A rock uniaxial compression failure short-term prediction method based on machine learning

Abstract

This invention provides a short-term prediction method for uniaxial compression failure of rocks based on machine learning, belonging to the technical field of uniaxial compression failure prediction. This method reconstructs the state space based on strain and acoustic emission data during uniaxial compression, constructing a multidimensional state space. A local Jacobian matrix is ​​built using local linear modeling. After determining that the rock has entered the accelerated deformation stage, the time and displacement parameters within this stage are dimensionless. A power function model is used to calculate the dimensionless trend index, achieving sub-stage division of the accelerated failure stage. A failure time prediction model based on a long short-term memory neural network is constructed to dynamically predict the rock failure time. Combined with the rock's current deformation stage, a short-term prediction of the failure process is achieved. This invention can dynamically correlate the failure time prediction results with the rock deformation evolution stage, improving the accuracy and reliability of failure prediction and possessing significant application value.

Description

Technical Field

[0001] This invention relates to the field of rock uniaxial compression failure prediction technology, and in particular to a short-term prediction and forecasting method for rock uniaxial compression failure based on machine learning. Background Technology

[0002] Rocks are widely found in nature and are common materials in fields such as water conservancy and hydropower engineering, mining engineering, tunnel transportation engineering, and geological engineering. Rock deformation and failure are frequent engineering problems encountered in construction, safe operation and maintenance, and disaster prevention and mitigation of natural geological disasters. They are one of the main controlling factors inducing rock mass system disasters such as earthquakes, landslides, collapses, rock bursts, rockbursts, tunnel roof falls, and gas outbursts. Their suddenness and destructive force often lead to huge losses, posing a serious threat to human life, safety, and socio-economic development. Therefore, how to effectively prevent and control rock deformation and failure has always been a key and challenging issue in rock mechanics research. Rock failure prediction and forecasting is one of the effective means of mitigating disaster risks. The core issue in conducting rock failure prediction and forecasting is how to accurately identify precursors of rock failure and conduct reliable short-term forecasts, thereby providing sufficient lead time for formulating and implementing disaster prevention and mitigation measures.

[0003] Currently, short-term rock failure prediction mainly employs traditional empirical models, numerical simulations, and machine learning methods, combined with monitoring techniques such as acoustic emission, strain measurement, infrared imaging, and digital image correlation (DIC) to predict rock failure time. However, existing studies tend to use single-parameter predictions, neglecting the coupling relationships between multi-source data and the evolution of temporal characteristics, leaving significant room for improvement in prediction accuracy and stability. Therefore, it is urgent to conduct research on the microscopic failure evolution processes of different lithological materials, systematically analyze the precursor characteristics of various rock types during loading failure, and explore reliable and operable failure prediction and forecasting methods to enhance the scientific rigor and effectiveness of geological disaster monitoring and early warning. Summary of the Invention

[0004] The purpose of this invention is to provide a short-term prediction method for uniaxial compression failure of rocks based on machine learning, which solves the technical problem that the prediction accuracy and stability of existing short-term rock prediction methods still have considerable room for improvement.

[0005] By introducing dynamic eigenvalues ​​and dimensionless trend indices, the rock deformation and failure process is identified in stages. The stage identification results are then combined with a failure time prediction method based on long short-term memory neural networks to achieve short-term prediction of rock failure processes, thereby improving the accuracy and reliability of failure prediction.

[0006] By constructing a short-term prediction model for rock failure, rock strain data, acoustic emission data, and stress-time data collected during uniaxial compression loading in the laboratory are input into the short-term prediction model for rock failure, thereby enabling the identification of rock failure evolution stages and the short-term prediction of failure time.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A short-term prediction method for uniaxial compression failure of rocks based on machine learning, the method comprising the following steps: Step 1: Under uniaxial compression loading conditions in the laboratory, load tests were conducted on the rock samples, and strain data, acoustic emission data, and stress-time data of the rock were collected simultaneously. Step 2: The state space of the collected strain data and acoustic emission data is reconstructed to construct a multi-dimensional state space that reflects the rock deformation evolution process. Based on the local linear modeling method, the local Jacobian matrix of the state space is constructed, and its principal eigenvalues ​​are extracted to obtain the sequence of dynamic eigenvalues ​​DEV changing with time, which is used to identify whether the rock has entered the accelerated deformation stage. Step 3: After determining that the rock has entered the accelerated deformation stage, the time and displacement parameters within the accelerated deformation stage are dimensionless, and the evolution process of the accelerated deformation stage is fitted using a power function model to calculate the dimensionless trend index m, so as to further divide the accelerated failure stage of the rock. Step 4: Construct a failure time prediction model based on acoustic emission characteristic parameters, dynamically predict the rock failure time, and combine the predicted failure time with the accelerated deformation stage and its sub-stages to achieve short-term prediction of the rock failure process.

[0008] Furthermore, in step 1, the loading test of the rock sample involves the construction of uniaxial compression test data for at least three different rock types, including sandstone, marble, and granite. Furthermore, in step 2, the dynamic eigenvalue DEV is used as the threshold for identifying critical changes in rock structure stability when its absolute value equals 1. Furthermore, in step 2, the local linear modeling method is the S-map method, which constructs a local Jacobian matrix by locally linearly fitting the current state vector in the state space. Furthermore, in step 3, the dimensionless processing uses the start and end times of the accelerated deformation stage as normalization boundaries to normalize the time and displacement parameters.

[0009] Furthermore, in step 3, the dimensionless trend index m is obtained by fitting the power function model using the sliding least squares method.

[0010] Furthermore, based on the magnitude of the dimensionless trend index m, the accelerated deformation stage of the rock is divided into three sub-stages: a, b, and c.

[0011] Furthermore, in step 4, the destruction time prediction model is a Long Short-Term Memory (LSTM) neural network model.

[0012] Furthermore, in step 4, as the rock evolves from stage a to stage b or c, the error in predicting the failure time gradually decreases.

[0013] Under uniaxial compression loading conditions in the laboratory, loading tests were conducted on rock samples, and strain data, acoustic emission data, and stress-time data of the rock were collected simultaneously. Based on the acquired rock strain data and acoustic emission data, the data is reconstructed into a state space to build a multidimensional state space that reflects the rock deformation evolution process. The local Jacobian matrix of the state space is constructed using a local linear modeling method. The principal eigenvalues ​​of the local Jacobian matrix are extracted to form a sequence of dynamic eigenvalues ​​DEV changing over time, which is used to identify the evolution state of rock structural stability. Based on the changes in dynamic characteristic value DEV, the rock deformation and failure process is staged. When the absolute value of the dynamic characteristic value approaches or reaches a preset threshold, the rock is determined to have entered the accelerated deformation stage from a stable state. After determining that the rock has entered the accelerated deformation stage, the time and displacement parameters within the accelerated deformation stage are made dimensionless, and the evolution process of the accelerated deformation stage is fitted using a power function model. The dimensionless trend index m is calculated, and the accelerated deformation stage is further divided into multiple sub-stages of failure based on the numerical variation characteristics of the dimensionless trend index m. Based on the sub-stage division of the accelerated deformation stage, a failure time prediction model based on a long short-term memory neural network is constructed using acoustic emission characteristic parameters as input. This model dynamically predicts the failure time of rocks and combines the predicted failure time with the accelerated deformation stage and its sub-stages to achieve short-term prediction of the rock failure process. The present invention, by adopting the above-described technical solution, has the following beneficial effects: This invention introduces the Dynamic Eigenvalue (DEV) to monitor the evolution of system stability during rock deformation and failure in real time, enabling dynamic identification of the transition process from a stable state to a critical failure state, thus improving the accuracy and physical consistency of failure stage identification. Through dimensionless processing and the introduction of a dimensionless trend index m, the accelerated deformation stages of rocks are further subdivided, achieving hierarchical identification of failure precursor information and transforming the rock failure evolution process from a single-stage description to a multi-stage refined characterization. By combining the rock failure evolution stage identification results with a failure time prediction method based on a long short-term memory neural network, the failure time prediction results are dynamically correlated with the current deformation state of the rock, realizing a short-term prediction mechanism where the reliability of failure time prediction gradually increases as the failure stage progresses. This invention is applicable to the failure process analysis of rocks of different lithologies under uniaxial compression conditions, possessing good versatility and engineering application value, and can provide technical support for failure early warning and risk assessment in rock mechanics experiments. Attached Figure Description

[0014] Figure 1 This is a flowchart of the present invention; Figure 2 This is a schematic diagram illustrating the construction process of the LSTM model dataset in this invention; Figure 3 This invention provides a sample stress curve, relative energy, and its cumulative value over time. Figure 4 This is a diagram identifying the deformation stages of sandstone in this invention; Figure 5 This is a diagram identifying the deformation stages of marble in this invention; Figure 6 This is a diagram identifying the deformation stages of granite in this invention; Figure 7 This is a comparison diagram of the fitting result of the cumulative energy m value of acoustic emission provided by the present invention and the sub-stage division of the tangent angle acceleration stage; Figure 8 This is a comparison diagram of the fitting result of the cumulative ringing count of acoustic emission provided by the present invention and the sub-stage division of the tangent angle acceleration stage; Figure 9 This is a comparison diagram of the fitting results of the m-value of rock strain provided by the present invention and the sub-stage division of the tangent angle acceleration stage; Figure 10 This is a framework diagram for rock failure prediction in this invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and preferred embodiments. However, it should be noted that many details listed in the specification are merely to provide the reader with a thorough understanding of one or more aspects of the present invention, and these aspects of the invention can be implemented even without these specific details.

[0016] like Figure 1 As shown, a method for identifying rock failure evolution stages and predicting short-term failures under uniaxial compression conditions includes the following steps: Step 1: Under uniaxial compression loading conditions in the laboratory, a loading test is conducted on the rock specimen. Strain data and acoustic emission data of the rock specimen are acquired simultaneously, and the corresponding stress-time data are recorded. Step 2: Predict the rock failure time based on the long short-term memory neural network model; Step 3: Propose a new method for identifying deformation stages during rock deformation and failure, including identification of all stages of deformation and detailed division of accelerated deformation stages; Step 4: Construct a short-term rock failure prediction framework that integrates "deformation stage identification, failure precursor identification, and failure time prediction"; The following description uses more specific examples.

[0017] A machine learning-based method for short-term prediction of uniaxial compression failure in rocks includes the following steps: 1. Uniaxial loading tests were conducted using the RMT-301 rock and concrete mechanics testing system, with a maximum vertical test pressure of 1500 kN. The mechanical parameters and deformation characteristics of the specimens under uniaxial loading were primarily obtained using this testing machine. The data acquisition system consisted of stress-strain and AE data acquisition devices. The preamplifier's amplification factor, threshold, and sampling frequency were set to 40 dB, 40 dB, and 3 MSPS, respectively, to obtain the most complete AE signal possible and reduce noise interference from machine operation. The AE system was set to impact-driven acquisition mode, with the relevant timing parameters for PDT, HDT, and HLT set to 200 μs, 800 μs, and 1000 μs, respectively.

[0018] 2. The LSTM model's ability to resolve long-term dependencies hinges on a key component in the hidden layers called a "memory cell." Each "memory cell" connects three gating components in a specific way, using "gates" to process information that needs to be remembered and forgotten. During computation, candidate memories are evaluated... a t Selective storage, neuron input gate i t Forgotten Gate ft Output gate o t Primarily responsible for regulating and transmitting information flow, i t It will add the required information. f t Selective forgetting occurs to eliminate irrelevant data. o t Output the useful information to obtain the required output value.

[0019] (1) Model building: Candidate memory: Input Gate: Forgotten Gate: Output gate: Where σ(·) represents the sigmoid function; tanh(·) represents the tangent function value; and the symbol (·) represents the multiplication operation for each element. X t-1 Indicates input data; h t-1 Indicates the hidden state of the preneuron; b c , b i , b f , b o These are the adjustment parameters for each gate, which are mainly obtained during the model training process.

[0020] Regarding the current state C t Activate the function, combined with the current o t The output of the LSTM neuron is obtained h t .

[0021] (2) Model data preprocessing: Data preprocessing was performed on the time series data for different AE features. First, the sequence for each AE feature was transformed into the sequence prepared for the model using a time sliding window. The transformation process using the time sliding window is as follows: Figure 2 As shown.

[0022] Then, the prepared model data is normalized using the following formula: in, dat* i Represents normalized data for time series. dat i The raw data representing the time series. S express dat i The mean of all subscripts i in the set. S Indicates all dat i The standard deviation.

[0023] The entire model's dataset, including normalized input features and output targets, is divided into training datasets and test datasets.

[0024] (3) Model output target and input parameters This invention defines the stress peak point on the stress-time curve under uniaxial compression loading as the failure point. Since various precursory phenomena are observed in mechanical tests around the stress peak, the time from failure to the stress peak is considered the output target of the machine learning model in this study.

[0025] Table 1 AE Features The dataset input to the model is derived from the accumulation of acoustic emission (AE) features, and the data is divided into five groups, as shown in Table 1. Taking some AE features as examples, Figure 3 The stress curve of a sample and its relative energy change and accumulation over time are shown, with a red circle indicating the stress peak. After feature sorting and combination selection, the first 12 features in Table 2 are used as input features. Table 2 shows the selected feature combinations used as input features. 3. Deformation Stage Identification in Rock Deformation and Failure Processes: Focusing on the overall dynamic characteristics of the approaching failure stage, a short-term forecasting framework integrating "deformation stage identification—failure precursor identification—failure time prediction" is constructed. Specifically, firstly, the dynamic eigenvalue (DEV) and trend index (m) are used to dynamically divide the entire rock deformation process; then, based on acoustic emission signals, stress-strain curves, and DIC surface strain field data, accelerated evolution characteristics are extracted and failure precursors are identified; finally, combined with existing failure time prediction methods, reliable predictions of rock failure within the short-term phase are achieved.

[0026] (1) Identification of Deformation Stages Throughout the Process: To achieve early identification of rock failure and deformation stages, dynamic indicators that reflect the evolution of system stability are introduced. First, a state-space reconstruction method is used to reconstruct the one-dimensional time series data into a multi-dimensional state space through time delay embedding. Then, the S-map method is used to analyze the current time point... t The Jacobian matrix is ​​estimated by fitting a local linear model. For each time point... t Construct a linear model to predict the state at the next time point: In the formula, C It is a linear coefficient matrix that represents the relationship between the current state and the future state. X t It is a point in time. t The state vector, v It is the offset.

[0027] In S-map, the local Jacobian matrix J coefficient C These coefficients represent the linear relationships between the various dimensions of the state space. Using these coefficients, a local Jacobian matrix can be constructed: This local Jacobian matrix J t Reflects the system's time t Local stability in the vicinity.

[0028] By calculating the eigenvalues ​​of the Jacobian matrix, the stability information of the system can be obtained. For each local Jacobian matrix... J t Calculate its principal eigenvalues: At the critical inflection point, the absolute value of the system's principal eigenvalues ​​will approach 1, which means that the system's stability has changed significantly.

[0029] Over time, the principal eigenvalues λ dominant These characteristics change with the system state. Dynamic eigenvalues ​​(DEVs) are the time series of these principal eigenvalues. When the absolute value of DEV is close to 1, it indicates that the system is approaching a critical inflection point. The critical inflection point is the moment when the system transitions from one stable state to another or enters an unstable state.

[0030] Figure 4-6The sandstone (a), marble (b), and granite (c) shown in the figure demonstrate the deformation and failure stages of the three rocks identified by strain. As can be seen from the figure, the method based on dynamic eigenvalues ​​can accurately identify the acceleration stage of rock strain.

[0031] (2) Division of accelerated deformation stages: Once the accelerated rock failure stage is identified, it should be further refined into multiple sub-stages (III(a), III(b), III(c)). For ease of comparison and analysis, these sub-stages will be remapped to a suitable dimensionless space. Time and displacement are normalized for this purpose. in t 0、 t f These represent the start and end times of the accelerated rock fracturing stage. d 0 and d f This represents the cumulative displacement value at the corresponding time. Through this dimensionless mapping, the landslide displacement processes under different landslides, scales, and triggering mechanisms can be uniformly projected onto a standardized two-dimensional model. D T In space, it is convenient to perform trend fitting, classification and comparative analysis within that space.

[0032] Subsequently, a power function model was adopted based on the normalized curve: This model fits the evolution trend of the accelerated failure stage of rocks and classifies different types of dynamic behaviors according to the magnitude of the trend index m, thereby achieving a unified expression of the evolution process from complex and diverse to finite modes.

[0033] To further improve the accuracy of our understanding of the accelerated rock failure stage, this section, based on the aforementioned normalized trend fitting method, attempts to further subdivide the evolutionary process within the accelerated stage into more detailed sub-stages, in order to characterize its continuous evolutionary features of "gradual change – abrupt change".

[0034] To obtain the accelerated evolution characteristics of rock failure, cumulative acoustic emission energy, cumulative ringing count, and rock strain were selected as feature parameters. First, the parameters were normalized. Then, using the minimum three data points as the initial window, a power-law fitting was performed using the moving least squares method to extract the dimensionless exponent *m* corresponding to each time stage. Finally, the sub-stages of the accelerated rock failure process were divided according to the range of the dimensionless parameter *m* defined in Table 3. The results of dividing the sub-stages using cumulative acoustic emission energy, acoustic emission ringing count, and rock strain are shown below. Figures 5 to 6As shown.

[0035] Table 3. Range of dimensionless parameter m Statistical analysis of the sub-stage division results for all samples revealed that the three samples exhibited a shorter duration in stage III(b), but their corresponding trend exponential growth rate was the fastest, indicating a highly sensitive system. The results of dividing stages III(a), III(b), and III(c) using the dimensionless parameter m value are consistent with those using the tangent angle division; the duration of ∠45°–∠80° is significantly longer than that of ∠80°–∠85°, and the time difference between III(b) and ∠80°, and between III(c) and ∠85°, is ≤3 seconds. The sub-stage division of accelerated rock failure using the dimensionless parameter m calculated by acoustic emission cumulative ringing count is as follows: Figure 7 As shown, the partitioning results are quite similar to those obtained using the cumulative energy of acoustic emission.

[0036] The DEV method emphasizes the integrity of system stability and the dynamic characteristics of continuous evolution over time, while the dimensionless trend fitting method is suitable for detailed decomposition of the accelerated failure stage. Combining the two methods, a joint forecasting framework is constructed: first, the DEV method is used to identify whether the system has entered the instability transition zone; then, the dimensionless m-value is used to refine the specific deformation stages of the accelerated failure stage, achieving multi-level and stage-by-stage accurate identification of the entire rock failure process.

[0037] 4. Short-Term Rock Failure Prediction Framework: This framework combines the predicted rock failure time with the three stages of accelerated rock failure to provide real-time predictions of the failure time at each stage. It uses a dynamic eigenvalue method to identify system stability changes throughout the entire rock failure process. A state space is constructed by analyzing characteristic parameters such as strain and acoustic emission, and principal eigenvalues ​​are extracted to monitor whether the system is approaching a critical state. Figure 9 As shown, when the rock strain |DEV| approaches 1, it can effectively indicate that the rock has entered the accelerated deformation stage. Matters not covered in this invention are common knowledge.

[0038] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A machine learning-based short-term prediction method for rock uniaxial compression failure, characterized in that, The method includes the following steps: Step 1: Under uniaxial compression loading conditions in the laboratory, load tests were conducted on the rock samples, and strain data, acoustic emission data, and stress-time data of the rock were collected simultaneously. Step 2: The state space of the collected strain data and acoustic emission data is reconstructed to construct a multi-dimensional state space that reflects the rock deformation evolution process. Based on the local linear modeling method, the local Jacobian matrix of the state space is constructed, and its principal eigenvalues ​​are extracted to obtain the sequence of dynamic eigenvalues ​​DEV changing with time, which is used to identify whether the rock has entered the accelerated deformation stage. Step 3: After determining that the rock has entered the accelerated deformation stage, the time and displacement parameters within the accelerated deformation stage are dimensionless, and the evolution process of the accelerated deformation stage is fitted using a power function model to calculate the dimensionless trend index m, so as to further divide the accelerated failure stage of the rock. Step 4: Construct a failure time prediction model based on acoustic emission characteristic parameters, dynamically predict the rock failure time, and combine the predicted failure time with the accelerated deformation stage and its sub-stages to achieve short-term prediction of the rock failure process.

2. The method of claim 1, wherein the method is characterized by: In step 1, loading tests were conducted on rock samples, and uniaxial compression test data were constructed for at least three different rock types, including sandstone, marble, and granite.

3. The machine learning based rock uniaxial compression failure short-term prediction method according to claim 1, characterized in that: In step 2, the dynamic eigenvalue DEV is used as the threshold for identifying the critical change in rock structure stability when its absolute value equals 1.

4. The machine learning based rock uniaxial compression failure short-term prediction method according to claim 1, characterized in that: In step 2, the local linear modeling method is the S-map method, which constructs a local Jacobian matrix by locally linearly fitting the current state vector in the state space.

5. The machine learning based rock uniaxial compression failure short-term prediction method according to claim 1, characterized in that: In step 3, the dimensionless processing uses the start and end times of the accelerated deformation stage as normalization boundaries to normalize the time and displacement parameters.

6. The machine learning based rock uniaxial compression failure short-term prediction method according to claim 1, characterized in that: In step 3, the dimensionless trend index m is obtained by fitting the power function model using the sliding least squares method.

7. The machine learning based rock uniaxial compression failure short-term prediction method according to claim 6, characterized in that: Based on the magnitude of the dimensionless trend index m, the accelerated deformation stage of rocks is divided into three sub-stages: a, b, and c.

8. The machine learning based rock uniaxial compression failure short-term prediction method according to claim 1, characterized in that: In step 4, the destruction time prediction model is a Long Short-Term Memory (LSTM) neural network model.

9. The machine learning based rock uniaxial compression failure short-term prediction method according to claim 1, characterized in that: In step 4, as the rock evolves from stage a to stage b or c, the error in predicting the failure time gradually decreases.

Images