An intelligent rock mass strength parameter analysis method and system based on a data-mechanics-model-driven model

Abstract

This invention belongs to the field of tunnel engineering, specifically relating to a data-mechanism-model-driven intelligent analysis method and system for rock mass strength parameters. It includes: generating a simulation dataset adapted to the current engineering rock mass strength and geological conditions through high-fidelity drilling numerical simulation, and training a multilayer perceptron as a physical information proxy model to achieve physical mapping of the rock drilling process. Secondly, the three-dimensional parameters are processed by three-dimensional spatial aggregation and segmentation. Combining graph neural network analysis and transformer model information extraction, a hybrid loss function for multi-dimensional data is defined. This hybrid loss function, together with the physical loss function, forms an optimized loss function with physical information and drilling data. An automatic optimization process for the loss function is then performed to achieve rock mass strength prediction. This invention can effectively analyze rock mass strength parameters.

Description

Technical Field

[0001] This invention belongs to the field of tunnel engineering, specifically relating to a data-mechanics-model-driven intelligent analysis method and system for rock mass strength parameters. Background Technology

[0002] In increasingly complex geological environments, the types and distribution of disasters occurring during tunnel construction exhibit strong suddenness and unpredictability. The physical and mechanical parameters of the rock mass itself, combined with the cutting and weakening effects of complex structural surfaces and creep effects, collectively lead to damage and failure of the complex rock mass, further resulting in deformation and instability of the surrounding rock at the tunnel face. The analysis and evaluation of surrounding rock parameters during construction are limited by technical and time constraints, and the experiential application of surrounding rock classification in engineering practice remains limited. Taking the drill-and-blast method as an example, the design methods for key processes such as excavation and blasting rely on manual judgment supplemented by on-site testing, making it impossible to quickly and accurately classify the surrounding rock, which indirectly exacerbates the waste of support stiffness. On the other hand, traditional manual drilling using pneumatic drills, with isolated drilling mechanisms, cannot achieve the conversion between physical parameters and real-time data, and the high-dimensional connection with geological parameters remains unclear. In recent years, as the drill-and-blast method for tunneling has undergone a system transformation from manual to mechanized operations, domestic and foreign equipment manufacturers such as Ampert and CRCC Heavy Industry have developed a series of automated and intelligent complete sets of technical equipment. Among them, the rock drilling rig, which is in direct contact with the surrounding rock, can extract a large amount of drilling data from the drilling process thanks to the development of automation technology and measurement-while-drilling technology, making it possible to analyze the rock mass parameters in the tunnel face.

[0003] The physical and mechanical changes that occur during the contact between the drill bit and drilling tools vary depending on the rock conditions. While the drilling process through intact rock can allow for observation of rock strength using theoretical-empirical models, these models lack adaptability to complex and variable geological environments. This has led to a shift in research focus from the rock fracturing mechanism itself to the relationship between drilling data and rock properties. The key characterizer of "abnormal behavior" during drilling is the data itself; focusing on the correlation between changes in drilling parameters and different hardness levels and rock types forms the basis of many studies. The amount of drilling parameter data obtained during engineering operations is enormous. The accuracy of single-hole data is often correlated with time and depth, resulting in cyclical data from different spatial locations and time sets. Furthermore, the types of data obtained are complex, including drill bit penetration speed, forward and tangential forces, and information such as recorded time and location. This requires analyzing both temporal and spatial dual-series drilling data.

[0004] Considering the uncertainty and sparsity of parameters in complex engineering geology, under such conditions, the specially designed neural network architecture, facing the massive amount of data and the implicit complex physical laws, requires higher quality data to train machine learning models with superior feature engineering and stronger generalization ability. Therefore, it is necessary to develop a physical information fusion model for refined analysis of surrounding rocks to realize geological informatization while enhancing the relationship between feature parameters, in order to serve as an effective means to help understand, evaluate and predict geological conditions. Summary of the Invention

[0005] The present invention provides an intelligent analysis method and system for rock mass strength parameters based on a data-mechanism-model dual-driven approach. This method overcomes the problems of poor transfer learning ability and insufficient model generalization ability in deep learning by integrating a physical information proxy model constructed from high-fidelity numerical simulation with a deep learning feature analysis model. It achieves intelligent analysis of tunnel drilling parameters, thereby predicting the strength parameters of the rock mass ahead of the tunnel face in real time, accurately, and reliably, and providing support for construction decisions.

[0006] According to the present invention, a data-mechanics-model-driven intelligent analysis method for rock mass strength parameters includes the following steps: S1: Constructing a high-fidelity numerical simulation dataset for the drilling process: A simulation model of the drilling process reflecting the interaction mechanism between the drill bit and the rock mass is established based on the finite element method; different combinations of rock mass strength parameters are set in the model for simulation, and the corresponding virtual drilling parameter sequence is output; a simulation dataset covering engineering geological conditions is generated, which contains physical information of the drilling process. S2: Training the physical information proxy model: A multilayer perceptron was selected as the surrogate model to train the acquired virtual drilling parameter sequence, thereby obtaining a fast physical information surrogate model for the drilling process. S3: Input for 3D spatial data clustering: Multi-source time-series data during the drilling process are collected in real time and synchronously. After spatial alignment, the data is sampled at fixed depth intervals to form a continuous dataset. The KNN clustering algorithm is used to cluster the three-dimensional drilling data to obtain the nearest neighbor relationship of the parameters within the space. S4: Physical data mixed data loss: A Transformer encoder is used as the core feature extractor. After high-dimensional mapping, a set of deep spatiotemporal feature vectors representing the comprehensive state of the rock mass in the current drilling section are obtained. Spatial neighborhood information and borehole time-series information are fused and input into a GNN graph neural network to obtain a comprehensive feature matrix containing neighborhood information, which is used as the data loss function. At the same time, a physical information proxy model is fused as the physical loss, which together constitutes a hybrid loss function for physical data mixed data loss. S5: Strength inversion prediction under physical constraints: The hybrid loss function obtained in S4 is trained iteratively using deep learning, and the hyperparameters are optimized using the GA genetic algorithm to construct an intensity prediction model that can extract data from drilling data that conforms to both statistical and physical laws.

[0007] Preferably, in step S1, the virtual drilling parameter sequence includes drilling speed, torque, and axial thrust; the rock strength parameter combination includes the rock's internal friction angle, cohesion, elastic modulus, and Poisson's ratio.

[0008] Preferably, step S1 specifically includes: S1.1: Determine the key simulation parameters that affect drilling response and rock mass strength, including drill pressure, rotation speed, and drill string diameter; for rock masses, it is necessary to determine the internal friction angle, cohesion, elastic modulus, and Poisson's ratio. S1.2: The simulation model is established using the finite element method; the interaction between the drill bit and the rock mass is analyzed using explicit dynamics to simulate the dynamic drilling process of cutting, crushing, extrusion, and friction; the simulation output includes time-series data of drilling speed, torque, and axial thrust during the drilling process. S1.3: Perform data augmentation on the original data generated by the simulation, including adding Gaussian white noise, impulse interference, and simulating data loss; finally, standardize or normalize all input parameters and output responses to form a structured simulation dataset.

[0009] Preferably, step S2 specifically includes: S2.1: The input layer is designed as a parallel channel to receive the timing signals of the timing data, and the output layer is the rock compressive strength; S2.2: For time-series data, a one-dimensional convolutional neural network (1D-CNN) is used to extract local fluctuation features, which are then connected to a gated recurrent unit (GRU) to capture long-range dependencies. The output is a drilling dynamic feature vector. For real-time engineering parameters, a fully connected encoder is used for fusion, and the output is a state feature vector. S2.3: Randomly divide the simulation samples according to the ratio of training set:test set = 7:3, set the hyperparameter range for optimization, and complete the surrogate model construction after multiple training sessions.

[0010] Preferably, in step S3, the parameters of the drilling process in the 3D spatial data clustering input need to be standardized and preprocessed to conform to the format of the Transformer sample set: (1) Slice the continuous data stream into equal-length sequence samples with a fixed length, and perform independent normalization on each parameter; (2) Imput missing and outlier values ​​in the sequence and inject positional codes into the data to preserve temporal information; (3) Divide the processed sequence samples into training set, validation set and test set, and organize the input in batch processing form; After extracting sample data, the KNN network is used to perform unsupervised aggregation of data from different nodes, providing the model with spatial attention labels. At the same time, the Transformer model is responsible for extracting deep and abstract temporal features from the drilling time series data. The features of both are combined and input into the GNN graph neural network as one of the data loss functions for training.

[0011] Preferably, step S3 specifically includes: S3.1: Receive standardized sample data from the drilling rig on site. After data cleaning based on drilling time and drilling trend, extract the drilling time at different drilling depths, calculate the drilling depth increment corresponding to each time point in real time, and encode it into an additional position vector; then concatenate them to form a matrix D containing drilling data at different positions, D=[B1B2…B ... n ] T B n This represents the drilling data for the nth borehole;

[0012] in x,y Represents position, p i,1 ~p i,6 Represents different drilling parameters; S3.2: Using the borehole data structure described in S3.1, borehole clustering is performed based on KNN. First, the distance metric function between borehole data vectors is defined. :

[0013] Indicates the first The first drilling vector One component; Indicates the first The first drilling vector One component; The weight coefficients for each dimension satisfy... ; Clustering label decision function for boreholes: For the target borehole Its K-nearest neighbor set Defined as:

[0014] in This represents the distance sorting function; The adaptive formula for determining the value of K is:

[0015] in: This represents the total number of boreholes. For drilling The parameter dispersion; These are the average drilling parameters; drilling Clustering label decision function for:

[0016] in: For the set of all possible clustering categories; The current specific cluster category; For smoothing terms, when When the denominator is zero, the numerical value is incorrect.

[0017] Preferably, in step S4, the hybrid loss function includes physical loss and data loss. The physical loss is replaced by a physical information proxy model to inject the physical laws during the drilling process, while the data loss is the strength loss of engineering drilling parameters and on-site strength. The two are combined to form a hybrid loss function that provides physical constraints, which serves as the basis for the training and convergence of the subsequent neural network model.

[0018] Preferably, in step S4, the hybrid loss function includes three parts: a data loss term, a physical loss term, and a regularization term.

[0019] This represents the parameters to be optimized in the prediction model; These are the weighting coefficients for each loss term, satisfying... ; For data loss items; This is a physical loss item; This is the regularization loss term; The data loss term is expressed in the form of mean squared error:

[0020] n is the number of samples measured on-site; The primary prediction model is based on the original drilling parameters. The output; For the first The field-measured surrounding rock strength label for each borehole; The physical loss item The calculation formula is:

[0021] in: The primary prediction model is based on the original drilling parameters. The output; For MLP physical proxy models based on physical feature vectors The output; The original drilling parameters are mapped to a physical relation function. The obtained physical feature vector; The regularization loss term adopts the L2 regularization form:

[0022] in For model parameters Dimensions.

[0023] Preferably, in step S5, during the initial training phase, the physical loss is increased. The weights are used to strongly guide the network to learn basic physical relationships; in the later stages of training, the data loss is gradually increased. The weights are adjusted to make the network converge in a direction that aligns with the actual drilling parameters in the engineering process.

[0024] This invention provides an intelligent analysis system for rock mass strength parameters based on data-mechanics-model driven approach, which adopts the aforementioned intelligent analysis method for rock mass strength parameters based on data-mechanics-model driven approach.

[0025] The beneficial effects of this invention are as follows: 1) This invention is based on a physical information proxy model, which directly relates the target to key parameters (torque, thrust) in the mechanical-surrounding rock interaction, ensuring that the inversion results have the physical mechanism basis of the drilling process and avoiding the black box absurd solutions that may occur in pure data models.

[0026] 2) This invention is the first to combine the prediction of surrounding rock strength with high-fidelity numerical simulation and its physical information. It generates a set of geological parameters covering the geological conditions of surrounding rock in real engineering projects for large-scale sample calculations, and combines on-site drilling parameters with surrounding rock strength labels to improve the prediction generalization performance.

[0027] 3) This invention establishes a sample set based on drilling rig parameters, field test data, and a simulation database. Except for initial field testing, the entire process requires no manual intervention or experience-based interpretation. Traditional machine learning methods require a large amount of labeled data for training in the target domain, which is impractical in engineering. This invention, through simulation pre-training, requires only a few field-calibrated samples in practical applications, achieving rapid deployment with zero or small samples, making it highly practical for engineering applications. Simultaneously, the introduction of simulation data reduces over-reliance on expert experience, enabling frontline technicians who are not top experts to obtain reliable analysis results. Attached Figure Description

[0028] Figure 1 This is a flowchart of a data-mechanics-model-driven intelligent analysis method for rock mass strength parameters provided in Example 1; Figure 2 This is a schematic block diagram of the method in Example 1; Figure 3 The simulation model and stress diagram of the drilling process are shown in the embodiment. Figure 4 This is a schematic diagram of the drilling parameter data distribution curve in the embodiment; Figure 5 This is a schematic diagram of 1D-CNN convolutional feature extraction in Example 1; Figure 6 This is a schematic diagram of the Physical Information Proxy Model (MLP) architecture in Example 1; Figure 7 This is a schematic diagram of the Transformer model architecture in Example 1.

[0029] Figure 8 This is a schematic diagram of the actual sample intensity prediction in Example 2. Detailed Implementation

[0030] To further understand the content of this invention, a detailed description of the invention will be provided in conjunction with the accompanying drawings and embodiments. It should be understood that the embodiments are merely illustrative and not limiting of the invention.

[0031] Example like Figure 1 As shown, this embodiment provides an intelligent analytical method for rock mass strength parameters based on data-mechanics-model driven approach, which includes the following steps: S1: Constructing a high-fidelity numerical simulation dataset for the drilling process: A simulation model of the drilling process reflecting the interaction mechanism between the drill bit and the rock mass is established based on the finite element method. Different combinations of rock mass strength parameters are set in the model for simulation, and the corresponding virtual drilling parameter sequence is output. A simulation dataset covering engineering geological conditions is generated, which contains physical information of the drilling process.

[0032] The virtual drilling parameter sequence includes drilling speed, torque, and axial thrust; the rock strength parameter combination includes the rock's internal friction angle, cohesion, elastic modulus, and Poisson's ratio.

[0033] In step S1, specifically: S1.1: Determine the key simulation parameters that affect drilling response and rock mass strength, including drill pressure, rotation speed, and drill string diameter; for rock masses, it is necessary to determine the internal friction angle, cohesion, elastic modulus, and Poisson's ratio. S1.2: The simulation model is established using the finite element method; the interaction between the drill bit and the rock mass is analyzed using explicit dynamics to simulate the dynamic drilling process of cutting, crushing, extrusion, and friction; the simulation output includes time-series data of drilling speed, torque, and axial thrust during the drilling process. S1.3: Perform data augmentation on the original data generated by the simulation, including adding Gaussian white noise, impulse interference, and simulating data loss; finally, standardize or normalize all input parameters and output responses to form a structured simulation dataset.

[0034] S2: Training the physical information proxy model: A multilayer perceptron (MLP) was selected as the surrogate model to train a fast physical information surrogate model for the drilling process on the acquired virtual drilling parameter sequence. A schematic diagram of the MLP architecture is shown below. Figure 6 As shown.

[0035] In step S2, specifically: S2.1: The input layer is designed as a parallel channel to receive time-series data (such as torque, drilling pressure, and rotational speed) and the output layer is the rock compressive strength. S2.2: For time-series data, a one-dimensional convolutional neural network (1D-CNN) is used to extract local fluctuation features. A schematic diagram of 1D-CNN convolutional feature extraction is shown below. Figure 5 As shown; then, the gated loop unit (GRU) is connected to capture long-range dependencies, and the output is a drilling dynamic feature vector; real-time engineering parameters are fused through a fully connected encoder, and the output is a state feature vector; S2.3: Randomly divide the simulation samples according to the ratio of training set:test set = 7:3, set the hyperparameter range for optimization, and complete the surrogate model construction after multiple training sessions.

[0036] S3: Input for 3D spatial data clustering: Multi-source time-series data are collected in real time and synchronously during the drilling process. After spatial alignment, a continuous dataset is formed by sampling at fixed depth intervals. The KNN clustering algorithm is used to cluster the three-dimensional drilling data to obtain the nearest neighbor relationship of the parameters within the space.

[0037] In step S3, the parameters of the drilling process need to be standardized and preprocessed in the 3D spatial data clustering input to conform to the format of the Transformer sample set: (1) Slice the continuous data stream into equal-length sequence samples with a fixed length, and perform independent normalization on each parameter; (2) Imput missing and outlier values ​​in the sequence and inject positional codes into the data to preserve temporal information; (3) Divide the processed sequence samples into training set, validation set and test set, and organize the input in batch processing form; After extracting sample data, the KNN network is used to perform unsupervised aggregation of data from different nodes, providing the model with spatial attention labels. At the same time, the Transformer model is responsible for extracting deep and abstract temporal features from the drilling time series data. The features of both are combined and input into the GNN graph neural network as one of the data loss functions for training.

[0038] In step S3, specifically: S3.1: Receive standardized sample data from the drilling rig on site. After data cleaning based on drilling time and drilling trend, extract the drilling time at different drilling depths, calculate the drilling depth increment corresponding to each time point in real time, and encode it into an additional position vector; then concatenate them to form a matrix D containing drilling data at different positions, D=[B1B2…B ... n ] T B n This represents the drilling data for the nth borehole;

[0039] in x,y Represents position, p i,1 ~p i,6 Represents different drilling parameters; S3.2: Using the borehole data structure described in S3.1, borehole clustering is performed based on KNN. First, the distance metric function between borehole data vectors is defined. :

[0040] Indicates the first The first drilling vector One component; The weight coefficients for each dimension satisfy... ; Clustering label decision function for boreholes: For the target borehole Its K-nearest neighbor set Defined as:

[0041] in This represents the distance sorting function; The adaptive formula for determining the value of K is:

[0042] in: This represents the total number of boreholes. For drilling The parameter dispersion; These are the average drilling parameters; drilling Clustering label decision function for:

[0043] in: For the set of all possible clustering categories; For the current specific cluster category; For smoothing terms, when When the two borehole vectors are exactly the same, the denominator is zero, which leads to numerical errors.

[0044] S4: Physical data mixed data loss: The Transformer encoder is used as the core feature extractor. After high-dimensional mapping, a set of deep spatiotemporal feature vectors representing the overall state of the rock mass in the current drilling section are obtained. The schematic diagram of the Transformer model architecture is shown below. Figure 7 As shown, spatial neighborhood information and borehole time-series information are fused and input into a GNN graph neural network to obtain a comprehensive feature matrix containing neighborhood information, which is used as the data loss function. At the same time, a physical information proxy model is fused as the physical loss, and together they form a hybrid loss function for physical data hybrid loss.

[0045] In step S4, the hybrid loss function includes physical loss and data loss. The physical loss is replaced by a physical information proxy model to inject the physical laws during the drilling process, while the data loss is the strength loss of engineering drilling parameters and on-site strength. The two are combined to form a hybrid loss function that provides physical constraints, which serves as the basis for the training and convergence of the subsequent neural network model.

[0046] In step S4, the hybrid loss function consists of three parts: a data loss term, a physical loss term, and a regularization term.

[0047] This represents the parameters to be optimized in the prediction model; These are the weighting coefficients for each loss term, satisfying... ; For data loss items; This is a physical loss item; This is the regularization loss term; The data loss term is expressed in the form of mean squared error:

[0048] n represents the number of samples measured on-site (i.e., the number of labeled boreholes). The primary prediction model is based on the original drilling parameters. The output; For the first The field-measured surrounding rock strength label for each borehole; The physical loss item The calculation formula is:

[0049] in: The primary prediction model is based on the original drilling parameters. The output; For MLP physical proxy models based on physical feature vectors The output; The original drilling parameters are mapped to a physical relation function. The obtained physical feature vector; The regularization loss term adopts the L2 regularization form:

[0050] in For model parameters Dimensions.

[0051] S5: Strength inversion prediction under physical constraints: The hybrid loss function obtained in S4 is trained iteratively using deep learning, and the hyperparameters are optimized using the GA genetic algorithm to construct an intensity prediction model that can extract data from drilling data that conforms to both statistical and physical laws.

[0052] In step S5, during the initial training phase, the physical loss is increased. The weights are used to strongly guide the network to learn basic physical relationships; in the later stages of training, the data loss is gradually increased. The weights are adjusted to make the network converge in a direction that aligns with the actual drilling parameters in the engineering process.

[0053] like Figure 2 As shown, the aforementioned steps can be summarized into a simplified technical roadmap. First, a high-fidelity numerical simulation dataset of the drilling process needs to be constructed. Based on the obtained dataset, a training physical information proxy model is established. A three-dimensional spatial data clustering input method (GNN) based on the on-site drilling process is established to form a training loss function for physical data mixing. After training through a neural network, intensity inversion prediction based on physical constraints is achieved.

[0054] This embodiment provides a data-mechanism-model-driven intelligent analysis system for rock mass strength parameters, which adopts the aforementioned data-mechanism-model-driven intelligent analysis method for rock mass strength parameters.

[0055] Example 2 This embodiment provides a data-mechanism-model-driven intelligent analysis method for rock mass strength parameters, which includes the following steps: S1: Constructing a high-fidelity numerical simulation dataset for the drilling process: First, determine the geological conditions of the engineering project and collect data to obtain data related to the simulation model and establish a simulation calculation database.

[0056] Specifically, the collected data includes the type of surrounding rock, the lithology of the surrounding rock, and the uniaxial compressive strength of the rock. UCS internal friction angle φ Cohesion c Drill bit size, drill bit operating pressure W Drill bit operating speed R Model calculations include acquiring the force variation curves of the model during drilling and organizing them into a database to serve as the data foundation for large-scale sample machine learning models.

[0057] S101: Using a full factorial design method, 1000 different parameter combinations are generated within the above parameter space. UCS , φ , c , W , R For each set of parameters, a high-fidelity numerical simulation was run to simulate the complete drilling process.

[0058] S102: Establish a numerical simulation model of rock based on the finite element method ABAQUS. In the model, import the drill bit model of the rock drilling rig used in the project, perform mesh generation, and place it in a unified computational space with the generated rock mass mesh. The constitutive model is selected as an elastoplastic model.

[0059] S103: Extract the time-varying drilling response sequence from each simulation result as label data. Key response features include, but are not limited to: time-varying instantaneous drilling speed, axial force, radial force, and torque. Figure 3 As shown, a stress cloud map of a certain time step is obtained. The stress (0.8~2.4MPa) of the current time step is extracted from the cloud map. The same applies to parameters such as speed and torque.

[0060] Constructing a simulation dataset: For each set of input parameters [ UCS , φ , c , W , R [and its corresponding multi-channel drilling response time series data] V(t) , F axial (t) , F rotaryl (t) , T rotaryl (t) [Linking] to form a high-fidelity simulation dataset ,in This is a concatenated vector of rock mass parameters and drilling parameters. This is multidimensional response time-series data.

[0061] S2: Training the physical information proxy model: S201: For simulation datasets The response time series data in the dataset undergoes standardized data preprocessing, and the sampling length is standardized to obtain a uniform-length dataset. The time-series data, along with the recorded rock mass strength under the working condition, are used as output labels to complete the assembly of the training sample library.

[0062] S202: Model Architecture Design: Construct an MLP model as a proxy model. Its core function is to establish drilling data obtained from the input layer (different rock mass parameters). An end-to-end mapping from the input layer to the output layer (rock mass strength). The model consists of an input layer, multiple hidden layers, and an output layer connected sequentially.

[0063] S203: Establishment Basic statistical characteristics of time series data include mean, variance, and extreme values. These are used to describe the basic distribution and anomalies of the data. It is necessary to filter and integrate the values ​​of the same parameter at different locations and times, and calculate the frequency of occurrence within different value intervals to obtain the data distribution curves for each drilling parameter, such as... Figure 4 As shown, data distribution curves based on six parameters, including rotational speed and rotational pressure, can be obtained. At this point, basic statistical feature values ​​are extracted based on the characteristics of the curves.

[0064] Specifically, the calculation methods for the basic statistical characteristics of time series data: In the formula Indicates the first One parameter channel, This represents the total number of time steps in the current time series data; S204: Next, feature extraction is performed on the time-series data, and a lightweight 1D-CNN (such as...) is constructed. Figure 5 As shown), the time step is 5000. A suitable sliding window size can be selected, such as 100. Its convolution kernel automatically slides on the original time series data, learns and extracts local morphological patterns (such as pulses, steps, and trend segments of different shapes), and obtains a combination of statistical and temporal surrogate model features.

[0065] Specifically, the 1D-CNN feature extraction structure for temporal data uses temporal data with drill-down parameters as the input layer, and the shape of the input data is as follows: ,in Batch size: This indicates the number of samples processed at one time. : Sequence length, representing the number of time steps for each sample; 1: Number of channels, representing single-channel time series data.

[0066] The key convolution in the structure relies on the convolution kernel sliding sequentially from left to right across the temporal data, moving one stride at a time (default is 1). Padding controls the length of the output sequence. The formula is as follows:

[0067] In the formula Indicates the first The output value at each position, i.e. the feature value after the convolution operation; Indicates the convolution kernel number 1 j Each weight parameter can be used to learn filter weights; Indicates the (th)th input sequencei + j ) elements of original time series data points; This represents the learnable bias parameter; This indicates the kernel size.

[0068] The drilling parameters include pressure, velocity, and other parameters, which are multi-channel data. Based on the above, the convolution formula is extended as follows:

[0069] The number of input channels is 6 in this example; Indicates the first c The convolution kernel in the nth input channel j Each weight parameter can be used to learn filter weights; Indicates the first c The (th)th sequence in the input channel i + j () elements of original time series data points.

[0070] S205: The extracted features are used as input to the training set to begin training. During training, not only is Mean Squared Error (MSE) used as the base loss, but Mean Absolute Percentage Error (MAPE) and weighted MSE for key regions are also introduced. MAPE balances the error contributions from displacement predictions of different magnitudes, while weighted MSE imposes a greater penalty on prediction errors in deformation-sensitive regions that are of most concern in engineering, thus establishing... UCS and The rock mass strength training error function is used to iterate until convergence.

[0071] S3: Input for 3D spatial data clustering: S301: During drilling operations, determine the sampling frequency, such as collecting data once per second. During the operation, synchronously and continuously collect the timing data from the aforementioned sensors. After drilling operations are completed at the construction site, package the multi-sensor timing data from the entire operation.

[0072] S302: After data acquisition, standardization and cleaning of the data are required. For the current sample label, a uniaxial compression test can be performed on the rock in the current cycle to obtain the rock mass strength, and this strength can be associated with the current drilling data to form the sample label.

[0073] S303: The KNN clustering algorithm was used to perform unsupervised clustering of the samples in terms of spatial location; For new drilling Calculate its distance to all training samples; Select the closest One neighbor (in this embodiment, we take) ); A weighted voting method was used to determine the strength classification, resulting in a qualitative classification label for weak rock strength.

[0074] S4: Physical data mixing loss function: S401: Based on the KNN clustering labels obtained in step S3, convert them into category embedding vectors:

[0075] In this embodiment That is, 5-dimensional category embedding.

[0076] S402: The drilling time sequence data structure obtained in step S3 is as follows: in: Time steps (in this embodiment) ); : Parameter dimensions while drilling (such as drilling pressure, rotational speed, torque, etc.) ); .

[0077] Then, the time-series data is decoded:

[0078] S403: For each drill node It integrates two features: Transformer temporal features: KNN qualitative clustering features: .

[0079] Node features are obtained through a feature fusion layer:

[0080] After inputting the vector fused with spatial features into the GNN graph neural network, for each borehole node, prediction is performed by combining node features with global graph features:

[0081]

[0082] in This represents the predicted intensity value.

[0083]

[0084] Substituting the loss function into the data yields the loss function based on the engineering drilling data.

[0085] S404: Combine the depth feature vector output from step S4 with the data loss. The data is fed into a fully connected neural network (called the data prediction branch) to pre-determine the layer strength. Simultaneously, the same set of drilling data is input into the pre-trained MLP physics model (called the physics prediction branch) from step S2 to obtain predictions based on physical mechanisms.

[0086] S405: The total loss function is composed of data loss. Physical loss With regularization loss The weighted composition aims to simultaneously minimize the deviation between the predicted value and the label, as well as the difference between the data prediction and the physical prediction.

[0087] S406: As Figure 6 As shown, the model can be divided into an input layer, a hidden layer, and an output layer, and the model can be further converged through training.

[0088] S5: Rock mass strength parameter inversion prediction under physical constraints: S501: Since the above optimization problem may be non-convex and has a large number of parameters, a genetic algorithm (GA) is used to globally optimize the model parameters of the data prediction branch. The genetic algorithm simulates the natural selection process, iteratively evolving the population through selection, crossover, and mutation operations to ultimately find the optimal model. The minimum optimal model parameters.

[0089] Specifically, a set of model parameters for the data prediction branch is randomly generated as the initial population.

[0090] Calculate the corresponding parameters for each individual (i.e., a set of model parameters). Fitness assessment is performed, and fitness is defined as follows:

[0091] Selection: A roulette wheel selection is performed based on the fitness of the model parameters to retain the best individuals.

[0092] Crossover: Randomly pair individuals, exchange some parameters, and generate new individuals.

[0093] Mutation: Randomly changing certain parameter values ​​with a small probability to increase population diversity.

[0094] Iteration: Repeat the steps of selection, crossover, and mutation until the fitness converges.

[0095] S502: After optimization, a set of intensity prediction machine learning models adapted to the current geological conditions (such as...) Figure 7As shown in the figure, based on this, cyclic drilling data can be entered for real-time strength prediction. The comparison between predicted and actual strength is as follows. Figure 8 As shown, it can achieve high-precision intensity prediction.

[0096] This embodiment provides a rock mass strength analysis system, which includes: High-fidelity drilling process numerical simulation module: used to obtain the calculation parameter range from the project and carry out calculations.

[0097] Physical Information Agent Model Training Module: Used for extracting and training physical information from computational samples obtained from engineering projects.

[0098] On-site drilling parameter acquisition module: used for acquiring on-site drilling data and sample labels during the drilling process.

[0099] The spatiotemporal feature extraction module based on multidimensional data is used for feature extraction and sample pair construction of drill data.

[0100] Rock mass strength parameter inversion and prediction module under physical constraints: used to load the sample set under physical constraints for training and prediction.

[0101] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it performs the above-described normalization processing based on drilling parameters.

[0102] The present invention and its embodiments have been described above illustratively, and this description is not restrictive. The accompanying drawings are merely one embodiment of the present invention; actual application scenarios, model structures, feature combinations, or deployment architectures are not limited thereto. Therefore, if those skilled in the art, inspired by this description, design a structure, method, or system substantially identical to this technical solution without departing from the spirit of the invention and without creative effort, such design should fall within the protection scope of the present invention.

Claims

1. A data-mechanism-model-driven intelligent analytical method for rock mass strength parameters, characterized in that, Includes the following steps: S1: Constructing a high-fidelity numerical simulation dataset for the drilling process: A simulation model of the drilling process reflecting the interaction mechanism between the drill bit and the rock mass is established based on the finite element method; different combinations of rock mass strength parameters are set in the model for simulation, and the corresponding virtual drilling parameter sequence is output; a simulation dataset covering engineering geological conditions is generated, which contains physical information of the drilling process. S2: Training the physical information proxy model: A multilayer perceptron was selected as the surrogate model to train the acquired virtual drilling parameter sequence, thereby obtaining a fast physical information surrogate model for the drilling process. S3: Input for 3D spatial data clustering: Multi-source time-series data during the drilling process are collected in real time and synchronously. After spatial alignment, the data is sampled at fixed depth intervals to form a continuous dataset. The KNN clustering algorithm is used to cluster the three-dimensional drilling data to obtain the nearest neighbor relationship of the parameters within the space. S4: Physical data mixed data loss: A Transformer encoder is used as the core feature extractor. After high-dimensional mapping, a set of deep spatiotemporal feature vectors representing the comprehensive state of the rock mass in the current drilling section are obtained. Spatial neighborhood information and borehole time-series information are fused and input into a GNN graph neural network to obtain a comprehensive feature matrix containing neighborhood information, which is used as the data loss function. At the same time, a physical information proxy model is fused as the physical loss, which together constitutes a hybrid loss function for physical data mixed data loss. S5: Strength inversion prediction under physical constraints: The hybrid loss function obtained in S4 is trained iteratively using deep learning, and the hyperparameters are optimized using the GA genetic algorithm to construct an intensity prediction model that can extract data from drilling data that conforms to both statistical and physical laws.

2. The intelligent analytical method for rock mass strength parameters based on data-mechanics-model driven according to claim 1, characterized in that, In step S1, the virtual drilling parameter sequence includes drilling speed, torque, and axial thrust; the rock strength parameter combination includes the rock's internal friction angle, cohesion, elastic modulus, and Poisson's ratio.

3. The intelligent analytical method for rock mass strength parameters based on data-mechanics-model driven according to claim 2, characterized in that, In step S1, specifically: S1.1: Determine the key simulation parameters that affect drilling response and rock mass strength, including drill pressure, rotation speed, and drill string diameter; for rock masses, it is necessary to determine the internal friction angle, cohesion, elastic modulus, and Poisson's ratio. S1.2: A simulation model is established using the finite element method; the interaction between the drill bit and the rock mass is analyzed using explicit dynamics to simulate the dynamic drilling process of cutting, crushing, extrusion, and friction. The simulation output includes time-series data of drilling speed, torque, and axial thrust during the drilling process; S1.3: Perform data augmentation on the original data generated by the simulation, including adding Gaussian white noise, impulse interference, and simulating data loss; finally, standardize or normalize all input parameters and output responses to form a structured simulation dataset.

4. The intelligent analytical method for rock mass strength parameters based on data-mechanics-model driven according to claim 3, characterized in that, In step S2, specifically: S2.1: The input layer is designed as a parallel channel to receive the timing signals of the timing data, and the output layer is the rock compressive strength; S2.2: For time-series data, a one-dimensional convolutional neural network (1D-CNN) is used to extract local fluctuation features, which are then connected to a gated recurrent unit (GRU) to capture long-range dependencies. The output is a drilling dynamic feature vector. For real-time engineering parameters, a fully connected encoder is used for fusion, and the output is a state feature vector. S2.3: Randomly divide the simulation samples according to the ratio of training set:test set = 7:3, set the hyperparameter range for optimization, and complete the surrogate model construction after multiple training sessions.

5. The intelligent analytical method for rock mass strength parameters based on data-mechanics-model driven according to claim 4, characterized in that, In step S3, the parameters of the drilling process need to be standardized and preprocessed in the 3D spatial data clustering input to conform to the format of the Transformer sample set: (1) Slice the continuous data stream into equal-length sequence samples with a fixed length, and perform independent normalization on each parameter; (2) Imput missing and outlier values ​​in the sequence and inject positional codes into the data to preserve temporal information; (3) Divide the processed sequence samples into training set, validation set and test set, and organize the input in batch processing form; After extracting sample data, the KNN network is used to perform unsupervised aggregation of data from different nodes, providing the model with spatial attention labels. At the same time, the Transformer model is responsible for extracting deep and abstract temporal features from the drilling time series data. The features of both are combined and input into the GNN graph neural network as one of the data loss functions for training.

6. The intelligent analytical method for rock mass strength parameters based on data-mechanics-model driven according to claim 5, characterized in that, In step S3, specifically: S3.1: Receive standardized sample data from the drilling rig on site. After data cleaning based on drilling time and drilling trend, extract the drilling time at different drilling depths, calculate the drilling depth increment corresponding to each time point in real time, and encode it into an additional position vector; then concatenate them to form a matrix D containing drilling data at different positions, D=[B1B2 … B n ] T B n This represents the drilling data for the nth borehole; ； in x,y Represents position, p i,1 ~p i,6 Represents different drilling parameters; S3.2: Using the borehole data structure described in S3.1, borehole clustering is performed based on KNN. First, the distance metric function between borehole data vectors is defined. : ； Indicates the first The first drilling vector One component; Indicates the first The first drilling vector One component; The weight coefficients for each dimension satisfy... ; Clustering label decision function for boreholes: For the target borehole Its K-nearest neighbor set Defined as: ； in This represents the distance sorting function; The adaptive formula for determining the value of K is: ； in: This represents the total number of boreholes. For drilling The parameter dispersion; These are the average drilling parameters; drilling Clustering label decision function for: ； in: For the set of all possible clustering categories; The current specific cluster category; For smoothing terms, when When the denominator is zero, the numerical value is incorrect.

7. The intelligent analytical method for rock mass strength parameters based on data-mechanics-model driven according to claim 6, characterized in that, In step S4, the hybrid loss function includes physical loss and data loss. The physical loss is replaced by a physical information proxy model to inject the physical laws during the drilling process, while the data loss is the strength loss of engineering drilling parameters and on-site strength. The two are combined to form a hybrid loss function that provides physical constraints, which serves as the basis for the training and convergence of the subsequent neural network model.

8. The intelligent analytical method for rock mass strength parameters based on data-mechanics-model driven according to claim 7, characterized in that, In step S4, the hybrid loss function consists of three parts: a data loss term, a physical loss term, and a regularization term. ； This represents the parameters to be optimized in the prediction model; These are the weighting coefficients for each loss term, satisfying... ; For data loss items; This is a physical loss item; This is the regularization loss term; The data loss term is expressed in the form of mean squared error: ； n is the number of samples measured on-site; The primary prediction model is based on the original drilling parameters. The output; For the first The field-measured surrounding rock strength label for each borehole; The physical loss item The calculation formula is: ； in: The primary prediction model is based on the original drilling parameters. The output; For MLP physical proxy models based on physical feature vectors The output; The original drilling parameters are mapped to a physical relation function. The obtained physical feature vector; The regularization loss term adopts the L2 regularization form: ； in For model parameters Dimensions.

9. The intelligent analytical method for rock mass strength parameters based on data-mechanics-model driven according to claim 8, characterized in that, In step S5, during the initial training phase, the physical loss is increased. The weights of these weights strongly guide the network to learn basic physical relationships; In the later stages of training, gradually increase the data loss. The weights are adjusted to make the network converge in a direction that aligns with the actual drilling parameters in the engineering process.

10. A data-mechanics-model-driven intelligent analysis system for rock mass strength parameters, characterized in that, It employs a data-mechanism-model-driven intelligent analysis method for rock mass strength parameters as described in any one of claims 1-9.

Images