Rock mass stability detection method and device, medium and equipment

Abstract

The invention belongs to the technical field of geological engineering and geotechnical engineering, and particularly discloses a rock mass stability detection method and device, a medium and equipment, and the method comprises the steps: collecting a monitoring signal of a to-be-detected rock mass; the monitoring signal is preprocessed; constructing a rock mass stability detection model, and training the model; and inputting the preprocessed monitoring signal into the trained rock mass stability detection model so as to detect the stability of the rock mass to be detected. The rock mass detection precision can be improved.

Description

Technical Field

[0001] The present application belongs to the field of geological engineering and geotechnical engineering technology, and specifically relates to a rock stability detection method, device, medium and equipment. Background Art

[0002] Traditional rock stability assessment methods mainly rely on empirical formulas, field surveys, and laboratory tests. These methods are often time-consuming and costly, and it is difficult to monitor rock changes in real time. Summary of the Invention

[0003] In response to the deficiencies in the prior art, the purpose of this application is to provide a rock stability detection method, device, medium and equipment. This application aims to provide a more accurate, timely and highly adaptable rock stability detection method.

[0004] To achieve the above objectives, this application provides the following technical solutions:

[0005] A rock mass stability detection method comprises: collecting monitoring signals of a rock mass to be detected, the monitoring signals including acoustic emission signals, displacement signals, stress and strain signals, inclination signals, and temperature and humidity signals around the rock mass to be detected; preprocessing the monitoring signals; constructing a rock mass stability detection model and training the model; the rock mass stability detection model introduces a multimodal data access layer to simultaneously process the monitoring signals; the multimodal data access layer also introduces a self-attention mechanism to dynamically calculate the importance of multimodal features at each time point; and inputting the preprocessed monitoring signals into the trained rock mass stability detection model to detect the stability of the rock mass to be detected.

[0006] Optionally, the monitoring signal is preprocessed, including: noise filtering, feature extraction and event identification of the acoustic emission signal; trend removal, outlier detection and correction of the displacement signal; temperature compensation and nonlinear correction of the stress-strain signal and incremental cumulative effect analysis; gravity compensation and posture fusion of the inclination signal and dynamic response adjustment.

[0007] Optionally, the rock stability detection model includes: an input unit for inputting the preprocessed monitoring signal; a feature extraction unit for extracting features from the input preprocessed monitoring signal; and an output unit for outputting a stability score of the rock mass to be tested based on the features extracted by the feature extraction unit.

[0008] Optionally, the input unit includes: a multimodal data access layer and a data integration and transmission layer, wherein the multimodal data access layer is used to standardize and clock synchronize the accessed monitoring signals to obtain multimodal data; and the data integration and transmission layer is used to fuse the multimodal data.

[0009] Optionally, the feature extraction unit includes: a spatial feature extraction module, a time series module and a physical rule embedding module connected in sequence.

[0010] Optionally, the rock stability detection model is trained through the following steps: obtaining historical monitoring data of the rock mass and preprocessing it, and dividing the preprocessed historical monitoring data into a training set and a validation set in proportion; setting training parameters, and using the training set to train the model. During the model training process, when the number of model iterations meets the preset value, the model training is completed; using the validation set to verify the trained model. During the verification process, if the mean absolute error, mean square error, and root mean square error, which are model performance evaluation indicators, are all less than the threshold, the model verification passes; otherwise, adjust the training parameters or expand the training set samples to retrain the model until the verification passes.

[0011] The present application also provides a rock stability detection device, which includes: an acquisition module for collecting monitoring signals of the rock mass to be tested, wherein the monitoring signals include acoustic emission signals, displacement signals, stress and strain signals, inclination signals, and temperature and humidity signals around the rock mass to be tested; a preprocessing module for preprocessing the monitoring signals; a model construction and training module for constructing a rock stability detection model and training the model; and a prediction module for inputting the preprocessed monitoring signals into the trained rock stability detection model to detect the stability of the rock mass to be tested.

[0012] Optionally, the preprocessing module includes: a first preprocessing submodule, which is used to filter noise, extract features and identify events on the acoustic emission signal; a second preprocessing submodule, which is used to remove trends, detect and correct outliers on the displacement signal; a third preprocessing submodule, which is used to perform temperature compensation and nonlinear correction on the stress-strain signal and perform incremental cumulative effect analysis; and a fourth preprocessing submodule, which is used to perform gravity compensation and posture fusion on the inclination signal and perform dynamic response adjustment.

[0013] The present application also provides a storage medium comprising instructions, which, when executed on a computer, enables the computer to execute the method as described in any of the preceding items.

[0014] The present application also provides an electronic device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements any of the above methods when executing the program.

[0015] Compared with the prior art, the present invention has the following beneficial effects:

[0016] This application introduces a multimodal data access layer and a data integration and transmission layer to achieve high-precision, real-time data acquisition and standardized processing from various sensors and environmental variables. A self-attention mechanism ensures temporal synchronization and global alignment of multi-source data, and dynamically adjusts the importance weights of each feature. Furthermore, this application combines physical rule embedding with intelligent diagnostic models to not only improve the accuracy and reliability of rock stability assessments but also enhance the system's interpretability and adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flow chart of a rock mass stability detection method provided by one embodiment of the present application;

[0018] Figure 2 is a structural diagram of a bunker stability detection model provided by another embodiment of the present application;

[0019] Figure 3 It is a structural schematic diagram of a rock stability detection device provided in another embodiment of the present application. DETAILED DESCRIPTION

[0020] Specific embodiments of the present application will be described in detail below with reference to the accompanying drawings. Although specific embodiments of the present application are shown in the accompanying drawings, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments described herein. Rather, these embodiments are provided to enable a more thorough understanding of the present application and to fully convey the scope of the present application to those skilled in the art.

[0021] Figure 1 FIG. 1 is a flow chart of a rock mass stability detection method provided by an exemplary embodiment of the present application. Figure 1 As shown, the method includes the following steps:

[0022] S100: Acoustic emission signals, displacement signals, stress and strain signals, inclination signals of the rock mass to be measured, and temperature and humidity signals around the rock mass to be measured are collected using an acoustic emission sensor, a laser rangefinder, a fiber Bragg grating (FBG) sensor, an inclinometer, and a temperature and humidity sensor; S200: Preprocessing the acoustic emission signals, displacement signals, stress and strain signals, inclination signals, and temperature and humidity signals; S300: Constructing a rock mass stability detection model and training the model; S400: Inputting the preprocessed acoustic emission signals, displacement signals, stress and strain signals, inclination signals, and temperature and humidity signals into the trained rock mass stability detection model to detect the stability of the rock mass to be measured.

[0023] In another exemplary embodiment, in step S200, pre-processing the monitoring signal includes the following steps:

[0024] S201: Preprocessing the acoustic emission signal, specifically including:

[0025] S2011: performing noise filtering on the acoustic emission signal;

[0026] In this step, the acoustic emission signal may be subjected to noise filtering by any of the following methods, for example, frequency domain filtering (converting the time domain signal to the frequency domain through Fourier transform, then using a bandpass filter to remove components outside the expected frequency range, and then converting the signal back to the time domain through inverse Fourier transform), adaptive filtering (using an adaptive filtering algorithm, such as the LMS (least mean square) algorithm or the RLS (recursive least squares method), which can dynamically adjust the filter parameters according to environmental changes to better eliminate background noise), and wavelet transform (using the wavelet transform with multi-resolution analysis characteristics to effectively separate signals and noise at different scales, and selecting appropriate wavelet basis functions and threshold processing to achieve denoising).

[0027] S2012: performing feature extraction and event recognition on the acoustic emission signal after noise filtering;

[0028] In this step, feature extraction is the extraction of information that can represent the characteristics of potential events from the acoustic emission signal that has been preprocessed (such as noise filtering). These features can be time domain features, such as amplitude, duration, rise time, etc.; they can also be frequency domain features, such as center frequency, bandwidth, and spectral energy distribution. In addition to time domain and frequency domain features, wavelet transform can also be used to obtain features at multiple resolutions, which helps to capture information at different scales. In addition, high-order statistics or model-based methods (such as AR model parameters) can also be used as features. Once the features are extracted, the next step is to identify specific events based on these features. This can be done by setting threshold rules. For example, when one or more features exceed a predefined threshold, it is considered that a specific type of event has occurred.

[0029] S202: Preprocessing the displacement signal, specifically including:

[0030] S2021: performing trend removal on the displacement signal;

[0031] In this step, the long-term average displacement change rate is first calculated, and a linear or nonlinear trend model is constructed based on it. Secondly, the trend model is subtracted from the original displacement data to obtain the net displacement change series after removing the long-term trend.

[0032] Specifically, calculating the long-term average displacement rate of change involves: first, determining an appropriate window period for calculating the long-term average displacement rate of change. The choice of this window period depends on the characteristics of the displacement data in the specific application scenario, such as the type of rock mass, the detection time span, and the expected frequency of geological activity. For rock mass stability testing, data spanning several days to several weeks can be selected as the window. Secondly, cumulative displacement is calculated, that is, the cumulative displacement at each time point is calculated based on the original displacement data. Then, using a sliding window technique, the average cumulative displacement corresponding to each time point is calculated within the selected time window (such as 7 days or 30 days). This will produce a series of average displacement values ​​that represent the long-term trend. Finally, linear regression can be applied to the entire detection period or each fixed-size sliding window to fit a straight line representing the long-term trend. The slope of this line is the long-term average displacement rate of change.

[0033] Based on the long-term average displacement change rate calculated above, we need to build a mathematical model that can describe these long-term change trends. A simple linear model can be used here, namely: ,in, represents the long-term average displacement change rate, represents the intercept term.

[0034] The last step is to subtract the result of the trend model prediction from the original displacement data to obtain the net displacement change sequence after removing the influence of the long-term trend. The direct subtraction method can be used, that is, for each time point , subtract the trend model prediction value corresponding to that time point from the original displacement value, and the result is the net displacement change after detrending .

[0035] Through the above steps, the short-term fluctuation components in the displacement signal can be effectively separated, thereby identifying potential safety risks and evaluating rock stability.

[0036] S2022: Detecting and correcting abnormal points on the displacement signal after trend removal;

[0037] In this step, statistical methods (such as the Z-score method and box plot rule) can be used to identify and mark data points that significantly deviate from the normal range. For data points marked as abnormal, interpolation or neighboring point replacement can be used for correction.

[0038] S203: Preprocessing the stress-strain signal, specifically including:

[0039] S2031: Performing temperature compensation and nonlinear correction on the stress-strain signal;

[0040] In this step, because the physical properties of the rock mass (such as elastic modulus, thermal expansion coefficient, etc.) change with temperature, the measured strain will be different under different temperature conditions even if the same stress is applied. In order to eliminate the influence of this temperature effect, temperature compensation must be implemented, including:

[0041] First, we need to understand the temperature dependence of the rock mass, which can be obtained through laboratory testing. For example, for some rocks or concrete rock masses, its elastic modulus can be obtained through experiments. and thermal expansion coefficient With temperature Secondly, according to the thermal expansion coefficient of the rock mass , it can be calculated that due to temperature change Thermal strain caused , which is the part of the strain caused purely by temperature change without involving external forces. Finally, the above thermal strain is deducted from the raw strain reading to obtain the corrected mechanical strain , thus eliminating the influence of temperature factors on strain measurement.

[0042] Rock masses exhibit significant nonlinear behavior during loading, meaning the stress-strain relationship is not a simple linear proportionality. As the load increases, the rock mass undergoes complex phenomena such as plastic deformation and crack propagation, which render traditional linear assumptions inapplicable. Therefore, appropriate mathematical models are needed to describe this nonlinear stress-strain response and to modify the original readings accordingly.

[0043] Specifically, first, according to the characteristics of the rock mass and the expected deformation mode, a suitable constitutive model can be selected, such as the hyperbolic model, the power law model, the Mohr-Coulomb criterion or other more complex elastic-plastic models. Each model has its own specific set of parameters to characterize different properties of the rock mass. Secondly, the parameters of the selected nonlinear model are estimated using experimental data under known conditions (such as uniaxial compression tests, triaxial compression tests). This step can be completed by the least squares method or other optimization algorithms, with the aim of finding a set of optimal parameters so that the model predictions are as close as possible to the measured values. Finally, once the nonlinear model and its parameters are determined, they can be applied to the original stress-strain readings to achieve nonlinear correction. Specifically, if the original readings are and , then the “real” stress can be recalculated based on the nonlinear model or strain For example, in a typical elastic-plastic model, after exceeding the yield point, the stress will no longer increase according to the initial elastic modulus, but will follow a more gentle slope; at this time, the stress-strain expression provided by the model needs to be used for correction.

[0044] S2032: performing incremental cumulative effect analysis on the stress-strain signal after temperature compensation and nonlinear correction;

[0045] This step analyzes the amount of permanent deformation remaining after each loading and unloading cycle and calculates the cumulative effect of the increments. This helps us understand the plastic deformation process of the rock mass and its contribution to its overall stability. Through differential calculations, we obtain the net deformation increments within successive loading cycles, providing more accurate basic data for subsequent modeling.

[0046] S204: Preprocessing the tilt signal, specifically including:

[0047] S2041: Performing gravity compensation and posture fusion on the tilt signal;

[0048] In this step, the gravity compensation algorithm is used to correct the tilt measurement error caused by the earth's gravitational field. Specifically, it includes: first, ensuring that the inclinometer used (such as an accelerometer or gyroscope) has been accurately calibrated to reduce inherent hardware errors. Secondly, the static component caused by gravity is separated using the data of the three-axis accelerometer. By calculating the sum of the three axial acceleration vectors and normalizing it to the gravitational acceleration g, the component of gravity on each axis is obtained. Finally, the angular deviation caused by gravity is subtracted from the original tilt angle reading. For example, if the accelerometer detects a non-zero z-axis acceleration value, it means that the device is not placed completely horizontally, and the tilt angles on the x and y axes need to be adjusted based on this information.

[0049] In addition, multiple sensor data (such as accelerometer, gyroscope, magnetometer, etc.) can be combined to provide more accurate attitude estimation.

[0050] S2042: Dynamically adjusting the inclination signal after gravity compensation and posture fusion, specifically including:

[0051] Frequency response optimization: Based on the motion characteristics in different frequency ranges, the filter parameters are adjusted so that the system can quickly respond to high-frequency vibrations while filtering out unnecessary low-frequency interference.

[0052] Adaptive filtering: Adaptive filtering algorithms (such as LMS or RLS) are introduced. These algorithms can automatically adjust the filter coefficients according to the actual input signal to better track dynamically changing postures.

[0053] Inertial Navigation System Assistance: For high-speed movement or severe shaking, the short-term high-precision positioning and orientation information provided by the Inertial Navigation System (INS) can help stabilize the tilt measurement results.

[0054] Temperature compensation: Considering that temperature changes may affect the performance of sensors, especially accelerometers and gyroscopes, temperature compensation measures are implemented to ensure good measurement accuracy in various environments.

[0055] Delay Correction: Identify and correct measurement lags caused by signal processing delays, ensuring that the output tilt data reflects the current true status in a timely manner.

[0056] Through the above steps, not only can the influence of the gravity field on the inclination measurement be effectively eliminated, but also high measurement accuracy can be maintained in complex and changeable actual application scenarios, providing reliable data support for rock stability detection.

[0057] In another exemplary embodiment, Figure 2As shown, the rock stability detection model includes: an input unit for inputting the preprocessed monitoring signal; a feature extraction unit for extracting features from the input preprocessed monitoring signal; and an output unit for predicting the stability of the rock mass to be tested based on the features extracted by the feature extraction unit.

[0058] In this embodiment, the input unit includes a multimodal data access layer and a data integration and transmission layer. The multimodal data access layer first defines a unified data access standard for each input signal. Specifically, each sensor is connected to an independent data receiving node, which is responsible for reading the input data and converting it into a standardized digital format. After completing the standardization conversion, the converted signals are further globally synchronized to ensure that the timestamps of all sensors are aligned, allowing data from different sensors to be compared and analyzed for correlation at the same time point. Furthermore, the multimodal data access layer incorporates a self-attention mechanism. This mechanism dynamically calculates a weight vector for each multimodal feature at each time point based on the current time point and specific environmental conditions. This weight vector reflects the relative importance of each sensor signal at that specific moment, allowing the model to more accurately capture changes in rock mass conditions. Specifically, the self-attention mechanism automatically learns and assigns appropriate weights to each input signal by evaluating the interrelationships between sensor signals and their correlations with environmental factors. These weights adjust accordingly as time and environmental conditions change, ensuring that the model always focuses on the most relevant data sources, improving the accuracy and reliability of predictions. For example, in some cases, acoustic emission signals may better reflect potential instability than displacement signals; at other times, changes in humidity or temperature may have a more significant impact on the rock mass. Therefore, the presence of the self-attention mechanism gives the model greater adaptability and higher sensitivity. Furthermore, to further enhance the robustness of the model, this application also incorporates acquired environmental variables (temperature and humidity signals) as auxiliary information into the self-attention mechanism. These environmental variables are processed together with the raw sensor data to ensure that the model maintains good performance under different geological conditions and can respond promptly to any abnormalities.

[0059] The data integration and transmission layer includes a feature vector assembly layer and a data cache and streaming processing layer, wherein the feature vector assembly layer is used to combine the acoustic emission signal, displacement signal, stress strain signal, inclination signal, and temperature and humidity signals that have been pre-processed into a comprehensive feature vector, and serve as the final input of the rock stability detection model. Specifically, the feature vector assembly layer includes a multimodal fusion module, a feature alignment module, and a feature splicing module, wherein the multimodal fusion module includes a linear transformation layer, an embedding layer, and a fully connected layer connected in sequence, and each sensor signal is adjusted to a unified dimension through the linear transformation layer. ,Right now: ,in, Indicates the The original feature vector of a sensor (such as acoustic emission sensor, displacement sensor, stress strain sensor or tilt sensor) at a certain time point; W i Represents a linear transformation matrix used to transform the features of different sensors into a unified dimension ; After linear transformation, The feature vector of each sensor; Represents the acoustic emission signal, represents the displacement signal, represents the stress-strain signal, Indicates temperature.

[0060] For environmental variables, they are converted through the embedding layer, and the converted environmental variables are combined with the transformed sensor signals through the fully connected layer.

[0061] The feature alignment module includes a time series alignment layer, a spatial relationship modeling layer, and a local-global interaction layer connected in sequence, wherein the time series alignment layer can ensure that the timestamps of all sensors are consistent. The time series alignment layer includes an input layer, a feature extraction layer, a dynamic time warping layer, an adaptive interpolation module, a noise suppression and anomaly detection module, and an output layer connected in sequence, wherein the input layer is used to receive time series data sampled at different frequencies from multiple sensors. And the corresponding timestamp The feature extraction layer uses a one-dimensional convolutional neural network to capture the local and global specific features of each time series. , output a series of feature vectors ,in, Represents the feature dimension. The dynamic time warping layer finds the best matching path by calculating the DTW distance matrix between any two time series to help determine how to adjust different time series to achieve synchronization. The adaptive interpolation module is used to input the feature vector , and use the multi-head attention mechanism to calculate the similarity weight matrix between different time series , and output the updated feature representation , specifically expressed as:

[0062]

[0063] in, 、 、 represents the learnable parameter matrix, 、 、 represents query, key, and value vectors, 、 Respectively represent Hedi The original feature representation of a time series, Represents the key vector dimension.

[0064] The noise suppression and anomaly detection module introduces an autoencoder, which can learn the distribution pattern of normal signals, thereby identifying and filtering out abnormal points or noise components in the signal.

[0065] The output layer includes a fully connected layer and an activation function (such as ReLU) to output the updated feature representation , thus obtaining the final aligned time series .

[0066] The spatial relationship modeling layer is used to capture possible spatial correlations between sensors, especially when the sensors are distributed over a large area. Specifically, assuming there are n sensors distributed at different locations in the rock mass, forming an undirected graph with n nodes, where each node represents the location of a sensor and the edge weights reflect the physical distance or mechanical interaction strength between two sensors. The spatial relationship modeling layer can be used to capture possible spatial correlations between sensors. The spatial relationship modeling layer is specifically expressed as:

[0067]

[0068] in, Indicates the The node feature matrix of the layer, represents the adjacency matrix plus self-loops ( , represents the original adjacency matrix, which represents the connection relationship between sensors. is the identity matrix, ensuring that each node is connected to itself), represents the degree matrix, which indicates the number of neighbors of each node, Indicates the The weight matrix of the layer is used to control the importance of information transmission. represents the activation function, Indicates the Node feature matrix after layer update.

[0069] The local-global interaction layer includes a multi-scale perception module, a graph construction and embedding module, a self-attention mechanism enhancement module, a cross-scale fusion module and a dynamic context aggregation module. The multi-scale perception module performs parallel convolution operations by using multiple convolution kernels of different sizes (for example, 3×3, 5×5, 7×7), and a batch normalization layer and ReLU activation function are set after each convolution layer. Finally, the results of all convolution layers are merged into a comprehensive feature representation by splicing or adding.

[0070] The graph construction and embedding module can construct an undirected graph based on the physical location of the sensor. , where the node Indicates sensor, edge Represents the connection weight between nodes.

[0071] The self-attention enhancement module introduces local and global attention mechanisms to strengthen the connection between local features and global context. The local attention mechanism applies self-attention to the features of each node, calculates the importance of nodes in its neighborhood, and updates the node features accordingly. The global attention mechanism applies self-attention to the entire undirected graph, allowing each node to consider information from all other nodes.

[0072] The cross-scale fusion module is used to fuse feature maps from different scales to ensure that the model can learn both fine-grained and coarse-grained spatial patterns. The cross-scale fusion module is expressed as:

[0073]

[0074] in, represents the final feature map after fusion, Indicates the The weight corresponding to the scale feature map, Indicates the Feature maps at different scales, Represents the original input feature map used in the residual connection or the feature map after a simple transformation, which is used to prevent the original information from being completely covered.

[0075] The cross-scale fusion module integrates feature information at different scales, enabling the model to capture various patterns in the rock mass, from tiny cracks to large-scale deformation. That is, the multi-scale feature fusion enables the model to not only identify subtle local changes (such as early crack propagation and displacement anomalies), but also understand the changing trends of the overall rock mass structure (such as cumulative displacement).

[0076] The dynamic context aggregation module introduces a gated unit (GRU) to automatically adjust the ratio of local and global features according to the current task requirements, so that the final output feature representation It contains rich local details while retaining global background information.

[0077] The feature splicing module includes a direct splicing layer and a depth splicing layer, wherein the direct splicing layer is used to sequentially splice all preprocessed feature vectors to form a long vector, which is specifically expressed as: ,in, 、 、 、 、 Respectively represent the acoustic emission, displacement, stress strain, inclination and environmental characteristic vectors after preprocessing, Represents a long vector formed by concatenating the above feature vectors in sequence.

[0078] The deep splicing layer uses the embedding layer in deep learning to create a specific embedding representation for each type of sensor and then splice them together. Specifically, it is expressed as:

[0079]

[0080]

[0081]

[0082]

[0083]

[0084]

[0085] in, Represents the embedding layer, which maps high-dimensional sparse features to low-dimensional dense space to better capture the similarities and differences between features. 、 、 、 and They are the representations of acoustic emission, displacement, stress and strain, inclination and environmental characteristics after being transformed by the embedding layer. Represents the comprehensive feature vector after deep concatenation, including the embedded representation of each type of sensor.

[0086] The direct splicing layer simply concatenates all pre-processed feature vectors in sequence to form a long vector, which can retain the complete information of the original features and is suitable for tasks that need to maintain the relative position relationship between features. The deep splicing layer uses a more complex embedding representation method to create a specific embedding representation for each type of sensor and splices them together, so that the model can learn the deep correlation between different types of signals, thereby enhancing the understanding of changes in rock state. Using a combination of these two splicing methods, the model can not only rely on the data of a single sensor, but also capture the interaction between multimodal data. For example, through the direct splicing layer, the model can directly access the original features provided by each sensor, while through the deep splicing layer, it can mine the high-level abstract features hidden behind these features. This two-layer structure enables the model to perform feature fusion at multiple levels, improving the quality of the final output results.

[0087] In summary, the feature vector assembly layer achieves time synchronization, global clock alignment and self-attention mechanism weighting by fusing multimodal data and combining environmental variables, ensuring data consistency and accurate capture of rock state changes; at the same time, this layer uses spatial relationship modeling and local-global interaction enhancement technology to understand the intrinsic spatial correlation and overall deformation trend of the rock mass, and ensures the quality of input data and the real-time response capability of the system through noise suppression, anomaly detection, lightweight design and streaming processing, ultimately providing high-quality, highly interpretable and computationally efficient feature representation for rock stability detection.

[0088] The data cache and streaming processing layer is used to establish a data cache area to temporarily store newly arrived but unprocessed data, support streaming processing mode, and enable the system to respond to the latest detection results in a timely manner. Specifically, the data cache and streaming processing layer first uses an efficient FIFO memory queue to temporarily store newly arrived but unprocessed sensor data (i.e., the acoustic emission signal, displacement signal, stress strain signal, inclination signal, and environmental variables mentioned above). Each new data point When data enters the system, it is added to the end of the queue. When data needs to be processed, the earliest arriving data point is taken from the head of the queue for processing. To ensure data security and long-term storage needs, the data caching and streaming processing layer also provides disk-level persistent storage options (for example, using a database (such as SQLite, MySQL) or a file system (such as CSV, Parquet format) to regularly back up data in memory to the database or file system.

[0089] Next, the stream processing engine uses a micro-batch processing method to divide the continuously flowing data into small batches (each small batch contains a certain number of time point data, for example, data collected per second is regarded as a batch, and the size of the small batch can be flexibly adjusted according to the throughput and response time requirements of the system) to balance real-time response and computing efficiency, and maintains a fixed-size time window through a sliding window mechanism (the window sliding step can be fixed or dynamically adjusted according to actual conditions. For example, a 5-minute sliding window is set, which slides forward once a minute to update the latest 5 minutes of data) to continuously update the latest data segments. The data cache and stream processing layer also introduces an incremental update strategy, which ensures that as new data arrives, only the necessary parts are updated, rather than recalculating the entire feature vector (that is, for each new time point t, only the features related to that time point are updated, while other unchanged parts are retained), saving computing resources. At the same time, to deal with emergencies, the data cache and streaming processing layer also sets up a fast channel to prioritize key events, and integrates online anomaly detection algorithms (for example, using statistical methods such as the Z-score method and machine learning models such as the Isolation Forest) to monitor data streams in real time and immediately trigger an alarm mechanism when anomalies are found.

[0090] Finally, the data caching and streaming processing layer also introduces a model inference interface. When enough data is ready, inference requests are managed by using a message queue (such as RabbitMQ, Kafka) or a task scheduler (such as Celery) to avoid blocking the main data flow.

[0091] The above design not only enhances the robustness and flexibility of the model, but also ensures the efficient operation and timely response capabilities of the rock stability detection model, providing decision makers with reliable data support and early warning information.

[0092] The feature extraction unit includes a spatial feature extraction module, a time series module and a physical rule embedding module. Below, this application describes the network structure of each module in detail.

[0093] The spatial feature extraction module comprises a sequentially connected multi-scale convolutional network, an enhanced convolutional layer, an adaptive activation function, a spatial pyramid layer, an attention mechanism layer, and a lightweight fully connected layer. The multi-scale convolutional network first captures local spatial correlations at different scales using convolutional kernels of various sizes (e.g., 3×3, 5×5, and 7×7), ensuring that the model can simultaneously learn both fine-grained (e.g., small cracks) and coarse-grained (e.g., large-scale deformation) spatial information. Furthermore, the model is combined with dilated convolution to expand the receptive field, further enhancing its ability to understand complex geological structures. The enhanced convolutional layer utilizes depthwise separable convolution and a dynamic convolution kernel generation mechanism (a technique used in convolutional neural networks (CNNs) to improve model flexibility and adaptability. This mechanism allows the convolution kernel (i.e., filter or weight matrix) to be dynamically adjusted based on input data or contextual information, rather than remaining fixed throughout training). This not only reduces computational effort and the number of parameters, but also improves the model's adaptability to diverse scenarios, enabling the model to adaptively adjust its internal structure and parameter settings based on changes in actual geological conditions and physical context, thereby better understanding and predicting rock mass behavior. The spatial pyramid pooling layer extracts spatial features at different levels through multi-level pooling operations (such as max pooling and average pooling). This effectively collects information from the entire input region, helping the model understand the relationship between global context and local details and increasing robustness to scale changes. The attention mechanism layer introduces self-attention or channel attention mechanisms, enabling the model to focus on the most representative feature regions, strengthen the importance of important features, and filter out irrelevant noise, thereby improving the quality of feature representation, especially when identifying early signs of instability. The lightweight fully connected layer uses sparse connections or factorization techniques to reconstruct the fully connected layer, reducing the number of parameters while maintaining sufficient expressive power, preventing overfitting, and optimizing computing resource utilization.

[0094] Based on the above structure, the spatial feature extraction module can extract high-level features from rock monitoring data, including but not limited to micro-crack extension, stress and strain changes, displacement trends, etc., which can not only enhance the model's understanding of local details and overall background information, but also improve the accuracy and reliability of the model's prediction results.

[0095] The time series module includes a multi-scale temporal enhancement layer, an adaptive attention mechanism layer, a graph neural network layer, and a dynamic gating unit, which are connected in sequence. The time series module also uses residual connections and skip channels. Specifically, the multi-scale temporal enhancement layer is used to capture features at different time scales to ensure that the model can understand short-term and long-term dependencies. Specifically, it includes:

[0096] Multi-resolution decomposition: Wavelet transform or empirical mode decomposition (EMD) is used to decompose the original time series data into multiple subsequences with different frequencies.

[0097] Independent processing path: Each subsequence is processed by an independent bidirectional recurrent neural network (Bi-RNN) to generate the corresponding hidden state representation.

[0098] Fusion layer: Use weighted average or convolution operations to fuse the hidden states of all subsequences together to form the final multi-scale feature representation.

[0099] The adaptive attention mechanism layer is used to dynamically adjust the attention weight to highlight important time points. The adaptive attention mechanism layer introduces environmental variables (such as temperature and humidity) as additional inputs and uses them together with time series features to generate queries. ,key Sum , expressed as: , , ,in, Indicates time The hidden state of Indicates time A vector of environmental variables, 、 、W v Represents the linear transformation matrix, which is used to generate queries ,key Sum , Indicates that the state will be hidden and environment variables Concatenate into a longer vector as the input of the query.

[0100] The adaptive attention mechanism layer combines the context-aware query to calculate the importance weight of each time point , expressed as:

[0101]

[0102] in, Represents the query vector at time The expression, Represents the key vector at time The expression, Represents the dimension of the key vector, used to scale the dot product result to stabilize the gradient, represents the Kullback-Leibler divergence.

[0103] The graph neural network layer is used to model the spatial correlation between sensors to capture complex spatial-temporal interactions.

[0104] First, each sensor on the rock mass is regarded as a node, and the edge weight is defined according to the physical distance or mechanical interaction strength between them to form a spatiotemporal graph. ; Secondly, a graph convolutional network (GCN) layer is applied to update the state of each node to capture the spatial correlation between sensors; finally, the node features processed by GCN are fused with the hidden state output by Bi-RNN to form a comprehensive feature representation, which is expressed as:

[0105]

[0106] in, Represents multi-scale features, Represents the graph feature representation after GCN processing, Represents a fusion operation, which is used to integrate spatiotemporal features.

[0107] The dynamic gating unit is used to control the information flow, selectively retain useful information, and filter noise.

[0108] The physical rule embedding module includes an adaptive loss function layer, a physical perception regularization layer, a meta-learning layer, an interpretability enhancement layer and a physical rule embedding network connected in sequence.

[0109] The adaptive loss function layer is used to construct a loss function that can dynamically adjust weights based on data characteristics to ensure that the model can effectively comply with physical rules in different scenarios. The adaptive loss function layer specifically includes a basic loss term, a physical constraint term, and an adaptive weight mechanism. The basic loss term is used to measure the difference between the model's predicted value and the true label to ensure that the model can fit the observed data well. The basic loss term is expressed as follows:

[0110]

[0111] The physical constraint term adopts, for example, the Mohr-Coulomb criterion to ensure that the model output complies with the basic laws of rock mechanics. The physical constraint term is expressed as follows:

[0112]

[0113] in, represents the physical constraint loss term based on the Mohr-Coulomb criterion, ensuring that the model output conforms to the basic laws of rock mechanics. Represents the weight coefficient, which is used to control the importance of the physical constraint term. By adjusting this coefficient, a balance can be found between data fitting and physical consistency. represents the maximum value of the principal stress, represents the minimum value of the principal stress, Indicates cohesion, indicating the cohesive strength of rock, Represents the internal friction angle, which is used to describe the friction characteristics of rock materials.

[0114] The adaptive weight mechanism can dynamically adjust the importance of physical constraints according to training progress, data characteristics or specific conditions, so that the model can better comply with physical rules in different stages and scenarios. Specifically, it includes time-dependent weights and conditional dependency weights , where the time-dependent weight gradually increases the influence of physical constraints as training progresses to guide the model to better follow physical rules. The time-dependent weight is expressed as:

[0115]

[0116] in, Represents the weight function over time, dynamically adjusting the importance of physical constraints as training progresses. Represents the initial weight, which defines the relative importance of the physical constraints in the early stages of training. Indicates the magnitude of weight change, which determines the speed at which the weight of the physical constraint item increases as training progresses. Represents a time function, which can be linear growth or exponential decay, and is designed according to specific needs.

[0117] The conditional dependency weight can dynamically adjust the importance of the physical constraint item according to the characteristics of the input data (such as stress level, environmental variables, etc.). The conditional dependency weight is expressed as:

[0118]

[0119] in, Represents the weight function since the condition, dynamically adjusts the importance of the physical constraint terms according to the characteristics of the input data (such as stress level, environmental variables, etc.), represents the input feature vector, which contains sensor data and environmental variables, Represents a parameterized function, which is learned through training and can be used according to the input features Dynamically adjust weights, Representation function A collection of parameters.

[0120] The total loss function is expressed as:

[0121]

[0122] in, represents the total loss function that comprehensively considers data fitting error and physical constraints, represents the data-driven base loss, ensuring that the model can fit the observed data well, Represents the weight function over time, dynamically adjusting the importance of physical constraints. represents the physical constraint loss term based on the Mohr-Coulomb criterion, The weight function that represents the conditional dependency adjusts the importance of the physical constraint terms according to the characteristics of the input data. Represents other possible physical consistency constraints, such as stress balance, energy conservation, etc.

[0123] The physical-aware regularization layer includes a physical consistency condition definition module, an adaptive regularization strength module, a regularization term construction module, and a feedback mechanism. The physical consistency condition definition module is used to define and calculate various physical rules (such as stress balance, energy conservation, and the Mohr-Coulomb criterion) to ensure that the model output conforms to the basic laws of rock mechanics. Specifically, the physical consistency condition definition module includes a stress analysis submodule for evaluating the mechanical behavior of rock mass under different stress states; an energy analysis submodule for evaluating energy changes within the system to ensure compliance with the law of energy conservation; and a failure criterion submodule for assessing the potential for failure of rock mass under different stress conditions, such as based on the Mohr-Coulomb criterion.

[0124] The adaptive regularization strength module can automatically adjust the regularization strength applied to the physical rules according to the specific requirements of the task and the characteristics of the input data, which helps to balance the strictness of the model for physical constraints, thereby ensuring the physical rationality of the prediction results and avoiding excessive restrictions that reduce the flexibility of the model. Specifically, first, for each physical rule , set an initial regularization weight ; Second, use the task difficulty or expected learning difficulty to initialize the learning rate of each task During the training process, based on the performance on the validation set (such as the prediction error and the degree of satisfaction of physical rules ), and input data characteristics such as stress levels and environment variables ), update the regularization weights:

[0125]

[0126] in, Represents a function that receives multiple input parameters. Represents the regularization weight for the current time step.

[0127] The regularization term building module is responsible for defining and calculating various physical consistency conditions, converting them into mathematical expressions and adding them to the loss function as regularization terms. These regularization terms are used to penalize prediction results that violate physical rules, thereby guiding the model output to be closer to real physical phenomena. Specifically, this module defines a set of physical consistency conditions, such as stress balance, energy conservation, Mohr-Coulomb criterion, etc. For each physical rule , construct the corresponding mathematical expression ,in, Represents the model parameters. All physical rule regularization terms and basic loss terms Combined to form the total loss function:

[0128]

[0129] in, is the regularization weight provided by the adaptive regularization strength module, represents the sum of all physical regularization terms.

[0130] The feedback mechanism is used to continuously monitor the performance of the model and continuously improve the model's behavior through a cyclical iterative process. This ensures that the model can maintain its physical consistency and predictive accuracy even when faced with new or unseen data. Specifically, the feedback mechanism implements the monitoring model output Whether it meets the preset physical conditions , and record the deviation If the deviation is found to be beyond the allowable range, the relevant parameters (such as regularization weights ) to make corrections, as shown below:

[0131]

[0132] in, represents the corrected regularization weight, Represents the scaling function.

[0133] The meta-learning layer includes a multi-task dataset, a fast adaptation mechanism, a task diversity generator, an adaptive learning rate mechanism, a task representation module, and a meta-optimizer. The multi-task dataset is responsible for preparing a dataset for meta-learning, including samples of multiple different tasks, specifically including:

[0134] Meta-training set : Contains multiple training tasks, each consisting of a small number of samples, designed to simulate the diversity and complexity of the real world;

[0135] Meta-validation set : Used to evaluate the performance of the model on new tasks, usually also contains a small number of samples to test the generalization ability of the model.

[0136] The rapid adaptation mechanism may adopt, for example, MAML (Model-Agnostic Meta-Learning) or the Reptile algorithm, which enables the model to quickly learn new physical rules from a small number of samples to improve its adaptability to new tasks.

[0137] The task diversity generator is used to generate tasks covering different geological conditions and physical backgrounds to enhance the generalization ability of the model.

[0138] The adaptive learning rate mechanism dynamically adjusts the learning rate based on task difficulty, ensuring effective model convergence across different tasks. Specifically, an initial learning rate is set for each task, based on the task's complexity or expected learning difficulty. As training progresses, the learning rate is dynamically adjusted based on the model's performance on the validation set (e.g., prediction error, compliance with physical rules, etc.). For example, if the model performs poorly on a particular task, the learning rate can be increased to accelerate convergence; otherwise, it can be decreased to refine the adjustment.

[0139] The task representation module is used to extract key task-related features from the input data (such as time series features of sensor readings, environmental variables, etc.), and use an encoder network (such as LSTM, Transformer, etc.) to convert these features into fixed-length task representation vectors to help the model better understand the similarities and differences between tasks.

[0140] The meta-optimizer is responsible for updating model parameters, enabling it to find optimal solutions across different tasks while maintaining good generalization. Specifically, it uses optimization algorithms suitable for meta-learning (such as Adam and SGD) combined with the results of the fast adaptation mechanism to update model parameters. By training on multiple tasks, the meta-optimizer can capture commonalities between different tasks, helping the model adapt to new tasks more quickly.

[0141] The interpretability enhancement layer includes a sequentially connected feature importance analysis module, a rule-driven learning path module, a local explanatory model module, and an uncertainty estimation module. The feature importance analysis module calculates the impact of each feature on the prediction results using SHAP values ​​or LIME, generating a feature importance chart. The rule-driven learning path module, incorporating expert knowledge, guides the learning of model parameters, ensuring that the model output conforms to the fundamental laws of rock mechanics. The local explanatory model module provides explanations for individual predictions using techniques such as LIME (which explains complex model predictions by fitting a simple, easily interpretable model (such as linear regression) to a local scale) or SHAP (which provides the specific contribution of each feature to the prediction, helping users understand the importance of different features). The uncertainty estimation module assesses the confidence of model predictions, identifies high-risk areas, and provides users with more decision-making support information. Specifically, the uncertainty estimation module can employ Bayesian Neural Networks (BNNs), which introduce probability distributions to represent weights and assess the uncertainty of model predictions.

[0142] The physical rule embedding network includes an input layer, a physical perception convolution layer, a physical consistency check layer, and an adaptive fusion layer, which are connected in sequence. The input layer is used to receive multimodal feature vector sequences, such as sensor data (acoustic emission, displacement, stress and strain, and tilt signals) and environmental variables (temperature, humidity, etc.).

[0143] The physics-aware convolution layer introduces convolution operations with physical constraints to ensure that the design of the convolution kernel and the weight update process take into account the basic laws of rock mechanics, thereby capturing spatial patterns related to physical rules and improving the accuracy of model prediction. The physics-aware convolution layer can be expressed as follows:

[0144]

[0145] in, represents the convolution kernel with physical constraints, represents the bias term, represents the convolution operation, represents a multimodal feature vector sequence, Represents the activation function.

[0146] The physical consistency check layer is used to verify in real time whether the model output meets the preset physical conditions, such as stress balance, energy conservation and Mohr-Coulomb criterion, and automatically correct the output when necessary to ensure the physical rationality of the prediction results. The adaptive fusion layer is used to dynamically adjust the importance of different source features and flexibly respond to different geological conditions according to specific task requirements, thereby enhancing the generalization ability and adaptability of the model, so that the model can flexibly respond to different geological conditions. The physical consistency detection layer can be expressed as follows:

[0147]

[0148] in, represents the correction term calculated according to physical rules.

[0149] The adaptive fusion layer is used to dynamically adjust the importance of features from different sources and flexibly respond to different geological conditions according to specific task requirements, thereby enhancing the generalization and adaptability of the model. The adaptive fusion layer is represented as:

[0150]

[0151] in, Indicates the Characteristics in time The weight of Indicates feature representations from different sources, Indicates the number of feature sources.

[0152] The output unit includes a regressor that directly outputs a continuous value representing the stability score of the rock mass. The regressor is expressed as:

[0153]

[0154] in, represents the input feature vector, represents the model parameters, Represents a regression function.

[0155] Below, this application provides an illustrative example of the above technical solution through the specific data collected.

[0156] The data collected includes:

[0157] Acoustic emission signal (AE Signal):

[0158] Peak amplitude: 50 dB, duration: 2 ms, center frequency: 150 kHz;

[0159] Displacement Signal:

[0160] X-axis displacement: +3.2 mm, Y-axis displacement: -1.8 mm, Z-axis displacement: +0.5 mm;

[0161] Stress-Strain Signal:

[0162] Normal stress: 4 MPa, shear stress: 2 MPa, strain rate: 0.001 s^-1;

[0163] Tilt Signal:

[0164] X-axis tilt angle: 0.7°, Y-axis tilt angle: -0.3°;

[0165] Environmental Variables:

[0166] Temperature: 18℃, humidity: 65%.

[0167] After preprocessing the above data and inputting it into the model, the following prediction results can be obtained:

[0168] The rock mass stability score is 0.87 (ranging from 0 to 1, with larger values ​​representing more stability).

[0169] The above data means that the rock mass to be tested is assessed to be relatively stable, but still requires continuous monitoring.

[0170] In another exemplary embodiment, in step S300, the rock mass stability detection model is trained by the following steps:

[0171] S301: Obtain and preprocess historical monitoring data of the rock mass, and divide the preprocessed historical monitoring data into a training set and a validation set in a ratio of 7:3;

[0172] In this step, the historical monitoring data is preprocessed according to the above-mentioned preprocessing method, which will not be described in detail here.

[0173] S202: Set training parameters. For example, booster selects the tree-based model gbtree by default, sets learning_rate to 0.01, max_depth to 3, and the maximum number of iterations to 300. The model is trained using the training set. During the model training process, when the number of model iterations meets the preset value, the model training is completed.

[0174] S203: Use the validation set to validate the trained model. During the validation process, if the mean absolute error (MAE), mean square error (MSE), and root mean square error (RMSE), which are the model performance evaluation indicators, are all less than the threshold (the MAE threshold is set to 0.05, and the MSE threshold is set to 0.001), the model validation passes; otherwise, adjust the training parameters or expand the training set samples (for example, adjust learning_rate to 0.005 and max_depth to 5) and retrain the model until the validation passes.

[0175] In another exemplary embodiment, Figure 3 As shown, the present application also provides a rock stability detection device, which includes: an acquisition module 100 for acquiring monitoring signals of the rock mass to be tested; a preprocessing module 200 for preprocessing the monitoring signals; a model construction and training module 300 for constructing a rock stability detection model and training the model; and a prediction module 400 for inputting the preprocessed monitoring signals into the trained rock stability detection model to detect the stability of the rock mass to be tested.

[0176] In another exemplary embodiment, the present application further provides a storage medium comprising instructions, which, when executed on a computer, enables the computer to execute any of the methods described above.

[0177] In another exemplary embodiment, the present application also provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements any of the methods described above when executing the program.

[0178] The above embodiments are intended only to illustrate the technical concepts and features of this application. Their purpose is to enable those familiar with the art to understand the content of this application and implement it accordingly. They are not intended to limit the scope of protection of this application. Any equivalent changes or modifications made in accordance with the spirit of this application shall be included in the scope of protection of this application.

Claims

1. A rock mass stability detection method, characterized in that: The detection method comprises: Collecting monitoring signals of the rock mass to be measured, wherein the monitoring signals include acoustic emission signals, displacement signals, stress and strain signals, inclination signals, and temperature and humidity signals around the rock mass to be measured; Preprocessing the monitoring signal; A rock mass stability detection model is constructed and trained; the rock mass stability detection model introduces a multimodal data access layer to simultaneously process the monitoring signals; the multimodal data access layer also introduces a self-attention mechanism to dynamically calculate the importance of multimodal features at each time point; The pre-processed monitoring signal is input into the trained rock mass stability detection model to detect the stability of the rock mass to be tested.

2. A rock mass stability detection method according to claim 1, characterized in that: Preprocessing the monitoring signal includes: performing noise filtering, feature extraction and event recognition on the acoustic emission signal; Detrending the displacement signal and detecting and correcting abnormal points; Performing temperature compensation and nonlinear correction on the stress-strain signal and performing incremental cumulative effect analysis; Gravity compensation and posture fusion are performed on the inclination signal, and dynamic response adjustment is performed.

3. A rock mass stability detection method according to claim 1, characterized in that: The rock mass stability detection model includes: An input unit, used for inputting the pre-processed monitoring signal; A feature extraction unit, configured to extract features from the input pre-processed monitoring signal; The output unit is used to output the stability score of the rock mass to be measured based on the features extracted by the feature extraction unit.

4. A rock mass stability detection method according to claim 3, characterized in that: The input unit includes: Multimodal data access layer and data integration and transmission layer, where: The multimodal data access layer is used to standardize and synchronize the clock of the accessed monitoring signals to obtain multimodal data; The data integration and transmission layer is used to fuse the multimodal data.

5. A rock mass stability detection method according to claim 3, characterized in that: The feature extraction unit includes: Spatial feature extraction module, time series module and physical rule embedding module.

6. A rock mass stability detection method according to claim 1, characterized in that: The rock mass stability detection model is trained by the following steps: Obtain and preprocess historical monitoring data of the rock mass, and divide the preprocessed historical monitoring data into a training set and a validation set in proportion; Set the training parameters and use the training set to train the model. During the model training process, when the number of model iterations meets the preset value, the model training is completed; The trained model is verified using the validation set. During the verification process, if the mean absolute error, mean square error, and root mean square error, which are the model performance evaluation indicators, are all less than the threshold, the model verification passes; otherwise, the training parameters are adjusted or the training set samples are expanded to retrain the model until the verification passes.

7. A rock mass stability detection device, characterized in that: The device comprises: An acquisition module is used to acquire monitoring signals of the rock mass to be measured, wherein the monitoring signals include acoustic emission signals, displacement signals, stress and strain signals, inclination signals, and temperature and humidity signals around the rock mass to be measured; A preprocessing module, used for preprocessing the monitoring signal; Model building and training module, used to build rock stability detection model and train the model; The prediction module is used to input the pre-processed monitoring signal into the trained rock mass stability detection model to detect the stability of the rock mass to be tested.

8. A rock mass stability detection device according to claim 7, characterized in that: The pre-processing module comprises: A first preprocessing submodule is used to filter noise, extract features and identify events on the acoustic emission signal; A second preprocessing submodule performs trend removal on the displacement signal and performs outlier detection and correction; A third preprocessing submodule performs temperature compensation and nonlinear correction on the stress-strain signal and performs incremental cumulative effect analysis; The fourth pre-processing submodule performs gravity compensation and posture fusion on the inclination signal and performs dynamic response adjustment.

9. A storage medium, characterized in that: The method comprises instructions which, when executed on a computer, cause the computer to execute the method according to any one of claims 1 to 6.

10. An electronic device, characterized in that: The electronic device comprises: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the method according to any one of claims 1 to 6 is implemented.

Images