Drilling rock hardness detection method based on topology-aware cross-modal machine learning

Abstract

This invention discloses a drilling rock hardness detection method based on topology-aware cross-modal machine learning, belonging to the fields of intelligent mining equipment state perception and industrial data analysis. First, vibration and acoustic signals synchronously acquired under the same drilling condition are preprocessed and feature extracted to construct paired bimodal feature samples. Candidate granular divisions are generated using the vibration mode as the main view, and the compactness of the divisions is verified using the acoustic mode as an auxiliary view. Bimodal granular units are obtained synchronously through sample indexing. Subsequently, a density-weighted joint granular map and a graph Laplacian matrix are constructed based on the granular center, sample density, and bimodal local radius to form topological divergence constraints. These topological divergence constraints are combined with the bimodal covariance to establish a topology-aware cross-modal subspace learning model. The bimodal projection matrix is ​​obtained through generalized eigenvalue decomposition or whitened singular value decomposition. Finally, the projected vibration and acoustic features are fused and input into a supervised classifier to output soft rock, medium-hard rock, or hard rock identification results. This invention can reduce the impact of isolated noise and local anomalous samples on the neighborhood structure, adapt to non-uniform density distribution and cross-modal local structural inconsistencies, and improve the stability and accuracy of drilling rock hardness identification under small sample and high-dimensional conditions.

Description

Technical Field

[0001] This invention relates to drilling rock hardness detection technology, specifically a method based on topology-sensing cross-modal machine learning, belonging to the field of intelligent mining equipment sensing and industrial data analysis technology. This method is applicable to tunneling-support-anchoring equipment, drilling-anchoring equipment, and other underground engineering drilling equipment. It utilizes paired vibration and acoustic characteristics acquired under the same drilling conditions to classify and detect soft, medium-hard, and hard rocks. Background Technology

[0002] With the increasing digitalization of smart mines, smart tunneling, and underground engineering equipment, tunneling, temporary support, anchoring, and drilling operations are gradually shifting from decentralized control to collaborative operation, real-time sensing, and adaptive decision-making. Rock hardness is a crucial physical property affecting drilling resistance, feed rate, drill bit wear, equipment vibration, rock fragmentation patterns, anchoring parameters, and support quality. Failure to obtain reliable rock hardness information in a timely manner during drilling can easily lead to unreasonable settings for drilling thrust, rotation speed, and feed rate, resulting in reduced drilling efficiency, excessive drill bit wear, equipment load fluctuations, decreased borehole quality, or mismatch between support parameters and surrounding rock conditions.

[0003] Multi-source industrial monitoring data are simultaneously generated when the drill bit contacts, cuts, and breaks the rock. Vibration signals reflect mechanical impact, cutting resistance, contact state, and drilling stability; acoustic signals reflect crack propagation, local fracturing, resonance, and energy release. Both types of signals originate from the same drill bit-rock interaction process, possessing a natural sample-level correspondence, while simultaneously describing rock hardness from different physical perspectives. Therefore, cross-modal fusion of vibration and acoustic information can provide a more comprehensive data foundation for online detection of drilled rock hardness.

[0004] Currently, rock hardness testing in drilling operations mostly relies on single drilling parameters or single modal signals. Some methods directly stitch together vibration and acoustic features before processing by a classifier. While these approaches can achieve some classification results, their utilization of the structural features of drilling site data remains limited.

[0005] In actual operation, mechanical shocks, hydraulic fluctuations, differences in sensor installation, and environmental noise can introduce isolated outliers and local fluctuation fragments into the samples. If a fixed neighborhood is constructed directly from a single sample, the local geometric relationships can easily be altered by a small number of unstable samples. At the same time, the response regions formed by rocks of different hardness often vary in density, making it difficult to balance compact and sparse regions with a uniform neighborhood scale and equal-weighted connectivity.

[0006] Although vibration and acoustic signals originate from the same drilling event, their local topologies are not always consistent due to the influence of mechanical transmission paths and acoustic propagation conditions. Independently dividing neighborhoods in the two modes may cause semantic misalignment of paired samples in subsequent modeling; building maps based solely on one mode or directly splicing features cannot fully utilize the verification and supplementation of the local structure by the other mode.

[0007] Canonical correlation analysis and other cross-modal subspace methods primarily rely on global cross-modal covariance and global intra-view covariance to establish projections, but they do not adequately consider local structure, density reliability, and non-uniform distribution. Existing local preservation or graph regularization methods typically still use point-level K-nearest neighbor relationships, making them sensitive to noise and cross-modal neighborhood mismatches. Under conditions of small samples and high-dimensional features, the intra-view covariance matrix may also be ill-conditioned or near-singular, further affecting the stability of the projection solution.

[0008] Therefore, it is necessary to characterize local sample groups with stable granular units while maintaining the correspondence between paired vibration-acoustic samples and global statistical correlation. A collaborative granular topology is constructed through main view guidance, auxiliary view verification, and index synchronization. Density weight, joint granular map, and local topological divergence are introduced into the in-view constraints to improve the stability of rock hardness detection under complex drilling conditions. Summary of the Invention

[0009] The purpose of this invention is to provide a drilling rock hardness detection method based on topology-aware cross-modal machine learning, addressing the problems of non-uniform density distribution of vibration-acoustic multimodal industrial data, noise-sensitive samples, cross-modal local topological inconsistencies, and instability in high-dimensional solutions for small samples in existing rock hardness detection methods. This method enables the learned low-dimensional representation to simultaneously maintain the global statistical correlation of paired vibration-acoustic samples, local topological stability within modes, and the separability of different rock hardness categories. This invention utilizes pre-obtained vibration feature matrices paired one-to-one according to the same drilling state. Harmony and acoustic feature matrix The input is the vibration and acoustic features, which can be obtained by filtering, segmenting, normalizing, and extracting features from synchronously acquired vibration and acoustic signals. The core technology of this invention does not limit the specific sensor model or front-end feature extraction method. The method of this invention includes the following steps: 1. Constructing a collaborative granular topology Vibration modes are used as the primary view to guide the partitioning of candidate granular units, and acoustic modes are used as the secondary view to verify the compactness improvement of the candidate partitioning. A synchronous granular unit set is formed through cross-modal indexing.

[0010] (1a) Paired feature matrices and initial granular units The vibration characteristic matrix and the acoustic characteristic matrix are respectively represented as follows: in, The vibration characteristic matrix, This is the acoustic feature matrix; For the first One vibration feature vector, To and The corresponding number One acoustic feature vector; The total number of paired samples; and These are the vibration characteristic dimension and the acoustic characteristic dimension, respectively. Each pair Corresponding to the same drilling condition and the same rock hardness label; Initialize all training samples into a single granular unit: in, For the initial set of granular units, This is the initial granular unit containing the indices of all training samples.

[0011] (1b) Main view guides candidate granular division For any candidate granular unit Calculate its average compactness in the vibration mode and acoustic mode respectively: The granular centers of the two modes are defined as follows: Represents granular unit The number of samples in; and These represent the compactness of the vibrational modes and the acoustic modes, respectively. and These represent the centers of the granular unit in the two modes, respectively. Represents the Euclidean norm; In the vibration feature space, the sample pair with the largest distance is selected as the dividing pole: in, and These are the sample indices corresponding to the two dividing poles. and For the corresponding vibration characteristic vector; Candidate granular units are selected based on the distance of each sample to the two dividing poles. Divided into two candidate sub-units: G1 and G2 are two candidate sub-elements obtained from G. Vibration modes are used as the main view because drilling vibrations have a direct response to changes in cutting resistance, mechanical impact, and drilling stability, and can be used to determine the local spatial division direction.

[0012] (1c) Auxiliary view verification and minimum sample number constraint The candidate sub-element indices obtained from the vibration modes are synchronously applied to the acoustic modes, and the bimodal weighted compactness after candidate partitioning is calculated: The candidate partition passes the bimodal compactness verification when any of the following conditions are met: in, To define the threshold, Using "OR" logic allows vibration modes and acoustic modes to complement each other: when one mode is significantly affected by noise, while the other mode can still reflect the effective class structure, the classification can be retained as long as the local compactness of at least one mode is significantly improved. To suppress noise-induced overly fine subdivisions, candidate sub-units must also satisfy: in, This represents the minimum number of samples allowed in a candidate subunit. If either the compactness condition or the minimum sample number condition is not met, the candidate subunit is rejected and the original granular unit is retained. .

[0013] (1d) Synchronization of recursive refinement and cross-modal indexing Repeat sub-steps (1b) and (1c) until no candidate partitions satisfy the acceptance criteria are found, or the preset maximum number of granular units is reached. For each granular unit obtained, maintain the same sample index in both the vibration and acoustic modes: in, and The first The sample index set of each granular unit in vibration mode and acoustic mode, with double arrows indicating that they have the same sample index; This represents the final total number of granular units. This yields the set of synchronized granular units: Cross-modal index synchronization can prevent semantic misalignment of samples caused by independent partitioning of two modalities, ensuring that subsequent granular centers, density weights, and joint topology are all built on the same set of drilling events.

[0014] 2. Construct density-weighted joint granularity map and topological divergence matrix (2a) Dual-modal granular center For the Synchronous granular units Calculate the granular centers of the vibration mode and acoustic mode respectively: Arrange the granular centers in columns to form a bimodal granular center matrix: Granular centers, as topological surrogate representations of local sample clusters, are less sensitive to transient disturbances and isolated anomalies compared to single point-level samples.

[0015] (2b) Density weight Calculate the density weight based on the sample proportion of each synchronous granular unit: And form a density weight matrix: in, For the first Density weights of synchronous granular units; This is the density weight matrix. Granular units with more samples usually correspond to more stable local regions, and therefore contribute more to subsequent industrial data analysis and subspace learning; sparse regions with fewer samples or noise-sensitive regions are relatively suppressed.

[0016] (2c) Bimodal normalized distance, K-nearest neighbor joint granular graph and graph Laplacian matrix Calculate the first one respectively The mean square radius of each synchronous granular element in the vibration and acoustic modes: in, It is a numerically stable term; For any two synchronous granular units and Construct the bimodal normalized distance: in, The modal fusion weights are used to adjust the contributions of vibration and acoustic modes to the joint distance; Construct an initial similarity matrix by combining density weights, local neighborhoods, and exponential kernels: in, For scaling parameters; granular unit of Nearest neighbor index set; The nearest neighbor number; granular unit with granular units The initial similarity; Symmetricize the similarity matrix and construct the degree matrix and the symmetric normalized graph Laplacian matrix: in, To symmetricize the similarity matrix; This is the degree matrix; for An identity matrix of order 1; The Laplacian matrix is ​​a symmetric normalized graph. This joint granular graph is not a simple single-view fixed neighborhood, but simultaneously utilizes vibrational and acoustic granular centers, local radii, density reliability, and K-nearest neighbor relationships to describe multimodal local topology.

[0017] (2d) Density-weighted topological divergence matrix Using the bimodal granular center matrix, density weight matrix, and graphical Laplacian matrix, the density-weighted topological divergence matrices of the vibrational and acoustic modes are constructed: in, and These are the density-weighted topological divergence matrices for the vibration modes and the acoustic modes, respectively. This is the square root of the density weight matrix. The divergence matrix is ​​used to embed granular topology, density reliability, and local structural relationships into in-view constraints.

[0018] 3. Establish topology-aware cross-modal subspace learning objectives (3a) Calculation of Centralized Covariance Build Order-centered matrix: in, for An identity matrix of order 1. for A column vector of all 1s; Center the paired feature matrices and calculate the point-to-point cross-modal covariance and the bimodal view intra-covariance: in, and It is a cross-modal covariance matrix used to preserve the direct physical correspondence between vibration samples and acoustic samples under the same drilling event; and These are the covariance matrices of the global view of vibration modes and acoustic modes, respectively, calculated from all training samples.

[0019] (3b) Optimization of constraints and matrix trace within topology-aware view The global view covariance, local topological divergence, and ridge regularization term are fused: in, and The constraint matrix within the topology-aware view; is the topological regularization coefficient, used to adjust the relative contribution of covariance and local topological divergence within the global view; is the ridge regularization parameter, used to improve the numerical stability of matrix inversion and eigenvalue decomposition under small sample and high-dimensional conditions; and These are the identity matrices of the corresponding dimensions; Establish a matrix-form topology-aware cross-modal machine learning objective: Its equivalent constraint form is: in, and These are the projection matrices for vibration modes and acoustic modes, respectively; Represents the trace of a matrix; for An identity matrix of order 1; Let be the dimension of the objective subspace. The numerator of the objective function maintains cross-modal point-to-point correlation, while the constraint terms simultaneously incorporate statistical information within the global view and density-weighted granular topological information, thereby obtaining a low-dimensional representation that balances global correlation and local spatial stability.

[0020] 4. Solve for the projection matrix, fuse projection features, and perform rock hardness detection. (4a) Solving by generalized eigenvalue decomposition The coupled generalized characteristic equation is obtained from the above constraint optimization: Further written in the form of block generalized eigenvalues: in, For the first The topological sensing canonical correlation coefficients are selected. The largest one is chosen. The non-negative generalized eigenvalues ​​and their corresponding eigenvectors constitute the vibration mode projection matrix. Harmony acoustic modal projection matrix .

[0021] (4b) Solving whitening singular value decomposition To improve stability in ill-conditioned covariance and small-sample, high-dimensional scenarios, a whitening cross-modal matrix can also be constructed: And perform singular value decomposition: Therefore, we can conclude that: in, and These are the whitening cross-modal matrices. The left singular vector matrix and the right singular vector matrix; It is a singular value diagonal matrix. In the actual solution, it is chosen to be similar to the previous... The column corresponding to the largest singular value.

[0022] (4c) Projection, fusion and rock hardness testing The centered bimodal projection features are obtained using the projection matrix: in, and These are the centered low-dimensional projective feature matrices of all training samples in the vibration and acoustic modes, respectively, both containing... The first sample; its first The columns correspond to the first one. The vibrational and acoustic projection features of paired samples, with the dimension of a single projection feature being... .

[0023] The following methods can be used for fusion: Weighted summation fusion: Feature splicing and fusion: Sum and difference integration: in, To integrate weights, This indicates taking the absolute value of each element; all the above fusion matrices represent samples in columns. (Selection) , or One of them is a fusion feature Input a K-nearest neighbor classifier, support vector machine, random forest, linear discriminant classifier or other supervised machine learning classifier, and output the soft rock, medium hard rock or hard rock category; For online samples to be tested, the paired vibration and acoustic features are processed using the centralized or normalized parameters saved during the training phase, and the processed features are denoted as follows: and Then calculate: And obtained according to the fusion method determined in the training phase. As an online fusion feature, the rock hardness detection result is output by the classifier. The detection result can be transmitted to the drilling control, anchoring parameter recommendation, or support decision-making modules to assist in adjusting the thrust, rotation speed, feed rate, and anchoring strategy.

[0024] Compared with the prior art, the present invention has the following beneficial effects: This invention uses a vibration master view to determine the granular partitioning direction and an acoustic auxiliary view to verify the partitioning results. Both modes always use the same sample index. Replacing point-level neighborhoods with synchronous granular units and their centers reduces the impact of isolated anomalies, transient impacts, and local noise on topological relationships, while avoiding semantic misalignment caused by independent partitioning of the two modes.

[0025] This invention sets density weights based on the proportion of granular unit samples and constructs a joint granular map by combining bimodal granular center, local radius and K-nearest neighbor relationship. It can provide different reliability characteristics for compact high-density regions and sparse irregular regions, thereby adapting to the non-uniform density distribution of drilling data.

[0026] This invention, while maintaining the cross-modal covariance of paired samples, incorporates the global in-view covariance, local topological divergence, and ridge regularization into the in-view constraints, so that the projection results take into account both global statistical correlation and local topological stability. Combined with whitened singular value decomposition, it can alleviate the instability of the solution caused by ill-conditioned or near-singular covariance matrix under small sample and high-dimensional conditions.

[0027] like Figure 2 As shown, the standardized parameters, projection matrix, fusion rules, and classifier obtained from training can be distributed to the edge computing unit from the server or industrial internet platform. The device only needs to complete feature processing, projection fusion, and rock hardness determination. The detection results can also be uploaded to the platform and transmitted to the drilling control, anchoring parameter recommendation, or support decision modules, facilitating real-time on-site detection, remote analysis, and model updates. Attached Figure Description

[0028] Figure 1 This is a flowchart of the overall process of the drilling rock hardness detection method based on topology-aware cross-modal machine learning of the present invention.

[0029] Figure 2 This is a schematic diagram illustrating the deployment and application of the present invention in an edge computing unit or industrial internet platform.

[0030] Figure 3 This is a black-and-white normalized confusion matrix diagram of the rock hardness test results of this invention. Detailed Implementation

[0031] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The embodiments are used to illustrate the technical principles and possible implementations of the present invention, and are not intended to limit the scope of protection of the present invention. Those skilled in the art, without departing from the core concept of the present invention, may make equivalent substitutions for sensor types, feature configurations, parameter values, solvers, fusion methods, and classifiers, all of which should fall within the scope of the technical concept of the present invention.

[0032] 1. Vibration acoustic drilling data acquisition and paired sample formation This embodiment employs a tunneling-supporting-anchoring experimental platform to conduct drilling tests. The platform includes a support and anchoring system, a travel system, a protection system, a drilling and anchoring system, a hydraulic drive system, an electrical control system, and a multimodal sensing system. The platform provides repeatable drill bit-rock contact, cutting, and fracturing conditions, and simultaneously acquires mechanical vibration and acoustic responses.

[0033] Rock specimens with different hardness grades were arranged on the roof of a simulated tunnel, and classified into three categories according to their hardness: soft rock, medium-hard rock, and hard rock. The specimen area could be divided into a formal drilling zone and a pre-drilling zone. The pre-drilling zone was used for equipment preheating, parameter calibration, drill bit stabilization, and data acquisition system checks; the formal drilling zone was used to obtain training and testing data. The rock hardness category corresponding to each drilling record was determined by the known category of the specimen.

[0034] The vibration sensor is fixed to the drilling anchoring mechanism or a location with strong mechanical coupling to the drill pipe, while the acoustic sensor is placed near the drilling area. Both types of sensors use a unified clock or synchronous triggering method and can be set to the same sampling frequency. In this embodiment, the sampling frequency of both the vibration signal and the acoustic signal is 10000 Hz to maintain a time correspondence.

[0035] The original vibration and acoustic signals are filtered to remove low-frequency drift, background interference, and frequency components unrelated to drilling. The signals are then segmented according to fixed-length time windows, fixed drilling displacements, or drilling stages, so that each vibration segment is paired with an acoustic segment within the same time interval.

[0036] Vibration characteristics may include one or more of the following: root mean square value, peak value, peak-to-peak value, variance, kurtosis, skewness, waveform factor, impulse factor, spectral centroid, dominant frequency, bandwidth energy, spectral entropy, and wavelet packet energy; acoustic characteristics may include one or more of the following: sound pressure level, root mean square sound pressure level, dominant frequency, bandwidth, spectral centroid, spectral energy, spectral entropy, Mel-frequency cepstral coefficients, or time-frequency energy. After normalization, the vibration and acoustic characteristics corresponding to the same drilling segment are arranged in columns to form... and .

[0037] It should be noted that the aforementioned synchronous acquisition, filtering, segmentation, and feature extraction are one implementation method for obtaining the input data of this invention. For vibration feature matrices and acoustic feature matrices that have already been formed by other acquisition systems and paired one-to-one according to drilling states, the subsequent collaborative granular topology and cross-modal machine learning steps of this invention can be directly executed.

[0038] In the specific dataset used to verify the present invention, both vibration and acoustic modes employed 27-dimensional features, resulting in 1785 paired samples, with 595 sample pairs each for soft rock, medium-hard rock, and hard rock. Each sample pair originated from the same drilling interval and shared the same rock hardness label. The aforementioned dimensions and sample numbers are merely illustrative of the feasible implementation of this embodiment and do not constitute a limitation on the input dimensions and sample size of the present invention.

[0039] 2. Collaborative Granular Topology Modeling During the training phase, vibration and acoustic features are standardized separately to ensure that each dimension has a comparable scale. The standardized parameters are calculated and stored only from the training data for subsequent testing or online sample processing, preventing leakage of test information.

[0040] Initialize all training samples as granular units For candidate granular units to be judged, the bimodal compactness is calculated according to the aforementioned formula, and the farthest sample pair is selected as the dividing pole in the vibration feature space. Since the vibration signal directly reflects the changes in mechanical impact and cutting resistance, a clearer initial local boundary can be obtained using the vibration master view.

[0041] Formed according to the distance of the sample from the two poles and Then, the same index is synchronized to the acoustic modes. The bimodal weighted compactness after partitioning is calculated; a partition is accepted when the compactness of at least one view of the vibration mode or acoustic mode meets the threshold improvement and both sub-units meet the minimum sample number constraint. This process is repeated recursively until no valid partitions are available.

[0042] In this embodiment, the segmentation threshold This can be determined through cross-validation on the training set; The settings can be adjusted based on sample size, noise level, and classifier requirements. If the sample size is small, the settings can be increased. To prevent overfitting; if the sample size is large and the local structure is complex, the granularity constraint can be appropriately reduced to obtain a more detailed topological representation.

[0043] Unlike traditional clustering methods, the granular partitioning of this invention is not used to directly determine rock categories, but rather to extract stable local topological proxies. Each granular unit may contain samples of the same category or samples in the category transition zone, and the final rock hardness is still determined by a supervised classifier on the projected features.

[0044] 3. Solving using a combination of granular graph and topologically aware subspace. For synchronous granular units, calculate the bimodal granular center, sample proportion density weight, and mean square radius. Construct a K-nearest neighbor graph based on the bimodal normalized distance, and multiply the density weights by the graph edge weights. Symmetrically transform the similarity matrix to obtain the symmetric normalized graph Laplacian matrix. .

[0045] Modal fusion weights This parameter is used to adjust the influence of vibration and acoustic modes on the combined distance. When the mechanical impact at the drilling site is strong and the acoustic propagation conditions are poor, the vibration mode weight can be appropriately increased; when the mechanical installation structure has a strong attenuation effect on vibration, but the acoustic signal is clear, the vibration mode weight can be decreased. (Scale adjustment parameter) and nearest neighbor number This can be determined through cross-validation.

[0046] use , , and calculate and Then, it is combined with the covariance and ridge regularization term within the global view to form and Topological regularization coefficients Controlling the constraint strength of local granular topology; when When the model degenerates into cross-modal correlation learning without granular topological divergence, when When the scale is too large, it may overemphasize local structure and weaken the original cross-modal relevance. Therefore, it is advisable to select from the log-scale candidate set through validation on the training set. .

[0047] Ridge regularization parameters Used to improve matrix conditionality. In this embodiment, a moderately strong regularization parameter can obtain more stable detection results, approximately... The nearby settings performed well; this value is only an example and does not constitute a limitation on the scope of protection.

[0048] In the solution phase, whitening singular value decomposition is preferred. First, eigenvalue decomposition or Cholesky decomposition is performed on the constraint matrices within the two topology-aware views to calculate the inverse square root, and then the whitening matrix is ​​formed. And perform singular value decomposition. Before selection... The singular vectors are obtained and When the constraint matrices within the two views are relatively small and the numerical conditions are favorable, block generalized eigenvalue decomposition can be used to solve the problem directly.

[0049] To ensure reproducibility of the results, the splitting threshold, minimum number of samples, number of nearest neighbors, scale adjustment parameter, topological regularization coefficient, ridge regularization parameter, projection dimension, and fusion method were all determined through cross-validation within the training data. Test samples and their labels were not involved in parameter selection, projection dimension selection, or fusion rule selection; they were only used for a one-time evaluation after the model was fixed.

[0050] 4. Integration of classification and online detection After training, use and Vibrational and acoustic features are mapped to a low-dimensional subspace. If the projection dimensions of the two modes are the same, weighted summation fusion can be used; if it is necessary to fully preserve the complementary information of the two modes, feature concatenation can be used; if it is desired to utilize both common information and modal differences, sum-difference fusion can be used.

[0051] The fused features are input into a supervised classifier for rock hardness classification. In one specific implementation, the K-nearest neighbor algorithm can be used as the classifier, searching for nearest neighbor samples in the fusion space using Euclidean distance and outputting soft rock, medium-hard rock, or hard rock based on majority voting; alternatively, support vector machines, random forests, linear discriminant analysis, or lightweight neural networks can be used. The classifier is trained using only the fused features from the training set and the corresponding hardness labels.

[0052] In the comparative experiments of this embodiment, all methods used the same K-nearest neighbor classifier and the same data partitioning. During weighted fusion, the fusion coefficients were determined within the training set from candidate weights such as 0.3, 0.5, and 0.7; the number of nearest neighbors and the projection dimension of the classifier were also determined through validation within the training set. This setup allows for comparison of different subspace learning methods under consistent evaluation conditions.

[0053] During online detection, paired vibration and acoustic features are first obtained from the newly drilled segment, and then processed using normalized parameters, projection matrices, fusion rules, and classifiers saved during the training phase. Stiffness results for a single new sample can be obtained without rebuilding the entire training topology, making it suitable for real-time operation on equipment controllers, industrial PCs, or edge computing units.

[0054] The output results can be used as input for adaptive adjustment of equipment parameters. For example, when hard rock is detected, the drilling thrust can be increased and more stable rotation parameters can be adopted according to the equipment capacity; when soft rock is detected, excessive thrust can be reduced or the feed rate can be adjusted to avoid over-cutting, borehole wall damage, and equipment impact. Specific control strategies can be set by the upper-level control system according to the equipment structure and safety constraints.

[0055] 5. Industrial Internet Deployment Methods like Figure 2 As shown, vibration sensors and acoustic sensors constitute the field perception layer; the synchronous acquisition device completes data acquisition, time alignment and preprocessing; the edge computing unit calls the standardized parameters, projection matrix, fusion rules and classifier obtained in the training stage to complete feature generation, projection fusion and rock hardness determination; the detection results and necessary intermediate statistics can be uploaded to the industrial Internet platform for industrial data analysis, model version management, cross-device updates and remote diagnosis.

[0056] The training process can be completed on a server or industrial internet platform, and the obtained standardized parameters, projection matrix, and classifier are then distributed to the equipment. The equipment only needs to perform matrix multiplication and classification operations, which is suitable for resource-constrained environments. Data from multiple drilling rigs can also be aggregated on the platform for cross-equipment model updates, but differences in sensor installation between different devices should be addressed through calibration, domain adaptation, or retraining.

[0057] In this embodiment, the edge computing and industrial internet platform are optional deployment methods and do not change the core technical solution of the present invention, which uses collaborative granular topology and topology-aware cross-modal subspace learning to achieve rock hardness detection.

[0058] 6. Experimental verification The effectiveness of this invention was further verified through the following experiments: During the drilling process of the self-developed tunneling-supporting-anchoring experimental platform, vibration and acoustic signals were simultaneously collected. After filtering, segmentation, and feature extraction, three types of vibration-acoustic paired samples were constructed: soft rock, medium-hard rock, and hard rock, totaling 1785 groups, with 595 groups in each type. The experiment selected 80, 150, 220, 290, 360, 430, 500, and 570 samples from each type as training samples, with the remainder used as test samples. The results were compared with principal component analysis, linear discriminant analysis, generalized canonical correlation analysis, local preserving projection, and orthogonal canonical correlation analysis under the same data partitioning and classifier conditions. The results show that the detection accuracy of this invention is approximately 83% with 80 training samples per type and close to 90% with 570 training samples per type. The accuracy generally increases with the increase of training samples, and the detection performance is relatively stable. Figure 3 In the normalized confusion matrix shown, the detection results for the three types of rocks are mainly concentrated in the diagonal positions, with fewer direct misclassifications between soft and hard rocks. Experimental results demonstrate that the method disclosed in this invention is an accurate, stable, and effective method for detecting rock hardness.

Claims

1. A method for detecting the hardness of drilled rocks based on topology-aware cross-modal machine learning, characterized in that, Includes the following steps: (1) Obtain vibration characteristic matrices paired one-to-one according to the same drilling state Harmony and acoustic feature matrix ; (2) The vibration mode is used as the main view to determine the division direction of the candidate granular unit, and the acoustic mode is used as the auxiliary view to verify the compactness improvement of the candidate division. The same sample index is used for the two modes to form a synchronous granular unit set. (3) Calculate the dual-mode particle center, mean square radius and density weight according to the synchronous particle unit, construct the joint particle map and the Laplace matrix of the symmetric normalized map based on the dual-mode normalized distance, and form the density-weighted topological divergence matrix of the vibration mode and the acoustic mode. (4) Calculate the global view covariance matrix and point-to-point cross-modal covariance matrix of vibration mode and acoustic mode using all paired samples, combine the global view covariance matrix, the density-weighted topological divergence matrix and the ridge regularization term into the topology-aware view constraint matrix, and establish the topology-aware cross-modal subspace learning objective. (5) Solve the vibration mode projection matrix by generalized eigenvalue decomposition or whitened singular value decomposition. Harmony acoustic modal projection matrix The dual-modal low-dimensional projection features are obtained using the projection matrix, the dual-modal low-dimensional projection features are fused, and the fused features are input into a supervised classifier to output the rock hardness detection results for soft rock, medium-hard rock, or hard rock.

2. The method according to claim 1, characterized in that, Step (2) includes the following sub-steps: (2a) Vibration characteristic matrix Harmony and acoustic feature matrix They are represented as follows: in, For the first One vibration feature vector, To and The corresponding number Each acoustic feature vector ; The total number of paired samples; and These are the vibration characteristic dimension and the acoustic characteristic dimension, respectively. Represents the real number field; each pair For samples with the same drilling state and the same rock hardness label, initialize the index of all training samples as a single granular unit: in, For the initial set of granular units, This is the initial granular unit containing the indices of all training samples; (2b) For any candidate granular unit Calculate its center and average compactness in the vibration mode and acoustic mode, respectively: in, This is the set of sample indices corresponding to the candidate granular units. express The number of samples in; and They are respectively At the center of the vibration and acoustic modes; and They are respectively Average compactness in vibration and acoustic modes; Represents the Euclidean norm; the sample pair with the largest distance in the vibration characteristic space is selected as the dividing pole: in, and To traverse candidate granular units Temporary sample index used at that time; and In order to make The sample index corresponding to the maximum value. and For the corresponding vibration feature vector; when there are multiple maximum values, randomly select one pair of indices as the dividing poles; based on the distance of each sample to the two dividing poles, Divided into two candidate sub-units: in, and For candidate granular units The two candidate sub-units obtained from the division; (2c) The sample index of the candidate sub-unit is synchronously applied to the acoustic modality, and the bimodal weighted compactness after candidate partitioning is calculated: in, and The weighted compactness of the vibration modes and acoustic modes after the candidate partitioning are respectively; the candidate partitioning is accepted when at least one of the following compactness conditions is met, and the minimum sample size constraint is also met: in, To define the threshold, ; This represents the minimum number of samples allowed in a candidate subunit; if the above acceptance criteria are not met, the candidate subunit is rejected and the original candidate granular unit is retained. ; (2d) Repeat sub-steps (2b) and (2c) until no candidate partitions satisfy the acceptance criteria are found or the preset maximum number of granular units is reached; for the final obtained... Each synchronous granular unit denotes the sample index set from both modes as follows: And obtain a set of synchronous granular units: in, and They represent the first A set of sample indices for each granular unit in vibration and acoustic modes; This is the set of sample indices used by both. For granular unit indexing; This represents the final total number of synchronized granular units.

3. The method according to claim 2, characterized in that, Step (3) includes the following sub-steps: (3a) For the first Synchronous granular units Calculate the grain centers of the vibration mode and the acoustic mode, respectively: And the granular centers are arranged in columns to form a bimodal granular center matrix: in, Indicates the first Number of samples in each synchronous granular unit; and The first The granular center of a synchronous granular unit in both vibration and acoustic modes; and These are the granular center matrices for vibrational modes and acoustic modes, respectively. (3b) Calculate the density weights based on the sample proportions of each synchronous granular unit, and form a density weight matrix: in, For the first Density weights of synchronous granular units; This is a density weight matrix with each density weight as the main diagonal element; (3c) Calculate the first one respectively The mean square radius of each synchronous granular element in the vibration and acoustic modes: in, and The first The mean square radius of each synchronous granular unit in both vibration and acoustic modes; A numerically stable term greater than 0; for any two distinct synchronous granular units and Construct the bimodal normalized distance: in, and All are granular unit indexes, and ; For the first The and the first The bimodal normalized distance between synchronous granular units; The modality fusion weights are used; an initial similarity matrix is ​​constructed by combining density weights, local neighborhoods, and exponential kernels. : when ,and or season: In other cases, let ;in, A scale adjustment parameter greater than 0; For the first Synchronous granular units Nearest neighbor index set; Let the nearest neighbor number be denoted; the initial similarity matrix is ​​symmetricized, and a degree matrix and a symmetric normalized graph Laplacian matrix are constructed: in, To symmetricize the similarity matrix; for The corresponding degree matrix; for An identity matrix of order 1; The Laplace matrix of the symmetric normalized graph; (3d) Using the bimodal granular center matrix, density weight matrix, and symmetric normalized graph Laplacian matrix, construct the density-weighted topological divergence matrix: in, and These are the density-weighted topological divergence matrices for the vibration modes and the acoustic modes, respectively. Density weight matrix The square root matrix; superscript This indicates the matrix transpose.

4. The method according to claim 3, characterized in that, Step (4) includes the following sub-steps: (4a) Construct the centered matrix and calculate the point-to-point cross-modal covariance matrix and the bimodal global view intra-covariance matrix: in, for Order-centered matrix; for An identity matrix of order 1; for A column vector of all 1s; and These are the point-to-point cross-modal covariance matrices from vibration mode to acoustic mode and from acoustic mode to vibration mode, respectively; and These are the in-field covariance matrices of the vibration modes and acoustic modes calculated using all training samples, respectively. (4b) Combining the global view covariance matrix, density-weighted topological divergence matrix, and ridge regularization term yields the topology-aware view constraint matrix: And establish a topology-aware cross-modal subspace learning objective: Its equivalent constraint form is: in, and These are the constraint matrices within the topology-sensing view for the vibration and acoustic modes, respectively; and These are the projection matrices for vibration modes and acoustic modes, respectively; A topology regularization coefficient greater than or equal to 0, used to adjust the relative contribution of global view covariance to local topology divergence; The ridge regularization parameter is greater than 0; and They are respectively order and An identity matrix of order 1; Represents the trace of a matrix; for An identity matrix of order 1; Let be the dimension of the target subspace.

5. The method according to claim 4, characterized in that, In step (5), the projection matrix is ​​solved using any of the following methods, and the projection features are fused and classified: (5a) When using generalized eigenvalue decomposition, solve the following block generalized eigenvalue problems: in, For the reason before A diagonal matrix composed of topologically sensed canonical correlation coefficients; For the first Topological sensing canonical correlation coefficients Select the largest The non-negative generalized eigenvalues ​​and their corresponding eigenvectors constitute the vibration mode projection matrix. Harmony acoustic modal projection matrix ; (5b) When using whitening singular value decomposition, construct the whitening cross-modal matrix and perform singular value decomposition: in, For whitening cross-modal matrix; and These are the inverse square root matrices of the constraint matrices within the two topology-aware views, respectively. and They are respectively The left singular vector matrix and the right singular vector matrix; It is a singular value diagonal matrix; and Each by and Middle and front It consists of columns corresponding to the largest singular values; (5c) Obtain the centered bimodal low-dimensional projection feature matrix using the projection matrix: in, and Let be the centered low-dimensional projective feature matrices of all training samples in vibration and acoustic modes, respectively, both having a dimension of 1. ; and The The columns are respectively the first Vibrational projection features and acoustic projection features of paired samples; fused features are formed using any of the following methods: in, For fusion weights; , and These are weighted summation fusion features, splicing fusion features, and sum-difference fusion features, respectively. Indicates taking the absolute value of each element; select , or One of them is a fusion feature The fusion features Input one of the following: K-nearest neighbor classifier, support vector machine, random forest, or linear discriminant classifier, and output the rock hardness detection result for soft rock, medium-hard rock, or hard rock.

Images