A method and system for predicting rock strength and failure mode

Abstract

This invention relates to the field of rock mechanics and intelligent detection technology, specifically a method and system for predicting rock strength and failure modes. The method includes: obtaining a multi-source feature set based on data acquisition from a rock sample; constructing a feature coefficient prediction model, inputting the multi-source feature set into the feature coefficient prediction model to obtain feature coefficients; filtering the multi-source feature set based on the feature coefficients to obtain an optimal feature subset; constructing a rock strength and failure mode prediction model, and optimizing the hyperparameters of the rock strength and failure mode prediction model using an automated hyperparameter optimization framework; inputting the optimal feature subset into the optimized rock strength and failure mode prediction model to obtain the predicted rock strength value and failure mode. This invention solves the problems of insufficient accuracy and inability to predict rock strength and failure modes simultaneously in existing technologies. Through multi-source data collaborative prediction, the accuracy of the prediction is greatly improved.

Description

Technical Field

[0001] This invention relates to the field of rock mechanics and intelligent detection technology, specifically a method and system for predicting rock strength and failure modes. Background Technology

[0002] The strength parameters and failure modes of rocks are crucial for stability analysis and design in geotechnical engineering projects such as tunnels, slopes, and mines. Currently, uniaxial compression tests in the laboratory are the most direct method for obtaining these parameters, but they suffer from problems such as long processing time, high cost, and large dispersion. Empirical formula prediction methods based on single parameters (such as point load strength and wave velocity), while simple, have low accuracy and poor generalization ability, making it difficult to reflect the complex physical mechanisms of rock failure.

[0003] In recent years, machine learning and deep learning technologies have been introduced into the field of rock mechanics. However, existing technologies are mostly limited to applying a single data mode (such as using only rock images or acoustic emission signals), failing to fully utilize the multi-physics field information (deformation field, acoustic field, visual field, etc.) generated by rocks during stress. In addition, most existing models can only predict intensity and lack the ability to automatically identify failure modes (such as tensile failure, shear failure, and combined failure), while failure modes are crucial for assessing the catastrophic mechanisms of rock mass engineering.

[0004] Therefore, there is an urgent need for a method to predict rock strength and failure mode by integrating information from multiple sources, so as to achieve high-precision prediction of rock strength and failure mode. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for predicting rock strength and failure modes, which solves the problems of insufficient accuracy in predicting rock strength and failure modes and the inability to predict them simultaneously in existing technologies.

[0006] To achieve the above objectives, one aspect of the present invention provides a method for predicting rock strength and failure mode. The method includes: obtaining a multi-source feature set based on data acquisition of a rock sample to be tested; constructing a feature coefficient prediction model, inputting the multi-source feature set into the feature coefficient prediction model to obtain feature coefficients; filtering the multi-source feature set based on the feature coefficients to obtain an optimal feature subset; constructing a rock strength and failure mode prediction model, and optimizing the hyperparameters of the rock strength and failure mode prediction model using an automated hyperparameter optimization framework; and inputting the optimal feature subset into the optimized rock strength and failure mode prediction model to obtain a predicted rock strength value and a failure mode.

[0007] This invention constructs a multi-source feature set by collecting multi-source data from rock samples under test, then outputs feature coefficients through a feature coefficient prediction model to select the optimal feature subset. Simultaneously, it constructs a rock strength and failure mode prediction model and optimizes hyperparameters using an automated hyperparameter optimization framework. Finally, it inputs the optimal feature subset to obtain the prediction result. This invention comprehensively captures multi-dimensional information of the rock, eliminates redundant features to reduce model interference, quickly locates the optimal hyperparameters to improve model training efficiency, and effectively enhances the model's fitting and generalization capabilities. As a result, it achieves the joint prediction of rock strength and failure mode and improves the accuracy of rock strength and failure mode prediction.

[0008] Optionally, the multi-source feature set obtained based on the data acquisition of the rock sample to be tested includes: obtaining multi-source data based on the data acquisition of the rock sample to be tested; and performing feature extraction on the multi-source data to obtain a multi-source feature set including strain time series data, internal damage characterization time series data, surface feature data, basic physical feature data, and spectral feature data.

[0009] This invention first collects multi-source data from the rock sample to be tested, and then selectively extracts strain time series, internal damage characterization time series, surface features, basic physical features and spectral features to construct a multi-source feature set. This can comprehensively cover the dynamic process of rock deformation, internal damage state, surface structure, inherent physical properties and lithological correlation information, avoid the problem of one-sided information from single data types, and improve the data quality of the multi-source feature set.

[0010] Optionally, the multi-source data includes speckle images, acoustic emission impact signals, surface image data, basic physical parameters, and visible-near-infrared spectral data. The step of extracting features from the multi-source data to obtain a multi-source feature set including strain time-series data, internal damage characterization time-series data, surface feature data, basic physical feature data, and spectral feature data includes: processing the speckle images using image correlation to obtain initial strain time-series data; segmenting and processing the acoustic emission impact signals based on a pre-set time window to obtain initial internal damage characterization time-series data; performing dimensionality reduction processing on the initial strain time-series data and the initial internal damage characterization time-series data to obtain strain time-series data and internal damage characterization time-series data; extracting features from the surface image data using a pre-trained feature extractor to obtain surface feature data; performing vector processing on the basic physical parameters to obtain basic physical feature data; and performing denoising and dimensionality reduction processing on the visible-near-infrared spectral data to obtain spectral feature data.

[0011] This invention extracts initial strain time-series data from speckle images using image correlation, and combines time window segmentation to process acoustic emission impact signals to obtain initial time-series data for internal damage characterization. Then, dimensionality reduction processing is used to simplify redundancy and retain core dynamic information. A pre-trained feature extractor is used to mine the structural features of surface image data, and vector processing is used to transform basic physical parameters into regular feature vectors to optimize visible-near-infrared spectral data through noise reduction and dimensionality reduction, thereby improving the data quality of multi-source feature sets.

[0012] Optionally, the step of performing dimensionality reduction processing on the initial strain time series data and the initial internal damage characterization time series data to obtain strain time series data and internal damage characterization time series data includes: performing length normalization processing on the initial strain time series data and the initial internal damage characterization time series data; and performing principal component analysis on the length-normalized initial strain time series data and the initial internal damage characterization time series data to obtain strain time series data and internal damage characterization time series data, respectively.

[0013] This invention first normalizes the length of two types of initial time series data by using dimensionality reduction processing, which solves the problem of inconsistent time series data length caused by the difference in loading time of different rock samples, ensuring data comparability. Then, it further processes the data through principal component analysis, which can effectively remove redundant data while retaining key dynamic information, thereby improving the data quality of strain time series data and internal damage characterization time series.

[0014] Optionally, the denoising and dimensionality reduction processing of the visible-near-infrared spectral data to obtain spectral feature data includes: denoising the visible-near-infrared spectral data based on Savitzky-Golay smoothing to obtain denoised spectral data; optimizing the denoised spectral data using standard normal variable transformation to obtain optimized spectral data; and reducing the dimensionality of the optimized spectral data based on principal component analysis to obtain spectral feature data.

[0015] This invention effectively filters out random noise such as instrument errors and environmental interference through Savitzky-Golay smoothing, while preserving characteristic peak information related to rock mineral composition and weathering state in the spectrum to the greatest extent. Then, standard normal variable transformation eliminates spectral baseline drift and intensity fluctuations caused by differences in sample surface roughness and particle size, optimizing spectral data consistency. Finally, principal component analysis simplifies data dimensions and removes redundant information. These three progressively advanced processes improve spectral data quality and achieve efficient data compression, thereby enhancing the quality of spectral feature data.

[0016] Optionally, the construction of the feature coefficient prediction model includes: determining the optimal regularization strength parameter through cross-validation; constructing an objective function based on the optimal regularization strength parameter; and constructing the feature coefficient prediction model using the objective function based on minimum absolute shrinkage and selection operator regression.

[0017] This invention determines the optimal regularization strength parameter through cross-validation, which can accurately adapt to multi-source feature data and avoid blind parameter selection. Then, an objective function is constructed based on this parameter, providing a scientific basis for model optimization. Finally, the model is constructed by combining minimum absolute shrinkage and selection operator regression, which can effectively screen key features, suppress redundant features, and improve the performance of the feature coefficient prediction model.

[0018] Optionally, the objective function satisfies the following formula:

[0019]

[0020] in, Let be the objective function. Let be the vector of regression coefficients to be solved. The number of samples for the feature coefficient prediction model. For the first The true rock strength values ​​of each training sample. For the first The feature vectors of each training sample The regularization strength parameter is... For the number of features, Let the first of the regression coefficients in the vector to be solved be the first... Each component.

[0021] The objective function formula of this invention minimizes the combination of prediction error and regularization term. While fitting the relationship between features and target values, it controls the regularization strength using a regularization strength parameter to constrain the regression coefficient β. This ensures the model's fitting accuracy to the training samples while enabling feature selection through L1 regularization, compressing the coefficients of unimportant features to zero, thereby filtering out key features, simplifying the model structure, avoiding overfitting, and improving the model's generalization ability and interpretability in predicting rock strength and failure modes.

[0022] Optionally, the rock strength and failure mode prediction model satisfies the following formula:

[0023]

[0024] in, For loss function, As weight, The sample size for the rock strength and failure mode prediction model. No. The true value of rock strength for each sample. For the first Predicted rock strength values ​​for each sample The number of categories of destruction modes, For indicator functions, For the Realistic Destruction mode category, For the first Each sample is predicted as a category. The probability, As weight, This is the proportionality coefficient. For the first The longitudinal wave velocity of each sample.

[0025] This invention improves the performance of rock strength and failure mode prediction models by fusing the mean square error term and the cross-entropy term, which can accurately fit the continuous numerical prediction of rock strength and effectively handle the classification task of failure mode.

[0026] Optionally, the rock strength and failure mode prediction model includes a spectral correction submodule.

[0027] This invention, by setting a spectral correction submodule in the rock strength and failure mode prediction model, can specifically correct prediction biases caused by differences in rock lithological composition and weathering state, make up for the insufficient adaptation of lithological information in multi-source features, and effectively improve the generalization ability of the rock strength and failure mode prediction model.

[0028] In another aspect, the present invention provides a rock strength and failure mode prediction system, comprising: a processor, an input device, an output device, and a memory, wherein the processor, the input device, the output device, and the memory are interconnected, wherein the memory is used to store a computer program, the computer program including program instructions, and the processor is configured to invoke the program instructions to execute a rock strength and failure mode prediction method according to any of the preceding aspects of the present invention.

[0029] The rock strength and failure mode prediction system of the present invention has a compact structure, stable performance, high integration and simple composition. It can stably execute the rock strength and failure mode prediction method provided in the preceding aspect of the present invention, further improving the overall applicability and practical application capability of the present invention. Attached Figure Description

[0030] Figure 1 This is a flowchart of a method for predicting rock strength and failure mode according to an embodiment of the present invention;

[0031] Figure 2 This is a schematic diagram of a rock strength and failure mode prediction system according to an embodiment of the present invention. Detailed Implementation

[0032] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0033] Throughout this specification, references to "an embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, the phrases "in an embodiment," "in an embodiment," "an example," or "an example" appearing in various places throughout the specification do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in one or more embodiments or examples in any suitable combination and / or sub-combination. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0034] Please see Figure 1 To address the shortcomings of existing technologies, in an alternative implementation, such as Figure 1 The rock strength and failure mode prediction method shown includes the following steps:

[0035] Step S1: Obtain a multi-source feature set based on the data acquisition of the rock sample to be tested.

[0036] The process of obtaining a multi-source feature set based on data acquisition from the rock sample to be tested includes the following sub-steps:

[0037] Step S101: Obtain multi-source data based on the data acquisition of the rock sample to be tested.

[0038] In this embodiment, the multi-source data includes speckle images, acoustic emission impact signals, surface image data, basic physical parameters, and visible-near-infrared spectral data.

[0039] Based on the loading process of uniaxial compression test on the rock sample to be tested, a uniform speckle field is first prepared on the side of the rock sample. Then, two or more high-speed cameras are used to synchronously capture the speckle changes on the sample surface at a fixed acquisition frequency (e.g., 10fps). The capturing process continues until the sample is completely destroyed. The final raw data is a complete time sequence of speckle images, i.e. speckle images.

[0040] Acoustic emission hit signals are raw acoustic data of micro-fractures inside the rock collected by an acoustic emission (AE) system. The acquisition is synchronized with the entire loading process of the uniaxial compression test. During acquisition, two acoustic emission sensors need to be coupled to the side of the rock sample. The signal acquisition threshold is set to 40dB and the sampling frequency is 1MHz. The system continuously captures all acoustic emission hit signals during the loading process. The acoustic emission hit signals include parameters such as the arrival time, ring count, energy, amplitude and duration of each hit signal.

[0041] Surface image data are raw visual data recording the macroscopic structure of the rock sample surface. Acquisition is performed when the uniaxial compression test is loaded to approximately 50% of the estimated peak stress (loading can be paused or continuous capture can be performed using a separate high-resolution camera). The raw data obtained by directly photographing the rock sample surface with a high-resolution camera is a single high-resolution macroscopic image of the rock surface.

[0042] The fundamental physical parameters are the raw measurement data reflecting the inherent static properties of the rock, collected before the uniaxial compression test and measured using standardized experimental methods. Specifically, they include: rock density measured using the Archimedes displacement method; rock porosity measured using the saturation method or mercury intrusion porosimetry; and longitudinal and transverse wave velocities measured at both ends of the sample using an ultrasonic transmitter and receiver. The fundamental physical parameters are the specific measured values ​​of the above parameters.

[0043] Visible-near-infrared spectral data are raw spectral data characterizing rock lithology and weathering state. They are collected before the uniaxial compression test by scanning the side of the rock sample with a visible-near-infrared spectrometer.

[0044] Step S102: Perform feature extraction on the multi-source data to obtain a multi-source feature set including strain time series data, internal damage characterization time series data, surface feature data, basic physical feature data, and spectral feature data.

[0045] The process of extracting features from the multi-source data to obtain a multi-source feature set, including strain time-series data, internal damage characterization time-series data, surface feature data, basic physical feature data, and spectral feature data, specifically includes the following sub-steps:

[0046] Step S10201: The speckle image is processed using image correlation to obtain initial strain time series data.

[0047] In this embodiment, commercial or open-source digital image correlation (DIC) analysis software (such as GOMCorrelate or Ncorr) is used to process the temporal sequence of speckle images. Based on the principle of image correlation, the DIC analysis software obtains the full-field displacement field and strain field of the rock sample surface by matching the positional changes of pixels in adjacent frames of speckle images. On this basis, initial temporal information that can characterize the dynamic process of rock deformation is directly extracted from the processed full-field strain field data. This includes the sequence of the average axial strain of the sample surface changing with time, the average width of the transverse strain localization zone before failure, and the average expansion rate of the strain localization zone. The extracted information together constitutes the initial strain temporal data.

[0048] Step S10202: The acoustic emission impact signal is segmented and processed based on a pre-set time window to obtain the initial time series data for internal damage characterization.

[0049] In this embodiment, a pre-set time window of 1 second is defined. This window duration ensures the dynamic capture accuracy of internal micro-fracture signals during rock loading. Subsequently, using this 1-second time window as a unit, the original acoustic emission impact signals acquired by the acoustic emission system during the uniaxial compression test are time-series segmented to ensure coverage of the entire process from loading to rock failure. Within each segmented time window, three core indicators—ringing count, energy release, and event count—are statistically analyzed and calculated. These three indicators from all time windows throughout the loading process are then arranged chronologically to form corresponding ringing count time sequences, energy release time sequences, and event count time sequences. Simultaneously, based on the global data of the acoustic emission impact signals, the Gutenberg-Richter formula (…) is used to… ,in For the amplitude of acoustic emission events, For amplitude greater than or equal to The cumulative number of events is used to perform linear regression on the observed event amplitude and the corresponding cumulative number of events. The absolute value of the slope of the fitted line is taken as the b value. This value can characterize the scale distribution characteristics of micro-fractures inside the rock. The smaller the b value, the higher the proportion of large-scale micro-fractures. Finally, the above three types of time series are integrated with the calculated acoustic emission event b values ​​to obtain the initial data of internal damage characterization time series.

[0050] Step S10203: Perform dimensionality reduction processing on the initial strain time series data and the initial internal damage characterization time series data to obtain strain time series data and internal damage characterization time series data.

[0051] The dimensionality reduction processing of the initial strain time series data and the initial internal damage characterization time series data to obtain the strain time series data and internal damage characterization time series data includes:

[0052] Step S1020301: Perform length normalization processing on the initial strain time series data and the initial internal damage characterization time series data.

[0053] In this embodiment, because different rock samples have different loading times from loading to failure in uniaxial compression tests, the original lengths of the initial strain time series data and the initial internal damage characterization time series data will differ. To address the inconsistency in the original lengths (number of data points) of these two types of data caused by the difference in sample loading time, a standardized data length unification strategy is adopted. Specifically, based on the length distribution of the two types of time series data in the sample set, a unified target length that can completely preserve the dynamic evolution characteristics is determined (e.g., referring to the time series length of most samples, the target length is set to 200 data points). Then, commonly used engineering time series data resampling algorithms such as linear interpolation or cubic spline interpolation are used to interpolate and complete the time series data whose original length is shorter than the target length, and to sample and truncate the time series data whose original length is longer than the target length at equal intervals. Finally, the initial strain time series data and the initial internal damage characterization time series data of all samples are unified into time series sequences of the same length.

[0054] Step S1020302: Perform principal component analysis on the strain time series initial data and the internal damage characterization time series initial data after length normalization to obtain strain time series data and internal damage characterization time series data, respectively.

[0055] In this embodiment, the length-normalized initial strain time series data is organized into a sample-time series feature matrix. Each row of the matrix represents one rock sample, and each column represents the eigenvalue of one time series point. Then, the covariance matrix of this matrix is ​​calculated. By solving the eigenvalues ​​and corresponding eigenvectors of the covariance matrix, the top principal components with a cumulative variance contribution rate exceeding 99% are selected. These principal components together constitute the dimensionality-reduced strain time series data, realizing the elimination of redundancy and the preservation of core features of the original strain time series information. Then, the same PCA operation logic is applied to the length-normalized initial internal damage characterization time series data to obtain the internal damage characterization time series data.

[0056] Step S10204: Perform feature extraction on the surface image data according to the pre-trained feature extractor to obtain surface feature data.

[0057] In this embodiment, the feature extractor is a ResNet-50 deep convolutional neural network model pre-trained on the ImageNet dataset. After being trained on large-scale image data, this network has excellent spatial feature extraction capabilities and can effectively capture structural information related to the failure mode, such as rock surface texture, joints, and initial defects.

[0058] The surface image data is preprocessed and scaled to a fixed size (e.g., 224×224 pixels) to meet the input format requirements of the pre-trained feature extractor. Then, the preprocessed surface image is input into the ResNet-50 network for forward propagation calculation. Without changing the network's pre-training parameters, only the output of the global average pooling layer in the network is extracted. This output is a 2048-dimensional high-dimensional numerical vector, which is the surface feature data obtained after feature extraction from the surface image data.

[0059] Step S10205: Perform vector processing on the basic physical parameters to obtain basic physical feature data.

[0060] In this embodiment, the basic physical parameters include rock density, rock porosity, and rock longitudinal wave velocity. The specific measured values ​​of the above three basic physical parameters of a single rock sample are structured and integrated, and directly spliced ​​to form a 3-dimensional numerical vector, namely, basic physical feature data.

[0061] Step S10206: Denoise and dimension reduction are performed on the visible-near infrared spectral data to obtain spectral feature data.

[0062] The spectral feature data obtained by denoising and dimensionality reduction of the visible-near-infrared spectral data includes:

[0063] Step S1020601: Denoise the visible-near infrared spectral data based on Savitzky-Golay smoothing to obtain spectral denoised data.

[0064] In this embodiment, visible-near-infrared spectral data generally contains random noise introduced by instrument measurement errors, ambient light interference, and reflections from tiny impurities on the sample surface. The Savitzky-Golay smoothing algorithm is used to denoise the original spectral reflectance data. During the operation, the smoothing window size (usually 5-15 consecutive data points are selected to form a sliding window to ensure coverage of local spectral features while avoiding over-smoothing) and the polynomial fitting order (2nd-3rd order polynomials are commonly used to balance noise reduction and spectral feature preservation) need to be reasonably set according to the number of bands and the distribution of characteristic peaks in the spectral data. The algorithm performs polynomial fitting on the spectral data based on the least squares method within each sliding window, and replaces the values ​​of the original spectral data within the window with the values ​​of the fitted curve, thereby effectively filtering out random noise while preserving the characteristic peak information related to rock and mineral composition and weathering state in the visible-near-infrared spectrum to the greatest extent. After the above processing, denoised spectral data with eliminated random interference and clearer spectral features is obtained.

[0065] Step S1020602: Optimize the spectral denoising data using standard normal variable transformation to obtain optimized spectral data.

[0066] In this embodiment, although the spectral denoising data has been denoised, non-lithological factors such as differences in sample surface roughness and uneven particle size may still cause spectral baseline drift or intensity fluctuations, interfering with subsequent lithological feature extraction. For the spectral denoising data of a single rock sample, the mean reflectance of all spectral bands of the sample is first calculated. with standard deviation Then, for each band, the reflectance value ( According to the formula By performing band-by-band calculations, the reflectance value of each band is converted into a standardized value. This process is repeated across all bands of the sample spectrum to complete the above calculations, ultimately yielding the optimized spectral data for the sample.

[0067] Step S1020603: Dimensionality reduction processing is performed on the spectral optimization data based on principal component analysis to obtain spectral feature data.

[0068] In this embodiment, a structured data matrix is ​​formed by taking the spectral optimization data of each rock sample as one row and the reflectance value of each wavelength band as one column. Then, the covariance matrix of this matrix is ​​calculated. By solving the eigenvalues ​​and corresponding eigenvectors of the covariance matrix, the principal components that can reflect the main information of the spectral data are obtained. The principal components are sorted in descending order of eigenvalues. The first K principal components are selected in sequence until the cumulative variance contribution rate of the selected principal components reaches a preset threshold. Finally, the scores of these K principal components are concatenated in sequence to form a K-dimensional numerical vector for each sample. This vector is the spectral feature data.

[0069] Step S2: Construct a feature coefficient prediction model by inputting the multi-source feature set into the feature coefficient prediction model to obtain feature coefficients.

[0070] The construction of the feature coefficient prediction model specifically includes the following sub-steps:

[0071] Step S201: Determine the optimal regularization strength parameter through cross-validation.

[0072] In this embodiment, the optimal regularization strength parameter is determined through cross-validation, which requires determining the optimal regularization strength parameter based on the general objective function form of LASSO regression. According to the dimensions and distribution of the multi-source feature set (including strain time-series data, internal damage characterization time-series data, surface feature data, basic physical feature data, and spectral feature data), a set of candidate values ​​covering a reasonable range is preset to ensure that it can cover potential values ​​that effectively compress redundant features and retain key prediction features. The training samples corresponding to the multi-source feature set are randomly divided into K folds (commonly 5 folds), with each fold sample being mutually exclusive and the overall data distribution consistent, avoiding the influence of sample partitioning bias on parameter evaluation. For each candidate, 1 fold is selected as the validation set, and the remaining K-1 folds are selected as the training set. The training set is substituted into the general objective function to train the LASSO regression model, and then the model performance is calculated using the validation set. The average performance of the corresponding K validations is recorded. Finally, the average performance of all candidates is compared, and the value with the smallest mean square error in rock strength prediction is selected as the optimal regularization strength parameter.

[0073] Step S202: Construct an objective function based on the optimal regularization strength parameter.

[0074] The objective function satisfies the following formula:

[0075]

[0076] in, Let be the objective function. Let be the vector of regression coefficients to be solved. The number of samples for the feature coefficient prediction model. For the first The true rock strength values ​​of each training sample. For the first The feature vectors of each training sample The regularization strength parameter is... For the number of features, Let the first of the regression coefficients in the vector to be solved be the first... Each component.

[0077] Step S203: Based on minimum absolute shrinkage and selection operator regression, construct a feature coefficient prediction model using the objective function.

[0078] In this embodiment, the training sample data used to construct the feature coefficient prediction model comes from multiple sets of standard rock samples (processed into cylinders with a diameter of 50 mm and a height of 100 mm according to the recommendations of the International Society for Rock Mechanics) with different lithologies (such as granite, sandstone, and limestone), different weathering states, and different water-bearing conditions). The raw data includes speckle images, acoustic emission impact signals, surface images, basic physical parameters (density, porosity, P-wave velocity, etc.), and visible-near-infrared spectral data, acquired simultaneously through uniaxial compression tests. The raw data is preprocessed (speckle images are subjected to DIC analysis, length normalization, and PCA dimensionality reduction to obtain strain time-series data; acoustic emission impact signals are subjected to time window segmentation, length normalization, and PCA dimensionality reduction to obtain internal damage characterization time-series data; surface images are processed using a pre-trained ResNet-50 to extract surface feature data; and basic physical parameters are processed using vector processing). Basic physical characteristic data were obtained. Visible-near-infrared spectral data were smoothed using Savitzky-Golay, transformed using standard normal variables, and reduced using PCA to obtain spectral characteristic data, forming a historical multi-source feature set. The historical multi-source feature data for each training sample was matched with the true label (the true rock strength value was taken from the peak stress of the force-displacement curve of a pressure testing machine). Due to the significant differences in the dimensions and numerical ranges of different features (e.g., wave velocity can reach several kilometers per second, while porosity is only a few tenths of a second), standardization was required to prevent certain features from dominating the model training. The Z-score standardization method was used, and each feature dimension was processed as follows: subtract the mean of that dimension across all training samples, and then divide by its standard deviation. The Z-score standardization method was applied to each feature dimension using the following formula: [Formula omitted for brevity]. This yielded the standardized features. :

[0079]

[0080] in, This is the mean of this feature dimension on the training set. This is the standard deviation of the feature dimension on the training set, ultimately resulting in a mean of 0 and a standard deviation of 1 for each feature dimension.

[0081] Finally, taking the objective function with the optimal regularization strength parameter as the core of optimization, the training samples composed of the above multi-source feature set and the corresponding real labels are input, and the regression coefficient vector that minimizes the objective function is solved by the optimization algorithm of minimum absolute shrinkage and selection operator regression (such as gradient descent), thus constructing a feature coefficient prediction model that can receive multi-source features and output the corresponding feature coefficients.

[0082] Step S3: Based on the feature coefficients, the multi-source feature set is filtered to obtain the optimal feature subset.

[0083] In this embodiment, when filtering the multi-source feature set (including strain time series data, internal damage characterization time series data, surface feature data, basic physical feature data, and spectral feature data) based on feature coefficients, the core principle is based on the numerical properties of the feature coefficients: retaining features with non-zero feature coefficients (due to L1 regularization, the coefficients of redundant features will approach 0, and features with non-zero coefficients represent features that effectively contribute to rock strength prediction), and eliminating redundant features with zero feature coefficients. The final integrated feature set is the optimal feature subset, which can be directly used for training subsequent rock strength and failure mode prediction models, ensuring the effectiveness and conciseness of the model input features.

[0084] Step S4: Construct a rock strength and failure mode prediction model, and optimize the hyperparameters of the model using an automated hyperparameter optimization framework. The loss function of the rock strength and failure mode prediction model consists of three parts: the mean squared error of strength prediction, the cross-entropy of failure mode classification, and a penalty term based on physical priors. This loss function satisfies the following formula:

[0085]

[0086] in, This is the total loss function; The weights are used to balance the rock strength prediction loss and the failure mode classification loss; The number of samples; No. The true value of rock strength for each sample; For the first Predicted rock strength values ​​for each sample; The number of categories of destruction modes; This is an indicator function; its value is 1 when the internal condition is true, and 0 otherwise. For the first The true damage pattern category of each sample; For the first The probability that each sample is predicted to be of destruction mode category c; The weighting coefficients for the physical prior penalty term; For the first P-wave velocity of each sample; The intensity-wave velocity ratio coefficient is obtained by fitting an empirical dataset; the third term is... This is a physical prior penalty term. Based on the widely accepted knowledge in rock mechanics that "rock strength and P-wave velocity are usually positively correlated," this penalty term constrains the model's strength predictions. Not related to the longitudinal wave velocity The estimated values ​​deviated too much, so physical knowledge was embedded into the data-driven model. This effectively prevented the model from making absurd predictions that violated physical laws, and improved the model's generalization ability and robustness in cases of scarce samples or high noise. The proportionality coefficient... It can be predetermined by linear regression on historical datasets.

[0087] The rock strength and failure mode prediction model is constructed using the optimal feature subset obtained by LASSO regression as the core of the training data. This subset is derived from the processing and screening of multi-source features, specifically including historical strain time series data reduced by PCA, historical internal damage characterization time series data, historical surface feature data extracted by ResNet-50, historical basic physical feature data processed by vector, and historical spectral feature data optimized by SNV and reduced by PCA, as well as the corresponding historical rock strength and historical failure modes. The historical rock strength is calculated from the force-displacement curve of the pressure testing machine, and serves as the true strength label of the sample. The failure modes include tensile failure, shear failure, and combined failure. The historical failure modes are determined by at least two experienced engineers based on the final failure morphology of the sample after the test, combined with the failure process recorded by DIC, and are used as classification labels. Tensile failure (mainly manifested as axial splitting), shear failure (mainly manifested as single or multiple shear slip surfaces), and combined failure (tensile-shear mixed failure).

[0088] The model architecture adopts a deep learning framework with multi-source feature fusion. First, adaptive feature mapping is performed on different types of features. Then, the feature fusion module integrates multi-source information and finally sets up two prediction branches: one outputs the predicted rock strength value through a regression layer, and the other outputs the rock failure mode category through a classification layer (such as a Softmax layer). During training, the sum of the mean square error of strength prediction and the cross-entropy of failure mode classification is used as the joint loss function. The model parameters are iteratively optimized based on rock sample data labeled with actual strength values ​​and failure modes to ensure that the model has both accurate strength regression and failure mode classification capabilities.

[0089] The rock strength and failure mode prediction model includes a long short-term memory network layer, a convolutional network layer, a fully connected layer, a feature fusion layer, and a dual-output layer.

[0090] The Long Short-Term Memory (LSTM) network layer is mainly used to process time-series data in multi-source features. By capturing the dynamic evolution of time-series features, it extracts key information in the time dimension. The Convolutional Neural Network (CNN) layer is used for data with spatial correlation or significant local features. It extracts key local information by sliding convolutional kernels and strengthens the feature representation in the spatial dimension. The fully connected layer serves as the core of feature preprocessing and adaptation. It performs nonlinear mapping and dimension unification on the input multi-type features, transforming features from different sources into feature vector formats suitable for subsequent fusion, laying the foundation for multi-source information integration. The feature fusion layer inherits the output of the fully connected layer. It eliminates the information differences between different types of features through feature concatenation and weight allocation, extracting globally correlated features that are valuable for both intensity prediction and damage mode classification, achieving effective complementarity and synergy of multi-source features. The dual-output layer is designed for the model's dual prediction objectives. One layer is a regression-type output layer, which outputs a continuous rock strength prediction value that is consistent with the actual peak rock strength dimension based on the fused global features. The other layer is a classification output layer, which outputs the probability of discrete categories of damage modes such as tensile damage, shear damage, and combined damage through the Softmax activation function. Finally, it simultaneously completes the accurate prediction of rock strength and damage mode.

[0091] The rock strength and failure mode prediction model includes a spectral correction submodule. The core function of this submodule is to utilize the spectral feature data of the rock to intrinsically characterize its lithological composition and weathering state, and to specifically calibrate the global feature vector output by the feature fusion layer to correct prediction biases caused by lithological differences. Its specific structure and workflow are as follows: For the input part, the spectral correction submodule uses the spectral feature data extracted from the visible-near-infrared spectral data as its sole input. For the structure part, this submodule is a lightweight neural network consisting of two fully connected layers, with a ReLU activation function used to introduce nonlinearity in between. For the correction parameter generation part, the spectral feature data undergoes a nonlinear transformation through this subnetwork, outputting two vectors: a correction weight vector. A corrected bias vector b, where the dimensions of both vectors are the same as the global feature vector output by the feature fusion layer. The dimensions are the same. This process can be represented as:

[0092] in, This is spectral feature data.

[0093] In the feature correction operation section, after the feature fusion layer and before the dual output layer, the spectral correction submodule performs its core correction operation. This operation modifies the global feature vector. Element-wise scaling and offsetting are performed to generate a lithology-calibrated corrected eigenvector. Its mathematical expression is:

[0094]

[0095] in, This represents the Hadamard product (i.e., element-wise multiplication). This operation is equivalent to learning a lithology-based, personalized scaling factor and offset for each dimension of the global feature vector, thereby fine-tuning the feature representation to better fit the specific lithological context of the current rock sample.

[0096] The dual-output layer is designed for models with two prediction targets, and its input is the spectral correction submodule. One layer is a regression-based output layer, which outputs continuous rock strength predictions based on calibrated features, consistent with the actual peak strength dimension of the rock. The other layer is a classification-based output layer, which outputs the probability of discrete failure modes such as tensile failure, shear failure, and combined failure through a softmax activation function. Ultimately, accurate predictions of both rock strength and failure mode are achieved simultaneously.

[0097] Step S5: Input the optimal feature subset into the optimized rock strength and failure mode prediction model to obtain the rock strength prediction value and failure mode.

[0098] In this embodiment, when the optimal feature subset is input into the optimized rock strength and failure mode prediction model, the model's adaptation layer first performs dimensional matching and format conversion on the feature subset to ensure that temporal features are smoothly input into the long short-term memory network layer, spatial correlation features are accurately input into the convolutional neural network layer, and static features are efficiently connected to the fully connected layer. After each layer completes feature extraction, the feature fusion layer eliminates the differences between multi-source features and integrates them into a global feature vector. Finally, the results are generated synchronously by the dual output layers: the regression output layer outputs a continuous rock strength prediction value with the same dimension as the actual peak rock strength value based on the optimized weight parameters, and the categorization output layer outputs the failure mode through the Softmax activation function, thus achieving synchronous prediction of rock strength and failure mode.

[0099] like Figure 2 As shown, in another aspect, the present invention also provides a rock strength and failure mode prediction system, including: a processor, an input device, an output device, and a memory, wherein the processor, the input device, the output device, and the memory are interconnected, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to call the program instructions to execute relevant steps of a relevant embodiment of the rock strength and failure mode prediction method of the present invention.

[0100] This invention provides a rock strength and failure mode prediction system. The functional components can be integrated into a single processing unit, or each component can exist independently, or two or more components can be integrated into one unit. The integrated components can be implemented in hardware or as software functions.

[0101] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A method for predicting rock strength and failure mode, characterized in that, The method includes: A multi-source feature set was obtained based on data acquisition from the rock samples to be tested; Construct a feature coefficient prediction model, and input the multi-source feature set into the feature coefficient prediction model to obtain feature coefficients; The construction of the feature coefficient prediction model includes: The optimal regularization strength parameter is determined through cross-validation; Construct an objective function based on the optimal regularization strength parameter; The objective function satisfies the following formula: in, Let be the objective function. Let be the vector of regression coefficients to be solved. The number of samples for the feature coefficient prediction model. For the first The true rock strength values ​​of each training sample. For the first The feature vectors of each training sample The regularization strength parameter is... For the number of features, Let the first of the regression coefficients in the vector to be solved be the first... Each component; Based on minimum absolute contraction and selection operator regression, a feature coefficient prediction model is constructed using the objective function. The optimal feature subset is obtained by filtering the multi-source feature set based on the feature coefficients. A rock strength and failure mode prediction model was constructed, and the hyperparameters of the rock strength and failure mode prediction model were optimized using an automated hyperparameter optimization framework. The rock strength and failure mode prediction model satisfies the following formula: in, For loss function, As weight, The sample size for the rock strength and failure mode prediction model. No. The true value of rock strength for each sample. For the first Predicted rock strength values ​​for each sample The number of categories of destruction modes, For indicator functions, For the Realistic Destruction mode category, For the first Each sample is predicted as a category. The probability, As weight, This is the proportionality coefficient. For the first The longitudinal wave velocity of each sample; The optimal feature subset is input into the optimized rock strength and failure mode prediction model to obtain the rock strength prediction value and failure mode.

2. The method for predicting rock strength and failure mode according to claim 1, characterized in that, The multi-source feature set obtained from the data acquisition of the rock sample to be tested includes: Multi-source data were obtained based on data acquisition from the rock samples to be tested; Feature extraction is performed on the multi-source data to obtain a multi-source feature set including strain time series data, internal damage characterization time series data, surface feature data, basic physical feature data, and spectral feature data.

3. The method for predicting rock strength and failure mode according to claim 2, characterized in that, The multi-source data includes speckle images, acoustic emission impact signals, surface image data, fundamental physical parameters, and visible-near-infrared spectral data. The multi-source feature set obtained by feature extraction from the multi-source data includes strain time-series data, internal damage characterization time-series data, surface feature data, fundamental physical feature data, and spectral feature data, comprising: The initial strain time series data are obtained by processing the speckle image using image correlation methods. The acoustic emission impact signal is segmented and processed based on a pre-set time window to obtain the initial time series data for internal damage characterization; The strain time series initial data and the internal damage characterization time series initial data are subjected to dimensionality reduction processing to obtain strain time series data and internal damage characterization time series data; Surface feature data is obtained by extracting features from the surface image data using a pre-trained feature extractor. The basic physical parameters are processed into vector data to obtain basic physical feature data; The visible-near-infrared spectral data are denoised and dimensionally reduced to obtain spectral feature data.

4. The method for predicting rock strength and failure mode according to claim 3, characterized in that, The dimensionality reduction processing of the initial strain time series data and the initial internal damage characterization time series data to obtain the strain time series data and the internal damage characterization time series data includes: The initial strain time series data and the initial internal damage characterization time series data are subjected to length normalization processing; Principal component analysis was performed on the length-normalized initial strain time series data and the initial internal damage characterization time series data to obtain strain time series data and internal damage characterization time series data, respectively.

5. The method for predicting rock strength and failure mode according to claim 3, characterized in that, The denoising and dimensionality reduction processing of the visible-near-infrared spectral data to obtain spectral feature data includes: The visible-near-infrared spectral data are denoised by Savitzky-Golay smoothing to obtain denoised spectral data. The spectral denoising data is optimized by using standard normal variable transformation to obtain spectral optimized data; Principal component analysis was used to perform dimensionality reduction on the optimized spectral data to obtain spectral feature data.

6. The method for predicting rock strength and failure mode according to claim 1, characterized in that, The rock strength and failure mode prediction model includes a spectral correction submodule.

7. A rock strength and failure mode prediction system, characterized in that, include: The system includes a processor, an input device, an output device, and a memory, all interconnected, wherein the memory stores a computer program comprising program instructions, and the processor is configured to invoke the program instructions to execute a rock strength and failure mode prediction method as described in any one of claims 1 to 6.

Images