Real-time data-driven geologic intelligence online perception method and system

Abstract

A real-time data-driven online intelligent geological sensing method and system includes the following steps: (1) establishing a database of MWD parameters and corresponding core samples for standardizing the data processing flow; (2) constructing a time series classification model to establish a mapping relationship between MWD parameters and the excavated lithology in a reasonable and reliable manner; (3) converting the time series classification algorithm into an online learning version to make it more efficient and scalable, and continuously learning from the input observations to iteratively optimize the classification model. By implementing the above method, the ultimate goal is to improve the efficiency and safety of drilling operations while enhancing the intelligence level of HDD.

Description

Technical Field

[0001] This invention relates to the field of geological exploration technology, and in particular to a real-time data-driven intelligent online geological sensing method and system. Background Technology

[0002] Geological exploration is a crucial foundation for tunnel engineering, aiming to ascertain geological conditions and provide essential information for tunnel design, construction, and operation. With the development of information technology, intelligent methods are playing an increasingly prominent role in improving exploration efficiency and accuracy, making intelligent technology-based geological exploration a hot topic. Geological exploration primarily obtains subsurface spatial information through geophysical exploration and drilling. Drilling can directly obtain geological information, providing core samples for mineralogical and geochemical analysis, and has advantages such as ease of operation and strong adaptability. In particular, horizontal directional drilling (HDD) is playing an increasingly important role in geological exploration. Compared to traditional drilling, HDD is more efficient, less disruptive, suitable for traversing complex geological formations, and can achieve long-distance underground exploration, which is of great significance for determining the geological conditions of tunnels along the route.

[0003] Despite its advantages in geological identification, HDD (High-Diameter Directional Drilling) still faces long-standing technical challenges, primarily including directional drilling, trajectory control theory, and MWD (Mechanical, Dynamic, and Data-Driven Drilling). Directional drilling uses high-precision positioning units to determine drill bit coordinates, ensuring the borehole advances along a predetermined route. The theoretical model for directional drilling trajectories is still incomplete, leading to uncertainties in trajectory control. MWD addresses the issue of real-time data transmission during drilling. All these technologies aim to find optimal strategies in highly uncertain subsurface environments, thus requiring accurate quantification of the lithology ahead of the drill bit. Furthermore, due to data update lags, risks such as stuck drill bits are prone to occur, necessitating real-time monitoring of the lithology at the drill bit's location to adjust drilling parameters accordingly. Therefore, conducting data-driven geological classification research is imperative.

[0004] Existing research largely focuses on geological identification in conventional vertical drilling, while real-time lithology classification under HDD conditions remains insufficient. This is because constructing a complete core database is difficult, requiring simultaneous on-site acquisition of drilling parameters and corresponding geological samples. However, this invention is of significant necessity: on the one hand, unknown geology forces the use of conservative parameters during construction, reducing efficiency and increasing the risk of equipment failure; on the other hand, the lack of an adaptive control framework makes it difficult to respond promptly to geological changes, increasing the probability of trajectory deviation. Summary of the Invention

[0005] This invention proposes a real-time data-driven intelligent online sensing method and system for geology, introducing online learning into time series classification algorithms. This method utilizes the large amount of sensor data generated during HDD operations, combined with corresponding core samples, to train an online model, thereby achieving continuous updates to observational data.

[0006] A real-time data-driven intelligent online geological sensing method includes the following steps:

[0007] (1) Establish a database of MWD parameters and corresponding core samples for standardizing data processing procedures;

[0008] (2) Construct a time series classification model to establish a reasonable and reliable mapping relationship between the MWD parameters and the lithology of the excavated rocks;

[0009] (3) Transform the time series classification algorithm into an online learning version to make it more efficient and scalable, and continuously learn from the input observations to iteratively optimize the classification model.

[0010] Through the above methods, the level of intelligence in HDD is improved, while the efficiency and safety of drilling operations are enhanced.

[0011] Furthermore, it also includes an online framework for geological condition detection, which comprises four steps:

[0012] (1) Acquire measurement-while-drilling (MWD) data from horizontal directional drilling (HDD);

[0013] (2) Multidimensional feature extraction is performed using the method based on random convolution kernel transformation (ROCKET).

[0014] (3) Use a ridge regression classifier for classification;

[0015] (4) Realize online monitoring of lithology / geological conditions.

[0016] Furthermore, the ROCKET random convolution kernel transformation is an efficient time series classification method that uses random convolution kernels to extract discriminative features. Each convolution kernel is applied to each sample to generate feature maps, from which two types of features are extracted: the maximum value and a new feature—positive proportion PPV. The positive proportion PPV represents the proportion of sequence segments that are positively correlated with the convolution kernel, thereby improving classification accuracy.

[0017] The random convolution kernel transformation ROCKET transforms the time series using a large number of random convolution kernels, that is, it randomly sets the kernel length, dilation, kernel weights, bias, and padding parameters, as shown in formula (1).

[0018]

[0019] in Indicates the length of the input time series. The arithmetic mean;

[0020] For an input time series Each convolution kernel Multi-scale patterns are captured through dilated convolutions; the extended kernel is fused with the dilated dk.

[0021]

[0022] in ,and Convolution output Calculate according to formula (8):

[0023]

[0024] Rocket, a random convolution kernel transform, computes two aggregated features from each feature map, producing two real numbers. and As each Features:

[0025]

[0026] in It is an indicator;

[0027] For the Each kernel, with random convolution kernel transformation, generates two features for each time series (i.e., ...). and ), represented as :

[0028]

[0029] Where T is the transpose operator. This refers to the total number of cores. In this invention, K = 10000 (default value).

[0030] Connection feature matrix used for training samples for:

[0031]

[0032] in, It represents the number of training samples.

[0033] Furthermore, in the ridge regression classifier, the feature matrix generated by the random convolution kernel transformation ROCKET is shown in equation (12); for the defined four types of problems A ridge regression model for each class i was constructed; the feature subset matrix corresponding to the training samples belonging to class i is defined as:

[0034]

[0035] in, This represents the row vector of the evaluation features of the k-th sample in category i. It is the number of samples in category i;

[0036] Given a sample to be classified This represents its feature vector that has undergone random convolution kernel transformation and flattening, assuming it can be categorized. Feature matrix The sample linear approximation in:

[0037]

[0038] in, It is the vector of regression coefficients to be solved. Represents the approximate error vector;

[0039] To obtain a stable solution with good generalization ability, a ridge regression objective function is introduced:

[0040]

[0041] in This is the regularization parameter, which controls model complexity and balances the trade-off between training error and model variance; increasing... Will control more significantly The magnitude of the amplitude makes its coefficients tend to 0, thereby reducing the model variance but potentially increasing bias; reducing This reduces the penalty, allowing the model to fit the training data more closely;

[0042] By analyzing the objective function (15) with respect to... Differentiate and set it to zero to derive the closed-form solution. :

[0043]

[0044] in It is the identity matrix; solving formula (18) yields the optimal estimate of the regression coefficient vector:

[0045]

[0046] Using the estimated Test samples can be used Projected onto the feature subspace defined by the i-th class, its projection The calculation is as follows:

[0047]

[0048] in Defined as a category-specific projection matrix:

[0049]

[0050] Projection matrix It has symmetry And idempotency It is a sub-adjoint projection factor;

[0051] The classification decision is based on the original test samples. Its projection on each category subspace The principle of minimizing the square of the Euclidean distance between them; if If the sample truly belongs to class i, then the distance should be minimized; for a four-class classification task, the final predicted class of the test sample... The following discrimination rule is given:

[0052]

[0053] in For category indexing; This is the original measurement sample vector; For the test sample at the 1st Projection onto a subspace.

[0054] Furthermore, the time series classification algorithm employs a dynamic learning framework, which updates the model parameters gradually through batches of arriving data from the sequence.

[0055] The dataset is described by formula (23) as a multivariate time series dataset:

[0056]

[0057] in, Indicates the first A multivariate time series sample, The four sample categories are: crystalline limestone, nodular limestone, and fractured dolomite; N represents the number of samples, C represents the feature dimension, and T represents the time step.

[0058] The formal representation of the online learning process is as follows:

[0059]

[0060] in, It is the Mini Random Convolution Kernel Transform (Mini Random Convolution Kernel Transform) ROCKET convolution module. It uses fixed kernel parameters. It is an eigenvector;

[0061]

[0062] in, It is an online learnable weight matrix. It is an online learnable bias vector;

[0063] In each learning step In this process, samples arrive in batches; the model training process can be represented by formula (26):

[0064]

[0065] in For trainable parameters, For learning rate, For learning steps The batch size, L is the cross-entropy loss function:

[0066]

[0067] Furthermore, this invention employs four mature metrics: Accuracy (ACC), Precision (PRC), Recall (REC), and F1 Score to evaluate predictive performance and determine the optimal classifier. Accuracy (ACC) measures the ratio of correctly predicted positive samples to the total number of predicted positive samples; Recall (REC) is the ratio of correctly predicted positive samples to the total number of actual positive samples; Precision (PRC) is the ratio of correctly predicted positive samples to all samples predicted as positive. Precision (PRC) and Recall (REC) are often inversely related, reflecting the classifier's ability to distinguish between positive and negative classes. The F1 Score integrates Recall (REC) and Precision (PRC) into a composite metric, overcoming the limitations of either metric. The specific calculation method is as follows:

[0068]

[0069] Where N is the total number of samples. The number of sample categories, For category The number of false positives, For category The number of false positives.

[0070] Furthermore, the raw downhole ranging data collected by the horizontal directional drilling HDD includes records of the drilling rig's operation and shutdown phases; to ensure that the analysis focuses on relevant information, a state discrimination function is first constructed as shown in formula (33) to filter out data that corresponds only to the drilling phase;

[0071] This function identifies and extracts the periods when the drilling rig is actively drilling, excluding non-operational periods;

[0072] This step effectively reduces interference from irrelevant data and improves the quality of subsequent dataset processing; after filtering, the dataset only contains regular drill-down data, laying the foundation for subsequent preprocessing.

[0073]

[0074] in, Indicates time (1 indicates drilling in progress, 0 indicates shutdown) Indicates thrust. Represents the rotational speed; applying formula (33) to the entire dataset, only retaining those that satisfy... Rows with =1; after filtering, the dataset only contains normal drill-up data, which forms the basis for subsequent preprocessing;

[0075] The missing values ​​in the filtered dataset were then processed; these gaps originated from sensor interruptions or temporary recording failures during drilling operations. Missing values ​​were imputed using linear interpolation. Outlier removal was performed after interpolation to improve data reliability. All residual NaN values ​​were removed during statistical calculations. For each feature... Its mean with standard deviation Calculate using the following formula:

[0076]

[0077] in, Indicates the first The number of non-NaN values ​​in each feature Indicates the first The first of the features There are 1 non-NaN data value; outliers are defined as those falling within the interval. Values ​​other than those mentioned above;

[0078] For each value in the 𝑗th feature Make a judgment: If This process is performed iteratively on all features, while simultaneously identifying, counting, and replacing outliers; and calculating the proportion of outliers for each feature to the total sample size. This step can effectively eliminate extreme anomalies that are detrimental to model performance, accurately reflect drilling conditions, and avoid distortions caused by sensor errors or instantaneous events.

[0079] Furthermore, it also includes T-distributed random neighborhood embedding (t-SNE), a nonlinear dimensionality reduction technique used to visualize high-dimensional data in a low-dimensional space; applying T-distributed random neighborhood embedding (t-SNE) to measurement-while-drilling (MWD) data to explore spatial distribution characteristics under different geological conditions; this algorithm can capture complex manifolds and local similarities, and can effectively identify subtle changes in drilling parameters affected by geological conditions; since lithological changes and drilling interactions often lead to nonlinear behavior in MWD data, this characteristic makes it particularly suitable for this type of data analysis;

[0080] Furthermore, T-distributed random neighborhood embedding (t-SNE) employs a modified Kullback-Leibler divergence-optimized embedding, ensuring that similar data points in the high-dimensional space remain close in the low-dimensional representation, while dissimilar points are spaced out. This property helps to clearly visualize the clustering of corresponding rock categories. The mathematical formula for T-distributed random neighborhood embedding (t-SNE) begins with calculating pairwise similarity in the high-dimensional space; for data points in the original space... and Conditional probability Defined as:

[0081]

[0082] in, It is aimed at The Gaussian variance of the surrounding local density is adjusted; this variance is controlled by a binary search to satisfy a given perplexity, thus controlling the number of effective neighbors. Then the joint probability is... The matchup becomes:

[0083]

[0084] in The number of data points; in low-dimensional space, for a point and The Student's distribution with 1 degree of freedom is used for modeling to alleviate the "crowding" problem and represent distance; joint probability Defined as:

[0085]

[0086] The goal is to minimize the cost function. That is, the Kullback-Leibler divergence between the high-dimensional distribution and the low-dimensional distribution:

[0087]

[0088] By reducing the dimensionality of five key horizontal directional drilling (HDD) parameters, T-distributed random neighborhood embedding (t-SNE) helps identify patterns that support geological detection.

[0089] According to another aspect of the present invention, a real-time data-driven intelligent online geological sensing system based on the above-described method is provided.

[0090] The beneficial effects of this invention are: (1) a database of MWD parameters and corresponding core samples is established to standardize the data processing flow; (2) a time series classification model is constructed to establish the mapping relationship between MWD parameters and the excavated lithology in a reasonable and reliable manner; (3) the time series classification algorithm is transformed into an online learning version, making it more efficient, scalable, and capable of continuous learning from input observations, thereby iteratively optimizing the classification model. By implementing the above methods, the ultimate goal is to improve the efficiency and safety of drilling operations while enhancing the intelligence level of HDD. Attached Figure Description

[0091] Figure 1 This is the overall structure of the online geological condition detection framework.

[0092] Figure 2 It is a major component of the HDD system.

[0093] Figure 3 For coring and marking during horizontal directional drilling.

[0094] Figure 4 This is a simplified diagram of the Rocket.

[0095] Figure 5 This is an overview map of the geological conditions of the experimental area of ​​the Nanhe River crossing section of the South-to-North Water Diversion Project in Xiangyang, China, as an example.

[0096] Figure 6 This represents the changes in key parameters within a single day.

[0097] Figure 7 This represents the spatial distribution of MWD in t-SNE space, categorized by rock hardness type.

[0098] Figure 8 This represents the statistical correlation of the MWD parameters.

[0099] Figure 9 Rank the importance of the MWD parameters.

[0100] Figure 10 The statistical distribution of the five MWD parameters in the four formations is shown. Detailed Implementation

[0101] A real-time data-driven intelligent online geological sensing method includes the following steps:

[0102] (1) Establish a database of MWD parameters and corresponding core samples for standardizing data processing procedures;

[0103] (2) Construct a time series classification model to establish a reasonable and reliable mapping relationship between the MWD parameters and the lithology of the excavated rocks;

[0104] (3) Transform the time series classification algorithm into an online learning version to make it more efficient and scalable, and continuously learn from the input observations to iteratively optimize the classification model.

[0105] An online framework for online geological condition detection is provided, comprising four steps:

[0106] (1) Acquire measurement-while-drilling (MWD) data from horizontal directional drilling (HDD);

[0107] (2) Multidimensional feature extraction is performed using the method based on random convolution kernel transformation (ROCKET).

[0108] (3) Use a ridge regression classifier to perform classification and realize online detection of lithology / geological conditions;

[0109] (4) Obtain the geological condition test results. Figure 1 An overall overview of the proposed framework is given.

[0110] Data obtained from horizontal directional drilling:

[0111] The HDD system used in this invention consists of several key components, including drill bit, drill pipe, mud circulation system and MWD unit. Figure 2 The main components of an HDD system are described. The drill bit is equipped with cutting tools that can effectively drill into various formations; the drill string connects the drill bit to the surface and transmits torque; the mud circulation system is responsible for delivering drilling fluid to the drill bit to cool the tools and remove cuttings from the hole; the MWD unit integrates multiple sensors for real-time data acquisition during drilling operations.

[0112] The sensor layout in the HDD system has been optimized to maximize data acquisition efficiency. Figure 2The sensor layout, including accelerometers, gyroscopes, and pressure sensors, is shown. These sensors are arranged along the drill string to monitor parameters such as well inclination (dip), azimuth, and pressure in real time. This deployment enables precise tracking of the drill string's position and attitude, which is crucial for effective geological analysis. The MWD parameters of the HDD system are equally critical for effective detection of geological conditions, and specifications include thrust (TF), mud pump pressure (MPP), revolutions per minute (RPM), rate of penetration (ROP), and power head position (PHP). A summary of these parameters is shown in Table 1.

[0113] The drilling process comprises several stages. First, the drill bit is positioned at the target entry point; then, drilling fluid is circulated to lubricate the drill bit and remove cuttings. As drilling progresses, the MWD unit continuously acquires data and transmits it back to the surface for analysis. This real-time data acquisition allows for immediate adjustments to operational parameters, thereby improving operational efficiency and safety. Core samples are also collected during drilling to obtain geological condition information. Figure 3 Core samples are displayed. Each core sample is accompanied by specific identification information (including depth and location) to facilitate accurate geological analysis. This identification is crucial for correlating MWD data with the physical properties of the encountered strata.

[0114] Figure 2 Key equipment components of the HDD system: (a) Schematic diagram of the drill pipe assembly, including the motor unit, drive shaft, bypass valve, universal joint, and drill bit; (b) Front view of the tracked fully hydraulic directional drilling rig; (c) MWD data acquisition system for monitoring operating parameters and sensor data; (d) Schematic diagram of downhole instruments and tools in the HDD system, covering wired drill pipe assembly, non-magnetic drill pipe, measurement-while-drilling (MWD) probe, mud motor assembly, and drill bit.

[0115] Figure 3 Core sampling and labeling during horizontal directional drilling: (a) limestone; (b) breccia; (c) dolomite; (d) core sample label.

[0116] Table 1 Technical Specifications of Key MWD Parameters for HDD Systems

[0117] Randomized convolution kernel transformation (ROCKET)

[0118] The Rocket algorithm, proposed by Dempster et al., is an efficient time series classification method that utilizes random convolutional kernels to extract discriminative features. Each convolutional kernel operates on each sample, generating feature maps from which two types of features are extracted: maximum values ​​and a new feature—Proportion of Positive Values ​​(PVV). PPV represents the proportion of sequence segments positively correlated with the given convolutional kernel and has been shown to improve classification accuracy.

[0119] Unlike traditional deep learning methods that require high computational resources to optimize convolutional kernels, ROCKET transforms time series using a large number of random convolutional kernels. Specifically, it randomly sets parameters such as kernel length, dilation, kernel weights, bias, and padding, as shown in formula (1).

[0120]

[0121] in Indicates the length of the input time series. The arithmetic mean.

[0122] For an input time series Each convolution kernel Capture multi-scale patterns through dilated convolutions.

[0123]

[0124] in ,and Convolution output Calculate according to formula (8):

[0125]

[0126] Rocket computes two aggregated features from each feature map, generating two real numbers. and As each Features:

[0127]

[0128] in It is an indicator

[0129] For the Each kernel, Rocket generates two features for each time series (i.e., ...). and ), represented as :

[0130]

[0131] Where T is the transpose operator. This refers to the total number of cores. In this invention, K = 10000 (default value).

[0132] Connection feature matrix used for training samples for:

[0133]

[0134] in It represents the number of training samples.

[0135] Figure 4: A simplified diagram of Rocket. The example described has k random kernels and generates 2k random features for each input dimension.

[0136] Ridge Regression Classifier

[0137] The features generated by Rocket are used as input to train a linear classifier. Ridge regression is typically chosen as this classifier because optimizing its regularization hyperparameters through cross-validation is computationally efficient. To improve classification performance within this framework, it is assumed that the feature vectors transformed by Rocket occupy a linear subspace within each class. Therefore, a new test sample belonging to that class can be approximated as a linear combination of training samples from the same class. This structural assumption necessitates the development of a specific ridge regression-based linear model within the context of the feature space generated by Rocket.

[0138] The feature matrix generated by Rocket is shown in equation (12). For the four types of problems defined in this invention... We constructed a ridge regression model for each class i. The feature subset matrix corresponding to the training samples belonging to class i is defined as:

[0139]

[0140] in This represents the row vector of the evaluation features of the k-th sample in category i. It is the number of samples in category i.

[0141] Given a sample to be classified (This represents its feature vector after Rocket transformation and flattening), we assume it can be categorized. Feature matrix The sample linear approximation in:

[0142]

[0143] in It is the vector of regression coefficients to be solved. This represents the approximation error vector.

[0144] To obtain a stable solution with good generalization ability, we introduce a ridge regression objective function:

[0145]

[0146] in This is the regularization parameter. This parameter controls model complexity, balancing the trade-off between training error and model variance. (Increase...) Will control more significantly The magnitude of the amplitude is increased, causing its coefficients to approach 0, thereby reducing model variance but potentially increasing bias. Lowering... This reduces the penalty, allowing the model to fit the training data more closely.

[0147] By analyzing the objective function (15) with respect to... By differentiating and setting it to zero, we derive the closed-form solution. :

[0148]

[0149] in It is the identity matrix. Solving formula (18) yields the optimal estimate of the regression coefficient vector:

[0150]

[0151] Using the estimated Test samples can be used Projected onto the feature subspace defined by the i-th class, its projection The calculation is as follows:

[0152]

[0153] in Defined as a category-specific projection matrix:

[0154]

[0155] It is important to note that the projection matrix It has symmetry And idempotency It is a sub-adjoint projection factor.

[0156] The classification decision is based on the original test samples. Its projection on each category subspace The principle of minimizing the square of the Euclidean distance between them. If If the sample truly belongs to class i, then this distance should be minimized. For a four-class classification task, the final predicted class of the test sample... The following discrimination rule is given:

[0157]

[0158] in For category indexing; This is the original measurement sample vector; For the test sample at the 1st Projection onto a subspace.

[0159] Online geological condition monitoring:

[0160] The core objective of this invention is to deploy a real-time geological condition monitoring system capable of processing continuous drilling measurement data streams. Traditional batch processing methods, which involve retrospective analysis after large-scale data collection, are inherently unsuitable for scenarios requiring immediate operational decisions. Timely identification of geological conditions is crucial in contexts such as well drilling optimization and early hazard detection. Batch processing introduces significant delays, hindering proactive adjustments to drilling parameters based on the actual formation being drilled at any given moment. Therefore, online analysis is not merely desirable but a fundamental requirement for realizing the intended practical impact of this invention. It enables an immediate feedback loop, transforming raw sensor data into actionable geological insights within the operational timeframe, thereby improving drilling efficiency, accuracy, and safety.

[0161] While the Rocket algorithm demonstrates exceptional capabilities in time series classification tasks, its applicability in constrained online environments varies significantly. MiniRocket addresses this key limitation in online scenarios through strategic optimization. Similar to Rocket, MiniRocket employs a convolutional kernel and trains a linear classifier on the transformed features. However, unlike Rocket, which uses random kernels, MiniRocket uses a fixed, small convolutional kernel, exhibiting high determinism. This algorithm preserves Rocket's PPV features while adjusting key parameters: the kernel length is fixed at 9, and the kernel weights are limited to values ​​between -1 and 2, forming the kernel function through combination. Furthermore, MiniRocket derives a bias term from the convolutional output. For selected training samples and a given convolutional kernel, MiniRocket computes the convolutional output and extracts the [0.25, 0.5, 0.75] quantiles. These quantiles are used as optional bias values. See the table for a detailed comparison.

[0162] Table 2. Comparison of key parameters between Rocket and MiniRocket.

[0163]

[0164] The online multivariate time series classification algorithm based on MiniRocket employs a dynamic learning framework, progressively updating model parameters through batches of arriving data. The task of this invention can be described by formula (23) for a multivariate time series dataset:

[0165]

[0166] in Indicates the first A multivariate time series sample, The samples correspond to four categories: crystalline limestone, nodular limestone, and fractured dolomite. N represents the number of samples, C represents the feature dimension, and T represents the time step.

[0167] The formal representation of the online learning process is as follows:

[0168]

[0169] in It's the MiniRocket convolution module. It uses fixed kernel parameters. It is an eigenvector.

[0170]

[0171] in It is an online learnable weight matrix. It is an online learnable bias vector.

[0172] In each learning step In this process, samples arrive in batches. The model training process can be represented by formula (26):

[0173]

[0174] in For trainable parameters, For learning rate, For learning steps The batch size, L is the cross-entropy loss function:

[0175]

[0176] This invention employs four established metrics: Accuracy (ACC), Precision Ratio (PRC), Recall (REC), and F1 Score to evaluate predictive performance and determine the optimal classifier. ACC measures the ratio of correctly predicted positive samples to the total number of predicted positive samples and is the most widely used metric. However, in imbalanced classification tasks, ACC cannot reliably reflect classifier performance. Recall (REC) is the ratio of correctly predicted positive samples to the total number of actual positive samples. Positive Prediction Ratio (PRC) is the ratio of correctly predicted positive samples to all samples predicted as positive. PRC and REC often show an inverse relationship, reflecting the classifier's ability to distinguish between positive and negative classes. The F1 score integrates REC and PRC into a composite metric, overcoming the limitations of either. The specific calculation method is as follows:

[0177]

[0178] Where N is the total number of samples. The number of sample categories, For category The number of false positives, For category The number of false positives.

[0179] Experimental research

[0180] The data used in this invention originates from the South-to-North Water Diversion Project (Middle Route). The geological background of the water conveyance tunnels is extremely complex, presenting various engineering and environmental geological challenges. The tunnel route traverses fault zones and intersects with karst cave sections within soluble rock strata, a combination that significantly increases the risk of geological hazards. To mitigate these risks and comprehensively investigate the geological conditions along the tunnel route, horizontal directional drilling (HDD) technology was employed for detailed geological exploration, while minimizing surface disturbance. An overview of the experimental scope can be found in [link to experimental scope]. Figure 5 .

[0181] The operating parameters of the horizontal directional drilling rig are collected in real time by sensors installed at various locations on the equipment. These sensors operate at a frequency of 1 Hz to ensure the continuity and accuracy of data acquisition. Each data point monitors five key operating parameters: thrust, mud pump pressure, revolutions per minute, drilling speed, and power head position. These parameters reflect the interaction between the equipment and the formation during drilling, providing important indicators for detecting changes in rock properties and potential risks. Figure 6 The diagram shows the typical changes in the aforementioned key MWD parameters over a single day.

[0182] To facilitate the analysis and classification of geological conditions, a lithology-based rock hardness classification system was adopted. This method categorizes rocks into different types based on their lithological characteristics, which directly affect drilling response and equipment behavior. As shown in Table 3, the following classifications were used: extremely hard rocks (e.g., dolomite), hard rocks (e.g., crystalline limestone), relatively soft rocks (e.g., nodular limestone), and soft rocks (e.g., fractured rocks). The construction log detailed the lithology and rock mass classification of the rocks along the drilling route, ensuring a consistent mapping between the recorded strata and the established classification system. Each rock type was accompanied by a representative sample illustration to aid visual identification.

[0183] Implementation details

[0184] In this invention, the raw downhole ranging data collected from the horizontal directional drilling system includes records of both the drilling rig's operational and shutdown phases. To ensure that the analysis focuses on relevant information, a state discrimination function, as shown in formula (33), is first constructed to filter out data corresponding only to the drilling phase. This function identifies and extracts the periods when the drilling rig is in an active drilling state, excluding non-operational periods. This discrimination is based on key operational indicators that exhibit significant characteristics during active drilling, such as... Figure 6 As shown. For example, relevant parameters remain non-zero during the active drilling phase, while they show zero values ​​or slight fluctuations during the shutdown phase. This step effectively reduces interference from irrelevant data and improves the quality of subsequent dataset processing. After filtering, the dataset contains only regular drilling data, laying the foundation for subsequent preprocessing.

[0185]

[0186] in Indicates time (1 indicates drilling in progress, 0 indicates shutdown) Indicates thrust. Represents the rotational speed. Applying formula (33) to the entire dataset, only those satisfying the condition are retained. Rows with a value of 1. After filtering, the dataset contains only normal drill-up data, which forms the basis for subsequent preprocessing.

[0187] Missing values ​​in the filtered dataset were then addressed. These gaps primarily stemmed from sensor interruptions or temporary recording failures during drilling operations. Missing values ​​were imputed using linear interpolation. Outlier removal was then performed after interpolation to improve data reliability. All residual NaN values ​​were removed during statistical calculations. For each feature... Its mean with standard deviation Calculate using the following formula:

[0188]

[0189] in Indicates the first The number of non-NaN values ​​in each feature Indicates the first The first of the features There are [number] non-NaN data values. Outliers are defined as those falling within the range [range]. Values ​​other than those listed above.

[0190] For each value in the 𝑗th feature Make a judgment: If This process iteratively executes on all features, simultaneously identifying, counting, and replacing outliers; and calculating the proportion of outliers for each feature relative to the total sample size. This step effectively eliminates extreme anomalies that may negatively impact model performance, accurately reflecting drilling conditions and avoiding distortions caused by sensor errors or transient events.

[0191] T-distributed random neighborhood embedding (t-SNE) is a nonlinear dimensionality reduction technique widely used to visualize high-dimensional data in a low-dimensional space. This method preserves local structure and reveals clustering patterns that may not be easily perceived in the original high-dimensional space. This invention applies t-SNE to MWD data to explore the spatial distribution characteristics under different geological conditions. The algorithm captures complex manifolds and local similarities, which is crucial for identifying subtle changes in drilling parameters influenced by geological conditions. Since lithological variations and drilling interactions often lead to nonlinear behavior in MWD data, this characteristic makes it particularly suitable for this type of data analysis. Furthermore, t-SNE employs a modified Kullback-Leibler divergence-optimized embedding, ensuring that similar data points in the high-dimensional space remain close in the low-dimensional representation, while dissimilar points are spaced out. This characteristic helps to clearly visualize clusters corresponding to rock categories. The mathematical formula for t-SNE begins with calculating pairwise similarities in the high-dimensional space. For data points in the original space... and Conditional probability Defined as:

[0192]

[0193] in It is aimed at The Gaussian variance of the surrounding local density is adjusted. This variance is determined through a binary search to achieve a fixed perplexity, thereby controlling the number of effective neighbors. The joint probability is then... Symmetricalization:

[0194]

[0195] in This represents the number of data points. In low-dimensional space, for a point... and A Student's distribution with 1 degree of freedom is used for modeling to alleviate the "crowding" problem and represent distance. Joint probability Defined as:

[0196]

[0197] The goal is to minimize the cost function. That is, the Kullback-Leibler divergence between the high-dimensional distribution and the low-dimensional distribution:

[0198]

[0199] By reducing the dimensionality of five key HDD parameters, t-SNE helps identify patterns that support geological detection. Its visualization results are shown below. Figure 7 The spatial distribution of data points reveals a significant separation trend between different geological conditions. For instance, in Figure 7 In (a), dolomite data points are mainly clustered in the low-value region of "Dimension 1," while nodular limestone is located in the high-value region of "Dimension 3." However, crystalline limestone and breccia exhibit overlapping distributions in some dimensions (such as "Dimension 2"), suggesting that they may have similar MWD characteristics due to geological continuity or measurement noise. Overall, Figure 7 The MWD characteristics are clearly presented in the reduced-dimensional space, providing an important reference for geological classification.

[0200] Figure 7 : Spatial distribution of MWD by rock hardness category in t-SNE space; (a) 3D representation; (b) Dimension 1 and Dimension 2; (c) Dimension 2 and Dimension 3; (d) Dimension 1 and Dimension 3

[0201] Distribution and correlation analyses were performed to visualize the preprocessed data. Figure 8 Pearson correlation analysis was performed to demonstrate the characteristics of different downhole measurement parameters in the dataset. Figure 9 The study presents the ranking of the characteristic importance of five downhole measurement parameters and their cumulative contribution. Regarding the MWD parameters, TF (thrust) contributes the most, followed by RPM (rotation speed); in contrast, ROP (rate of penetration) has the lowest characteristic importance. Overall, the contribution distribution of each parameter is relatively balanced. This analysis provides crucial evidence for the relative importance of each MWD parameter in geological condition monitoring. Figure 10 Histograms are used to illustrate the distribution range and frequency of each MWD parameter in the preprocessed HDD dataset, highlighting the variability of parameters such as torque and thrust at different drilling stages. (Variation at different drilling stages.)

[0202] Figure 10 Statistical distribution of five MWD parameters in four formations: (a) thrust; (b) mud pump pressure; (c) rotational speed; (d) penetration depth; (e) power head position.

[0203] The experimental setup was designed to rigorously evaluate the proposed online learning framework for geological condition detection. A comprehensive dataset containing 1000 MWD samples was used, with balanced representativeness across four geological categories (250 samples per category). The dataset was split in a 7:3 ratio, resulting in 700 samples for training and 300 samples for testing. All experiments were conducted on a computing system equipped with an Intel i5-13600KF processor and an NVIDIA RTX4060, running Ubuntu 20.04 and Python 3.7.12. Hyperparameter configuration followed established practices for time series classification using random convolutional networks (see Table 4), employing 10,000 random kernels to achieve optimal feature diversity while maintaining computational efficiency. Model training used a batch size of 20 samples, a parameter carefully chosen to balance memory constraints and gradient stability during online learning. The learning rate was set to 1e-4 to ensure stable convergence when solving the ridge regression objective, and a random seed of 42 ensured complete reproducibility of all experimental runs. This configuration achieves an optimal balance between computational efficiency and model performance, making it suitable for real-time geological condition detection applications. This training process enables the model to detect geological changes in real time based on cleaned MWD data patterns.

[0204] Table 4. Hyperparameter configuration for the proposed online MiniRocket framework

[0205] Results Analysis

[0206] This section provides a comprehensive evaluation of the proposed online MiniRocket framework through systematic analysis, covering online learning performance, accuracy comparison with benchmark models, robustness to missing data, and resistance to noise interference. Experimental results show that the method can achieve stable geological classification under various complex constraints, providing substantial validation for its practical application in real drilling environments. Specific analysis results are as follows:

[0207] (1) The online MiniRocket framework achieved stable and accurate geological condition detection on practical training samples. The online classification performance was systematically evaluated through an incremental learning process: the model was updated and evaluated every 20 samples in 5 independent runs, thus establishing a 95% confidence interval. All performance indicators showed a consistent improvement pattern: rapid improvement in the initial training phase (first 200 samples), gradually converging to a stable value after about 400-500 training samples. The overall accuracy reached about 0.95, and the confidence interval narrowed significantly, indicating that the method has strong robustness to random initialization effects. It is worth noting that dolomite and breccia both obtained near-perfect F1 scores (>0.98) with minimal variation in their respective runs, while crystalline limestone performed slightly worse (F1≈0.92) and had a slightly wider confidence interval, indicating that it is relatively more sensitive to the composition of training data. The spheroidal limestone category performed moderately, reaching stable performance after about 300 samples (F1≈0.94). The online MiniRocket framework significantly improves accuracy in capturing vertical sequences of complex geological structures, especially by effectively reducing the classification error of spheroidal limestone compared to the benchmark Rocket model.

[0208] (2) Among the six machine learning methods, the online MiniRocket model performed best in the geological condition detection task. As shown in Table 4, this model achieved the highest scores in both overall accuracy (0.957) and macro F1 score (0.957), surpassing all benchmark algorithms. Its excellence is further reflected in the consistently high F1 scores across various rock types: dolomite achieved a perfect score of 1.000, crystalline limestone 0.918, nodular limestone 0.942, and breccia as high as 0.968. Although the LSTM-FCN model also achieved a perfect F1 score of 1.000 in dolomite classification, and the Transformer model also achieved a high score of 0.993, its performance was not stable enough—in the crystalline limestone classification, the scores of LSTM-FCN and Transformer dropped significantly to 0.803 and 0.817, respectively. The Rocket model performed reasonably well but ranked second, with both overall accuracy and macro F1 score slightly below 0.900. In stark contrast, traditional methods such as DTW-Kmeans and CNNs show a significant performance decline, with macroscopic F1 scores falling below 0.770, highlighting their inherent limitations in handling such complex pattern recognition tasks. The confusion matrix of the online MiniRocket model shows that the predictions are most concentrated along the main diagonal. This clear main diagonal contrasts sharply with the matrices of other models—which exhibit a gradual increase in off-diagonal elements, corresponding to a significant increase in confusion rates between specific rock categories.

[0209] Table 5. Performance evaluation of different algorithms on four types of geological condition classification tasks

[0210]

[0211] (3) The proposed online MiniRocket method and the Rocket method demonstrate significant advantages in handling time series data with varying degrees of missing values, thus exhibiting strong robustness in classifying geological conditions under high uncertainty. To evaluate the model's performance under incomplete input conditions, we conducted a series of comparative experiments using incomplete time series input data with different missing rates of 10%, 20%, 30%, 40%, 50%, and 60%. These missing values ​​simulate the uncertainty in real datasets. For each missing rate, the data deletion process was repeated 100 times to generate diverse datasets. Subsequently, both the online MiniRocket method and the Rocket method were tested on 100 new datasets to evaluate their robustness to missing data. The results under different missing data rates are shown in Table 5. As the proportion of missing data increases, the mean and median of the accuracy (ACC) and macro F1 score gradually decrease, while the corresponding standard deviation increases slightly. These findings indicate that dataset incompleteness reduces classification performance and introduces additional uncertainty to some extent. However, the negative impact of missing data on the classification performance of the online MiniRocket method and the Rocket method remains relatively limited. Notably, under the same missing rate, the dispersion of the results distribution of the online MiniRocket method is consistently smaller than that of the Rocket method, indicating that the online MiniRocket method exhibits higher stability when handling incomplete time series inputs. Therefore, the developed online classification method can effectively extract meaningful geological information from incomplete time series datasets in a flexible manner, demonstrating high robustness and the potential to be extended to more complex engineering scenarios—scenarios typically characterized by high uncertainty and randomness. Furthermore, these experiments confirm that the classification performance of the online MiniRocket method consistently outperforms the Rocket method in this case, with its average accuracy (ACC) and macroscopic F1 score being at least 0.061 and 0.062 higher, respectively.

[0212] Table 6 Comparative performance evaluation as the proportion of incomplete input increases (based on the average of more than 100 experiments)

[0213]

[0214] (4) The proposed online MiniRocket method also demonstrates excellent robustness to noise interference, enabling it to reliably extract meaningful geological information from noise-contaminated, high-uncertainty MWD data. To simulate the noise environment in the MWD data, multiplicative Gaussian noise was applied to the original MWD dataset. The noise addition process involves generating a zero-mean Gaussian noise matrix E with the same dimensions as the original data, where each element... Follows independent and identically distributed: ,in This represents the noise standard deviation (noise level). Noise is then injected via element-wise multiplication (Hadamard product). That is, make the following modifications to each element: ,in In the experiment, The noise level was increased from 0.2 to 1.0 in increments of 0.2, corresponding to noise levels of 20%, 40%, 60%, 80%, and 100%. For each noise level, a random dataset was generated and repeated 100 times to ensure variability. The results show that at a 20% noise input level, the online MiniRocket method maintains high accuracy comparable to the noise-free data in Table 5, with its accuracy (ACC) and macro F1 score only 0.004 and 0.027 lower, respectively. In contrast, the Rocket method performs poorly: its accuracy and macro F1 score are 0.054 and 0.056 lower, respectively, than those of the noise-free experiments in Table 5. As the noise level increases, the decrease in accuracy and macro F1 score of the online MiniRocket method is significantly less than that of the Rocket method. Even at 100% noise levels, the online MiniRocket method achieves an average accuracy (ACC) of 0.877 and an average macro F1 score of 0.852, while the Rocket method drops to an average accuracy of 0.567 and an average macro F1 score of 0.548. This superior performance is likely attributed to the characteristics of the online learning mechanism: by incrementally processing the data stream and adaptively updating the classification centers, it effectively prevents the accumulation of noise-induced errors. Overall, the online MiniRocket method demonstrates high robustness to noise, highlighting its superior reliability in addressing the challenges of downhole measurement data quality.

[0215] Imbalanced sample distribution significantly impacts the performance of classification tasks, especially in geological condition detection, where rock type distribution is often naturally skewed. This section investigates extreme imbalance phenomena—two specific imbalance scenarios: one where one class has an extremely small number of samples, and the other where one class has an abnormally large number of samples. Specifically, we will explore how severe undersampling and oversampling scenarios affect the performance of the Rocket algorithm and the online minirocket algorithm.

[0216] (1) Category undersampling significantly reduces the algorithm's ability to identify target categories. However, the online MiniRocket model successfully mitigated this adverse effect, maintaining near-stable performance even with a sampling rate reduced to 95.5%. To investigate the performance of severely undersampled categories, we artificially reduced the number of samples for the dolomite category, which originally had the highest accuracy. To ensure statistical robustness, the experiment was repeated 100 times with different random seeds. The F1 score of dolomite in the Rocket model reached an inflection point at (70%, 0.879) as the reduction ratio increased: when the reduction ratio was ≤70%, the model performance was basically unaffected; however, after exceeding this threshold, the performance deteriorated sharply, and the Rocket model could not guarantee the classification accuracy of a few categories. In contrast, the online MiniRocket model exhibited stronger robustness. Its F1 score showed almost no decreasing trend before the sample reduction reached 95.5%, with the inflection point appearing at (95.5%, 0.962). This means that compared to the standard Rocket model, its tolerance for sample reduction was improved by 25.5%. The higher resistance of the online MiniRocket model to extremely small sample sizes likely stems from its incremental learning characteristics—which allow it to adapt more effectively to the distribution of available data during training.

[0217] (2) Conversely, the effect of overrepresentation was examined by increasing the sample size of the initially least accurate crystalline limestone. Synthetic Minority Oversampling Technique (SMOTE) was used to generate synthetic samples. This technique is achieved by interpolating existing minority class instances within the feature space: for each minority class sample, multiple synthetic instances are generated along the line segment connecting it to its k nearest neighbors of the same class. This method effectively improves the diversity and sample size of the minority class, avoiding simple duplication. After applying SMOTE to the crystalline limestone category, 100 repeated experiments were conducted.

[0218] This invention systematically developed and validated the online MiniRocket framework for real-time geological condition detection based on logging-while-drilling data. Through the following key innovations, this study established a reliable mapping relationship between drilling parameters and excavated rock strata types: creating a standardized database of drilling parameters and core samples; developing an adaptive geological condition detection model to ensure high accuracy under extreme conditions; and implementing an online learning architecture to continuously update the model with newly acquired logging-while-drilling data. This framework significantly improves the operational efficiency and safety of horizontal directional drilling, while also promoting the intelligentization of drilling operations in complex geological environments. Based on experimental research, the specific conclusions are as follows: (1) Online learning can quickly converge to a stable performance level of 0.95 accuracy within approximately 200 training samples. Comparative analysis with five benchmark methods shows that its classification performance is significantly better than other methods, with an overall accuracy improvement of 6.3% and a macroscopic F1 score improvement of 6.5%. (2) The online MiniRocket framework exhibits excellent missing data tolerance, maintaining stable classification performance even when the input data has up to 50% missing values. This confirms its robustness in handling incomplete MWD data streams during actual drilling operations. (3) The method exhibits excellent noise resistance, with performance degradation of less than 8.5% even under 100% Gaussian noise conditions, significantly outperforming the benchmark comparison method and ensuring reliable operation in high-noise drilling environments where sensor data quality is often affected. (4) The method demonstrates strong robustness under extreme class imbalance scenarios, with robustness 25.5% higher than the benchmark model. Even with a 95.5% reduction in minority class sample size, the method maintains stable performance, significantly outperforming the benchmark method—which exhibits severe performance degradation when the sample size is reduced by more than 70%. To further advance this research, several key tasks need to be carried out in the future. First, the framework should be extended to more complex geological environments, covering more rock types and higher degrees of heterogeneity, with particular attention to its transfer learning capabilities across different drilling projects and geological structures. Second, future research should explore integrating more sensor modalities—such as real-time vibration monitoring, acoustic emission data, and formation imaging—to construct a multimodal fusion system to further improve detection accuracy and achieve more comprehensive geological characterization. Finally, research should focus on developing an integrated real-time decision support system that combines detection algorithms with automatic drilling parameter optimization to achieve geological identification while adjusting operational parameters in real time, thereby improving drilling efficiency and reducing the risks brought about by uncertain formation conditions.

[0219] It will be apparent to those skilled in the art that this invention is not limited to the details of the exemplary embodiments described above, and that it can be implemented in other specific forms without departing from the spirit or essential characteristics of this invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of this invention is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included within this invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A real-time data-driven intelligent online geological sensing method, characterized in that, Includes the following steps: (1) Establish a database of MWD parameters and corresponding core samples for standardizing data processing procedures; (2) Construct a time series classification model to establish a reasonable and reliable mapping relationship between the MWD parameters and the lithology of the excavated rocks; (3) Transform the time series classification algorithm into an online learning version to make it more efficient and scalable, and continuously learn from the input observations to iteratively optimize the classification model.

2. The real-time data-driven intelligent online geological sensing method according to claim 1, characterized in that: It also includes an online framework for online geological condition detection, which comprises four steps: (1) Acquire measurement-while-drilling (MWD) data from horizontal directional drilling (HDD); (2) Multidimensional feature extraction is performed using the method based on random convolution kernel transformation (ROCKET). (3) Use a ridge regression classifier to perform classification and realize online detection of lithology / geological conditions; (4) Obtain the geological condition test results.

3. The real-time data-driven intelligent online geological sensing method according to claim 2, characterized in that: The random convolution kernel transformation ROCKET transforms the time series using a large number of random convolution kernels, that is, it randomly sets the kernel length, dilation rate, kernel weights, bias and padding parameters, as shown in formula (1). in Indicates the length of the input time series. The arithmetic mean; For an input time series Each convolution kernel Capture multi-scale patterns through dilated convolutions; in, ,and ; Convolution output Calculate according to formula (8): The random convolution kernel transform ROCKET computes two aggregated features from each feature map, producing two real numbers. and As each Features: in It is an indicator; For the Each kernel, the random convolution kernel transformation ROCKET generates two features for each time series (i.e., and ), representing : Where T is the transpose operator. It is the total number of nuclear components; Connection feature matrix used for training samples for: in, It represents the number of training samples.

4. The real-time data-driven intelligent online geological sensing method according to claim 3, characterized in that: In the ridge regression classifier, the feature matrix generated by the random convolution kernel transformation ROCKET is shown in equation (12); for the defined four classes of problems A ridge regression model for each class i was constructed; the feature subset matrix corresponding to the training samples belonging to class i is defined as: in, This represents the row vector of the evaluation features of the k-th sample in category i. It is the number of samples in category i; Given a sample to be classified This represents its feature vector after being transformed and flattened by a random convolution kernel (ROCKET), assuming it can be categorized. Feature matrix The sample linear approximation in: in, It is the vector of regression coefficients to be solved. Represents the approximate error vector; To obtain a stable solution with good generalization ability, a ridge regression objective function is introduced: in This is the regularization parameter, which controls model complexity and balances the trade-off between training error and model variance; increasing... Will control more significantly The magnitude of the amplitude makes its coefficients tend to 0, thereby reducing the model variance but potentially increasing bias; reducing This reduces the penalty, allowing the model to fit the training data more closely; By analyzing the objective function (15) with respect to... Differentiate and set it to zero to derive the closed-form solution. : in It is the identity matrix; solving formula (18) yields the optimal estimate of the regression coefficient vector: Using the estimated Test samples can be used Projected onto the feature subspace defined by the i-th class, its projection The calculation is as follows: in Defined as a category-specific projection matrix: Projection matrix It has symmetry And idempotency It is a sub-adjoint projection factor; The classification decision is based on the original test samples. Its projection on each category subspace The principle of minimizing the square of the Euclidean distance between them; if If the sample truly belongs to class i, then the distance should be minimized; for a four-class classification task, the final predicted class of the test sample... The following discrimination rule is given: in For category indexing; This is the original measurement sample vector; For the test sample at the 1st Projection onto a subspace.

5. The real-time data-driven intelligent online geological sensing method according to claim 2, characterized in that: The online geological condition detection system is a real-time geological condition detection system that processes continuous drilling measurement data streams, enabling instant feedback loops and transforming raw sensor data into actionable geological insights within the operating timeframe to improve drilling efficiency, accuracy, and safety.

6. The real-time data-driven intelligent online geological sensing method according to claim 1, characterized in that: The time series classification algorithm adopts a dynamic learning framework, which updates the model parameters step by step based on the batches of data arriving in the sequence. The dataset is described by formula (23) as a multivariate time series dataset: in, Indicates the first A multivariate time series sample, The four sample categories are: crystalline limestone, nodular limestone, and fractured dolomite; N represents the number of samples, C represents the feature dimension, and T represents the time step. The formal representation of the online learning process is as follows: in, It is the Mini Random Convolution Kernel Transform (Mini Random Convolution Kernel Transform) ROCKET convolution module. It uses fixed kernel parameters. It is an eigenvector; in, It is an online learnable weight matrix. It is an online learnable bias vector; In each learning step In this process, samples arrive in batches; the model training process can be represented by formula (26): in For trainable parameters, For learning rate, For learning steps The batch size, L is the cross-entropy loss function:

7. The real-time data-driven intelligent online geological sensing method according to claim 1, characterized in that: Four established metrics—Accuracy (ACC), Precision (PRC), Recall (REC), and F1 score—are used to evaluate predictive performance and determine the optimal classifier. Accuracy (ACC) measures the ratio of correctly predicted positive samples to the total number of predicted positive samples. Recall (REC) is the ratio of correctly predicted positive samples to the total number of actual positive samples. Precision (PRC) is the ratio of correctly predicted positive samples to all samples predicted as positive. Precision (PRC) and recall (REC) are often inversely related, reflecting the classifier's ability to distinguish between positive and negative classes. The F1 score integrates recall (REC) and precision (PRC) into a composite metric, overcoming the limitations of either metric alone. The specific calculation method is as follows: Where N is the total number of samples. The number of sample categories, For category The number of false positives, For category The number of false positives.

8. The real-time data-driven intelligent online geological sensing method according to claim 1, characterized in that: The raw downhole ranging data collected by the horizontal directional drilling HDD includes records of the drilling rig operation and shutdown stages; to ensure that the analysis focuses on relevant information, a state discrimination function is first constructed as shown in formula (33) to filter out data that corresponds only to the drilling stage; This function identifies and extracts the periods when the drilling rig is actively drilling, excluding non-operational periods; This step effectively reduces interference from irrelevant data and improves the quality of subsequent dataset processing; after filtering, the dataset only contains regular drill-down data, laying the foundation for subsequent preprocessing. in, Indicates time (1 indicates drilling in progress, 0 indicates shutdown) Indicates thrust. Represents the rotational speed; applying formula (33) to the entire dataset, only retaining those that satisfy... Rows with =1; after filtering, the dataset only contains normal drill-up data, which forms the basis for subsequent preprocessing; The missing values ​​in the filtered dataset were then processed; these gaps originated from sensor interruptions or temporary recording failures during drilling operations. Missing values ​​were imputed using linear interpolation. Outlier removal was performed after interpolation to improve data reliability. All residual NaN values ​​were removed during statistical calculations. For each feature... Its mean with standard deviation Calculate using the following formula: in, Indicates the first The number of non-NaN values ​​in each feature Indicates the first The first of the features There are 1 non-NaN data value; outliers are defined as those falling within the interval. Values ​​other than those mentioned above; For each value in the 𝑗th feature Make a judgment: If This process is performed iteratively on all features, while simultaneously identifying, counting, and replacing outliers; and calculating the proportion of outliers for each feature to the total sample size. This step can effectively eliminate extreme anomalies that are detrimental to model performance, accurately reflect drilling conditions, and avoid distortions caused by sensor errors or instantaneous events.

9. The real-time data-driven intelligent online geological sensing method according to claim 1, characterized in that: It also includes T-distributed random neighborhood embedding (t-SNE), a nonlinear dimensionality reduction technique used to visualize high-dimensional data in a low-dimensional space; applying T-distributed random neighborhood embedding (t-SNE) to measurement-while-drilling (MWD) data to explore the spatial distribution characteristics under different geological conditions; This algorithm can capture complex manifolds and local similarities, and can effectively identify subtle changes in drilling parameters affected by geological conditions. Since the interaction between lithological changes and drilling often leads to nonlinear behavior in measurement-while-drilling (MWD) data, this characteristic makes it particularly suitable for this type of data analysis. Furthermore, T-distributed random neighborhood embedding (t-SNE) employs a modified Kullback-Leibler divergence-optimized embedding, ensuring that similar data points in the high-dimensional space remain close in the low-dimensional representation, while dissimilar points are spaced out. This property helps to clearly visualize the clustering of corresponding rock categories. The mathematical formula for T-distributed random neighborhood embedding (t-SNE) begins with calculating pairwise similarity in the high-dimensional space; for data points in the original space... and Conditional probability Defined as: in It is aimed at The Gaussian variance of the surrounding local density is adjusted. This variance is determined through a binary search to achieve a fixed perplexity, thereby controlling the number of effective neighbors. The joint probability is then... Symmetricalization: in This represents the number of data points. In low-dimensional space, for a point and The Student's t distribution with 1 degree of freedom is used for modeling to alleviate the "crowding" problem and represent distance; Joint probability Defined as: The goal is to minimize the cost function. That is, the Kullback-Leibler divergence between the high-dimensional distribution and the low-dimensional distribution: By reducing the dimensionality of five key horizontal directional drilling (HDD) parameters, T-distributed random neighborhood embedding (t-SNE) helps identify patterns that support geological detection.

10. A real-time data-driven intelligent online geological sensing system, characterized in that: The sensing system employs the sensing method described in any one of claims 1-9.

Images