Intelligent optimization method and system for rock-soil drilling strategy

Abstract

The invention relates to the cross technical field of artificial intelligence and geotechnical engineering, discloses an intelligent optimization method and system for a geotechnical drilling strategy, and aims to solve the problems that a traditional drilling strategy depends on artificial experience, parameter adjustment is rigid, the multi-objective collaborative optimization capability is weak and the intelligent level is low. The method comprises the steps that bit pressure, rotating speed, torque, vibration signals and while-drilling rock debris image flow are collected in real time; constructing a stratum drillability index sequence through semantic segmentation and morphological analysis; extracting time-frequency characteristics of the vibration signals to form a drill bit-rock mass action state vector; calculating unit energy consumption efficiency and a torque fluctuation coefficient to perform efficiency evaluation; and inputting the multi-source information into a multi-target reinforcement learning model to generate a candidate parameter combination. By fusing multi-source heterogeneous data and machine intelligence, dynamic sensing and self-adaptive regulation and control in the drilling process are achieved, and the drilling efficiency, the equipment safety and the automation level are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and geotechnical engineering, specifically relating to an intelligent optimization method and system for geotechnical drilling strategies. Background Technology

[0002] With the continuous development of the geotechnical engineering field, drilling operations, as a key link in geological exploration, resource development and infrastructure construction, directly affect the overall progress and safety of the project through their construction efficiency and quality.

[0003] The drilling process involves complex geological conditions and variable process parameters, including drilling pressure, rotation speed, mud properties and drill bit type. The coordinated control of these parameters has a decisive impact on drilling efficiency, drill bit life and borehole stability.

[0004] Among them, the formulation and optimization of drilling strategy is the core task to improve drilling efficiency. Traditional drilling strategies mainly rely on engineers' experience and judgment and limited historical data for manual adjustment. They lack the ability to dynamically respond to real-time working conditions and are difficult to adapt to complex and ever-changing underground environments. Especially when facing heterogeneous rock formations, fractured zones or high-stress areas, serious problems such as improper matching of drilling parameters, low mechanical drilling speed, increased wear of drilling tools, and even stuck drills often occur.

[0005] Existing technologies for optimizing drilling strategies in geotechnical environments suffer from several key shortcomings: First, parameter adjustment mechanisms are rigid, with most systems only supporting preset mode switching and unable to adaptively adjust based on real-time drilling feedback data. Second, decision-making is often based on a single indicator, such as drilling speed or torque, neglecting the coupling relationship and trade-offs between multiple objectives. Third, there is a lack of in-depth analysis of historical successful drilling cases, with knowledge accumulation stored in a scattered document format, failing to construct a computable and reasonable strategy knowledge base. Finally, the level of intelligence is insufficient; existing methods have not yet incorporated machine intelligence for strategy generation and evolution, failing to achieve a fundamental shift from "experience-driven" to "model-driven + knowledge-guided." These problems severely restrict the realization of efficient and safe drilling under complex geological conditions, necessitating an intelligent drilling strategy generation mechanism with autonomous learning and dynamic optimization capabilities. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies by providing an intelligent optimization method and system for geotechnical drilling strategies, which can effectively solve the problems in the background technology. Currently, in geotechnical drilling operations, the setting of drilling parameters generally relies on manual experience or static models, lacking real-time perception and feedback capabilities for dynamic formation responses. This leads to large fluctuations in drilling efficiency, high equipment wear, and increased risk of drill string jamming. Furthermore, traditional optimization methods often adjust parameters based on a single objective function, failing to comprehensively consider the multi-objective coupling relationship between drilling rate, drill bit wear, torque fluctuations, and formation drillability, making it difficult to achieve globally optimal control. In addition, existing systems have low utilization of unstructured geological data (such as core images, drill cuttings morphology, and vibration signals) and weak information fusion capabilities, limiting the accuracy of intelligent decision-making. This invention constructs a machine-intelligent optimization architecture driven by multi-source heterogeneous data, achieving dynamic perception, state deduction, and adaptive strategy adjustment of the drilling process, significantly improving the intelligence level and engineering reliability of drilling operations.

[0007] To achieve the above objectives, the present invention provides the following technical solution: On one hand, an intelligent optimization system for rock and soil drilling strategies, comprising the following components: a data acquisition module, used to synchronously acquire multi-dimensional real-time working condition data during the drilling process, including drilling pressure, rotational speed, torque, feed rate, circulating fluid pressure, drill bit vibration acceleration signal, and drilling cuttings image stream; a geological information analysis module, used to perform pixel-level semantic segmentation on the acquired cuttings images, identify different mineral compositions and particle morphologies, and establish a local formation drillability index sequence in conjunction with the drilling depth; and a state feature extraction module, used to perform time-frequency domain joint analysis on the vibration acceleration signal, extracting the signal's envelope spectrum peak frequency, wavelet energy entropy, and Hilbert marginal spectrum dominant frequency offset as the drill bit-rock cuttings... The system comprises: a feature vector representing the interaction state of the drilling rig; a drilling efficiency evaluation module, used to construct a unit energy consumption drilling efficiency index based on drill pressure, rotation speed, and feed rate, and to calculate the stability score of the current drilling state in conjunction with the torque fluctuation coefficient; a strategy generation engine, used to receive geological information analysis results, state feature vectors, and efficiency evaluation scores, input them into a pre-trained multi-objective reinforcement learning model, and output a set of candidate drilling parameter combinations and their corresponding expected efficiency prediction values; a parameter optimization module, used to perform Pareto front search on multiple objective functions according to preset weight coefficients, and select the drilling parameter suggestion value with the highest comprehensive score; and a parameter execution and feedback module, used to send the optimized drilling parameters to the drilling rig control system and continuously monitor the actual response data after parameter execution, forming a closed-loop control circuit. Preferably, the geological information parsing module adopts a dual-branch convolutional neural network structure, wherein the main branch is used to perform the semantic segmentation task of the rock debris image and output the mineral category label of each pixel, and the auxiliary branch is used to extract the geometric feature parameters of the rock debris particles, including average particle size, roundness, aspect ratio and angularity. The geometric feature parameters and the mineral category distribution histogram together constitute the rock debris characterization vector. Furthermore, when performing semantic segmentation, the geological information parsing module adopts a hollow spatial pyramid pooling structure to expand the receptive field. At the same time, it introduces a channel attention mechanism to perform weighted fusion of feature maps output by different convolutional layers, thereby improving the recognition accuracy of tiny mineral particles. The loss function of the segmentation network is composed of a linear weighted sum of cross-entropy loss and Dice loss, with a weight ratio of 0.6:0.4. In addition, when performing wavelet energy entropy calculation, the state feature extraction module uses the db4 wavelet basis to decompose the vibration signal into 5 layers, calculates the normalized energy value of the detail coefficients of each layer, and uses its negative Shannon entropy as the wavelet energy entropy index. This index is highly sensitive to high-frequency energy diffusion caused by drill bit passivation. Preferably, when the state feature extraction module extracts the Hilbert marginal spectrum main frequency offset, it first performs empirical mode decomposition on the original vibration signal, selects the top 3 intrinsic mode function components related to the drilling main frequency, performs Hilbert transformation on them respectively and calculates the marginal spectrum, and counts the sum of the absolute values ​​of the offset of the main frequency components relative to the standard formation response spectrum as the state criterion. Furthermore, in the drilling efficiency evaluation module, the drilling efficiency index per unit energy consumption is defined as the drilling length achieved per kilojoule of energy input, in millimeters per kilojoule. The torque fluctuation coefficient is the ratio of the standard deviation of torque to the mean within the sliding time window, and the time window length is set to 10 seconds. Furthermore, the multi-objective reinforcement learning model used by the policy generation engine is built on the Deep Deterministic Policy Gradient (DDPG) framework. Its reward function consists of three weighted sub-items: the footage per unit time, the reciprocal of the drill bit wear per unit length, and the reciprocal of the torque fluctuation coefficient, with weight coefficients of 0.5, 0.3, and 0.2, respectively. A priority experience replay mechanism is introduced during model training to accelerate convergence. Preferably, when performing Pareto front search, the parameter optimization module uses a non-dominated sorting genetic algorithm (NSGA-II) with a population size of 100, an evolutionary generation of 200, a crossover probability of 0.9, a mutation probability of 0.1, and optimization variables including drilling pressure (range 5-30 kN), rotational speed (range 20-150 rpm), and propulsion speed (range 0.5-5.0 m / h). Furthermore, when the parameter optimization module selects the final recommended parameters, it introduces a safety constraint verification mechanism to compare the recommended parameters with the working condition range corresponding to historical stuck drill events. If there is an overlap, the recommended value of the advance speed is automatically reduced until it leaves the high-risk range. In addition, after receiving the optimization parameters, the parameter execution and feedback module adopts a phased and gradual adjustment strategy, decomposing the target parameters into 5 equally spaced adjustment steps, with an interval of 15 seconds between each step, to avoid mechanical shock caused by sudden parameter changes. On the other hand, a smart optimization method for rock and soil drilling strategies is proposed, the specific steps of which are as follows: Step S110: Real-time acquisition of working condition data during the drilling process through multiple types of sensors deployed on the drilling rig. The working condition data includes drilling pressure, rotational speed, torque, feed rate, circulating fluid pressure and triaxial vibration acceleration signal. Simultaneously, the drilling cuttings image stream captured by the image acquisition device is acquired. All data are aligned by timestamp and uploaded to the central processing unit. Step S120: Perform pixel-level semantic segmentation and particle morphology analysis on each frame of the rock cuttings image stream, identify the distribution ratio of quartz, feldspar, clay minerals and detrital particles, calculate the average particle size, roundness and angularity, and establish a formation drillability index sequence in combination with the current drilling depth. Step S130: Perform wavelet packet decomposition and Hilbert-Huang transform on the vibration acceleration signal to extract the wavelet energy entropy, envelope spectrum peak frequency and Hilbert marginal spectrum main frequency offset of the signal, and construct a feature vector characterizing the interaction state between the drill bit and the rock mass. Step S140: Calculate the unit energy consumption drilling efficiency based on real-time drilling pressure, rotation speed and propulsion speed. At the same time, calculate the ratio of the standard deviation and mean of the torque within the sliding time window as the torque fluctuation coefficient, and comprehensively generate the performance evaluation score of the current drilling state. Step S150: The formation drillability index sequence, state feature vector and performance evaluation score are input into the multi-objective reinforcement learning model, and the model infers and generates a set of candidate drilling parameter combinations and their corresponding multi-dimensional performance prediction values. Step S160: Perform Pareto front analysis on the candidate parameter combinations, screen out the recommended parameters that achieve the optimal balance between drilling efficiency, equipment wear and operation stability, and output them to the drilling rig control system for execution after safety constraint verification. Step S170: Monitor the actual drilling response data after the recommended parameters are executed, and feed the newly collected data back to the data acquisition module to form a closed-loop optimization cycle and continuously iterate and update the drilling strategy. Preferably, in step S120, the semantic segmentation of the rock debris image adopts a U-Net network structure, in which the encoder part introduces ResNet-34 pre-trained weights, the decoder part is equipped with skip connections and sub-pixel convolutional upsampling layers, and the network output is the class probability distribution of each pixel, with a maximum number of classes of 8. Furthermore, in step S120, when performing particle morphology analysis, morphological closing operation is used to eliminate image noise, independent particle contours are extracted through connected component analysis, and the minimum circumscribed ellipse is fitted to each contour to calculate the major axis, minor axis and orientation angle, thereby deriving the roundness and aspect ratio. In addition, in step S130, the wavelet packet decomposition uses the Symlets wavelet basis, with a decomposition layer of 6 layers and a frequency band covering 0-2000 Hz. The energy percentage of each layer is used to calculate the energy entropy index, which shows a monotonically increasing trend with the degree of drill bit wear. Preferably, in step S150, the multi-objective reinforcement learning model uses historical drilling data to construct a state transition sequence during the offline training phase. The state space includes the working condition feature vector and formation index of the first 5 seconds, and the action space is the incremental adjustment of drilling pressure, rotation speed and thrust speed. The adjustment range of each step is limited to within ±10%. Furthermore, in step S160, during the Pareto front analysis, a hypervolume indicator is calculated for each candidate solution to measure its dominance in the multi-objective space, and the final recommended solution is the non-dominated solution with the largest hypervolume value. In addition, in step S170, the closed-loop feedback cycle is set to 30 seconds, and a complete optimization process is triggered after each new data acquisition to ensure that the strategy update frequency matches the dynamic changes in drilling. Compared with the prior art, the present invention has the following beneficial effects: By integrating semantic information from rock cuttings images with time-frequency characteristics of vibration signals, dual dynamic perception of formation drillability and drill bit status is achieved. Compared with traditional single-parameter monitoring methods, the accuracy of status identification is improved to over 92%. A policy generation mechanism based on multi-objective reinforcement learning was constructed, which can automatically optimize the drilling efficiency, equipment protection and operational safety. Field tests show that the average drilling speed is increased by 18.7% and the drill bit replacement frequency is reduced by 23.4%. By introducing a closed-loop feedback and progressive parameter adjustment mechanism, mechanical stress shocks caused by sudden changes in control commands were effectively avoided, resulting in a 41.2% reduction in the failure rate of the drilling rig's transmission system. A dual screening strategy of non-dominated sorting and safety constraints is adopted to ensure that the recommended parameters meet both the requirements of optimal performance and are within the safe operating range, and the early warning response time for high-risk working conditions is shortened to less than 15 seconds. The system has strong generalization capabilities, can adapt to different geological conditions and drilling rig models, and has an adaptive parameter adjustment coverage of over 95%, significantly reducing reliance on operator experience. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the overall technical solution architecture of the "Intelligent Optimization Method and System for Rock and Soil Drilling Strategy" proposed in this invention. Figure 2 This is a schematic diagram of the core principle framework of {multi-objective reinforcement learning-driven drilling strategy generation and Pareto optimization} in this invention. Detailed Implementation

[0009] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0010] Example 1 Please refer to Figure 1 box Figure 2 In a deep mineral resource exploration project, drilling operations faced complex and varied composite rock formations, including alternating layers of hard granite, weak shale, and fractured zones. Traditional methods of adjusting drilling parameters based on manual experience resulted in drastic fluctuations in drilling efficiency, severe drill bit wear, and numerous instances of sudden torque increases and stuck drill bits. To address these issues, this embodiment deploys an intelligent optimization system for rock and soil drilling strategies. Integrated into the drilling rig control platform, this system dynamically generates and executes the optimal drilling strategy by real-time sensing of formation response and equipment status, achieving closed-loop control throughout the entire process.

[0011] After system startup, the data acquisition module begins synchronously collecting multi-dimensional operating condition data. This module consists of a high-precision torque sensor, a triaxial accelerometer, a laser rangefinder, an electromagnetic flowmeter, and an industrial-grade CMOS image acquisition unit, deployed near the drill bit end of the drill pipe and at the mud return outlet. The drill pressure sensor has a range of 0-50 kN, a resolution of 0.1 kN, and a sampling frequency of 100 Hz; the rotational speed is measured by an encoder, with a range of 0-200 rpm and an accuracy of ±0.5 rpm; the feed rate is measured by a displacement sensor, in m / h, with a minimum resolution of 0.01 m / h; the circulating fluid pressure sensor has a range of 0-20 MPa and an accuracy of ±0.1 MPa; the triaxial vibration accelerometer is installed on the drill pipe body, with a measurement frequency band of 0.1-5000 Hz, a dynamic range of ±50 g, and a sampling rate of 5120 Hz. The image acquisition device is located in the cuttings return channel and is equipped with a ring-shaped LED light source and a dustproof glass cover. It captures a 1920×1080 resolution color image every 5 seconds, with a stable frame rate of 2 fps, ensuring the formation of a cuttings image stream during continuous drilling. All sensors and imaging devices are connected to the central processing unit via industrial Ethernet and are synchronized using the IEEE 1588 precision time protocol, with timestamp alignment accuracy better than 1 millisecond, ensuring strict alignment of multi-source data in the time dimension.

[0012] In step S120, the geological information parsing module processes the rock debris image stream. Each frame of the image first undergoes white balance correction and gamma transformation to enhance contrast, and then is input into a semantic segmentation network based on the U-Net architecture. The encoder of this network uses ResNet-34 pre-trained weights for initialization and contains 5 downsampling levels. Each layer uses 3×3 convolutions and batch normalization operations, with ReLU as the activation function. After the 3rd, 4th, and 5th levels, a hollow spatial pyramid pooling (ASPP) structure is introduced, using convolutional kernels with dilation rates of 1, 6, 12, and 18 to extract multi-scale contextual information in parallel. The output feature maps are then upsampled after 1×1 convolution fusion. The decoder fuses the feature maps of the corresponding levels of the encoder through skip connections and uses sub-pixel convolution (Pixel Shuffle) to achieve high-fidelity upsampling, finally outputting the category probability distribution of each pixel. The network pre-defines 8 mineral labels: quartz, potassium feldspar, plagioclase, mica, calcite, clay minerals, rock debris fragments, and background. The loss function consists of a linearly weighted cross-entropy loss and a Dice loss, with weights of 0.6 and 0.4, respectively, to balance class imbalance and boundary segmentation accuracy. During training, a dataset containing 120,000 labeled images was used, covering different lighting, humidity, and particle stacking states. The model achieved an average crossover-union (mIoU) of 0.87 on the validation set.

[0013] After semantic segmentation, the system performs particle morphology analysis. The original image undergoes morphological closing operations (structuring element is a disk with a radius of 3 pixels) to eliminate porosity and noise, followed by binarization. The threshold is adaptively determined using the Otsu algorithm. Connected component analysis is used to extract independent particle contours, filtering target regions with an area greater than 50 pixels and less than 10,000 pixels. A minimum bounding ellipse is fitted to each contour to obtain the major axis length L, minor axis length S, and orientation angle θ. Roundness is defined as 4π × area / (perimeter)², ranging from 0 to 1; values ​​closer to 1 indicate a more circular particle. The aspect ratio is defined as L / S, reflecting particle extensibility. Angularity is obtained by calculating the ratio of the number of extreme points of contour curvature to the normalized perimeter, used to quantify the sharpness of particle edges. Distribution histograms for each mineral category are statistically analyzed, and combined with the average particle size (arithmetic mean of the equivalent diameters of all particles), average roundness, and average angularity, a 128-dimensional rock debris characterization vector is formed. This vector is bound to the current drilling depth and is input into the formation drillability index calculation unit.

[0014] The Drillability Index (DI) is constructed based on a database of cuttings characterization vectors and known lithology-drilling resistance mappings. This database is calibrated through laboratory core drilling tests and covers the drilling depth per unit energy consumption for 20 typical rock samples under standard parameters. The system weights the proportions of each mineral component in the current cuttings with grain morphology parameters to synthesize a comprehensive hardness factor H. The weights are dynamically adjusted based on the Mohs hardness of the minerals and the grain angularity. The DI calculation formula is as follows: in, This is the normalized comprehensive hardness factor, with a value of 0-1; The porosity estimate is derived from the clay mineral content and particle bulk density. and The empirical coefficients are set to 0.7 and 0.3, respectively. The DI sequence is generated with a sliding window step of 1 meter to form a local drillability curve that varies with drilling depth, which is used to characterize the vertical heterogeneity of the formation.

[0015] In step S130, the state feature extraction module processes the triaxial vibration acceleration signal. After removing DC offset and high-frequency noise through a 0.5-2000Hz bandpass filter, the original signal undergoes wavelet packet decomposition. The Symlets wavelet basis (sym4) is selected, with 6 decomposition levels, forming 64 equal-bandwidth sub-bands, each approximately 31.25 Hz wide. The energy proportion of the detail coefficients in each level is calculated and normalized. The wavelet energy entropy is defined as: in, For the first The energy percentage of each frequency band. This indicator is extremely sensitive to high-frequency energy diffusion caused by drill bit tooth wear; as the drill bit becomes dull, It exhibits a monotonically increasing trend. In this embodiment, when Wear warning is triggered when it exceeds 0.85.

[0016] Simultaneously, the system performs a Hilbert-Huang transform. First, empirical mode decomposition (EMD) is performed on the vibration signal to obtain several intrinsic mode functions (IMFs). The top three IMF components are selected, and their instantaneous frequency range matches the drilling rig's main rotational frequency (f_r) and its harmonics (2f_r, 3f_r). A Hilbert transform is performed on each selected IMF to construct a time-frequency amplitude spectrum, and then the Hilbert marginal spectrum—the integral of the amplitude over time—is calculated. The absolute values ​​of the offsets of the dominant frequency components (the frequencies corresponding to the maximum energy values) in each marginal spectrum relative to the dominant frequency of the standard granite formation response spectrum are calculated, and the sum of these three values ​​is taken as the Hilbert marginal spectrum dominant frequency offset Δf_hilbert. This index reflects the change in the contact stiffness between the drill bit and the rock mass; a significant increase in Δf_hilbert indicates a possible entry into a weak interlayer or eccentric vibration.

[0017] The peak frequency of the envelope spectrum is obtained through Hilbert envelope demodulation. The original vibration signal is bandpass filtered (center frequency 1200 Hz, bandwidth 400 Hz), and a Hilbert transform is performed to obtain an analytic signal. The magnitude of this analytic signal is then used to form the envelope signal. A Fast Fourier Transform (FFT) is performed on this envelope signal, and the frequency component with the highest amplitude is identified as the peak frequency f_envelope of the envelope spectrum. This frequency is typically related to the drill bit impact cycle or the characteristic frequency of bearing failure, and its stability can be used to determine the operating status of the drill bit.

[0018] The three indicators mentioned above—wavelet energy entropy, envelope spectrum peak frequency, and Hilbert marginal spectrum dominant frequency offset—constitute a 3-dimensional state feature vector, which reflects the dynamic characteristics of the drill bit-rock interaction in real time.

[0019] In step S140, the drilling efficiency evaluation module calculates the drilling efficiency per unit energy consumption and the torque fluctuation coefficient. Drilling efficiency per unit energy consumption Defined as: in, The drilling length increment (in mm) within a 10-second sliding window. This is the average value (unit: kW) of the product of drilling pressure and drilling speed within this window. =10 s. The result is in mm / kJ. This indicator directly measures energy utilization efficiency; a higher value indicates more efficient drilling.

[0020] Torque ripple coefficient Defined as: in, The standard deviation of torque within a 10-second window (unit: kN·m). This represents the average torque. It characterizes the stability of the drilling process. When it exceeds 0.15, it is considered a high-fluctuation state, which is prone to mechanical fatigue and vibration instability.

[0021] The efficiency score S_efficiency was synthesized using a weighted normalization method. and Z-score standardization is performed separately, and a comprehensive score is calculated by combining weights of 0.6 and 0.4. The score ranges from 0 to 1 and is output to the policy generation engine in real time.

[0022] In step S150, the policy generation engine receives three types of inputs: a formation drillability index sequence (DI values ​​for the first 10 meters), a current state feature vector (3-dimensional), and a performance evaluation score (1-dimensional), collectively forming a 14-dimensional state space. This engine is equipped with a pre-trained multi-objective reinforcement learning model, built on the Deep Deterministic Policy Gradient (DDPG) framework. The model includes an Actor network and a Critic network, both using fully connected structures. The Actor network takes the state vector as input and outputs adjustments (action space) for drilling pressure, rotational speed, and propulsion speed. Each adjustment is limited to ±10% of the current value, with absolute values ​​not exceeding ±3 kN, ±15 rpm, and ±0.5 m / h, respectively. The Critic network evaluates the value of state-action pairs, and its reward function R is composed of three weighted parts: in, The advance per unit time (m / h) is the amount of footage. The wear rate per unit length of drill bit (μm / m) is estimated using an integral relationship model between vibration energy entropy and drilling pressure. During training, 2000 hours of historical drilling data are used to construct state transition sequences. A prioritized experience replay mechanism is employed, sampling based on the magnitude of the TD error, significantly improving learning efficiency. After offline training, the model is deployed on an embedded GPU platform, achieving an inference latency of less than 50 ms.

[0023] The model generates 100 sets of candidate parameter combinations on each call. Each set includes the target drilling pressure, rotation speed, and propulsion speed values, and outputs corresponding multi-dimensional performance prediction values: expected footage, expected wear rate, and expected torque fluctuation coefficient.

[0024] In step S160, the parameter optimization module performs Pareto front analysis on 100 candidate solutions. The optimization objectives are to maximize the drilling depth, minimize the wear rate, and minimize the torque fluctuation coefficient. A non-dominated sorting genetic algorithm (NSGA-II) is used, with a population size of 100, 200 generations, a crossover probability of 0.9, and a mutation probability of 0.1. The decision variables are drill pressure (5-30 kN), rotational speed (20-150 rpm), and thrust rate (0.5-5.0 m / h), encoded using real numbers. Non-dominated sorting and crowding calculations are performed on each generation of the population, and the front solution set is preserved.

[0025] After obtaining the Pareto front, the system calculates the hypervolume indicator for each non-dominated solution, using a reference point (expansion rate = 0.8 m / h, wear rate = 80 μm / m, fluctuation coefficient = 0.25) as a benchmark to measure its dominance in the target space. The final recommended solution is the candidate parameter combination with the largest hypervolume value.

[0026] Then, a safety constraint verification is performed. The system has a built-in database of historical stuck drill events, recording the operating conditions during which stuck drill events occurred: drill pressure > 25 kN, feed rate > 3.5 m / h, and torque fluctuation coefficient > 0.18. If the recommended parameters fall into this high-risk range, a speed reduction mechanism is automatically activated: the feed rate is reduced in increments of 0.2 m / h until it leaves the risk range or drops to the minimum allowable value of 1.0 m / h. After the verification passes, the final optimized parameters are output.

[0027] In step S170, the parameter execution and feedback module receives the recommended parameters and adopts a phased, gradual adjustment strategy. The target parameter is decomposed into five equally spaced adjustment steps, each 15 seconds apart. For example, if the current drilling pressure is 18 kN and the target is 22 kN, then it increases by 0.8 kN every 15 seconds to avoid sudden pressure changes in the hydraulic system. The rotational speed and feed rate are adjusted synchronously and proportionally. The control system sends commands to the drilling rig's hydraulic pump station and motor driver via the CAN bus, with an execution accuracy error of less than ±2%.

[0028] The system is set to a closed-loop feedback cycle of 30 seconds. Each time new data is collected, a complete optimization process is triggered. The central processing unit initiates a new round of execution from S110 to S170, forming a continuously iterative closed-loop control. In this embodiment, during a 12-hour continuous drilling test, the system completed 1440 optimization cycles, with an average cycle time of 28.7 seconds, meeting real-time requirements.

[0029] Field application results show that, compared to manual control, this system increased the average mechanical drilling speed from 2.1 m / h to 2.5 m / h, an increase of 19.0%; the cumulative drill bit footage increased from 85 meters to 111 meters; and the wear rate decreased by 23.8%. The average torque fluctuation coefficient decreased from 0.16 to 0.11, and the RMS value of the transmission system vibration acceleration decreased by 37%. A total of 6 potential stuck drill risks were identified and avoided, with response times all less than 12 seconds. During stable system operation, no manual intervention was required, significantly reducing reliance on operator experience.

[0030] Example 2 To address the extreme geological conditions encountered in subsea natural gas hydrate drilling operations, characterized by alternating distributions of ultra-soft sedimentary layers and hard nodules, this embodiment proposes a modified intelligent optimization system for rock and soil drilling strategies with greater adaptability. Compared to Embodiment 1, the core difference lies in the reconstruction of the technical approach for the state feature extraction module and the reward function for the strategy generation engine, in order to address the problems of low signal-to-noise ratio of vibration signals and the failure of traditional feature extraction in low-intensity strata.

[0031] In this scenario, the vibration signal during drilling is mainly dominated by the elastic deformation of the drill pipe and the disturbance of mud flow, and the drill bit-rock contact signal is severely submerged, resulting in poor stability of wavelet energy entropy and envelope spectrum features. Therefore, the state feature extraction module in this embodiment abandons the traditional time-frequency analysis method and instead adopts an unsupervised feature learning architecture based on a deep autoencoder. Specifically, the original triaxial vibration signal is filtered against anti-aliasing and then segmented in 5-second windows at a sampling rate of 2048Hz, forming a 10240×3 time-series matrix. This matrix is ​​input to a stacked autoencoder network, which contains four encoding layers and four decoding layers, with 512, 256, 128, and 64 neurons per layer, respectively, and the activation function is LeakyReLU (slope 0.01). The network is unsupervised pre-trained by minimizing the mean square error between the input and the reconstructed output to learn a low-dimensional manifold representation of the vibration signal. After training, the encoder part is frozen, and the 64-dimensional vector of the last hidden layer is extracted as the data-driven state feature vector. This vector does not rely on manually designed features and can automatically capture nonlinear dynamic patterns during the switching between soft and hard formations.

[0032] Meanwhile, the geological information analysis module enhances its ability to identify fine particles. Due to the high water content and easy aggregation of rock fragments caused by hydrate decomposition, this embodiment adds an ultrasonic atomization dispersion unit during the image preprocessing stage. This unit performs 5 seconds of ultrasonic processing on the rock fragment samples (frequency 40 kHz, power 50 W), effectively separating aggregated particles. The semantic segmentation network adjusts its category system, adding "hydrate residual phase" and "ice crystal precipitation phase" labels, expanding the network output to 10 categories. Fractal dimension is introduced as a new feature in particle morphology analysis. The complexity of the particle outline is calculated using box counting to identify rock mass disintegration characteristics caused by hydrate decomposition.

[0033] The drilling performance evaluation module maintains the unit energy consumption efficiency calculation method, but adjusts the time window length of the torque fluctuation coefficient to 5 seconds to adapt to the rapidly changing torque response in soft formations. The performance score generation logic remains unchanged.

[0034] The multi-objective reinforcement learning model of the policy generation engine undergoes structural adaptation. Because the drill bit wear mechanism in soft formations changes from cutting to compression, the weights of the wear term in the original reward function become invalid. This embodiment reconstructs the reward function as follows: in, The axial force fluctuation amplitude is defined as the difference between the maximum and minimum drilling pressure within a 5-second window, in kN. This parameter is used to suppress axial impact caused by drill bit jumping between hard and soft interfaces, ensuring borehole integrity. The model was retrained using historical subsea drilling data, totaling 1200 hours of valid data covering hydrate, clay, and gravel layers. A curriculum learning strategy was introduced during training, first learning simple layer transitions and then gradually introducing complex working conditions.

[0035] The parameter optimization module maintains the NSGA-II algorithm framework, but the range of optimization variables has been adjusted: the drill pressure range is reduced to 3-20 kN (to prevent hydrate layer perforation), the rotation speed range is 40-120 rpm (to avoid excessive shear heat causing hydrate decomposition), and the propulsion speed is 0.3-3.0 m / h. The safety constraint library has been updated to a seabed-specific risk model: when the drill pressure is <5 kN and the propulsion speed is >2.0 m / h and When the load is greater than 4 kN, it is determined to be a high-risk state of "drill bit drift" and a speed reduction check is triggered.

[0036] The parameter execution strategy still employs a 5-step incremental adjustment, but the interval between each step is shortened to 10 seconds to accelerate the response to sudden changes in soft formations. The feedback cycle is also adjusted to 20 seconds to increase the control frequency.

[0037] In a field test at a pilot well in the South China Sea, the system successfully traversed a hydrate-rich layer up to 42 meters thick. During the test, seven hard nodules were identified, and the system was able to predict and slow down 1.5 meters in advance to avoid drill bit impact damage. The average drilling speed reached 1.8 m / h, a 21.3% improvement over traditional methods; the average axial force fluctuation decreased from 6.2 kN to 3.8 kN, and borehole collapse events were reduced by 80%. The system operated continuously for 76 hours without failure, verifying its robustness and adaptability under extreme geological conditions.

[0038] Example 3 To address the need for precise drilling in densely piped areas using micro-tunnel jacking machines in urban underground space development, this embodiment constructs an embedded version of a lightweight, low-power intelligent optimization system for geotechnical drilling strategies. Its substantial difference from the aforementioned embodiments lies in the modular integration of the system architecture and the lightweight design of the algorithm, focusing on solving the real-time optimization problem in scenarios with limited edge computing resources.

[0039] This embodiment eliminates the independent geological information analysis module, instead employing an end-to-end parameter prediction architecture based on a lightweight convolutional neural network. The image acquisition device is simplified to a low-resolution (640×480) CMOS sensor, capturing one frame of rock debris image every 10 seconds. Images are directly input into the MobileNetV3-small network; the backbone network's output feature map is globally averaged and then concatenated with the vibration signal features. Vibration signal processing uses a fast spectral kurtosis algorithm instead of wavelet packet decomposition: calculating the fourth-order cumulant spectrum, identifying the frequency band with the maximum kurtosis as the optimal demodulation frequency band, and then extracting the peak frequency and amplitude of the envelope spectrum. This method reduces the computational complexity from O(N squared) to O(N log N), making it suitable for embedded deployment.

[0040] The state feature vector is a 3-dimensional vector composed of the envelope spectrum peak frequency, axial acceleration RMS value, and rotational phase consistency index (calculated by cross-correlation function to determine the synchronicity of the three-axis signals). The performance evaluation module only retains the unit energy consumption efficiency calculation and omits the torque fluctuation coefficient evaluation to reduce the computational load.

[0041] The policy generation engine employs a compressed DDPG model: the Actor network is compressed to 3 fully connected layers (128-64-32 neurons), and the Critic network is similarly compressed. The model is obtained through knowledge distillation from the large-scale model in Example 1, reducing the input dimension to 6 dimensions (DI index, 3-dimensional state features, efficiency score, and previous action). The action space remains the adjustment of drilling pressure, rotation speed, and propulsion speed, but the adjustment granularity is widened to ±0.5 kN, ±5 rpm, and ±0.2 m / h, respectively.

[0042] The parameter optimization module simplifies the Pareto search process: instead of generating 100 candidate solutions, the Actor network directly outputs one optimal action, which is then executed after rapid verification by the safety rule engine. The safety rule is preset as follows: if a sudden change in the peak frequency of the envelope spectrum (±20% threshold) is detected, the advance speed is forcibly reduced to 1.0 m / h while maintaining the drilling pressure, and the response is observed for 15 seconds.

[0043] The parameters are adjusted in three progressive steps, with a 20-second interval, to accommodate the slow hydraulic response of the miniature pipe jacking machine. The feedback cycle is set to 60 seconds to balance real-time performance and power consumption.

[0044] The system is deployed on an embedded industrial computer with an ARM Cortex-A72 quad-core processor (1.8 GHz) and 4GB of memory. The entire algorithm takes an average of 820 ms to run and consumes less than 15 W of power. During the construction of an underground utility tunnel in Beijing, it successfully avoided three unknown obstacles. By identifying the geological strata and adjusting its attitude in advance, the drilling trajectory deviation was controlled within ±5 cm, achieving a 60% improvement in accuracy compared to manual operation, and no pipeline damage occurred.

[0045] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for intelligent optimization of rock and soil drilling strategies, characterized in that, Includes the following steps: Real-time operating data during the drilling process is collected synchronously by multiple types of sensors. The real-time operating data includes drilling pressure, rotational speed, torque, feed rate, circulating fluid pressure, triaxial vibration acceleration signal, and drilling cuttings image stream acquired by the image acquisition device. Pixel-level semantic segmentation is performed on each frame of the rock cuttings image stream to identify the distribution ratio of quartz, feldspar, clay minerals and detrital particles. The average particle size, roundness, aspect ratio and angularity are calculated by combining connected component analysis and minimum circumscribed ellipse fitting. A rock cuttings characterization vector is generated based on the mineral composition ratio and particle morphology parameters. A formation drillability index sequence is established in combination with the current drilling depth. The vibration acceleration signal is subjected to wavelet packet decomposition and empirical mode decomposition to extract wavelet energy entropy, envelope spectrum peak frequency and Hilbert marginal spectrum main frequency offset to form a three-dimensional state feature vector. The drilling efficiency per unit energy consumption is calculated based on real-time drilling pressure, rotation speed and feed rate. At the same time, the ratio of the standard deviation of torque to the mean within the sliding time window is used as the torque fluctuation coefficient. The drilling efficiency per unit energy consumption and the reciprocal of the torque fluctuation coefficient are weighted and normalized to generate a drilling performance evaluation score. The formation drillability index sequence, state feature vector, and performance evaluation score are input into a pre-trained multi-objective reinforcement learning model. The multi-objective reinforcement learning model is constructed based on a deep deterministic policy gradient framework and outputs a set of candidate drilling parameter combinations and their corresponding multi-dimensional performance prediction values.

2. The intelligent optimization method for rock and soil drilling strategy according to claim 1, characterized in that, The U-Net network structure is used for pixel-level semantic segmentation of rock debris images. The encoder part introduces ResNet-34 pre-trained weights, and the decoder part is equipped with skip connections and sub-pixel convolutional upsampling layers. The network output is the class probability distribution of each pixel.

3. The intelligent optimization method for rock and soil drilling strategy according to claim 1, characterized in that, In the particle morphology analysis process, morphological closing operations are used to eliminate image noise, and independent particle contours are extracted through connected component analysis. For each contour, the minimum bounding ellipse is fitted to calculate the major axis, minor axis and orientation angle, and then the roundness and aspect ratio are derived.

4. The intelligent optimization method for rock and soil drilling strategy according to claim 1, characterized in that, The wavelet packet decomposition uses the Symlets wavelet basis, with a decomposition layer of 6 layers and a frequency band covering 0-2000 Hz. The energy percentage of each layer is used to calculate the wavelet energy entropy index.

5. The intelligent optimization method for rock and soil drilling strategy according to claim 1, characterized in that, The extraction of the dominant frequency offset of the Hilbert marginal spectrum includes: performing empirical mode decomposition on the vibration signal, selecting the top three intrinsic mode function components related to the drilling dominant frequency, performing Hilbert transform on each component and calculating the marginal spectrum, and using the sum of the absolute values ​​of the offsets of the dominant frequency components relative to the standard formation response spectrum as the state criterion.

6. The intelligent optimization method for rock and soil drilling strategy according to claim 1, characterized in that, The multi-objective reinforcement learning model uses historical drilling data to construct a state transition sequence during the offline training phase. The state space includes the working condition feature vector and formation index for the first 5 seconds, and the action space consists of the incremental adjustment of drilling pressure, rotation speed, and propulsion speed. The adjustment range for each step is limited to within ±10%.

7. The intelligent optimization method for rock and soil drilling strategy according to claim 1, characterized in that, The non-dominated sorting uses the non-dominated sorting genetic algorithm NSGA-II, and the optimization variables include drilling pressure, rotational speed, and propulsion speed.

8. The intelligent optimization method for rock and soil drilling strategy according to claim 1, characterized in that, The safety constraint verification mechanism for recommended parameters compares the recommended parameters with the working condition ranges corresponding to historical stuck drill events. If there is an overlap, the recommended propulsion speed is automatically reduced until the drill is removed from the high-risk range.

9. The intelligent optimization method for rock and soil drilling strategy according to claim 1, characterized in that, The phased and gradual adjustment strategy breaks down the target parameters into five equally spaced adjustment steps, with each step having an interval of 15 seconds, and gradually adjusts the drilling pressure, rotation speed and propulsion speed proportionally.

10. An intelligent optimization system for rock and soil drilling strategies, characterized in that, include: The data acquisition module is used to synchronously acquire multi-dimensional real-time working condition data during the drilling process, including drill pressure, rotational speed, torque, feed rate, circulating fluid pressure, drill bit vibration acceleration signal, and drilling cuttings image stream. The geological information analysis module is used to perform pixel-level semantic segmentation on the acquired cuttings images, identify different mineral compositions and particle morphologies, and establish a local formation drillability index sequence based on the drilling depth. The state feature extraction module is used to perform time-frequency domain joint analysis on the vibration acceleration signal, extracting the signal's envelope spectrum peak frequency, wavelet energy entropy, and Hilbert marginal spectrum dominant frequency offset as feature vectors of the drill bit-rock interaction state. The drilling efficiency evaluation module is used to evaluate the drilling performance based on… Drilling pressure, rotational speed, and feed rate are used to construct unit energy consumption drilling efficiency indicators, and combined with torque fluctuation coefficients to calculate the stability score of the current drilling state; the strategy generation engine receives geological information analysis results, state feature vectors, and performance evaluation scores, inputs them into a pre-trained multi-objective reinforcement learning model, and outputs a set of candidate drilling parameter combinations and their corresponding expected performance prediction values; the parameter optimization module performs Pareto front search on multiple objective functions according to preset weight coefficients, and selects the drilling parameter suggestion values ​​with the highest comprehensive score; the parameter execution and feedback module sends the optimized drilling parameters to the drilling rig control system and continuously monitors the actual response data after parameter execution, forming a closed-loop control loop.

Images