characterized in that, The triaxial vibration signal, acoustic emission signal, and ambient temperature and humidity signal are respectively extracted and retained, including: Calculate the ratio of the short-time average to the long-time average of the acoustic emission signal. When the ratio exceeds a preset event trigger threshold, take the moment when the ratio exceeds the threshold as the trigger point, extract data from a range of a first preset time before the trigger point to a second preset time after the trigger point, and record this time range as the first time window. The triaxial vibration signal and the ambient temperature and humidity signal are simultaneously captured and retained within the first time window.

4. The method for identifying precursors of rockfalls based on acoustic vibration signals and self-supervised representation learning according to claim 1, characterized in that, The neural network model is pre-obtained through a training method based on self-supervised representation learning, specifically including: S01: Obtain a large number of unlabeled acoustic and vibration signal fragments of the target unstable rock mass during historical monitoring periods to construct a training dataset; S02, Generate positive sample pairs and negative sample pairs based on the acoustic and vibration signal segments in the training dataset; S03, Construct a pre-trained network architecture that includes an encoder and a projector head; S04, input each signal segment of the positive sample pair and the signal segment of the negative sample pair into the encoder respectively to extract the corresponding original features; perform nonlinear mapping on each of the original features through the projection head to obtain the corresponding feature vector; S05, construct a pre-training task based on the pre-constructed contrastive loss function, with the goal of maximizing the similarity between the feature vectors corresponding to the two signal segments in the positive sample pair and minimizing the similarity between the positive sample pair and the negative sample pair, update the parameters of the encoder and the projection head until convergence, and obtain the pre-trained encoder; The contrast loss function includes a temporal consistency constraint term, which is used to constrain the representation distance of signal segments collected by the same monitoring node in adjacent time windows in the embedding space and is positively correlated with the time interval between the adjacent time windows. S06, a classification head is connected to the output of the pre-trained encoder, and supervised fine-tuning of the pre-trained encoder and the classification head is performed using a small number of labeled samples with state labels to obtain a trained neural network model.

5. The method for identifying precursors of rockfalls based on acoustic and vibration signals and self-supervised representation learning according to claim 4, characterized in that, The step of generating positive sample pairs based on the acoustic signal segments in the training dataset specifically includes: Based on two enhanced signals generated by applying different data enhancement operations to the same acoustic and vibration signal segment, construct a first-class positive sample pair; Based on acoustic and vibration signal segments collected at the same monitoring node in adjacent time windows, a second type of positive sample pair is constructed. The second type of positive sample pair is used to constrain the model to learn the continuous evolution characteristics of the rupture signal in the time dimension. Based on acoustic and vibration signal fragments collected by different monitoring nodes in the same dangerous rock fracture zone at the same time period, a third type of positive sample pair is constructed. The third type of positive sample pair is used to constrain the model to learn the consistent representation of the same rupture event in the spatial dimension. The positive sample pair is composed of the first type of positive sample pair, the second type of positive sample pair, and the third type of positive sample pair.

6. The method for identifying precursors of rockfalls based on acoustic and vibration signals and self-supervised representation learning according to claim 5, characterized in that, The step of constructing a first-class positive sample pair based on two enhanced signals generated by applying different data enhancement operations to the same acoustic signal segment specifically includes: Obtain an acoustic vibration signal segment from the training dataset as the source signal segment; A combination of enhancement operations is randomly applied to the source signal segment to obtain a first enhanced signal; the combination of enhancement operations includes one or more of the following: noise addition based on the actual environmental noise spectrum of the unstable rock mass, time clipping based on the duration of the rock fracture signal, and frequency domain masking based on the main frequency band of the rock fracture. A second enhanced signal is obtained by randomly applying a combination of enhancement operations different from the first enhanced signal to the source signal segment. The first enhanced signal and the second enhanced signal are used to construct the first type of positive sample pair.

7. The method for identifying precursors of rockfalls based on acoustic and vibration signals and self-supervised representation learning according to claim 6, characterized in that, In step S03, a pre-trained network architecture including an encoder and a projection head is constructed; the encoder is a multi-scale temporal convolutional network, which includes multiple temporal convolutional kernels of different sizes arranged in parallel. The multiple temporal convolutional kernels of different sizes include a first-size convolutional kernel for extracting high-frequency transient features of acoustic emission signals and a second-size convolutional kernel for extracting low-frequency continuous features of vibration signals. The encoder also includes an attention module, which is used to perform weighted fusion of features at different time scales.

8. The method for identifying precursors of rockfalls based on acoustic and vibration signals and self-supervised representation learning according to claim 7, characterized in that, The step of comparing the average value and the ratio with preset warning trigger thresholds to generate a warning result specifically includes: When the average value exceeds a first threshold and the ratio exceeds a preset ratio threshold, a first-level warning is triggered. When the average value exceeds the second threshold and the ratio exceeds the preset ratio threshold, a second-level warning is triggered. The second threshold is greater than the first threshold.

9. The method for identifying precursors of rockfalls based on acoustic and vibration signals and self-supervised representation learning according to claim 8, characterized in that, Following step S4, an incremental learning and update step for the neural network model is also included: The inference results of the deployed neural network model on real-time data are correlated with the feedback information on whether a collapse has occurred, forming new labeled samples; When the number of newly labeled samples reaches a preset threshold or a fixed update time interval, the model parameters are updated using an elastic weight consolidation strategy.

10. The method for identifying precursors of rockfalls based on acoustic and vibration signals and self-supervised representation learning according to claim 9, characterized in that, The method of updating model parameters using an elastic weight consolidation strategy specifically includes: Get newly labeled samples; Obtain the labeled samples used in the fine-tuning stage in step S06, assign a higher sampling weight to the samples labeled as precursor states in the labeled samples used in the fine-tuning stage than to the samples in normal states, calculate the Fisher information matrix based on the weighted samples, and use the Fisher information matrix to measure the importance of each parameter of the neural network model to the task of identifying the state of dangerous rock mass. Based on the contribution of each network layer in the encoder to the identification results of the precursor state of unstable rock mass, different interlayer amplification factors are set for different network layers, with the amplification factor of deep convolutional layers being higher than that of shallow convolutional layers. Model training is performed based on the newly labeled samples. The product of the Fisher information matrix and the inter-layer amplification coefficient is used as the regularization constraint strength. Penalty terms are applied to parameters whose importance is higher than a preset threshold to constrain their deviation from historical values ​​during training and complete the model parameter update.

Images