Drill hole quality and stratum characteristic inversion method and device based on dynamics and rock debris form and medium

Abstract

The invention discloses a drilling quality and stratum characteristic inversion method and device based on dynamics and rock debris morphology and a medium, and relates to the field of artificial intelligence, and the method comprises the steps: synchronously collecting a drilling dynamics signal and a rock debris image sequence in a drilling process; extracting a drilling dynamics high-frequency characteristic spectrum based on the drilling dynamics signal; extracting a rock debris morphological parameter sequence based on the rock debris image sequence; inputting the drilling dynamics high-frequency characteristic spectrum and the rock debris morphological parameter sequence which are aligned in time into a pre-trained multi-task deep learning model; and a hole wall quality evaluation result and a stratum characteristic identification result synchronously generated by the multi-task deep learning model are obtained, the hole wall quality evaluation result is used for representing the hole wall state, and the stratum characteristic identification result is used for representing the stratum attribute. According to the method, the drilling quality and the stratum characteristics can be inversed in real time based on deep fusion and synchronization of the multi-source while-drilling information.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, equipment, and medium for inverting borehole quality and formation characteristics based on dynamics and rock cuttings morphology. Background Technology

[0002] In drilling operations in fields such as geotechnical engineering and geological exploration, obtaining real-time information on borehole quality and formation characteristics is crucial for guiding construction and ensuring project safety and efficiency.

[0003] In existing technologies, borehole wall quality assessment typically relies on post-drilling methods such as imaging or physical probing, which cannot achieve in-situ, real-time perception and evaluation during drilling. This results in the inability to promptly detect defects such as borehole wall roughness, fractures, or breakage. Furthermore, drilling cuttings, as the most direct reflection of the formation, contain rich information on lithology, mechanical properties, and fracturing mechanisms through their morphological characteristics (such as grain size, angularity, shape, and three-dimensional structure). However, existing technologies mainly rely on two-dimensional images for simple qualitative lithological identification, lacking precise quantitative analysis methods for the three-dimensional morphology of drilling cuttings. Moreover, existing drilling formation identification methods are mostly based on inversion using single-dimensional drilling dynamic parameters such as drilling rate and torque, resulting in limited model identification accuracy. Summary of the Invention

[0004] This invention provides a method, equipment, and medium for inverting borehole quality and formation characteristics based on dynamics and cuttings morphology. The technical problem it aims to solve is: how to provide a method that can synchronously and in real-time invert borehole quality and formation characteristics based on deep fusion of multi-source drilling information.

[0005] In a first aspect, embodiments of the present invention provide a method for inverting borehole quality and formation characteristics based on dynamics and cuttings morphology, including: Simultaneously acquire drilling dynamics signals and rock cuttings image sequences during the drilling process; High-frequency feature spectrum of drilling dynamics is extracted based on the drilling dynamics signal; Extract the rock debris morphology parameter sequence based on the rock debris image sequence; The time-aligned high-frequency feature spectrum of drilling dynamics and the sequence of cuttings morphology parameters are input into a pre-trained multi-task deep learning model. The borehole wall quality assessment results and formation characteristic identification results generated synchronously by the multi-task deep learning model are obtained. The borehole wall quality assessment results are used to characterize the borehole wall state, and the formation characteristic identification results are used to characterize the formation properties.

[0006] Optionally, the synchronous acquisition of drilling dynamics signals and cuttings image sequences during the drilling process includes: The acoustic emission source in the control drill assembly emits synchronous trigger pulse signals at a fixed period. The drilling dynamics signal acquisition device is equipped with a first acoustic receiving sensor, and the cuttings image acquisition device is equipped with a second acoustic receiving sensor. The first acoustic receiving sensor and the second acoustic receiving sensor receive the synchronous trigger pulse signals. The drilling dynamics signal acquisition device controls the first sampling point time when the first acoustic receiving sensor receives the synchronous trigger pulse signal, and calibrates it as the zero time reference of the drilling dynamics signal time axis; The control rock cuttings image acquisition device calibrates the first sampling point time when the second acoustic receiving sensor receives the synchronous trigger pulse signal as the zero time reference of the rock cuttings image time axis; The drilling dynamics signal acquisition device and the cuttings image acquisition device are based on a unified zero-time reference and perform data acquisition with equal time windows.

[0007] Optionally, the step of extracting the high-frequency feature spectrum of drilling dynamics based on the drilling dynamics signal includes: The triaxial acceleration signal and acoustic emission signal are obtained from the drilling dynamics signal; Cross-correlation calculations are performed on the three channels of the triaxial acceleration signal, and an enhanced signal beam is synthesized based on the maximum coherence. The enhanced signal beam points in the direction of contact between the drill bit and the borehole wall. Perform a short-time Fourier transform on the acoustic emission signal to generate the acoustic emission time spectrum; Calculate the number of newly appearing rock debris outlines in a preset unit time in a sequence of rock debris images; When the number of newly appearing rock fragment outlines exceeds a preset threshold within a unit of time, a rock breaking event is determined to have occurred, and the trigger time of the rock breaking event is recorded. Based on the triggering time, a first corresponding time segment of a preset duration is extracted from the enhanced signal beam, and a second corresponding time segment of a preset duration is extracted from the acoustic emission time spectrum; Calculate the total energy of the first signal, the mean of the first instantaneous frequency, the variance of the first instantaneous frequency, and the first spectral entropy within the first corresponding time segment; Calculate the total energy of the second signal, the mean of the second instantaneous frequency, the variance of the second instantaneous frequency, and the second spectral entropy within the second corresponding time segment; The total energy of the first signal, the mean of the first instantaneous frequency, the variance of the first instantaneous frequency, the first spectral entropy, the total energy of the second signal, the mean of the second instantaneous frequency, the variance of the second instantaneous frequency, and the second spectral entropy are combined to form the high-frequency characteristic spectrum of drilling dynamics.

[0008] Optionally, the rock debris image sequence includes consecutive frame images of the rock debris flow acquired using a binocular vision system, and the extraction of a rock debris morphology parameter sequence based on the rock debris image sequence includes: Based on the continuous frame images, a three-dimensional point cloud model of the rock debris flow is obtained through stereo matching and three-dimensional reconstruction. The 3D point cloud model is segmented to obtain multiple point cloud clusters, where each point cloud cluster represents a rock fragment. For each point cloud cluster, calculate the ratio of the surface area to the volume of the point cloud cluster, and use the ratio as the shape factor of the rock debris corresponding to the point cloud cluster; For each point cloud cluster, calculate the standard deviation of the distribution of normal vectors of all points on the surface of the point cloud cluster, and use the standard deviation as the surface roughness index of the rock debris corresponding to the point cloud cluster. For each point cloud cluster, calculate the average length-to-width ratio and length-to-height ratio of the smallest bounding cuboid of the point cloud cluster, and use the average value as the flatness index of the rock debris corresponding to the point cloud cluster. For the current time window, the mean, standard deviation, and distribution histogram information of the shape factor, surface roughness index, and flatness index of all rock fragments within the current time window are statistically analyzed to obtain the sequence of rock fragment morphology parameters for the current time window.

[0009] Optionally, the step of extracting the rock debris morphology parameter sequence based on the rock debris image sequence further includes: For each point cloud cluster, identify the fracture surface region in the point cloud cluster and extract the three-dimensional contour of the fracture surface region; Calculate the three-dimensional fractal dimension of the three-dimensional profile of the fracture surface region; Based on the estimated area of ​​the fracture surface region, the volume of the rock fragments corresponding to the point cloud cluster, the three-dimensional fractal dimension, and the pre-calibrated energy conversion coefficient, the equivalent compressive strength factor of the rock fragments corresponding to the point cloud cluster is calculated. For the current time window, calculate the average and variance of the equivalent compressive strength factor of all rock cuttings and add them to the rock cutting morphology parameter sequence for the current time window.

[0010] Optionally, the multi-task deep learning model includes a common feature encoder, a gated shared unit, a borehole wall quality assessment branch, and a formation characteristic identification branch; The common feature encoder performs primary fusion of the spliced ​​high-frequency feature spectrum of drilling dynamics and the sequence of rock cutting morphology parameters to obtain primary fused features; The gated shared unit contains a multi-head self-attention network. The multi-head self-attention network takes primary fusion features as input and outputs a first feature selection weight matrix and a second feature selection weight matrix. The first feature selection weight matrix corresponds to the borehole wall quality assessment branch, and the second feature selection weight matrix corresponds to the formation characteristic identification branch. The input features of the pore wall quality assessment branch are the element-wise product of the primary fusion features and the first feature selection weight matrix; The input features of the stratigraphic characteristic identification branch are the element-wise product of the primary fusion features and the second feature selection weight matrix.

[0011] Optionally, the steps for generating the borehole wall quality assessment results in the borehole wall quality assessment branch include: The input features of the hole wall quality assessment branch are simultaneously fed into the first one-dimensional convolutional path, the second one-dimensional convolutional path, and the third one-dimensional convolutional path. The first one-dimensional convolutional path, the second one-dimensional convolutional path, and the third one-dimensional convolutional path have different time scales, and the first scale features, the second scale features, and the third scale features are extracted respectively. The first scale feature, the second scale feature and the third scale feature are concatenated and calculated through the self-attention layer to obtain the first contribution weight corresponding to the first scale feature, the second contribution weight corresponding to the second scale feature and the third contribution weight corresponding to the third scale feature. Based on the first contribution weight, the second contribution weight, and the third contribution weight, the first-scale feature, the second-scale feature, and the third-scale feature are weighted and fused to obtain the fused feature. Based on the fusion features, the power spectral density integral value of the surface undulation of the hole wall is calculated through the first output submodule, where the power spectral density integral value is used as a roughness quantification index. Based on the fusion features, the coefficient of variation of the borehole diameter along the depth direction is calculated through the second output submodule, where the coefficient of variation serves as an indicator of borehole diameter stability. Based on the fusion features, the probability value of hole wall instability risk is calculated through the third output submodule. The probability value of hole wall instability risk is calculated based on the vibration energy dissipation rate and ranges from 0 to 1.

[0012] Optionally, the steps for generating stratigraphic characteristic identification results in the stratigraphic characteristic identification branch include: The input features of the formation characteristic identification branch are concatenated with the current drilling condition parameter vector and transformed through the fully connected layer to obtain the basic features. The drilling condition parameter vector includes drilling pressure, rotation speed and displacement. The basic features are transformed using a learnable working condition compensation matrix to obtain compensated features. The transformation operation of the working condition compensation matrix is ​​combined with the drilling working condition parameter vector. The compensated features are used to eliminate the influence of working condition changes. Based on the compensated features, the classification probability distribution of stratigraphic lithology is generated through the lithology classification output head; Based on the compensated features, a predicted value of dynamic rock hardness is generated through a hardness regression output head. Based on the compensated characteristics, a normalized index of the critical confining pressure for the brittle-plastic transition of rocks is generated through the critical index output head.

[0013] Optionally, before inputting the time-aligned high-frequency feature spectrum of drilling dynamics and the sequence of cuttings morphology parameters into a pre-trained multi-task deep learning model, a feature adaptive fusion step is also included: Real-time analysis of the variance of drilling pressure and torque in drilling dynamics signals within the sliding time window; If the variance is lower than the first preset threshold, the current state is determined to be a steady-state drilling state; If it is a steady-state drilling state, the high-frequency feature spectrum of drilling dynamics in the current time window is directly spliced ​​with the sequence of cuttings morphology parameters as the final input feature; If the variance is not lower than the first preset threshold, then the current drilling state is determined to be abnormal. If the drilling is in an abnormal state, the high-frequency characteristic spectrum of drilling dynamics and the sequence of cuttings morphology parameters from multiple consecutive time windows in the past are input into the recurrent neural network in chronological order. Historical context feature vectors are extracted using a recurrent neural network; The historical context feature vector is concatenated with the high-frequency feature spectrum of drilling dynamics and the sequence of cuttings morphology parameters of the current time window to serve as the final input feature.

[0014] Optionally, the steps to obtain a pre-trained multi-task deep learning model include: The multi-task deep learning model is initially trained on a source domain dataset containing various geological conditions to obtain the basic model; When drilling begins at a new target site, drilling dynamics signals and rock cuttings image sequences are collected within the initial footage at the target site, and local verification is performed to obtain verification results. Based on the collected drilling dynamics signals and rock cuttings image sequences within the initial advance of the target site and the verification results, a target domain dataset is constructed. Add a domain classifier after the common feature encoder of the base model; Perform adversarial training, which involves alternatingly repeating the following two optimization phases until the model converges: In the first stage, the parameters of the common feature encoder, the borehole wall quality assessment branch, and the formation characteristic identification branch are fixed, and the parameters of the domain classifier are updated. The source domain dataset and the target domain dataset are respectively input into the common feature encoder to obtain the source domain features and the target domain features. The domain classifier is trained using the source domain features, the target domain features, and the corresponding domain labels so that the domain classifier can distinguish whether the features come from the source domain dataset or the target domain dataset. In the second stage, the parameters of the fixed domain classifier are updated, and the parameters of the common feature encoder are updated. The source domain dataset and the target domain dataset are input into the common feature encoder, and the domain confusion loss is calculated through the domain classifier. The domain confusion loss is optimized through the backpropagation algorithm to update the parameters of the common feature encoder, so that the features extracted by the common feature encoder are difficult for the domain classifier to distinguish their source. During adversarial training, the parameters of the borehole wall quality assessment branch and the formation characteristic identification branch are updated using the drilling dynamics signals, cuttings image sequences and corresponding local verification results in the target domain dataset.

[0015] Optionally, the steps of training a multi-task deep learning model may also include introducing task orthogonality constraints: Convert the first feature selection weight matrix output by the gated shared unit into a first vector; Convert the second feature selection weight matrix output by the gated shared unit into a second vector; Calculate the cosine similarity between the first vector and the second vector. The cosine similarity constitutes the task orthogonality constraint term. Construct a total loss function, which includes the first loss term corresponding to the borehole wall quality assessment branch, the second loss term corresponding to the formation characteristic identification branch, and the task orthogonality constraint term. The multi-task deep learning model is trained by minimizing the total loss function, so that the first feature selection weight matrix and the second feature selection weight matrix are orthogonal in the vector space.

[0016] Optionally, the method further includes: The real-time output of borehole wall quality assessment results and formation characteristic identification results are rendered into a three-dimensional geological model associated with the borehole trajectory for dynamic visualization. It provides a human-computer interaction interface, which receives correction annotations from the operator on the inversion results. The correction annotations include confirmation, rejection or re-delineation of suspected defective segments or stratigraphic interfaces. The correction annotation is combined with the high-frequency characteristic spectrum of drilling dynamics and the sequence of rock cutting morphology parameters at the corresponding time of the correction annotation to form an online learning sample. Collect multiple online learning samples; Using an incremental learning approach, the parameters of the last two fully connected layers in the multi-task deep learning model are periodically fine-tuned using the online learning samples.

[0017] Secondly, embodiments of the present invention also provide a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0018] Thirdly, embodiments of the present invention also provide a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the above-described method.

[0019] This invention provides a method, equipment, and medium for inverting borehole quality and formation characteristics based on dynamics and cuttings morphology. The method includes: simultaneously acquiring drilling dynamics signals and cuttings image sequences during the drilling process; extracting high-frequency feature spectra of drilling dynamics based on the drilling dynamics signals; extracting cuttings morphology parameter sequences based on the cuttings image sequences; inputting the time-aligned high-frequency feature spectra of drilling dynamics and the cuttings morphology parameter sequences into a pre-trained multi-task deep learning model; and obtaining borehole wall quality assessment results and formation characteristic identification results synchronously generated by the multi-task deep learning model. The borehole wall quality assessment results characterize the borehole wall state, and the formation characteristic identification results characterize formation properties. This invention, by simultaneously acquiring drilling dynamics signals and cuttings image sequences and fusing and analyzing these two types of information based on a multi-task deep learning model, achieves real-time, in-situ quantitative assessment of borehole wall roughness, integrity, and potential defects during the drilling process, filling the technological gap in borehole wall quality perception. Furthermore, rock cuttings are transformed from qualitative identification materials into a quantitative sequence of morphological parameters and deeply integrated with dynamic signals, greatly mining and utilizing the information on formation mechanics and fracturing mechanisms contained in the rock cuttings. Moreover, the multi-task model employed can simultaneously complete borehole wall condition assessment and formation characteristic identification, overcoming the limitations of traditional single inversion models in terms of limited accuracy and task fragmentation. This allows for the efficient and collaborative acquisition of two types of key information in a single drilling operation, significantly improving the informatization level and decision support capabilities of the drilling process. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A flowchart illustrating a method for inverting borehole quality and formation characteristics based on dynamics and cuttings morphology, provided for an embodiment of the present invention; Figure 2 This is a schematic block diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0024] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0025] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0026] As used in this specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]."

[0027] Please see Figure 1 This invention provides a method for inverting borehole quality and formation characteristics based on dynamics and cuttings morphology. The method includes the following steps: S1, synchronously acquires drilling dynamics signals and rock cuttings image sequences during the drilling process.

[0028] In practice, drilling dynamics signals are acquired using various sensors mounted on the drill string system. Specifically, a drill pressure sensor is typically mounted on the top of the drill pipe to measure the axial force applied to the drill bit; a torque sensor is mounted in the rotary drive system or top drive to measure rotational resistance; a triaxial accelerometer is mounted near the drill bit or on the drill collar to measure the vibrational acceleration of the drill string in three orthogonal directions; and an acoustic emission sensor is also mounted on the drill collar to capture high-frequency elastic waves generated when the rock fractures. These sensors continuously acquire data at a uniform, high sampling frequency (e.g., 10 Hz).

[0029] Furthermore, the acquisition of the cuttings image sequence is accomplished using a binocular vision system (e.g., a pair of calibrated high-speed industrial cameras) installed at the vibrating screen outlet, drilling fluid return pipeline, or dedicated cuttings sampling device. The camera captures the upward flow of cuttings carried by the drilling fluid at a fixed frame rate, generating a continuous two-dimensional image sequence. To achieve time synchronization, the two acquisition systems can be connected to the same high-precision clock source, or each data sample can be timestamped with a precise timestamp based on the same reference during data recording.

[0030] In some preferred embodiments, the synchronous acquisition of drilling dynamics signals and cuttings image sequences during the drilling process includes: controlling the acoustic emission source in the drill string assembly to emit synchronous trigger pulse signals at a fixed period, wherein the drilling dynamics signal acquisition device is equipped with a first acoustic receiving sensor, and the cuttings image acquisition device is equipped with a second acoustic receiving sensor, and the first and second acoustic receiving sensors receive the synchronous trigger pulse signals; controlling the drilling dynamics signal acquisition device to calibrate the time of the first sampling point of the synchronous trigger pulse signal received by the first acoustic receiving sensor as the zero time reference of the drilling dynamics signal time axis; controlling the cuttings image acquisition device to calibrate the time of the first sampling point of the synchronous trigger pulse signal received by the second acoustic receiving sensor as the zero time reference of the cuttings image time axis; and controlling the drilling dynamics signal acquisition device and the cuttings image acquisition device to perform data acquisition with equal-length windows based on a unified zero time reference.

[0031] In practice, the step of controlling the acoustic emission source in the drill string assembly to emit synchronous trigger pulse signals at fixed intervals is accomplished by a dedicated module installed in the drill collar or near-bit sub. This module contains a high-temperature, high-pressure resistant piezoelectric ceramic transducer, a control circuit, and a battery. The control circuit drives the transducer to generate a brief, specific-frequency acoustic pulse at a predetermined, precise time interval.

[0032] Furthermore, the first acoustic receiving sensor in the drilling dynamics signal acquisition device is typically mounted on the drill table or ground riser, with good mechanical coupling to the drill string, and is used to receive acoustic synchronization pulses transmitted through the drill string. The second acoustic receiving sensor in the cuttings image acquisition device is mounted on the camera housing, exposed to the air, and is used to receive the same pulse signal propagating through the air medium. Both receiving sensors include signal conditioning circuitry to amplify and filter the received analog signals, converting them into level signals recognizable by the digital acquisition card.

[0033] Furthermore, the drilling dynamics signal acquisition device calibrates the moment when the first sampling point of the pulse signal received by the first acoustic receiving sensor is used as the zero-time reference for the drilling dynamics signal time axis. Specifically, the data acquisition card of the drilling dynamics signal acquisition system continuously monitors the input voltage of the first acoustic receiving sensor channel. When the voltage value exceeds a set trigger threshold, the acquisition card immediately records the value of its internal high-precision counter and marks this event as the time origin T0. Thereafter, the timestamps of all synchronously acquired sampling data from channels such as drill pressure, torque, acceleration, and acoustic emission are calculated based on T0; for example, the timestamp of the nth sampling point is T0 + n * sampling interval.

[0034] Furthermore, the rock cuttings image acquisition device performs similar operations. The image acquisition card or smart camera monitors the input of the second acoustic receiving sensor, and when a pulse trigger signal is detected, it records the current time as the zero point T0' of the image timeline. Each subsequent frame of the image is assigned a precise timestamp relative to T0'.

[0035] Furthermore, the two acquisition devices are controlled to acquire data within equal-length windows based on a unified zero-time reference. This means that after system initialization and successful zero-time calibration, the two systems segment and package the data according to the same time window length, for example, every 1000 milliseconds. The dynamics system packages all sampling point data acquired within each 1000 millisecond into a data block, while the imaging system packages all image frames captured within this 1000 millisecond period into an image sequence block. Because T0 and T0' of the two systems correspond to the arrival event of the same physical pulse, their generated k-th data block and k-th image block are strictly aligned in time, jointly representing the same 1000-millisecond drilling process downhole.

[0036] This embodiment employs active acoustic transmission and reception to achieve hardware-level synchronization of multi-source data, effectively solving the time inaccuracy problem caused by clock drift and software latency in distributed acquisition systems. In the complex electromagnetic environment of drilling sites, relying on network time synchronization or software time synchronization often fails to achieve sub-millisecond synchronization accuracy. This invention creates a common physical time reference signal propagating within the drilling system by installing an acoustic emission source on the drill string. The drilling dynamics acquisition system and the cuttings image acquisition system capture this reference signal through their respective independent sensors that respond to the same physical event, using the instant of signal capture as the starting point of the data stream. Since the propagation speed of sound waves in steel drill strings and air is known and stable, the pulse arrival time difference between the two receiving sensors due to the different media is a fixed system delay. During initial system calibration, this delay value τ is determined by measurement or calculation. During time axis calibration, the trigger time T0' recorded by the cuttings image acquisition device is subtracted (or added) from (or added to) the delay τ to align it with the zero-time reference T0 of the drilling dynamics signal acquisition device, thereby achieving precise time synchronization between the two systems based on the same physical event. Subsequently, data is collected in equal-length windows based on a unified zero-time reference. This method ensures that the uniformity of the time reference originates from a common physical event, rather than software coordination, thus possessing high reliability and anti-interference capability.

[0037] S2, extract the high-frequency feature spectrum of drilling dynamics based on the drilling dynamics signal.

[0038] In practice, the acquired raw drilling dynamics signals are processed. First, preprocessing is performed, including removing power frequency interference and sensor noise, and standardizing the signals. Further, time-frequency analysis algorithms are used. For example, short-time Fourier transforms are applied to the vibration acceleration signal and the acoustic emission signal respectively. This transform segments the signal on the time axis, and a Fourier transform is performed on each segment to obtain the distribution of signal energy in a two-dimensional plane of time and frequency, i.e., the time spectrum. Further, features within a preset high-frequency band are extracted from these time spectra, such as calculating the energy integral, spectral centroid frequency, and spectral width of a specific band within a sliding time window. These multiple high-frequency features extracted from different sensor signals are combined into a comprehensive feature vector, which is the high-frequency feature spectrum of the drilling dynamics.

[0039] In some preferred embodiments, the extraction of high-frequency feature spectrum of drilling dynamics based on the drilling dynamics signal includes: obtaining triaxial acceleration signal and acoustic emission signal from the drilling dynamics signal; performing cross-correlation calculation on the three channels of the triaxial acceleration signal, and synthesizing an enhanced signal beam based on maximum coherence, the enhanced signal beam pointing in the direction of contact between the drill bit and the borehole wall; performing short-time Fourier transform on the acoustic emission signal to generate an acoustic emission time spectrum; calculating the number of newly appearing rock cuttings contours in the rock cuttings image sequence within a preset unit time; when the number of newly appearing rock cuttings contours within a unit time exceeds a preset threshold, determining that a rock breaking event has occurred, and recording the trigger time of the rock breaking event; based on At the triggering time, a first corresponding time segment of a preset duration is extracted from the enhanced signal beam, and a second corresponding time segment of a preset duration is extracted from the acoustic emission time spectrum; the total energy of the first signal, the mean of the first instantaneous frequency, the variance of the first instantaneous frequency, and the first spectral entropy within the first corresponding time segment are calculated; the total energy of the second signal, the mean of the second instantaneous frequency, the variance of the second instantaneous frequency, and the second spectral entropy within the second corresponding time segment are calculated; the total energy of the first signal, the mean of the first instantaneous frequency, the variance of the first instantaneous frequency, the first spectral entropy, the total energy of the second signal, the mean of the second instantaneous frequency, the variance of the second instantaneous frequency, and the second spectral entropy are combined to form a high-frequency characteristic spectrum of drilling dynamics.

[0040] In practice, firstly, the three-channel signals from the triaxial accelerometer and the single-channel signal from the acoustic emission sensor are separated from the continuous drilling dynamics signal data stream. These signals are digital sequences that have undergone preliminary preprocessing.

[0041] Furthermore, cross-correlation calculations are performed on the three channels of the triaxial acceleration signal. Specifically, one channel's signal is selected as a reference, such as the Z-axis signal along the drill string axis. The X-axis and Y-axis signals are then cross-correlated with the Z-axis signal, respectively. The cross-correlation function measures the similarity between two signals at different time offsets. The cross-correlation coefficients between the X-axis and Z-axis signals, and between the Y-axis and Z-axis signals, are calculated at multiple possible time offsets. The time offset that maximizes the cross-correlation coefficient is found; this offset reflects the time lag between vibration signals in different directions. Based on the time delay corresponding to this maximum coherence, time alignment compensation is performed on the X-axis and Y-axis signals. Then, the aligned three channel signals are superimposed with specific weights to synthesize a new signal beam. This synthesis process aims to enhance vibration components that have specific spatial directionality and stable phase relationships between channels, such as vibration modes that may originate from periodic collisions between the drill bit and the borehole wall at specific angles.

[0042] Furthermore, a short-time Fourier transform is performed on the acoustic emission signal. The time-domain waveform of the acoustic emission signal is divided into a series of short time segments, with partial overlap between adjacent segments. The data within each short time segment is windowed and then subjected to a fast Fourier transform to obtain the spectrum of that segment. The spectra of all time segments are arranged in chronological order to form a two-dimensional matrix, i.e., the acoustic emission time spectrum. The horizontal axis of this matrix represents time, and the vertical axis represents frequency; the values ​​of the matrix elements represent the signal energy intensity at the corresponding time and frequency points.

[0043] Furthermore, the number of newly appearing rock debris contours per unit time in the rock debris image sequence is calculated. The image processing algorithm continuously analyzes the rock debris image stream. The algorithm detects the contours of all rock debris particles in each frame. By comparing the rock debris positions and contour features between consecutive frames, the algorithm can determine which rock debris is newly appearing in the current frame and which was previously present and has moved. The total number of newly identified independent rock debris contours in the image is counted within a very short statistical period, such as every 0.1 seconds.

[0044] Furthermore, when the quantity exceeds a preset threshold (which can be pre-set by those skilled in the art, and is not specifically limited in this invention), a rock breaking event is determined to have occurred, and the precise moment of occurrence of the event is recorded as the trigger moment. The threshold is determined experimentally based on drilling fluid flow rate, camera field of view size, and expected cuttings size. Exceeding the threshold means that a large number of new cuttings particles are generated in a very short time, which usually corresponds to a concentrated rock breaking process.

[0045] Furthermore, based on this trigger moment, a time segment of fixed length centered on the trigger moment is extracted from the previously synthesized enhanced signal beam. Simultaneously, a data block of the same time range is extracted from the acoustic emission time spectrum.

[0046] Furthermore, the characteristics of the signal beam within the time segment are calculated. The total energy of the time-domain signal within this segment is calculated, which is the sum of the squares of the amplitudes at all sampling points. The instantaneous frequency of the signal is calculated using methods such as Hilbert transform, and then the average and variance of the instantaneous frequency within this segment are obtained. The power spectrum of the signal in this segment is calculated and normalized to a probability distribution. The information entropy of this distribution is calculated to obtain the spectral entropy. These reflect the intensity of the vibration, the concentration and fluctuation of the frequency, and the complexity of the spectrum in the event, respectively.

[0047] Furthermore, the characteristics within the acoustic emission time-spectrum data block are calculated. For the truncated time-spectrum data block, the energy values ​​of all frequency points within the corresponding time period are summed to obtain the total energy. The average and variance of the spectral centroid frequency within this time period are calculated to describe the dominant frequency characteristics of acoustic emission. Similarly, the spectral entropy of the average power spectrum within this time period is calculated.

[0048] Furthermore, the eight eigenvalues ​​calculated from the vibration beam and acoustic emission are arranged in a fixed order and combined into a multi-dimensional vector. This vector is the high-frequency characteristic spectrum of drilling dynamics extracted for this rock fracturing event, which characterizes the dynamic response features accompanying the event from multiple dimensions.

[0049] This embodiment presents a dynamic feature extraction method guided by rock cuttings production events and integrating multiple physical signals, significantly improving the physical relevance and event specificity of the features. Traditional methods typically perform statistical analysis on continuous signals within fixed time windows, which may be mixed with fracturing events, frictional noise, and other interferences. This invention utilizes rock cuttings image information as an external trigger, extracting features from the corresponding time segment from the dynamic signal only when image analysis confirms significant rock fracturing. This ensures that the extracted dynamic features are directly related to the core physical process of rock fracturing, resulting in higher feature purity. Furthermore, this method does not analyze vibration or acoustic emission in isolation, but integrates both. By cross-correlation processing of triaxial acceleration signals, vibration modes aligned with the drill bit-hole wall interaction direction can be enhanced. Acoustic emission signals directly capture high-frequency elastic waves generated by microcracks within the rock. Extracting energy, frequency, and spectral complexity features from these two different but complementary signal sources, the constructed feature spectrum can simultaneously reflect macroscopic mechanical disturbances and microscopic fracturing processes, providing a more three-dimensional and refined dynamic description of rock fracturing events.

[0050] S3, extract the rock debris morphology parameter sequence based on the rock debris image sequence.

[0051] In practice, image preprocessing is first performed, including contrast adjustment, filtering and noise reduction, and using threshold segmentation or background subtraction methods to separate rock cuttings from the image background. For each segmented rock cutting region, its morphological descriptive parameters are calculated. For example, the equivalent diameter and area of ​​the particle are calculated through pixel statistics; its roundness is calculated through image moments, with lower roundness indicating more defined edges; and the complexity of its contour is calculated through boundary tracking. Furthermore, for a pre-defined time period corresponding to the drilling dynamics analysis window, the morphological parameters of all identified rock cuttings within that time period are statistically analyzed, and the mean and standard deviation of these parameters are calculated, and a grain size distribution curve is plotted. These statistics and distribution information are encoded into a feature vector, which is the sequence of rock cutting morphological parameters for that time window.

[0052] For example, in some preferred embodiments, the rock debris image sequence includes continuous frame images of the rock debris flow acquired using a binocular vision system. The extraction of rock debris morphology parameter sequences based on the rock debris image sequence includes: obtaining a three-dimensional point cloud model of the rock debris flow through stereo matching and three-dimensional reconstruction based on the continuous frame images; segmenting the three-dimensional point cloud model to obtain multiple point cloud clusters, where each point cloud cluster represents one rock debris; calculating the ratio of the surface area to the volume of each point cloud cluster, using this ratio as the shape factor of the rock debris corresponding to the point cloud cluster; calculating the standard deviation of the distribution of the normal vectors of all points on the surface of each point cloud cluster, using this standard deviation as the surface roughness index of the rock debris corresponding to the point cloud cluster; calculating the mean of the aspect ratio and height ratio of the minimum bounding cuboid of each point cloud cluster, using this mean as the flatness index of the rock debris corresponding to the point cloud cluster; and, for the current time window, statistically analyzing the mean, standard deviation, and distribution histogram information of the shape factor, surface roughness index, and flatness index of all rock debris within the current time window to obtain the rock debris morphology parameter sequence for the current time window.

[0053] In specific implementation, the steps for extracting the three-dimensional morphological parameters of rock cuttings based on binocular vision are as follows: The rock cuttings image sequence is acquired synchronously by a pair of precisely calibrated high-speed binocular cameras. The two cameras are installed at a certain horizontal distance and aimed at the detection area through which the rock cuttings flow from slightly different perspectives, thereby synchronously capturing the sequence of left and right view image pairs.

[0054] Furthermore, based on these consecutive frames of stereo images, a 3D point cloud model of the debris flow is obtained through stereo matching and 3D reconstruction. The processing procedure is as follows: First, the left and right images are corrected to achieve row alignment. Then, for each pixel in the left image, the corresponding pixel in the right image is searched for its best match. The matching is usually based on the brightness or texture similarity of the small area around the pixel, achieved by calculating a matching cost function such as normalized cross-correlation. After finding the matching point, the disparity value is calculated based on the difference in the horizontal position of the pixel in the left and right images. Based on intrinsic and extrinsic parameters such as the camera's focal length and baseline distance, the 2D coordinates of each pixel and its disparity value are converted into point coordinates in 3D space using the principle of triangulation. Performing this operation on all non-background areas in the image generates a dense 3D point cloud, where each point represents the spatial position of a point on the object's surface.

[0055] Further, the 3D point cloud model is segmented to obtain point cloud clusters representing individual rock fragments. Since the generated point cloud contains all objects within the field of view, each rock fragment needs to be separated. A spatial distance-based clustering algorithm, such as Euclidean clustering, is employed. This algorithm sets a distance threshold, grouping points within a certain spatial distance into the same set, while separating points that are farther apart into different sets. A minimum point count threshold is also set to filter out noise points. After this step, the point cloud is segmented into multiple independent point cloud clusters, each cluster corresponding to a single rock fragment.

[0056] Furthermore, for each segmented point cloud cluster, the ratio of its surface area to volume is calculated. The surface area can be estimated by calculating the convex hull surface area of ​​the point cloud cluster, or by summing the areas of all triangles after triangulating the point cloud. The volume can be estimated by calculating the convex hull volume, or by using a voxelization method for approximation. The ratio of surface area to volume is a shape descriptor; for objects of the same volume, the more complex, slender, or rougher the shape, the larger the ratio.

[0057] Furthermore, for each point cloud cluster, the standard deviation of the distribution of normal vectors across all points on its surface is calculated. First, the normal vector of each point in the point cloud needs to be estimated. This is typically achieved by analyzing the local plane formed by the point and its neighbors, using principal component analysis to obtain the normal direction of the local plane. For a rock debris point cloud cluster, the normal vectors of all its points are collected, and the statistical dispersion of these normal vectors in direction is calculated, for example, by calculating the standard deviation of their angles with the mean normal vector. This value reflects the overall roughness or undulation of the rock debris surface; the smoother the surface, the more consistent the normal vector directions, and the smaller the standard deviation.

[0058] Furthermore, for each point cloud cluster, calculate the mean of the aspect ratio and height ratio of its smallest circumscribed cuboid. Find a cuboid whose sides are parallel to the global coordinate axes, which completely encloses all points of the point cloud cluster and has the smallest volume. Determine the length, width, and height of this cuboid. Calculate the aspect ratio and height ratio, and then take the arithmetic mean of these two ratios. This mean reflects the flat or elongated nature of the rock fragments; for flaky rock fragments, this value is larger.

[0059] Furthermore, for the current processing time window, the system collects the shape factor, surface roughness index, and flatness index of all rock fragments obtained through the above process within this window. Then, it calculates the statistical characteristics of these three indicators within the current window, including their arithmetic mean and standard deviation. In addition, a distribution histogram is created for each indicator, dividing the indicator's value range into several intervals and counting the number of rock fragments falling within each interval. Finally, these statistical values ​​are combined into a structured feature vector, which is the sequence of rock fragment morphology parameters for the current time window.

[0060] This embodiment achieves a leap from two-dimensional planar observation to three-dimensional quantitative measurement of rock cutting morphology by introducing binocular stereo vision and three-dimensional point cloud analysis technology. Traditional rock cutting analysis based on monocular images can only obtain the projected contour information of particles, and cannot accurately determine their true three-dimensional size, volume, and surface spatial morphology, resulting in inherent errors in morphological description. This invention utilizes binocular vision to directly reconstruct a three-dimensional point cloud model of rock cuttings, providing real and complete spatial geometric data for morphological analysis. The shape factor, surface roughness index, and flatness index based on the three-dimensional point cloud can more fundamentally reflect the physical properties of rock cuttings than two-dimensional parameters. The shape factor is related to the energy consumption during rock fracturing; the surface roughness index is directly related to the microstructure of the fracture surface; and the flatness index reflects the development of bedding or joints in the rock. These three-dimensional morphological parameters establish a more direct and reliable theoretical connection with the mechanical properties and fracturing mechanism of rocks. Furthermore, this invention focuses on the statistical characteristics description of rock cutting groups over a time period, rather than individual rock cuttings. This group statistic has better stability, can smooth out the influence of individual outliers, and is more representative of the overall rock fracturing pattern under current drilling conditions. Using these high-fidelity three-dimensional group morphological features for inversion greatly enriches the stratigraphic response information that the model can utilize.

[0061] In some preferred embodiments, the step of extracting the rock debris morphology parameter sequence based on the rock debris image sequence further includes: for each point cloud cluster, identifying the fracture surface region in the point cloud cluster and extracting the three-dimensional contour of the fracture surface region; calculating the three-dimensional fractal dimension of the three-dimensional contour of the fracture surface region; calculating the equivalent compressive strength factor of the rock debris corresponding to the point cloud cluster based on the estimated area of ​​the fracture surface region, the volume of the rock debris corresponding to the point cloud cluster, the three-dimensional fractal dimension, and the pre-calibrated energy conversion coefficient; and calculating the average and variance of the equivalent compressive strength factors of all rock debris for the current time window, and appending them to the rock debris morphology parameter sequence of the current time window.

[0062] In practice, adding fracture energy characterization features to the rock fragment morphology parameter sequence includes the following steps: For each point cloud cluster that has undergone 3D reconstruction and basic morphological analysis, fracture surface regions need to be identified. Fracture surfaces are newly formed surfaces during rock fracturing, typically relatively flat and continuous, and differ geometrically from the original or weathered surfaces of the rock fragments. Further, identification can be achieved by analyzing the local curvature of the point cloud cluster; points in fracture surface regions often have lower principal curvature values. Alternatively, a region growing algorithm can be used, starting from a seed point, aggregating points with similar normal vector directions and spatial proximity to the seed point to form a candidate fracture surface region. All points in this region are extracted to constitute the 3D contour point set of the fracture surface region.

[0063] Furthermore, the three-dimensional fractal dimension of the fracture surface profile is calculated. The fractal dimension quantifies the irregularity and complexity of the profile. Box counting is used for calculation. A three-dimensional cubic mesh covers the bounding box space containing the entire fracture surface profile point set, with an initial large cube side length. The number of cubes in the mesh containing at least one profile point is counted. Then, the cube side length is gradually reduced, for example, halved each time, and the number of cubes containing points is recounted. In a log-log coordinate system, the relationship between the number of cubes and the reciprocal of the cube side length is plotted. The slope of this curve in the scale-free region approximates the three-dimensional fractal dimension. A higher fractal dimension indicates a more complex and irregular fracture surface profile.

[0064] Further, the equivalent compressive strength factor of the rock fragments corresponding to the point cloud cluster is calculated. This requires an estimated area of ​​the fracture surface region, the volume of the rock fragments, the calculated fractal dimension, and a pre-calibrated energy conversion coefficient. The fracture surface area can be obtained by summing the areas of the triangles after triangulating the point cloud of that region. The volume of the rock fragments was obtained earlier when calculating the shape factor. The energy conversion coefficient is obtained through calibration experiments in the laboratory: using standard rock samples with known uniaxial compressive strength, a crushing experiment is conducted, and the crushed rock fragments are collected and subjected to the same three-dimensional scanning and parameter calculations. The relationship between the conversion coefficient and the known strength and geometric parameters of the rock fragments is established through fitting multiple sets of data. In application, for each rock fragment from the field, the calibrated coefficient, combined with its own area, volume, and fractal dimension, is used to calculate an equivalent compressive strength factor according to a specific formula. This factor is a dimensionless index positively correlated with the energy density required to crush the rock fragment.

[0065] Furthermore, for the current analysis time window, the equivalent compressive strength factors of all rock fragments within that window are collected. The arithmetic mean and variance of these factors are calculated. The mean reflects the average energy consumption level of rock fracturing during that time period, while the variance reflects the fluctuations in energy consumption.

[0066] Furthermore, the calculated mean and variance, two new numerical features, are appended to the end of the previously generated feature vector of the rock fragment morphology parameter sequence, which contains basic morphological statistics. Thus, this sequence, in addition to describing the morphology, further incorporates information reflecting fracture energy.

[0067] This embodiment, based on three-dimensional morphological analysis, further introduces a mechanical energy characterization based on the fractal characteristics of the fracture surface, deepening the correlation between rock fragment information and formation mechanical properties. The rock fracturing process is essentially a process of energy dissipation and transformation; rock fragments, as the final product, contain information about energy dissipation through the geometric characteristics of their fracture surfaces. The relationship between simple macroscopic morphological parameters and rock strength can be influenced by various factors. This invention aims to construct a derived parameter—the equivalent compressive strength factor—more directly related to rock fracturing energy by calculating the three-dimensional fractal dimension of the fracture surface and introducing a laboratory-calibrated energy conversion coefficient. The fractal dimension is widely considered a quantitative indicator of fracture surface roughness, closely related to fracture toughness and the energy consumed in crack propagation; a higher fractal dimension usually indicates a more tortuous fracture path and higher energy dissipation. By linking the fractal dimension, fracture surface size, and rock fragment volume through calibration coefficients, the resulting factor attempts to explain the "origin" of rock fragments from an energy perspective. Adding this factor as a feature to the inversion model provides additional physical clues about the formation's rock resistance to fracturing or its energy requirements. This helps the model better distinguish rocks that are similar in traditional morphology but have different mechanical behaviors, such as distinguishing between hard and brittle rocks and rocks with high toughness, which is expected to improve the accuracy and physical interpretability of rock mechanical properties, especially strength and fracturing.

[0068] S4 inputs the time-aligned high-frequency feature spectrum of drilling dynamics and the sequence of cuttings morphology parameters into a pre-trained multi-task deep learning model.

[0069] In practice, since the acquisition system may start independently, the time axis of the drilling dynamics signal and the time axis of the cuttings image sequence need to be calibrated to ensure that their data segments represent the same drilling process. After calibration, the high-frequency feature spectrum of drilling dynamics and the sequence of cuttings morphology parameters corresponding to the same time window are concatenated along the feature dimension to form a longer joint feature vector. This joint feature vector is then input in real time into a multi-task deep learning model that has already been trained offline.

[0070] S5: Obtain the borehole wall quality assessment results and formation characteristic identification results generated synchronously by the multi-task deep learning model. The borehole wall quality assessment results are used to characterize the borehole wall state, and the formation characteristic identification results are used to characterize the formation attributes.

[0071] In practice, the pre-trained multi-task deep learning model receives the joint feature vector and performs forward propagation computation through its internal network structure. The model has two outputs. The first output produces a borehole wall quality assessment result, which can include quantified numerical indicators, such as a score representing borehole wall roughness and an indicator representing borehole wall integrity or the presence of significant fractures. The second output simultaneously produces a formation characteristic identification result, which can include a classification judgment of the current formation lithology, such as the probability of belonging to sandstone, mudstone, or limestone, and an estimated rock hardness or strength index. These results are output and can be used for real-time display or recording.

[0072] This invention achieves synchronous real-time perception of borehole engineering status and geological characteristics by combining drilling dynamics signals with rock cuttings image information and inputting them into a multi-task model for joint inversion. Drilling dynamics signals macroscopically reflect the entire process of the interaction between the drill bit and the formation, containing rich information generated by rock fracturing, drill string vibration, and friction between the drill string and the borehole wall. However, their ability to directly describe the local morphology of the borehole wall and rock fracturing products is limited. Rock cuttings image sequences provide intuitive visual evidence of the products after rock fracturing, and their morphological characteristics are directly related to the mechanical properties and fracturing mechanisms of the rock. Furthermore, this invention synchronously collects and characterizes these two complementary types of information in terms of physical source and form, and precisely aligns them in time to ensure the spatiotemporal consistency of the fused information. A pre-trained multi-task deep learning model processes these fused features, enabling the model to parse patterns related to borehole wall status and formation properties from the mixed information stream. This processing method allows for the parallel acquisition of two key results in a single analysis, overcoming the limitations of traditional methods such as lagging borehole quality evaluation and reliance on a single data source for formation identification. This provides a more immediate and comprehensive basis for drilling decisions, helping to optimize parameters in a timely manner during drilling to protect the borehole wall, identify potential risks, and accurately delineate formations.

[0073] Further, in some preferred embodiments, the multi-task deep learning model includes a common feature encoder, a gated sharing unit, a borehole wall quality assessment branch, and a formation characteristic identification branch; the common feature encoder performs primary fusion of the spliced ​​drilling dynamics high-frequency feature spectrum and the cuttings morphology parameter sequence to obtain primary fused features; the gated sharing unit includes a multi-head self-attention network, which takes the primary fused features as input and outputs a first feature selection weight matrix and a second feature selection weight matrix, the first feature selection weight matrix corresponding to the borehole wall quality assessment branch, and the second feature selection weight matrix corresponding to the formation characteristic identification branch; the input features of the borehole wall quality assessment branch are the element-wise product of the primary fused features and the first feature selection weight matrix; the input features of the formation characteristic identification branch are the element-wise product of the primary fused features and the second feature selection weight matrix.

[0074] In practice, the specific data processing flow of the multi-task deep learning model is as follows. The model first processes the input through a common feature encoder. The input is a one-dimensional feature vector sequence formed by directly concatenating the time-aligned high-frequency feature spectrum of drilling dynamics and the sequence of rock cutting morphology parameters along the feature dimension. The common feature encoder typically consists of several layers of one-dimensional convolutional neural networks. These convolutional layers use convolutional kernels of different sizes to perform sliding convolution operations on the input feature sequence along the time dimension. Each convolutional layer is usually followed by an activation function and possibly a pooling layer. The convolutional operation can extract local patterns and temporal dependencies in the input sequence. After multiple convolutions, the original concatenated features are transformed into a more abstract, higher-level feature representation, called the primary fusion feature. This feature integrates information from dynamics and rock cutting images and may contain latent patterns relevant to both tasks.

[0075] Furthermore, the initial fusion features are fed into the gate-controlled shared unit. The core of this unit is a multi-head self-attention network. For example, the multi-head self-attention network includes four attention heads, each with an output dimension of 64; the first feature selection weight matrix and the second feature selection weight matrix are generated through independent linear transformation layers, and their independence is optimized during training through orthogonality constraints.

[0076] Furthermore, the self-attention mechanism calculates the correlation strength between each position in the primary fused feature sequence and all positions in the sequence (including itself), i.e., the attention weight. The multi-head mechanism means running multiple independent attention calculations in parallel, each focusing on a different feature subspace or relational aspect. In this embodiment, the network is designed to output two independent feature selection weight matrices. During the learning process, the network dynamically generates these two matrices based on the specific content of the primary fused features. The weight distribution of the first matrix aims to identify feature elements or time locations that are more critical for evaluating borehole wall quality; the second matrix aims to identify parts that are more critical for identifying formation characteristics. These two matrices have dimensions compatible with the primary fused features.

[0077] Furthermore, the model performs feature routing. For the borehole quality assessment branch, the input features are not the original primary fusion features, but rather the result of element-wise multiplication (dot product) of the primary fusion features with the first feature selection weight matrix. This means that each value in the primary fusion features is scaled according to the weight at the corresponding position in the first matrix. Feature values ​​with high weights are enhanced, while those with low weights are suppressed. Similarly, for the formation characteristic identification branch, the input features are obtained by element-wise multiplication of the primary fusion features with the second feature selection weight matrix. Therefore, the features flowing into both branches are subsets of features that have been filtered and modulated with different task-specific weights; they originate from the same shared features but have different emphases.

[0078] This embodiment designs a multi-task network architecture with a gated sharing unit, achieving dynamic selection of shared features for task adaptation and optimizing feature utilization efficiency in multi-task learning. In traditional shared underlying networks, two task branches are forced to use the exact same underlying feature representations, which can lead to task conflicts, i.e., the needs of one task may interfere with the learning of another. The gated sharing unit introduced in this embodiment acts as an intelligent mediator. Its internal multi-head self-attention network can analyze the content of shared features and generate different attention filters for the two tasks. Through element-wise multiplication, these filters dynamically gate the shared features, thereby extracting and enhancing one set of relevant features for the borehole wall evaluation task, while extracting and enhancing another set of relevant features for the formation identification task. This mechanism allows the model to achieve soft task decoupling at the feature level while maintaining a shared feature extractor. It prompts the model to explore the shared feature space more finely, learning to route different information components to different tasks, thereby promoting knowledge sharing, reducing the number of parameters, and minimizing the negative impact between tasks. This helps each task branch obtain cleaner and more targeted inputs, thereby improving their respective learning effects and final performance.

[0079] In some preferred embodiments, the step of generating the pore wall quality assessment result by the pore wall quality assessment branch includes: simultaneously feeding the input features of the pore wall quality assessment branch into a first one-dimensional convolutional path, a second one-dimensional convolutional path, and a third one-dimensional convolutional path, wherein the first one-dimensional convolutional path, the second one-dimensional convolutional path, and the third one-dimensional convolutional path have different time scales, and extracting first-scale features, second-scale features, and third-scale features respectively; concatenating the first-scale features, the second-scale features, and the third-scale features, and calculating them through a self-attention layer to obtain the first contribution weight corresponding to the first-scale feature, the second contribution weight corresponding to the second-scale feature, and the third contribution weight corresponding to the third-scale feature. Based on the first contribution weight, the second contribution weight, and the third contribution weight, the first-scale features, the second-scale features, and the third-scale features are weighted and fused to obtain fused features. Based on the fused features, the power spectral density integral value of the surface undulation of the borehole wall is calculated through the first output submodule, where the power spectral density integral value is used as a roughness quantification index. Based on the fused features, the coefficient of variation of the borehole diameter along the depth direction is calculated through the second output submodule, where the coefficient of variation is used as a borehole diameter stability index. Based on the fused features, the probability value of borehole wall instability risk is calculated through the third output submodule, where the probability value of borehole wall instability risk is calculated based on the vibration energy dissipation rate and ranges from 0 to 1.

[0080] In practice, the process of generating specific evaluation results for the pore wall quality assessment branch is as follows: The input features of the branch are first simultaneously fed into three parallel processing pathways. The first one-dimensional convolutional pathway uses a small convolutional kernel, for example, with a width of 3, focusing on capturing fast, short-period fluctuation patterns in the input feature sequence. The second one-dimensional convolutional pathway uses a medium-width convolutional kernel, for example, with a width of 7, to extract trend changes at medium time scales. The third one-dimensional convolutional pathway uses a larger convolutional kernel or employs dilated convolution techniques to expand the receptive field, for example, with an equivalent width of 15, aiming to capture slowly varying, global background information. Each pathway typically contains multiple stacked one-dimensional convolutional layers, nonlinear activation layers, and possible downsampling layers. After processing by their respective pathways, the first-scale features, second-scale features, and third-scale features are output, representing the input features observed from different time perspectives.

[0081] Furthermore, the features at these three different scales are concatenated along the channel dimension to form a composite feature tensor that integrates multi-scale information. This composite feature is then fed into a self-attention layer. This self-attention layer calculates the correlation between the various parts within the composite feature and outputs a set of weight values. After normalization, these weights represent the relative importance or contribution of the first-scale, second-scale, and third-scale features to the final judgment of the pore wall quality. For example, when the input features suggest the presence of high-frequency, severe vibrations, the model may assign higher weights to features from shorter-scale channels.

[0082] Furthermore, based on the calculated contribution weights, the features at the three scales are weighted and fused. Specifically, the first-scale feature is multiplied by its corresponding contribution weight, the second-scale feature by its corresponding weight, and the third-scale feature by its corresponding weight. Then, these three weighted feature tensors are summed or their channels are concatenated to obtain the final fused feature. This feature dynamically emphasizes the most important time-scale information in the current input context.

[0083] Based on this fusion feature, the branch generates specific quantitative indicators in parallel through three independent output submodules. The first output submodule is a small neural network, such as a fully connected layer, that receives the fusion feature and outputs a scalar value. This value is interpreted as the power spectral density integral of the borehole wall surface undulations. During training, this submodule is supervised by the true power spectral density integral calculated from borehole imaging logging data. The larger this value, the rougher the borehole wall.

[0084] Furthermore, the second output submodule has a similar structure; it outputs another scalar value representing the coefficient of variation of the borehole diameter along the depth direction. The supervisory signal used to train this module comes from the true coefficient of variation calculated from the caliper logging data. This coefficient reflects the uniformity of the borehole diameter; a larger value indicates more irregular borehole diameter variations.

[0085] Furthermore, the third output submodule typically ends with a fully connected layer and a sigmoid activation function, outputting a scalar value between 0 and 1, representing the probability of hole wall instability. Training of this module relies on learning the association between labeled hole wall instability events and their corresponding features in historical data, with the goal of minimizing the difference between the predicted probability and the true event label.

[0086] This embodiment employs a multi-scale analysis and adaptive weighted fusion strategy within the borehole wall assessment branch, effectively enhancing the ability to distinguish complex borehole wall conditions. Borehole wall quality issues exhibit diverse characteristics, ranging from instantaneous impact damage to long-term wear and diameter expansion, as well as progressive instability risks. These phenomena possess different features at different time scales. By deploying convolutional pathways with different receptive fields in parallel, this branch can simultaneously capture complete signal patterns from transient to long-term. Furthermore, the introduction of a self-attention mechanism is crucial, enabling the model to intelligently determine which time scales of information are more critical based on the actual situation of the current input data and dynamically adjust the fusion weights. For example, when assessing pitting caused by drill bit bounce, short-scale features are given higher weights; while when assessing the overall borehole wall stability trend, long-scale features are more important. This data-driven, dynamically adjusted multi-scale information fusion approach allows the assessment model to more comprehensively and precisely analyze and quantify different quality dimensions of the borehole wall, thereby outputting assessment results that are more adaptive and accurate than single-scale analysis or fixed-rule fusion.

[0087] In some preferred embodiments, the step of generating formation characteristic identification results by the formation characteristic identification branch includes: concatenating the input features of the formation characteristic identification branch with the current drilling condition parameter vector, and transforming it through a fully connected layer to obtain basic features, wherein the drilling condition parameter vector includes drilling pressure, rotation speed, and displacement; transforming the basic features using a learnable condition compensation matrix to obtain compensated features, wherein the transformation operation of the condition compensation matrix is ​​combined with the drilling condition parameter vector, and the compensated features are used to eliminate the influence of condition changes; based on the compensated features, generating a classification probability distribution of formation lithology through a lithology classification output head; based on the compensated features, generating a predicted value of rock dynamic hardness through a hardness regression output head; and based on the compensated features, generating a normalized index of the critical confining pressure for the brittle-plastic transition of rock through a critical index output head.

[0088] In practice, the formation characteristic identification branch generates specific identification results as follows. First, the input features of the branch are concatenated with a parameter vector representing the current drilling conditions. This condition vector typically includes three key operational parameters: drilling pressure, rotary table speed, and drilling fluid discharge rate. This concatenation operation enables the subsequent network to simultaneously perceive both formation response and operating conditions.

[0089] Furthermore, the concatenated features are fed into one or more fully connected layers for nonlinear transformation and feature integration. The fully connected layers learn how to initially fuse and interact with formation response features and operating parameters, outputting a more abstract representation of the basic features.

[0090] Furthermore, a learnable working condition compensation matrix is ​​used to process the basic features. This compensation matrix is ​​a set of parameters learned during model training. Its operation aims to use working condition parameter information to correct systematic deviations in the basic features that may be caused by changes in working conditions. One implementation is to map the working condition parameters to a transformation vector or matrix through a small network, and then apply it to the basic features, for example, by performing a linear transformation of the feature space or bias adjustment. The goal is to generate a set of compensated features that primarily reflect the intrinsic properties of the formation itself, and are insensitive to changes in external operating conditions such as drilling pressure and rotation speed.

[0091] Furthermore, based on the compensated features, the branch generates the final result through three parallel output heads. The lithology classification output head typically consists of a fully connected layer and a softmax activation function, outputting a probability distribution vector. Each dimension of this vector corresponds to a preset lithology category, and the value represents the probability of belonging to that category.

[0092] Furthermore, the hardness regression output head is typically composed of a fully connected layer, which may use activation functions such as ReLU to ensure that the output is non-negative. It outputs a continuous scalar value representing the predicted dynamic rock hardness or strength index.

[0093] Furthermore, the critical index output head is also a regression head, outputting a scalar value. This value is interpreted as a normalized index related to rock mechanical behavior, such as indicating the tendency of a rock to transition from brittle fracture to plastic deformation. Training this output head requires constructing corresponding labels based on rock mechanics experimental data or expert knowledge.

[0094] This embodiment significantly improves the robustness and generalization ability across different drilling conditions by explicitly introducing a working condition compensation mechanism. In actual drilling, drilling dynamics and cuttings characteristics are not only determined by the formation but also strongly influenced by real-time drilling parameters. For example, high drilling pressure may produce signal characteristics similar to those of hard formations under low drilling pressure, leading to model misjudgment. This invention uses real-time working condition parameters as input and actively eliminates their influence on feature representation through a learnable compensation matrix. During training, this matrix learns how working condition parameters systematically "distort" features through a large amount of data covering different combinations of working conditions and learns how to correct this inversely. This allows the model to extract purer intrinsic formation features from mixed signals. Therefore, when the formation is the same but the drilling parameters change, the model can still provide stable identification results; when the formation is different, the model can more accurately attribute signal differences to the formation changes themselves. This design greatly enhances the practical reliability of this inversion method under different drilling strategies and operating conditions.

[0095] In some preferred embodiments, before inputting the time-aligned high-frequency feature spectrum of drilling dynamics and the cuttings morphology parameter sequence into a pre-trained multi-task deep learning model, the method further includes: real-time analysis of the variance of drilling pressure and torque in the drilling dynamics signal within a sliding time window; if the variance is lower than a first preset threshold, the current state is determined to be a steady-state drilling state; if it is a steady-state drilling state, the high-frequency feature spectrum of drilling dynamics and the cuttings morphology parameter sequence of the current time window are directly concatenated as the final input feature; if the variance is not lower than the first preset threshold, the current state is determined to be an abnormal drilling state; if it is an abnormal drilling state, the high-frequency feature spectrum of drilling dynamics and the cuttings morphology parameter sequence of multiple consecutive time windows in the past are input into a recurrent neural network in chronological order; historical context feature vectors are extracted through the recurrent neural network; and the historical context feature vectors are concatenated with the high-frequency feature spectrum of drilling dynamics and the cuttings morphology parameter sequence of the current time window as the final input feature.

[0096] In practice, the feature adaptive fusion step is implemented as follows: The system continuously analyzes drilling dynamics signals in real time. It maintains a sliding time window and calculates the variance of the drill pressure signal sequence and the variance of the torque signal sequence within this window. These two variance values ​​are combined, for example, by taking the average or the maximum value, as an indicator representing the stability of the current drilling process.

[0097] Furthermore, the calculated stability index is compared with a preset first threshold. This threshold is determined by analyzing historical stable drilling data. If the current index is below this threshold, the drilling process is considered to be in a steady state. In a steady state, the system believes that the current state is mainly determined by current information, with weak historical dependence. Therefore, the high-frequency feature spectrum of drilling dynamics and the sequence of cuttings morphology parameters extracted from the current time window are directly concatenated to form the joint feature that is finally input into the multi-task deep learning model.

[0098] Furthermore, if the current stability index is not lower than the first threshold, the drilling process is determined to be in an abnormal state. This abnormality may originate from encountering formation interfaces, fracture zones, or complex downhole conditions.

[0099] Furthermore, in abnormal conditions, the system not only uses the features of the current window, but also retrieves historical feature data corresponding to multiple consecutive time windows. For example, it retrieves the high-frequency feature spectrum of drilling dynamics and the sequence of cuttings morphology parameters from the five windows preceding the current window.

[0100] Furthermore, these historical features and current window features are arranged chronologically to form a feature sequence. This time sequence is then input into a recurrent neural network, such as a long short-term memory network. Recurrent neural networks have memory capabilities, enabling them to process sequential data and capture long-term temporal dependencies. The network processes the sequence progressively, and its final hidden state output encodes the entire sequence, i.e., the evolution of states from the past to the present and related information. This hidden state vector is called the historical context feature vector.

[0101] Furthermore, the historical context feature vector output by the recurrent neural network is concatenated with the original features of the current window representing the latest information to form the final joint features rich in contextual information, which are then input into the multi-task deep learning model for inversion.

[0102] This embodiment enhances the intelligent response capability of the inversion system to dynamic drilling processes by employing an adaptive information fusion strategy based on drilling stability criteria. Drilling is not a steady-state process; the importance of its data characteristics varies at different stages. In the stable phase, transient information is sufficient to reflect the state; in the changing phase, understanding the changing process is crucial. Instead of using a fixed, lengthy historical window, this invention designs an adaptive switch based on real-time signal variance. This allows the model to operate lightweightly during stable periods, focusing on the present; when an anomaly is detected, it automatically activates a context-aware mode, using a recurrent neural network to extract historical evolution trends. This enables the model to distinguish between transient disturbances and real formation changes, more accurately locate lithological interfaces, understand the process of borehole wall deterioration, and thus make more accurate judgments under complex conditions. This strategy balances computational efficiency and inversion accuracy.

[0103] In some preferred embodiments, the steps for obtaining a pre-trained multi-task deep learning model include: performing preliminary training on a source domain dataset containing various geological conditions to obtain a base model; when drilling begins at a new target site, collecting drilling dynamics signals and cuttings image sequences within the initial depth of the target site and performing local validation to obtain validation results; constructing a target domain dataset based on the collected drilling dynamics signals and cuttings image sequences within the initial depth of the target site and the validation results; adding a domain classifier after the common feature encoder of the base model; performing adversarial training, which includes alternating repetition of the following two optimization phases until the model converges: Phase 1: fixing the parameters of the common feature encoder, borehole wall quality assessment branch, and formation characteristic identification branch, and updating the parameters of the domain classifier; transferring the source... In the first stage, the source domain dataset and the target domain dataset are input into a common feature encoder to obtain source domain features and target domain features, respectively. A domain classifier is trained using these features and their corresponding domain labels, enabling it to distinguish whether features originate from the source or target domain dataset. In the second stage, the parameters of the domain classifier are fixed, and the parameters of the common feature encoder are updated. The source and target domain datasets are then input into the common feature encoder, and the domain confusion loss is calculated using the domain classifier. The domain confusion loss is optimized using backpropagation to update the parameters of the common feature encoder, making it difficult for the domain classifier to distinguish the source of features extracted by the common feature encoder. During adversarial training, the parameters of the borehole wall quality assessment branch and the formation characteristic identification branch are updated using the drilling dynamics signals, cuttings image sequences, and corresponding local validation results from the target domain dataset.

[0104] In practice, the training steps for obtaining a pre-trained multi-task deep learning model are as follows. First, preliminary training is performed on the source domain dataset. The source domain dataset consists of a large amount of historical data collected from work areas with multiple different geological backgrounds, including synchronized dynamic signals, rock cutting images, and corresponding borehole wall quality and formation characteristic labels. Using this dataset, the model is trained using standard supervised learning methods, enabling the model to learn the basic mapping relationship from features to labels, thus obtaining the basic model.

[0105] Furthermore, when the model is applied to a new work area, during the initial drilling phase, data from a certain drilling footage is collected according to this invention and locally validated. Validation may be accomplished through limited logging data, intensive cuttings sampling and analysis, or the experience of field engineers, obtaining a small amount of reliable labeled data specific to the new work area. Using this newly collected data and corresponding new labels, a small-scale target domain dataset is constructed.

[0106] It should be noted that the source domain dataset is derived from historical drilling data from multiple geological areas, including drilling dynamics signals, cuttings image sequences, and borehole wall quality and formation characteristic labels obtained through logging and core analysis. The validation of the target domain dataset was completed through logging-while-drilling and cuttings sampling laboratory analysis.

[0107] Furthermore, the base model is modified by inserting a new, simple domain classifier network, such as a two-layer fully connected network, after its common feature encoder to determine whether the input features originate from the source domain or the target domain.

[0108] Further, adversarial training is performed. This process alternates between two phases. In the first phase, the backbone network parameters of the model are fixed, and the domain classifier is trained. Source domain data and target domain data are input into the model, and the common feature encoder generates features. These features and their domain labels are used to train the domain classifier, with the goal of enabling the domain classifier to distinguish which domain the features come from as accurately as possible. In the second phase, the parameters of the domain classifier are fixed, and the common feature encoder is trained. The same data is input, but the training goal is to adjust the parameters of the common feature encoder so that the features it generates cannot be correctly distinguished by the fixed domain classifier, i.e., it confuses the two domains. Through this adversarial game, the common feature encoder is forced to learn to extract those universal features common to both domains and independent of the specific work area.

[0109] Furthermore, throughout the adversarial training process, the parameters of the borehole wall quality assessment branch and the formation characteristic identification branch are fine-tuned in a supervised learning manner using data and labels from the target domain dataset to adapt them to the label characteristics of the new work area.

[0110] Furthermore, the two stages described above are repeated alternately until the model performs satisfactorily on the target domain validation data. At this point, a pre-trained model adapted to the new work area is obtained.

[0111] This embodiment employs an adversarial domain adaptation strategy to effectively address the domain transfer problem of pre-trained models when applied to new work areas. Differences in geological conditions lead to varying data distributions, potentially causing performance degradation when directly applying the model. This invention utilizes adversarial training at the feature level, forcing the model to learn general feature representations independent of geological conditions, thereby effectively transferring knowledge learned from the source domain to the target domain. Furthermore, task branches are fine-tuned using a small amount of labeled data from the new work area, enabling the model to quickly adapt to the specific needs of the new work area. This method achieves rapid model adaptation and performance improvement in new environments with relatively low new data labeling costs, enhancing the practicality and universality of the approach.

[0112] In some preferred embodiments, the step of training a multi-task deep learning model further includes introducing task orthogonality constraints: converting the first feature selection weight matrix output by the gated shared unit into a first vector; converting the second feature selection weight matrix output by the gated shared unit into a second vector; calculating the cosine similarity between the first vector and the second vector, the cosine similarity constituting a task orthogonality constraint term; constructing a total loss function, the total loss function including a first loss term corresponding to the borehole wall quality assessment branch, a second loss term corresponding to the formation characteristic identification branch, and a task orthogonality constraint term; and training the multi-task deep learning model by minimizing the total loss function, such that the first feature selection weight matrix and the second feature selection weight matrix are orthogonal in the vector space.

[0113] In practice, during model training, each forward propagation involves obtaining the first feature selection weight matrix and the second feature selection weight matrix generated for the two tasks from the gated shared unit.

[0114] Furthermore, these two matrices are converted into one-dimensional vectors. Specifically, all elements in the matrices are arranged in row-major order to form two long vectors.

[0115] Further, the cosine similarity between the two vectors is calculated. Cosine similarity is obtained by calculating the inner product of the two vectors and then dividing by the product of their respective magnitudes. This value lies between -1 and +1, and its absolute value reflects how close the directions of the two vectors are. The closer the value is to zero, the closer the two vectors are to being orthogonal.

[0116] Furthermore, this cosine similarity value is treated as an additional loss term, called the task orthogonality constraint. To ensure that the two weight matrices are orthogonal, it is generally desirable to minimize the square of this similarity value.

[0117] Furthermore, when calculating the total loss, in addition to the losses from the borehole wall quality assessment branch and the formation characteristic identification branch, the orthogonality constraint term of this task is added and multiplied by a small coefficient to balance its effect. The total loss is a weighted sum of these three loss terms.

[0118] Furthermore, this total loss function is minimized through the backpropagation algorithm. The optimization process not only reduces the prediction errors of the two main tasks, but also adjusts the parameters of the gated shared unit through the gradient of the constraint terms, so that the vectors corresponding to the two feature selection weight matrices it generates gradually tend to be perpendicular in the vector space, that is, the dot product tends to zero, thereby achieving the trend of orthogonality.

[0119] This embodiment aims to optimize the behavior of the gated shared unit by adding task orthogonality constraints to the training objective, promoting a more complementary feature utilization pattern between the two tasks. Without guidance, the learned feature selection weight matrices of the two branches may be highly similar, resulting in high information redundancy and failure to fully utilize the diverse information in the shared features. The orthogonality constraint mathematically encourages the two weight matrices to focus on different feature directions. With this constraint applied, the model is guided to explore more dispersed subspaces in the shared feature space, ensuring that the feature sets relied upon by the borehole wall assessment branch and the formation identification branch are as non-overlapping and distinctive as possible. This helps reduce interference between tasks, prompting each branch to discover unique information more discriminative for its own task, thereby potentially improving the performance of both tasks and the overall learning efficiency of the model.

[0120] In some preferred embodiments, the method further includes: rendering the real-time output borehole wall quality assessment results and formation characteristic identification results into a three-dimensional geological model associated with the borehole trajectory for dynamic visualization; providing a human-computer interaction interface, which receives correction annotations from the operator on the inversion results, including confirmation, rejection, or re-delineation of suspected defective sections or formation interfaces; combining the correction annotations with the drilling dynamics high-frequency characteristic spectrum and cuttings morphology parameter sequence at the corresponding time of the correction annotations to form online learning samples; collecting multiple online learning samples; and periodically using the online learning samples in an incremental learning manner to fine-tune the parameters of the last two fully connected layers in the multi-task deep learning model.

[0121] In practice, the system binds the borehole wall quality indicators and formation characteristics obtained in real time with depth information and renders them in a 3D visualization software. Using the borehole design trajectory as the axis, the software visually displays results such as roughness, lithology, and risk probability along the depth direction using color or texture, forming a geological engineering profile that is updated during drilling.

[0122] Furthermore, the software provides an interactive interface that allows field engineers to view the profile. If an engineer believes, based on experience or other data (such as gas logging), that a certain assessment in the model is inaccurate, they can correct the inversion results for a specific depth segment by clicking or selecting with the mouse. For example, they can change the model's label from "high risk" to "low risk," or manually adjust automatically identified lithological boundaries. These manual corrections are recorded by the system as correction labels and associated with the specific depth interval.

[0123] Furthermore, the system automatically maps the depth interval corresponding to each correction annotation back to the data acquisition timeline. Then, it retrieves the original input features used for model inversion within that time interval from the stored historical data, namely the corresponding high-frequency feature spectrum of drilling dynamics and the sequence of cuttings morphology parameters.

[0124] Furthermore, the retrieved original features are paired with the corrected labels provided by engineers to form a "feature-new label" sample pair, which is then stored in an online learning sample pool.

[0125] Furthermore, as drilling progresses and engineers interact multiple times, several new samples will accumulate in the sample pool. The system periodically checks the sample pool, and when the number of samples reaches a certain threshold, it triggers a model fine-tuning.

[0126] Furthermore, the fine-tuning process employs an incremental learning approach. To avoid disrupting the extensive knowledge already learned by the model, this fine-tuning only updates the parameters of the last two fully connected layers in the multi-task deep learning model. Specifically, all other parameters in the model except for these two layers are frozen. Then, using samples from an online learning pool, the parameters of these two fully connected layers are trained on a small number of iterations at a very low learning rate. The training objective is to make the model's predictions on these new samples closer to the corrected labels provided by the engineer.

[0127] Furthermore, once the fine-tuning is complete, the updated model parameters take effect immediately for subsequent real-time inversion. This process can be repeated periodically, allowing the model to continuously incorporate engineers' domain knowledge as drilling operations progress, constantly optimizing its performance on the current well.

[0128] This embodiment introduces a human-computer interaction and online learning loop, organically integrating the real-time experience and knowledge of domain experts into the automatic inversion system, achieving continuous self-optimization and personalized adaptation. Automatic models may err when facing unknown or extreme situations; engineer corrections provide valuable, context-specific supervisory signals. By fine-tuning the model's terminal parameters through incremental learning, its decision boundaries can be quickly and lightweightly adjusted without significantly altering its core feature extraction capabilities, adapting it to the specific characteristics of the current well. This allows the model to "learn from ongoing work," and its performance is expected to continuously improve with increasing depth. Simultaneously, interactive visualization and correction functions enhance user trust and control over the system, transforming the method from a static tool into a dynamic, human-computer collaborative optimization system, ultimately improving the reliability of decision-making.

[0129] Please see Figure 2 , Figure 2 This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a terminal or a server, wherein the server can be a standalone server or a server cluster composed of multiple servers.

[0130] The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0131] The non-volatile storage medium 503 can store an operating system 5031 and a computer program 5032. When the computer program 5032 is executed, it enables the processor 502 to execute a method for inverting borehole quality and formation characteristics based on dynamics and cuttings morphology.

[0132] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0133] The internal memory 504 provides an environment for the operation of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute a method for inverting borehole quality and formation characteristics based on dynamics and cuttings morphology.

[0134] The network interface 505 is used for network communication with other devices. Those skilled in the art will understand that the above structure is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. A specific computer device 500 may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements.

[0135] The processor 502 is used to run a computer program 5032 stored in a memory to implement the steps of a borehole quality and formation characteristic inversion method based on dynamics and cuttings morphology provided in any of the above method embodiments.

[0136] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0137] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program may be stored in a storage medium, which is a computer-readable storage medium. The computer program is executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0138] Therefore, the present invention also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program. When executed by a processor, the computer program causes the processor to perform the steps of the borehole quality and formation characteristic inversion method based on dynamics and cuttings morphology provided in any of the above-described method embodiments.

[0139] The storage medium is a physical, non-transient storage medium, such as a USB flash drive, external hard drive, read-only memory (ROM), magnetic disk, or optical disk, or any other physical storage medium capable of storing program code. The computer-readable storage medium can be non-volatile or volatile.

[0140] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0141] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For example, the division of each unit is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0142] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the device of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0143] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0144] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0145] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Since these modifications and variations fall within the scope of the claims and their equivalents, this invention also intends to include these modifications and variations.

[0146] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for inverting borehole quality and formation characteristics based on dynamics and cuttings morphology, characterized in that, include: Simultaneously acquire drilling dynamics signals and rock cuttings image sequences during the drilling process; High-frequency feature spectrum of drilling dynamics is extracted based on the drilling dynamics signal; Extract the rock debris morphology parameter sequence based on the rock debris image sequence; The time-aligned high-frequency feature spectrum of drilling dynamics and the sequence of cuttings morphology parameters are input into a pre-trained multi-task deep learning model. The borehole wall quality assessment results and formation characteristic identification results generated synchronously by the multi-task deep learning model are obtained. The borehole wall quality assessment results are used to characterize the borehole wall state, and the formation characteristic identification results are used to characterize the formation properties.

2. The method for inverting borehole quality and formation characteristics based on dynamics and cuttings morphology according to claim 1, characterized in that, The synchronous acquisition of drilling dynamics signals and cuttings image sequences during the drilling process includes: The acoustic emission source in the control drill assembly emits synchronous trigger pulse signals at a fixed period. The drilling dynamics signal acquisition device is equipped with a first acoustic receiving sensor, and the cuttings image acquisition device is equipped with a second acoustic receiving sensor. The first acoustic receiving sensor and the second acoustic receiving sensor receive the synchronous trigger pulse signals. The drilling dynamics signal acquisition device controls the first sampling point time when the first acoustic receiving sensor receives the synchronous trigger pulse signal, and calibrates it as the zero time reference of the drilling dynamics signal time axis; The control rock cuttings image acquisition device calibrates the first sampling point time when the second acoustic receiving sensor receives the synchronous trigger pulse signal as the zero time reference of the rock cuttings image time axis; The drilling dynamics signal acquisition device and the cuttings image acquisition device are based on a unified zero-time reference and perform data acquisition with equal time windows.

3. The method for inverting borehole quality and formation characteristics based on dynamics and cuttings morphology according to claim 2, characterized in that, The extraction of high-frequency feature spectrum of drilling dynamics based on the drilling dynamics signal includes: The triaxial acceleration signal and acoustic emission signal are obtained from the drilling dynamics signal; Cross-correlation calculations are performed on the three channels of the triaxial acceleration signal, and an enhanced signal beam is synthesized based on the maximum coherence. The enhanced signal beam points in the direction of contact between the drill bit and the borehole wall. Perform a short-time Fourier transform on the acoustic emission signal to generate the acoustic emission time spectrum; Calculate the number of newly appearing rock debris outlines in a preset unit time in a sequence of rock debris images; When the number of newly appearing rock fragment outlines exceeds a preset threshold within a unit of time, a rock breaking event is determined to have occurred, and the trigger time of the rock breaking event is recorded. Based on the triggering time, a first corresponding time segment of a preset duration is extracted from the enhanced signal beam, and a second corresponding time segment of a preset duration is extracted from the acoustic emission time spectrum; Calculate the total energy of the first signal, the mean of the first instantaneous frequency, the variance of the first instantaneous frequency, and the first spectral entropy within the first corresponding time segment; Calculate the total energy of the second signal, the mean of the second instantaneous frequency, the variance of the second instantaneous frequency, and the second spectral entropy within the second corresponding time segment; The total energy of the first signal, the mean of the first instantaneous frequency, the variance of the first instantaneous frequency, the first spectral entropy, the total energy of the second signal, the mean of the second instantaneous frequency, the variance of the second instantaneous frequency, and the second spectral entropy are combined to form the high-frequency characteristic spectrum of drilling dynamics.

4. The method for inverting borehole quality and formation characteristics based on dynamics and cuttings morphology according to claim 1, characterized in that, The rock debris image sequence includes consecutive frame images of rock debris flow acquired using a binocular vision system. The extraction of a rock debris morphology parameter sequence based on the rock debris image sequence includes: Based on the continuous frame images, a three-dimensional point cloud model of the rock debris flow is obtained through stereo matching and three-dimensional reconstruction. The 3D point cloud model is segmented to obtain multiple point cloud clusters, where each point cloud cluster represents a rock fragment. For each point cloud cluster, calculate the ratio of the surface area to the volume of the point cloud cluster, and use the ratio as the shape factor of the rock debris corresponding to the point cloud cluster; For each point cloud cluster, calculate the standard deviation of the distribution of normal vectors of all points on the surface of the point cloud cluster, and use the standard deviation as the surface roughness index of the rock debris corresponding to the point cloud cluster. For each point cloud cluster, calculate the average length-to-width ratio and length-to-height ratio of the smallest bounding cuboid of the point cloud cluster, and use the average value as the flatness index of the rock debris corresponding to the point cloud cluster. For the current time window, the mean, standard deviation, and distribution histogram information of the shape factor, surface roughness index, and flatness index of all rock fragments within the current time window are statistically analyzed to obtain the sequence of rock fragment morphology parameters for the current time window.

5. The method for inverting borehole quality and formation characteristics based on dynamics and cuttings morphology according to claim 4, characterized in that, The step of extracting the rock debris morphology parameter sequence based on the rock debris image sequence further includes: For each point cloud cluster, identify the fracture surface region in the point cloud cluster and extract the three-dimensional contour of the fracture surface region; Calculate the three-dimensional fractal dimension of the three-dimensional profile of the fracture surface region; Based on the estimated area of ​​the fracture surface region, the volume of the rock fragments corresponding to the point cloud cluster, the three-dimensional fractal dimension, and the pre-calibrated energy conversion coefficient, the equivalent compressive strength factor of the rock fragments corresponding to the point cloud cluster is calculated. For the current time window, calculate the average and variance of the equivalent compressive strength factor of all rock cuttings and add them to the rock cutting morphology parameter sequence for the current time window.

6. The method for inverting borehole quality and formation characteristics based on dynamics and cuttings morphology according to claim 1, characterized in that, The multi-task deep learning model includes a common feature encoder, a gating shared unit, a borehole wall quality assessment branch, and a formation characteristic identification branch. The common feature encoder performs primary fusion of the spliced ​​high-frequency feature spectrum of drilling dynamics and the sequence of rock cutting morphology parameters to obtain primary fused features; The gated shared unit contains a multi-head self-attention network. The multi-head self-attention network takes primary fusion features as input and outputs a first feature selection weight matrix and a second feature selection weight matrix. The first feature selection weight matrix corresponds to the borehole wall quality assessment branch, and the second feature selection weight matrix corresponds to the formation characteristic identification branch. The input features of the pore wall quality assessment branch are the element-wise product of the primary fusion features and the first feature selection weight matrix; The input features of the stratigraphic characteristic identification branch are the element-wise product of the primary fusion features and the second feature selection weight matrix.

7. The method for inverting borehole quality and formation characteristics based on dynamics and cuttings morphology according to claim 6, characterized in that, The steps for generating borehole wall quality assessment results in the borehole wall quality assessment branch include: The input features of the hole wall quality assessment branch are simultaneously fed into the first one-dimensional convolutional path, the second one-dimensional convolutional path, and the third one-dimensional convolutional path. The first one-dimensional convolutional path, the second one-dimensional convolutional path, and the third one-dimensional convolutional path have different time scales, and the first scale features, the second scale features, and the third scale features are extracted respectively. The first scale feature, the second scale feature and the third scale feature are concatenated and calculated through the self-attention layer to obtain the first contribution weight corresponding to the first scale feature, the second contribution weight corresponding to the second scale feature and the third contribution weight corresponding to the third scale feature. Based on the first contribution weight, the second contribution weight, and the third contribution weight, the first-scale feature, the second-scale feature, and the third-scale feature are weighted and fused to obtain the fused feature. Based on the fusion features, the power spectral density integral value of the surface undulation of the hole wall is calculated through the first output submodule, where the power spectral density integral value is used as a roughness quantification index. Based on the fusion features, the second output submodule calculates the coefficient of variation of the borehole diameter along the depth direction, where the coefficient of variation serves as an indicator of borehole diameter stability. Based on the fusion features, the probability value of hole wall instability risk is calculated through the third output submodule. The probability value of hole wall instability risk is calculated based on the vibration energy dissipation rate and ranges from 0 to 1.

8. The method for inverting borehole quality and formation characteristics based on dynamics and cuttings morphology according to claim 6, characterized in that, The steps for generating stratigraphic characteristic identification results in the stratigraphic characteristic identification branch include: The input features of the formation characteristic identification branch are concatenated with the current drilling condition parameter vector and transformed through the fully connected layer to obtain the basic features. The drilling condition parameter vector includes drilling pressure, rotation speed and displacement. The basic features are transformed using a learnable working condition compensation matrix to obtain compensated features. The transformation operation of the working condition compensation matrix is ​​combined with the drilling working condition parameter vector. The compensated features are used to eliminate the influence of working condition changes. Based on the compensated features, the classification probability distribution of stratigraphic lithology is generated through the lithology classification output head; Based on the compensated features, a predicted value of dynamic rock hardness is generated through a hardness regression output head. Based on the compensated characteristics, a normalized index of the critical confining pressure for the brittle-plastic transition of rocks is generated through the critical index output head.

9. A computer device, characterized in that, The computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, can implement the method as described in any one of claims 1-8.

Images