An intelligent identification method for drilling ahead lithology and geological anomaly body in a coal mine underground

Abstract

The application discloses a kind of coal mine underground advanced drilling lithology and geological anomaly body intelligent identification method, first respectively collect while drilling time series data and drilling field image data and pre-process;Then the while drilling time series data and drilling field image data are high-precision synchronization and alignment;Then the while drilling time series data is converted into multi-channel two-dimensional image using selected conversion mode;Using different feature extractors, respectively extract the features of time series conversion image and drilling field image;Using a multi-modal fusion model, the features of time series conversion image and drilling field image are deeply fused;Finally, the fused features are analyzed using the constructed identification model, and the identification results containing the lithology category and the geological anomaly body category at the corresponding position are output, achieving intelligent and high-precision identification of lithology and various geological anomaly bodies in coal mine advanced drilling, and improving the accuracy and timeliness of geological structure detection, thereby ensuring mine production safety.

Description

Technical Field

[0001] This invention belongs to the field of mine borehole detection technology, specifically a method for intelligent identification of lithology and geological anomalies in underground coal mine advance drilling. Background Technology

[0002] Advanced drilling in coal mines is a crucial method for detecting geological structures and preventing various disasters. Currently, advanced drilling primarily relies on measurement-while-drilling (MWD) data, including time-series parameters such as drilling rate (ROP), weight on bit (WOB), rotational speed (RPM), torque (TORQUE), pump pressure, and natural gamma. These data are typically presented as curves, and experts analyze the curve characteristics based on experience to identify various geological anomalies encountered during drilling, such as lithology, faults, fracture zones, and collapse columns.

[0003] However, existing technologies have the following prominent problems in identifying lithology and geological anomalies during advanced drilling in coal mines: ① Insufficient accuracy in lithology identification: It is challenging to distinguish complex lithologies (such as mudstone, sandstone, carbonaceous mudstone, coal seams, and conglomerate) solely based on MWD curves, especially for thin coal seams, false roofs, and false bottoms, where curve characteristics may be unclear or have multiple interpretations. ② Limited information in geological anomaly identification: Weak interlayer anomalies such as faults, fracture zones, and collapse columns often exhibit multi-dimensional characteristics. For example, faults and fracture zones may be accompanied by drastic fluctuations in drilling pressure and torque, sudden increases in drilling speed, and changes in drill cuttings morphology (such as a large number of small, irregular, or clayey drill cuttings) and mud changes (such as abnormal return volume and turbidity). Existing methods mainly rely on a few parameters in the MWD data, resulting in limited ability to capture these multi-dimensional anomalies and a high risk of misjudgment or omission. ③ Subjectivity and Lag in Identification: Expert judgment relies on manual interpretation of complex MWD curves, which is easily influenced by subjective factors. Furthermore, analysis is often conducted only after a misjudgment has occurred, making early and accurate warnings difficult. ④ Challenges in Heterogeneous Data Fusion: There is a lack of effective methods to deeply fuse abstract, high-dimensional time-series data with intuitive, unstructured drilling site image data, thereby improving the accuracy and robustness of lithology and geological anomaly identification.

[0004] Therefore, there is an urgent need for a method that can deeply integrate drilling time-series data with drilling site image data, and achieve intelligent and high-precision identification of lithology and various geological anomalies (especially weak interlayers such as faults and fracture zones) in coal mine advanced drilling, so as to improve the accuracy and timeliness of geological structure detection and thus ensure mine production safety. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides an intelligent identification method for lithology and geological anomalies in underground pre-drilling in coal mines. By deeply fusing drilling time-series data with drilling site image data through multimodal data fusion, it enables intelligent and high-precision identification of lithology and various geological anomalies in pre-drilling in coal mines, and improves the accuracy and timeliness of geological structure detection.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: an intelligent identification method for lithology and geological anomalies in underground coal mine pre-drilling, comprising the following steps: Step 1: Time series data acquisition and preprocessing during drilling: Time series data of various key parameters during drilling are acquired in real time and continuously through the measurement while drilling (MWD) system, and the acquired data is preprocessed.

[0007] Step 2: On-site image data acquisition and preprocessing at the drilling site: Cameras are deployed at different locations in the underground coal mine drilling site (such as borehole exit, drill cuttings return area, mud pit, etc.) to acquire on-site images or video streams in real time; and each frame of the acquired images or video streams is preprocessed.

[0008] Step 3: High-precision synchronization and alignment of multimodal data: The time series data of each key parameter obtained in Step 1 and the image data obtained in Step 3 are sequentially synchronized with timestamps, aligned with well depth-time, synchronized with key events, and unified with sampling frequency to achieve high-precision synchronization and alignment of multimodal data.

[0009] Step 4: Converting Drilling Time Series Data into Two-Dimensional Images: The time series data of each key parameter after preprocessing in Step 2 are processed by one or more combinations of Gram angle field transformation, Markov transform field transformation, wavelet transform spectrum, and short-time Fourier transform spectrum to convert them into single-channel two-dimensional images corresponding to each key parameter. Finally, these single-channel two-dimensional images are stacked to form a multi-channel two-dimensional image with spatial structure features.

[0010] Step 5: Modal Feature Extraction S1. Time series conversion image feature extraction: The multi-channel two-dimensional image or single-channel two-dimensional image obtained in step four is processed by the first feature extractor to extract the lithology (such as regular texture corresponding to uniform rock layers, and irregular texture corresponding to fracture zones) and geological anomalies (such as periodic abrupt changes in drilling parameters indicating faults, and sudden drops in pump pressure indicating fracture zones) features contained in the current image.

[0011] S2. Drilling site image feature extraction: The drilling site image preprocessed in step two is processed using the second feature extractor to extract the lithological and geological anomaly features contained in the current image.

[0012] Step Six: Multimodal Deep Fusion and Feature Learning: A multimodal fusion model is constructed using deep learning techniques to transform the time-series image features extracted in Step Five (F... TS ) and drilling site image features (F IMG Deep and efficient integration of different modal data is carried out to fully explore the complementary information of different modal data for the identification of lithology and geological anomalies.

[0013] Step 7, Intelligent Identification: First, construct an identification model. Input the feature vector fused in Step 6 into the identification model for analysis and processing, and then output the geological identification results corresponding to the location of the currently collected data, including lithology category and geological anomaly category.

[0014] Furthermore, the key parameters in step one include drilling engineering parameters, formation physical parameters, and well depth / time. Drilling engineering parameters include weight on bit (WOB), rotational speed (RPM), torque (TORQUE), pump pressure (PUMPPRESSURE), drilling speed (ROP), and footage. These parameters reflect the mechanical response of the drill bit's interaction with the formation and are closely related to lithological hardness, structure, and fracturing. Formation physical parameters include natural gamma ray, resistivity, acoustic wave velocity, and density. These parameters directly reflect the physical properties of the formation and are important bases for lithological identification. Well depth / time includes recording the well depth value / timestamp for each data point, serving as the basis for subsequent data synchronization and correlation with geological stratigraphy.

[0015] Furthermore, the preprocessing in step one includes data cleaning, denoising, missing value imputation, and normalization / standardization; the preprocessing in step two includes denoising and enhancement, distortion correction, region of interest (ROI) extraction, and target detection.

[0016] Furthermore, the specific conversion process of Gram angle field conversion in step four is as follows: for the time series data of each key parameter after preprocessing (i.e., drilling pressure, rotation speed, drilling speed, natural gamma), calculate the GASF and GADF matrices respectively within a sliding window of a fixed length to generate multiple grayscale images.

[0017] The specific conversion process of Markov transformation field conversion is as follows: the time series data of each key parameter after preprocessing is converted by MTF within a sliding window to generate the corresponding two-dimensional grayscale image; it can capture the Markov properties of the time series and reflect the frequency and regularity of data transition between different states, which is very effective for identifying periodic patterns and state changes (such as changes in drilling resistance and sudden changes in pump pressure indicating fracture zones).

[0018] The specific conversion process of wavelet transform spectrum or short-time Fourier transform spectrum is as follows: Perform wavelet transform (such as continuous wavelet transform CWT) or short-time Fourier transform on the time series data of each key parameter after preprocessing to obtain the corresponding time spectrum. Use the energy or amplitude spectrum of the time spectrum as a two-dimensional grayscale image to capture the local features of the time series data at different frequency scales, such as high-frequency vibration and drill bit wear, low-frequency changes and formation properties, etc.

[0019] Furthermore, the first feature extractor in step five is obtained by training an image containing spatial features, texture features, and high-level semantic information using a deep convolutional neural network architecture. The second feature extractor is obtained by training an image containing lithology-related features and geological anomaly-related features. It adopts a deep convolutional neural network architecture and is trained on an image containing lithology-related features and geological anomaly-related features.

[0020] Furthermore, step six specifically includes: A. Construct a multi-stream neural network structure: This structure contains at least two independent branches, one of which receives time-series transformed image features F. TS Another branch receives drilling site image features F IMG Each branch, after acquiring its corresponding features, passes those features to the fusion layer.

[0021] B. Feature Interaction and Fusion Strategy: Employ attention-based or gating mechanisms to fuse data from different modalities; I. Fusion based on attention mechanism: Introducing the cross-attention mechanism from the Transformer architecture to allow features from two modalities to interact; F TS and F IMG The query is transformed into query Q, key K, and value V respectively through linear mapping; query Q uses the features of the drilling site image. TS Focus on key K of time series transformed image features IMG Sum V IMG Generate a time-series fusion feature enhanced with image information; use the time-series transformed image feature query Q IMG Key K for focusing on the features of drilling site images TS Sum V TS A time-series enhanced image fusion feature is generated; these two enhanced features are concatenated or summed to obtain the final fusion feature vector F. FUSION .

[0022] II. Gating Mechanism Fusion: Drawing inspiration from the gating concept of Long Short-Term Memory (LSTM) networks, a gating fusion unit is designed to control the inflow, retention, and integration of information from different modalities. Through learnable gating parameters, the contribution weights of different modal features to the final identification are adaptively adjusted. The gating fusion unit is represented as follows: Where g is the gate vector, F FUSION It is a fusion of feature vectors, F TS It is a time-series transformed image feature, F IMG These are the features of the drilling site image, σ is the Sigmoid activation function, and W... g and b g These are learnable weights and biases, where ⊙ represents element-wise multiplication; the gating vector g determines the fusion ratio of the two modal features.

[0023] III. Multi-scale feature fusion: Using step I or II, fusion is performed at multiple deep learning layers, that is, fusion is performed not only at the final feature layer, but also at intermediate layers, to achieve progressive fusion from low-level features to high-level semantics.

[0024] Furthermore, in step seven, the identified model is a KAN (Kolmogorov-Arnold Network), which, according to the Kolmogorov-Arnold representation theorem, represents a multidimensional function as a sum of univariate functions, i.e.: in, These are input features. It's weight. It is a univariate activation function (spline function). It is the activation function of the output layer.

[0025] Furthermore, the lithological categories in step seven include coal seams, mudstone, sandstone, carbonaceous mudstone, argillaceous sandstone, limestone, conglomerate, and limestone; the geological anomaly categories include normal strata (background type), faults, fracture zones, collapse columns, and weak interlayers.

[0026] Compared with existing technologies, this invention first collects and preprocesses drilling time-series data and drilling site image data respectively; then, it performs time stamp synchronization, well depth-time joint alignment, key event synchronization, and sampling frequency unification on the drilling time-series data and drilling site image data in sequence to achieve high-precision synchronization and alignment of multimodal data; then, it converts the drilling time-series data into two-dimensional images using a selected conversion method, thereby forming multi-channel two-dimensional images with spatial structure features; it uses a first feature extractor and a second feature extractor to extract features from the converted time-series images and the drilling site images respectively; it uses deep learning technology to construct a multimodal fusion model to deeply and efficiently fuse the features of the converted time-series images and the features of the drilling site images, so as to fully explore the complementary information of different modal data for lithology and geological anomaly identification; finally, it uses the constructed identification model to analyze and process the fused feature vectors, and outputs geological identification results containing lithology category and geological anomaly category at the corresponding location of the collected data. Through the above process, intelligent and high-precision identification of lithology and various geological anomalies in coal mine advanced drilling can be achieved, and the accuracy and timeliness of geological structure detection can be improved, thereby ensuring the safety of mine production. Attached Figure Description

[0027] Figure 1 This is the overall flowchart of the present invention. Detailed Implementation

[0028] The present invention will be further described below.

[0029] like Figure 1 As shown, the present invention includes the following steps: Step 1: Real-time acquisition and preprocessing of multi-source data: Step 1: Time-Series Data Acquisition and Preprocessing During Drilling: Time-series data of key parameters during drilling are acquired in real-time and continuously using a Measurement While Drilling (MWD) system. Key parameters include drilling engineering parameters, formation physical parameters, and depth / time. Drilling engineering parameters include weight on bit (WOB), rotational speed (RPM), torque (TORQUE), pump pressure (PUMPPRESSURE), drilling speed (ROP), and footage. These parameters reflect the mechanical response of the drill bit's interaction with the formation and are closely related to lithology, hardness, structure, and fracturing. Formation physical parameters include natural gamma ray, resistivity, acoustic wave velocity, and density. These parameters directly reflect the physical properties of the formation and are important bases for lithology identification. The depth / time parameter includes recording the depth value / timestamp of each data point, serving as the basis for subsequent data synchronization and geological stratigraphic correlation. The acquired data undergoes preprocessing, including data cleaning: outliers are identified and removed using a 3σ criterion-based method. For example, if a data point exceeds three times the standard deviation of the mean of its N preceding and following data points, it is considered an outlier.

[0030] Noise reduction: For parameters with large fluctuations such as drilling pressure and torque, a 5-point median filter and a smoothing window of 3 moving average are used to eliminate high-frequency noise.

[0031] Missing value filling: For short-term (less than 5 seconds) missing values ​​caused by signal interruption, linear interpolation is used to fill them.

[0032] Normalization / Standardization: Z-score standardization is used to normalize all parameters to a distribution with a mean of 0 and a variance of 1. The formula is: Where x is the original data, μ is the mean of the data, and σ is the standard deviation of the data.

[0033] Step 2: Drilling Site Image Data Acquisition and Preprocessing: Cameras are deployed at different locations in the coal mine drilling site (such as borehole exit, drill cuttings return area, mud pit, etc.). These cameras are industrial-grade high-definition cameras with underground adaptability such as explosion-proof, dustproof, waterproof, shockproof, low-light imaging, and long-distance transmission. They are used to acquire drilling site images or video streams in real time. Each frame of the acquired images or video streams is preprocessed, including denoising and enhancement: Adaptive histogram equalization (CLAHE) is used to enhance image contrast, and non-local means algorithm is used for denoising.

[0034] Distortion correction: Calibration is completed before camera installation to obtain distortion parameters, and real-time correction is performed when acquiring images.

[0035] Region of Interest (ROI) Extraction and Object Detection: A detector based on a YOLO pre-trained model is used to identify "drill cuttings piles" in borehole exit images and "mud flow" regions in mud pool images, and these regions are then cropped. For drill cuttings piles, image segmentation algorithms (such as U-Net) are further used to extract drill cuttings particles, and their average particle size and shape factor are calculated.

[0036] Step 3: High-precision synchronization and alignment of multimodal data: The time series data of each key parameter obtained in Step 1 and the image data obtained in Step 3 are sequentially synchronized with timestamps, aligned with well depth-time, synchronized with key events, and unified with sampling frequency to achieve high-precision synchronization and alignment of multimodal data.

[0037] Step 2: Converting Drilling Time Series Data into Two-Dimensional Images: The time series data of each key parameter after preprocessing in Step 2 are processed by one or more combinations of Gram angle field transformation, Markov transform field transformation, wavelet transform spectrum, and short-time Fourier transform spectrum, and then converted into single-channel two-dimensional images corresponding to each key parameter.

[0038] This embodiment employs Gram-angle field transformation: the time series data of each key parameter is converted into a GAF image to capture its temporal patterns and correlations.

[0039] a. Parameter selection: WOB (Weight on Drill), ROP (Rate on Drill), Natural Gamma (GR).

[0040] b. Sliding window: A sliding window with a length of 60 seconds (i.e., 60 data points) and a step size of 1 second is used.

[0041] c. GAF Transformation: For WOB, ROP, and GR data within each window, perform 0-1 normalization respectively, and then calculate their GASF and GADF matrices. For example, for the normalized time series... The polar coordinate transformation formula is: GASF matrix elements: GADF matrix elements: Each GASF and GADF matrix is ​​a 60×60 two-dimensional grayscale image.

[0042] Multi-channel combination: The GASF images from WOB, ROP, and GR are stacked into a 3-channel "fused time-series image" (similar to an RGB image). Simultaneously, the GADF images from WOB, ROP, and GR are stacked into another 3-channel image. Finally, these two 3-channel images are used together as input to the time-series modality.

[0043] Step 3: Modal Feature Extraction S1. Time Series Transformation Image Feature Extraction: The two 3-channel "fused time series images" generated in Step 2 are used as input. The first feature extractor processes these images using a SwinTransformer model pre-trained on the ImageNet dataset. The output after analysis is: for each input 3-channel image, the SwinTransformer outputs a 1024-dimensional feature vector. The feature vectors of the two images are concatenated to form a 2048-dimensional feature vector F. TS This allows for the extraction of lithological features (e.g., regular textures correspond to uniform rock layers, while irregular textures correspond to fracture zones) and geological anomalies (e.g., periodic abrupt changes in drilling parameters indicate faults, and sudden drops in pump pressure indicate fracture zones) contained in the current image.

[0044] S2. Drilling Site Image Feature Extraction: The preprocessed drilling site images from Step 1 are used as input and processed by a second feature extractor. The second feature extractor uses a VisionTransformer (ViT) model pre-trained on the ImageNet dataset. Output: For each input image, the ViT model outputs a 768-dimensional feature vector. The feature vectors of the two images are concatenated to form a 1536-dimensional feature vector F. IMG This allows for the extraction of lithological and geological anomaly features contained in the current image.

[0045] Step 4: Multimodal Deep Fusion and Feature Learning: A multimodal fusion model is constructed using deep learning techniques to transform the time-series image features extracted in Step 5 (F... TS ) and drilling site image features (F IMG Deep and efficient fusion will be carried out to fully explore the complementary information of different modal data for the identification of lithology and geological anomalies, specifically: A. Construct a multi-stream neural network structure: This structure contains at least two independent branches, one of which receives time-series transformed image features F. TS Another branch receives drilling site image features F IMG Each branch, after acquiring its corresponding features, passes those features to the fusion layer.

[0046] B. Feature Interaction and Fusion Strategy: Employ attention-based or gating mechanisms to fuse data from different modalities; I. Fusion based on attention mechanism: Introducing the cross-attention mechanism from the Transformer architecture to allow features from two modalities to interact; F TS and F IMG The query is transformed into query Q, key K, and value V respectively through linear mapping; query Q uses the features of the drilling site image. TS Focus on key K of time series transformed image features IMG Sum V IMG Generate a time-series fusion feature enhanced with image information; use the time-series transformed image feature query Q IMG Key K for focusing on the features of drilling site images TS Sum V TS This generates an image fusion feature enhanced with time-series information.

[0047] Calculate F TS For F IMG Attention enhancement features: Calculate F IMG For F TSAttention enhancement features: Where Q, K, V are the representations of the original features after linear transformation, and dk is the dimension of the Key vector.

[0048] Finally, the interactive and enhanced features are combined (e.g., concatenated or summed) and passed through a fusion layer to form a comprehensive fused feature vector F. FUSION This approach enables the model to adaptively integrate complementary information from different modalities, effectively capturing the multidimensional characteristics of lithology and geological anomalies.

[0049] II. Multi-scale feature fusion: Using step I, fusion is performed at multiple deep learning layers, that is, fusion is performed not only at the final feature layer, but also at intermediate layers, to achieve progressive fusion from low-level features to high-level semantics.

[0050] Step 5: Intelligent Identification: First, construct an identification model, which is a KAN (Kolmogorov-Arnold Network). Based on the Kolmogorov-Arnold representation theorem, it represents multidimensional functions as the sum of univariate functions, i.e.: in, These are input features. It's weight. It is a univariate activation function (spline function). It is the activation function of the output layer.

[0051] After the feature vector fused in step six is ​​input into the identification model for analysis and processing, the geological identification results corresponding to the current data collection location are output, including lithology category and geological anomaly category; among which, lithology category includes coal seam, mudstone, sandstone, carbonaceous mudstone, argillaceous sandstone, limestone, conglomerate, and limestone; geological anomaly category includes normal strata (background type), fault, fracture zone, collapse column, and weak interlayer.

[0052] After obtaining the identification result in this embodiment, it can be used in the following ways: a. Geological Anomaly Early Warning: When the identification results show the presence of a high-risk geological anomaly category (such as "fault", "fracture zone", "collapse column", "weak interlayer") and its confidence level exceeds the preset threshold, the system can trigger an early warning mechanism to promptly notify relevant personnel.

[0053] b. Drilling parameter optimization: Based on the identified lithology or anomaly type, drilling parameters (such as drilling pressure, rotation speed, pump pressure, etc.) can be intelligently adjusted to improve drilling efficiency and safety.

[0054] c. Geological data update: The identification results can be used to update geological profile maps and geological structure models in real time, improving the precision of geological exploration.

[0055] d. Visualization: The identification results can be visualized on the human-computer interaction interface in conjunction with the original data and multimodal transformation data, which facilitates expert analysis and decision-making.

[0056] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent identification of lithology and geological anomalies in underground coal mine advance drilling, characterized in that, Includes the following steps: Step 1: Time series data acquisition and preprocessing during drilling: Time series data of various key parameters during drilling are acquired in real time and continuously through the measurement while drilling system, and the acquired data is preprocessed. Step 2: Drilling Site Image Data Acquisition and Preprocessing: Cameras are deployed at different locations in the underground coal mine drilling site to acquire real-time images or video streams of the drilling site; and each frame of the acquired images or video streams is preprocessed. Step 3, High-precision synchronization and alignment of multimodal data: The time series data of each key parameter obtained in Step 1 and the image data obtained in Step 3 are sequentially synchronized with timestamps, aligned with well depth-time, synchronized with key events, and unified with sampling frequency to achieve high-precision synchronization and alignment of multimodal data; Step 4: Convert the time series data of each key parameter after drilling into a two-dimensional image: After preprocessing in Step 2, the time series data of each key parameter is processed by one or more combinations of Gram angle field transformation, Markov transformation field transformation, wavelet transform spectrum, and short-time Fourier transform spectrum, and then converted into a single-channel two-dimensional image corresponding to each key parameter. Finally, these single-channel two-dimensional images are stacked to form a multi-channel two-dimensional image with spatial structure features. Step 5: Modal Feature Extraction S1. Time series transformation image feature extraction: The multi-channel two-dimensional image or single-channel two-dimensional image obtained in step four is processed by the first feature extractor to extract the lithological and geological anomaly features contained in the current image. S2. Feature extraction of drilling site images: The drilling site images preprocessed in step two are processed using the second feature extractor to extract the lithological and geological anomaly features contained in the current images. Step Six: Multimodal Deep Fusion and Feature Learning: A multimodal fusion model is constructed using deep learning technology to deeply and efficiently fuse the time-series transformed image features extracted in Step Five with the drilling site image features, so as to fully explore the complementary information of different modal data for the identification of lithology and geological anomalies. Step 7, Intelligent Identification: First, construct an identification model. Input the feature vector fused in Step 6 into the identification model for analysis and processing, and then output the geological identification results corresponding to the location of the currently collected data, including lithology category and geological anomaly category.

2. The intelligent identification method for lithology and geological anomalies in underground coal mine advance drilling according to claim 1, characterized in that, The key parameters in step one include drilling engineering parameters, formation physical parameters, and well depth / time; among which, drilling engineering parameters include drilling pressure, rotation speed, torque, pump pressure, drilling speed, and footage; formation physical parameters include natural gamma, resistivity, acoustic wave, and density; well depth / time includes recording the well depth value / timestamp for each data point.

3. The intelligent identification method for lithology and geological anomalies in underground coal mine advance drilling according to claim 1, characterized in that, The preprocessing in step one includes data cleaning, denoising, missing value imputation, and normalization / standardization; the preprocessing in step two includes denoising and enhancement, distortion correction, region of interest extraction, and target detection.

4. The intelligent identification method for lithology and geological anomalies in underground coal mine advance drilling according to claim 1, characterized in that, The specific conversion process of Gram angle field conversion in step four is as follows: For the time series data of each key parameter after preprocessing, calculate the GASF and GADF matrices respectively within a sliding window of fixed length to generate multiple grayscale images; The specific conversion process of Markov transform field conversion is as follows: the time series data of each key parameter after preprocessing is converted by MTF within a sliding window to generate the corresponding two-dimensional grayscale image; The specific conversion process of wavelet transform spectrum or short-time Fourier transform spectrum is as follows: perform wavelet transform or short-time Fourier transform on the time series data of each key parameter after preprocessing to obtain the corresponding time spectrum. Use the energy or amplitude spectrum of the time spectrum as a two-dimensional grayscale image to capture the local features of the time series data at different frequency scales.

5. The intelligent identification method for lithology and geological anomalies in underground coal mine advance drilling according to claim 1, characterized in that, The first feature extractor in step five is obtained by training an image containing spatial features, texture features, and high-level semantic information using a deep convolutional neural network architecture. The second feature extractor is obtained by training an image containing lithology-related features and geological anomaly-related features. It adopts a deep convolutional neural network architecture and is trained on an image containing lithology-related features and geological anomaly-related features.

6. The intelligent identification method for lithology and geological anomalies in underground coal mine advance drilling according to claim 1, characterized in that, Step six specifically involves: A. Construct a multi-stream neural network structure: This structure contains at least two independent branches, one of which receives time-series transformed image features F. TS Another branch receives drilling site image features F IMG Each branch, after acquiring its corresponding features, passes those features to the fusion layer. B. Feature Interaction and Fusion Strategy: Employ attention-based or gating mechanisms to fuse data from different modalities; I. Fusion based on attention mechanism: Introducing the cross-attention mechanism from the Transformer architecture to allow features from two modalities to interact; F TS and F IMG The query is transformed into query Q, key K, and value V respectively through linear mapping; query Q uses the features of the drilling site image. TS Focus on key K of time series transformed image features IMG Sum V IMG This generates a time-series fusion feature enhanced with image information; Query Q using time series transformation of image features IMG Key K for focusing on the features of drilling site images TS Sum V TS This generates an image fusion feature enhanced with time-series information; The two enhanced features are concatenated or summed to obtain the final fused feature vector F. FUSION ; II. Gating Mechanism Fusion: A gating fusion unit is designed to control the inflow, retention, and integration of information from different modalities; through learnable gating parameters, the contribution weights of different modal features to the final identification are adaptively adjusted; the gating fusion unit is represented as follows: Where g is the gate vector, F FUSION It is a fusion of feature vectors, F TS It is a time-series transformed image feature, F IMG These are the features of the drilling site image, σ is the Sigmoid activation function, and W... g and b g These are learnable weights and biases, where ⊙ represents element-wise multiplication; the gating vector g determines the fusion ratio of the two modal features. III. Multi-scale feature fusion: Using step I or II, fusion is performed at multiple deep learning layers, that is, fusion is performed not only at the final feature layer, but also at intermediate layers, to achieve progressive fusion from low-level features to high-level semantics.

7. The intelligent identification method for lithology and geological anomalies in underground coal mine advance drilling according to claim 1, characterized in that, In step seven, the identified model is a KAN network, which, according to the Kolmogorov-Arnold representation theorem, represents a multidimensional function as the sum of univariate functions, i.e.: in, These are input features. It's weight. It is a univariate activation function (spline function). It is the activation function of the output layer.

8. The intelligent identification method for lithology and geological anomalies in underground coal mine advance drilling according to claim 1, characterized in that, The lithological categories in step seven include coal seams, mudstone, sandstone, carbonaceous mudstone, argillaceous sandstone, limestone, conglomerate, and limestone; the geological anomaly categories include normal strata, faults, fracture zones, collapse columns, and weak interlayers.

Images