A hard rock lithology identification and cutting parameter self-adaptive system based on machine vision

Abstract

The present application relates to the field of mine intelligent tunneling, and discloses a hard rock lithology identification and cutting parameter adaptive system based on machine vision, which comprises an image acquisition module, a lithology identification module, a rock hardness mapping module and an adaptive control module. The image acquisition module is integrated with air curtain cleaning and double-mode light supplementing devices to obtain rock surface images of the tunneling working face; the lithology identification module extracts lithology features by using an adaptive enhancement algorithm and an improved deep learning network; the rock hardness mapping module quantitatively outputs rock hardness values through a regression model; and the adaptive control module dynamically adjusts the rotating speed and feeding speed of the cutting head according to the rock hardness values, and smoothes mechanical impact through a parameter smoothing unit. In addition, a feedback optimization module uses cutting load data to perform closed-loop verification and Kalman filter fusion on the visual results. The present application solves the problems of difficult lithology identification and lagging cutting parameter adjustment in the high-dust underground environment.

Description

Technical Field

[0001] This invention relates to the field of intelligent tunneling in mining, specifically to a machine vision-based system for hard rock lithology identification and adaptive cutting parameters. Background Technology

[0002] my country is rich in coal resources, and with the advancement of mining technology, intelligent tunneling has become a key link in achieving efficient and safe coal mine production. In tunneling operations, real-time and accurate identification of the coal and rock lithology ahead of the working face, and adjusting the cutting parameters of the tunneling machine accordingly, is a core technology for improving operational efficiency and ensuring equipment safety.

[0003] Current coal and rock identification technologies mainly include X-ray detection, vibration detection, and image recognition. While X-ray detection can penetrate coal seams of a certain thickness, it poses radiation safety risks and its application in medium-thick coal seams is limited. Vibration detection identifies the cutting medium by analyzing vibration signals from the tunneling machine's body or cutting arm; however, this method is highly susceptible to interference from changes in the tunneling machine's own posture and complex underground mechanical vibration noise, leading to difficulties in signal feature extraction and insufficient identification stability and accuracy.

[0004] With the development of machine vision technology, image analysis-based lithology identification methods have gradually attracted attention. However, existing visual recognition technologies face severe challenges in practical applications in coal mines. First, the working face environment is harsh, with high concentrations of dust, water mist, and low or uneven lighting. Ordinary imaging equipment is prone to dust accumulation on the lens surface, affecting visibility, and the acquired images are often blurry, making it difficult to extract the subtle cracks and texture features of hard rock surfaces using conventional algorithms. Second, single visual recognition methods are prone to misjudgment when the rock surface is covered with coal slurry or dust, lacking an effective multi-source data verification mechanism, resulting in insufficient system reliability. Furthermore, existing adaptive control strategies often directly adjust parameters based on the recognition results, lacking a smoothing mechanism for changes in commands. When the lithology identification results change abruptly, the step change in control parameters can cause current surges in the cutting motor and overload oscillations in the mechanical transmission chain, accelerating equipment wear and affecting the service life of the tunneling machine. Therefore, there is an urgent need for an adaptive system for hard rock lithology identification and cutting parameters that can overcome the interference of extreme underground environments, possess multi-dimensional verification functions, and achieve stable control. Summary of the Invention

[0005] This invention provides a machine vision-based system for hard rock lithology identification and adaptive cutting parameters. The system includes an image acquisition module, a lithology identification module, a rock hardness mapping module, and an adaptive control module. These modules are sequentially connected in communication, forming a closed-loop architecture from environmental perception to execution control.

[0006] The image acquisition module is used to acquire high-resolution visible light images of the rock surface in front of the tunnel boring machine's working face. The lithology identification module is used to process the high-resolution visible light images to output lithology category, rock mass integrity score, and fracture density level. The rock hardness mapping module is used to determine the rock hardness value based on the parameter information output by the lithology identification module. The adaptive control module is used to generate cutting parameter commands based on the rock hardness value and send them to the tunnel boring machine's control system.

[0007] The image acquisition module is rigidly mounted on the cutting arm of the tunneling machine via a rigid mounting bracket. The module is located in the area slightly in front of the cutting head's rotation axis and within 50cm behind it, ensuring the field of view covers the rock surface to be cut in front of the cutting head. The main support and protective structure of the image acquisition module consists of a dustproof and waterproof shell, with an optical window made of high-strength quartz glass at the front end. The image acquisition module integrates a self-cleaning lens assembly and a supplementary lighting device. The self-cleaning lens assembly, located inside the dustproof and waterproof shell, includes a high-resolution industrial camera, a fixed-focus lens, and miniature air curtain nozzles. The miniature air curtain nozzles are arranged in a ring around the fixed-focus lens, receiving compressed air through a gas delivery pipeline and ejecting an airflow to remove dust and water mist adhering to the surface of the high-strength quartz glass. The supplementary lighting device consists of multiple independently controllable LED light sources arranged in a ring array, supporting both visible light and near-infrared dual-mode illumination.

[0008] The lithology identification module is configured in an edge computing device and communicates with the main control system of the tunneling machine via a controller area network bus. The lithology identification module runs a deep learning model, the input of which integrates an adaptive image enhancement unit. This adaptive image enhancement unit is configured to execute the Retinex dehazing algorithm, decomposing the high-resolution visible light image into illumination and reflection components and removing the illumination component; it is also configured to execute a contrast-limited adaptive histogram equalization algorithm, dividing the high-resolution visible light image into multiple non-overlapping small regions and performing histogram equalization independently. The backbone network of the deep learning model uses the GSConvV2 module, employing a multi-path parallel structure to perform convolution operations and concatenate and shuffle the output feature maps; the backbone network also incorporates a multi-scale detail fusion module, concatenating and fusing low-level and high-level features through cross-layer connection channels. The neck network of the deep learning model uses a weighted bidirectional feature pyramid network. The detection head of the deep learning model introduces an iRMB inverted residual block attention mechanism.

[0009] The rock hardness mapping module internally stores a preset lithology-uniaxial compressive strength correspondence table, used to query and obtain the estimated hardness range of the current rock based on the lithology category. The rock hardness mapping module also integrates a lightweight regression model, which uses the lithology feature vector as input data and directly predicts and outputs continuous uniaxial compressive strength values ​​through nonlinear mapping operations.

[0010] The invention also includes a feedback optimization module. This module is communicatively connected to a current sensor, a vibration sensor, a lithology identification module, and a rock hardness mapping module, all located at the cutting section of the tunneling machine. The feedback optimization module acquires the cutting motor current data collected by the current sensor and the cutting head vibration spectrum data collected by the vibration sensor, and calculates the physical inversion hardness value using a load and hardness inversion model. The feedback optimization module compares the predicted rock hardness value output by the rock hardness mapping module with the physical inversion hardness value. When the absolute value of the deviation exceeds a preset tolerance threshold, a Kalman filter algorithm is used to weight and fuse the visual prediction value and the physical inversion value to generate a comprehensive rock hardness value.

[0011] The adaptive control module has a pre-set control parameter mapping table for different rock hardness ranges. The module compares the rock hardness value with a preset hardness threshold to determine the cutting mode: when the cutting range is determined to be soft rock, a parameter combination of high rotational speed and high feed rate is generated; when the cutting range is determined to be hard rock, a parameter combination of low rotational speed and low feed rate is generated. The adaptive control module integrates a parameter smoothing transition unit. This unit uses an exponentially weighted moving average algorithm to calculate the current control output value based on the output value of the previous moment, the theoretical target control value of the current moment, and a preset smoothing coefficient. The unit also includes a rate-of-change limiting logic to calculate the absolute value of the difference between the currently calculated control output value and the output value of the previous moment. When the rate of change of the physical quantity corresponding to the absolute value of the difference exceeds a preset safety threshold, the control output value is forcibly truncated.

[0012] This invention provides a machine vision-based system for hard rock lithology identification and adaptive cutting parameters. It offers the following advantages:

[0013] 1. This invention solves the problem of difficult lithological identification in high-dust, low-illuminance environments in coal mines by combining hardware protection with enhanced algorithms. At the hardware level, a high-speed airflow barrier formed by micro-air curtain nozzles, combined with high-strength quartz glass, effectively blocks interference from dust and water mist on imaging. Furthermore, dual-mode illumination (visible and near-infrared) combined with a wide-band anti-reflective film enhances the penetration of optical imaging. At the algorithm level, an adaptive image enhancement unit removes the fogging effect and balances local contrast. Combined with the improved YOLOv8n model's GSConvV2 module and weighted bidirectional feature pyramid network, the ability to extract fine cracks and texture features from hard rock surfaces is enhanced, thus achieving accurate identification of lithological type and integrity under complex working conditions.

[0014] 2. This invention eliminates the impact of sudden changes in cutting parameters on the mechanical system through a parameter smoothing transition unit, extending the service life of the tunneling machine. Although the system can quickly generate target control commands based on rock hardness, in the actual execution phase, the parameter smoothing transition unit uses an exponentially weighted moving average algorithm and a rate of change limiting logic to forcibly constrain the variation range of rotational speed and feed rate within adjacent control cycles. This approach avoids current surges in the cutting motor and overload oscillations in the mechanical transmission chain caused by jumps in lithology identification results or sudden changes in rock hardness, ensuring the smoothness of the cutting head and cutting arm when switching between rock layers of different hardness.

[0015] 3. This invention constructs a closed-loop feedback mechanism based on the fusion of multiple physical quantities, improving the system's reliability and adaptability under extreme working conditions. The feedback optimization module uses the current data of the cutting motor and the vibration spectrum data of the cutting head to invert physical hardness and performs real-time verification of the visual prediction results. When visual recognition deviates due to factors such as coal slime covering the rock surface, the system can use the Kalman filter algorithm to correct the control strategy through physical feedback data to prevent misjudgment. In addition, the system can use the difficult sample data at this time for subsequent incremental training of the model, enabling the lithology identification model and the mapping relationship between hardness to continuously self-optimize as the geological environment changes. Attached Figure Description

[0016] Figure 1 This is a block diagram of the overall structure of the hard rock lithology identification and cutting parameter adaptive system of the present invention;

[0017] Figure 2 This is a schematic diagram showing the installation location of the image acquisition module of the present invention;

[0018] Figure 3 This is a cross-sectional view of the internal structure of the image acquisition module of the present invention;

[0019] Figure 4 This is a diagram of the improved YOLOv8n model network structure of the present invention;

[0020] Figure 5 This is a structural framework diagram of the GSConvV2 module of the present invention;

[0021] Figure 6 This is a schematic diagram of the MDFM module structure of the present invention;

[0022] Figure 7 This is a schematic diagram of the BiFPN feature fusion network of the present invention;

[0023] Figure 8 This is the adaptive control flowchart of the present invention.

[0024] The components include: 1. Tunneling machine; 2. Cutting head; 3. Cutting arm; 4. Image acquisition module; 41. Dustproof and waterproof housing; 42. High-resolution industrial camera; 43. Fixed focal length lens; 44. High-strength quartz glass; 45. Lighting device; 46. Miniature air curtain nozzle; 5. Gas delivery pipeline; 6. High-flexibility shielded cable; 7. Rock type identification module; 8. Compressed air source; and 9. Rigid mounting bracket. Detailed Implementation

[0025] The technical solutions in 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 embodiments of the present invention, and not all embodiments. 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.

[0026] Please see the appendix Figure 1 - Appendix Figure 3 This invention provides a machine vision-based hard rock lithology identification and cutting parameter adaptive system, including an image acquisition module 4, a lithology identification module 7, a rock hardness mapping module, and an adaptive control module.

[0027] The image acquisition module 4, lithology identification module 7, rock hardness mapping module, and adaptive control module are sequentially connected to form a closed-loop control architecture. Image acquisition module 4 acquires high-resolution visible light images of the rock surface ahead of the tunneling face. Lithology identification module 7 processes the high-resolution visible light images to output lithology category, rock mass integrity score, and fracture density level. Rock hardness mapping module determines the rock hardness value based on the parameters output by lithology identification module 7. Adaptive control module generates cutting parameter commands based on the rock hardness value and sends them to the control system of tunneling machine 1.

[0028] The image acquisition module 4 is fixedly mounted on the body of the cutting arm 3 of the tunneling machine 1 via a rigid mounting bracket 9. The image acquisition module 4 is located in the side-front region of the rotation axis of the cutting head 2, and within a range of 50cm behind the rotation axis of the cutting head 2. The field of view of the image acquisition module 4 covers the rock surface area to be cut in front of the cutting head 2.

[0029] The image acquisition module 4 includes a dustproof and waterproof housing 41, a self-cleaning lens assembly, and a supplementary lighting device 45. The dustproof and waterproof housing 41 is made of 316L stainless steel with an IP66 protection rating. An optical window is located at the front end of the dustproof and waterproof housing 41, and the optical window is composed of high-strength quartz glass 44. The high-strength quartz glass 44 is 5mm thick, has a light transmittance greater than 92%, and a compressive strength greater than or equal to 300MPa.

[0030] The self-cleaning lens assembly is located inside the dustproof and waterproof housing 41. The self-cleaning lens assembly includes a high-resolution industrial camera 42, a fixed-focus lens 43, and a miniature air curtain nozzle 46. The high-resolution industrial camera 42 uses a CMOS sensor with a resolution of 2448×2048, a frame rate of 30fps, and supports the GenICam protocol. The fixed-focus lens 43 is mounted at the front of the high-resolution industrial camera 42. The fixed-focus lens 43 has a focal length of 12mm, an aperture of F1.4, and is coated with a wide-band anti-reflective coating covering a wavelength range of 400nm to 1000nm.

[0031] The miniature air curtain nozzle 46 is arranged in a ring around the outer periphery of the fixed-focus lens 43, located on the outer side of the high-strength quartz glass 44. The miniature air curtain nozzle 46 is connected to a compressed air source 8 via a gas delivery pipeline 5. The compressed air source 8 is a downhole compressed air source. The miniature air curtain nozzle 46 is used to periodically spray clean airflow to remove dust and water mist adhering to the surface of the high-strength quartz glass 44.

[0032] The supplementary lighting device 45 consists of multiple independently controllable LED light sources, which are arranged in a ring array inside the front end of the dustproof and waterproof housing 41. The supplementary lighting device 45 includes a visible light illumination unit and a near-infrared illumination unit, supporting dual-mode illumination of visible light and near-infrared light. The color temperature range of the LED light sources is 5000K to 6500K, and it has PWM dimming function.

[0033] The image acquisition module 4 is connected to the lithology identification module 7 via a highly flexible shielded cable 6. The highly flexible shielded cable 6 is laid along the internal groove of the cutting arm 3, and a length with a bending radius greater than or equal to 200 mm is reserved at the movable joint of the cutting arm 3. The electrical interface of the image acquisition module 4 uses an explosion-proof connector.

[0034] The lithology identification module 7 is configured within the edge computing device. The edge computing device has an IP65 protection rating and is bolted to the main body of the tunneling machine 1. The edge computing device communicates with the main control PLC of the tunneling machine 1 via a CAN bus. The lithology identification module 7 runs on an improved YOLOv8n deep learning model to process the high-definition visible light images acquired by the image acquisition module 4 in real time.

[0035] The rock hardness mapping module is communicatively connected to the lithology identification module 7, receiving lithology type, rock mass integrity score, and fracture density level. The rock hardness mapping module is equipped with a preset lithology-uniaxial compressive strength correspondence table and a lightweight regression model for outputting rock hardness values.

[0036] The adaptive control module is communicatively connected to the rock hardness mapping module and also to the control system of the tunneling machine 1. The adaptive control module is used to dynamically adjust the rotational speed, feed rate, and upper limit of cutting power of the cutting head 2 according to the rock hardness value.

[0037] The machine vision-based hard rock lithology identification and cutting parameter adaptive system also includes a feedback optimization module. This module is communicatively connected to the current sensor, vibration sensor, lithology identification module 7, and rock hardness mapping module located at the cutting section of the tunneling machine 1. The feedback optimization module acquires the cutting motor current data collected by the current sensor, the cutting head vibration spectrum data collected by the vibration sensor, and the tool wear status data, and performs closed-loop correction on the identification results of the lithology identification module 7 and the mapping relationship of the rock hardness mapping module.

[0038] The image acquisition module 4 employs an explosion-proof and protective structure specifically designed for the complex working conditions of underground coal mines. The image acquisition module 4 includes a dustproof and waterproof housing 41, which constitutes the main support and protective structure of the image acquisition module 4. The main body of the dustproof and waterproof housing 41 is made of 316L stainless steel through precision machining. 316L stainless steel has corrosion resistance and high mechanical strength, capable of resisting corrosion from acidic water vapor underground and external physical impacts. The overall protection level of the dustproof and waterproof housing 41 reaches the IP66 standard, completely preventing the intrusion of foreign objects and withstanding strong water spray, thus ensuring the safe operation of internal precision electronic components in the high dust and spray dust suppression environment underground.

[0039] An optical window is provided on the front surface of the dustproof and waterproof housing 41, allowing light to pass through and enter the internal imaging components. The optical window is composed of embedded high-strength quartz glass 44. The high-strength quartz glass 44 is made of fused silica glass and has a thickness of 5 mm. The high-strength quartz glass 44 has excellent optical and mechanical properties. Specifically, its light transmittance is greater than 92%, ensuring effective transmission of light signals in low-light environments underground. Its compressive strength is greater than or equal to 300 MPa, effectively resisting the impact of flying rock fragments and coal chunks generated during tunneling operations, preventing the optical window from breaking and damaging the internal lens and sensors.

[0040] The electrical connections of the image acquisition module 4 employ a highly reliable explosion-proof design. All electrical interfaces on the dustproof and waterproof housing 41 use explosion-proof connectors to prevent electrical sparks from igniting underground gas or coal dust. The image acquisition module 4 transmits power and communicates with the external control unit via a highly flexible shielded cable 6. The highly flexible shielded cable 6 is laid along the internal cable channel of the cutting arm 3 of the tunneling machine 1 for mechanical protection. Considering that the cutting arm 3 needs to perform frequent lifting and swinging movements during operation, the highly flexible shielded cable 6 has a bending radius of greater than or equal to 200 mm at the moving joints of the cutting arm 3 to eliminate stress concentration and prevent fatigue fracture of the cable due to repeated bending. The highly flexible shielded cable 6 has an electromagnetic shielding layer, which can block electromagnetic interference generated by underground frequency converters and high-power motors, ensuring the stability and integrity of image data transmission. The image acquisition module 4 is mechanically connected to the cutting arm 3 via a rigid mounting bracket 9. The rigid mounting bracket 9 provides stable support and reduces the impact of tunneling vibration on imaging quality.

[0041] The self-cleaning lens assembly is located inside the dustproof and waterproof housing 41 of the image acquisition module 4, and is used to acquire high-quality rock surface image data in high dust and low light environments downhole. The self-cleaning lens assembly mainly consists of a high-resolution industrial camera 42, a fixed-focus lens 43, and a miniature air curtain nozzle 46.

[0042] The high-resolution industrial camera 42, as the core component for image acquisition, utilizes an industrial-grade CMOS image sensor. Specifically, the high-resolution industrial camera 42 is configured as the FLIR Blackfly S industrial camera (model BFS-U3-50S5C-C). The high-resolution industrial camera 42 has a resolution of 2448×2048 pixels, a frame rate of 30fps, and supports the GenICam universal camera interface protocol to ensure real-time image data transmission and compatibility with subsequent processing.

[0043] A fixed-focus lens 43 is mounted at the front end of the high-resolution industrial camera 42. The fixed-focus lens 43 has a fixed focal length of 12 mm, capable of covering the cutting area in front of the tunneling machine 1. The fixed-focus lens 43 features a large aperture of F1.4 to increase light transmission and adapt to the dim lighting conditions underground. The lens surface of the fixed-focus lens 43 is coated with a broadband anti-reflective film, which operates in the 400 nm to 1000 nm wavelength range, allowing both visible and near-infrared light to pass through simultaneously, thus enhancing image clarity in conjunction with subsequent image enhancement algorithms.

[0044] To address the issue of dust and water mist obstructing vision underground, the self-cleaning lens assembly integrates an active dust removal structure. Miniature air curtain nozzles 46 are arranged in a ring array, surrounding the fixed-focus lens 43 and located on the inner or outer edge of the high-strength quartz glass 44. The miniature air curtain nozzles 46 are physically connected to a compressed air source 8 underground via a gas delivery pipeline 5. The compressed air source 8 provides clean compressed air, which is delivered to the miniature air curtain nozzles 46 through the gas delivery pipeline 5. The miniature air curtain nozzles 46 can periodically eject high-speed airflow, forming a flowing air curtain on the surface of the high-strength quartz glass 44. This air curtain effectively blows away coal dust particles and water mist condensation adhering to the surface of the high-strength quartz glass 44, maintaining the cleanliness and light transmittance of the optical window and ensuring the imaging quality of the image acquisition module 4 during continuous operation.

[0045] The supplementary lighting device 45 is integrated and installed inside the front end of the dustproof and waterproof housing 41 of the image acquisition module 4, and is arranged in a ring array around the high-strength quartz glass 44. The supplementary lighting device 45 is used to provide a high-intensity and uniform illumination environment for the high-resolution industrial camera 42 in working conditions where there is no natural light and coal dust interference.

[0046] The supplementary lighting device 45 consists of multiple independently controllable LED light sources. In this embodiment, the supplementary lighting device 45 specifically consists of eight Osram (OSLON) SSL 150 LED light sources. The supplementary lighting device 45 supports dual-mode illumination of visible light and near-infrared light, and integrates a visible light illumination unit and a near-infrared illumination unit. The visible light illumination unit emits simulated sunlight with a color temperature of 5500K to reproduce the true color and texture of the rock surface. The near-infrared illumination unit emits light in the near-infrared band, utilizing the strong penetrating power of near-infrared light in dust particle media, combined with the wide-band anti-reflection characteristics of the fixed-focus lens 43, to enhance the imaging contrast of rock fissures and bedding features in heavily dusty environments.

[0047] The supplementary lighting device 45 features PWM (Pulse Width Modulation) dimming functionality. By receiving control signals, the device changes the pulse width duty cycle of the drive current, thereby achieving stepless adjustment of the output luminous flux. Based on the grayscale feedback from the image captured by the high-resolution industrial camera 42 or ambient lighting conditions, the device automatically adjusts its brightness to prevent overexposure due to proximity to the rock surface or underexposure due to excessive distance. The LED module of the supplementary lighting device 45 is encapsulated, giving it an IP68 protection rating, ensuring electrical insulation and luminous stability in high-humidity and water-spraying environments underground.

[0048] The lithology identification module 7 is deployed and operates within an edge computing device, which serves as the core computing platform for the machine vision-based hard rock lithology identification and cutting parameter adaptive system. To adapt to the harsh working environment of high humidity and dust in underground coal mines, the edge computing device adopts an industrial-grade rugged design, with an IP65 or higher protection rating to prevent the intrusion of external foreign objects and liquids. The edge computing device is mechanically fixed to the body of the tunneling machine 1 using high-strength bolts to ensure the stability of the connection between the edge computing device and the tunneling machine 1 during severe vibrations generated during cutting operations.

[0049] The edge computing device is equipped with a Controller Area Network (CAN) bus interface. The edge computing device establishes a bidirectional communication connection with the main control programmable logic controller (PLC) of the tunneling machine 1 via the CAN bus interface. The lithology identification module 7 uses the CAN bus to transmit the identified lithology category, rock mass integrity score, and fracture density level data to the main control system of the tunneling machine 1, and receives real-time operating status feedback from the tunneling machine 1.

[0050] In terms of hardware computing power configuration, the edge computing device integrates a high-performance central processing unit (CPU) and a graphics processing unit (GPU) to meet the performance requirements of deep learning algorithms for parallel computing and floating-point operations. In this embodiment, the CPU is configured as an AMD Ryzen 9 7945HX processor, and the GPU is configured as an NVIDIA GeForce RTX 4060 graphics card. The GPU is used to provide hardware acceleration for the deep learning model based on the improved YOLOv8n network running in the rock identification module 7, ensuring that the processing frame rate of the high-definition visible light images input from the image acquisition module 4 meets the real-time requirements.

[0051] In terms of the software operating environment, the edge computing device is built on the Windows 11 operating system. It integrates the PyTorch deep learning framework, specifically version 3.9.7. Simultaneously, the edge computing device is configured with a Computing Unified Device Architecture (CUDA) environment, specifically version 12.2, to schedule the underlying computing resources of the graphics processor. The rock identification module 7, based on the aforementioned hardware and software platform, performs the inference task of the object detection algorithm.

[0052] Please see the appendix Figure 4 The lithology identification module 7 integrates an adaptive image enhancement unit at the input of its improved YOLOv8n deep learning model. This adaptive image enhancement unit preprocesses the high-resolution visible light images acquired by the image acquisition module 4 to eliminate interference from the downhole environment on image quality and to meet the standardization requirements of the deep learning model for input data.

[0053] The adaptive image enhancement unit executes the Retinex dehazing algorithm. Given the presence of suspended dust at the tunneling face, which causes high-resolution visible light images to appear hazy or foggy, the Retinex dehazing algorithm, based on color constancy theory, decomposes the high-resolution visible light image into illumination and reflection components. The adaptive image enhancement unit estimates the illumination component using Gaussian filtering and removes it from the original image, thereby eliminating the uneven illumination and fogging effect caused by dust scattering and restoring the inherent reflective properties of the rock surface, i.e., the texture and color characteristics of the rock.

[0054] The adaptive image enhancement unit further incorporates the Contrast-Limited Adaptive Histogram Equalization (CLAHE) algorithm. For areas with localized shadows or low light that still exist despite the supplemental lighting provided by the supplemental lighting device 45, the CLAHE algorithm divides the high-resolution visible light image into multiple non-overlapping small regions. The adaptive image enhancement unit independently performs histogram equalization processing on each small region to enhance local contrast and improve the clarity of cracks and bedding details on the rock surface. Simultaneously, the CLAHE algorithm introduces a contrast threshold limitation to prevent excessive amplification of noise signals in flat areas of the rock surface.

[0055] The adaptive image enhancement unit also performs white balance correction to compensate for the color cast caused by the color temperature characteristics of the supplementary lighting device 45, ensuring accurate reproduction of lithological colors. After completing the above enhancement processing, the adaptive image enhancement unit performs image size normalization. The adaptive image enhancement unit uniformly adjusts the size of the processed image to 640×640 pixels to match the dimensionality requirements of the network input layer of the improved YOLOv8n deep learning model. The image data processed by the adaptive image enhancement unit is then input into the backbone network of the lithology identification module 7 for feature extraction.

[0056] The deep learning model running in lithology identification module 7 is built based on the YOLOv8n target detection network architecture, and its structure has been improved to address hard rock texture features and complex downhole backgrounds. The improved YOLOv8n deep learning model mainly consists of three parts: a backbone network, a neck network, and a detection head.

[0057] The backbone network is responsible for extracting features from the image output by the adaptive image enhancement unit. The backbone network employs lightweight, large-kernel, multi-path convolutional blocks, specifically the GSConvV2 module, such as... Figure 5 As shown, the GSConvV2 module replaces traditional convolutional layers to capture contextual information and high-resolution details from hard rock feature maps. The GSConvV2 module employs a multi-path parallel structure, performing standard convolution operations and depthwise separable convolution operations separately, and concatenating and shuffling the output feature maps from different paths. By introducing large-size convolutional kernels to expand the receptive field, the GSConvV2 module enhances the extraction capability of large-scale rock textures and localized fine fracture features while maintaining low computational cost.

[0058] To enhance the effectiveness of feature propagation, the improved YOLOv8n deep learning model incorporates a multi-scale detail fusion module (MDFM) after each GSConvV2 module in the backbone network, such as... Figure 6 As shown, the multi-scale detail fusion module is used to fuse high-level semantic features with low-level visual features. The multi-scale detail fusion module receives feature maps from the GSConvV2 module and, by establishing cross-layer connection channels, concatenates and fuses low-level features containing rich edge and texture information with high-level features containing category semantic information, thereby improving the detection rate and edge localization accuracy of the improved YOLOv8n deep learning model for hard rock targets.

[0059] The neck network, located between the backbone network and the detection head, is used to further integrate multi-scale features. The neck network employs a weighted bidirectional feature pyramid network (BiFPN), such as... Figure 7 As shown, a weighted bidirectional feature pyramid network replaces the original path aggregation network (PANet). The weighted bidirectional feature pyramid network introduces bidirectional cross-scale connections and a weighted feature fusion mechanism. It assigns different learnable weights to input features at different resolutions and performs a weighted summation based on the importance of each feature layer to the lithology identification task. In this way, the neck network can suppress the interference of complex backgrounds such as downhole cutting machine components, spray, or support structures on rock surface features, enhancing the expressive power of the feature pyramid.

[0060] The detection head is the output of the improved YOLOv8n deep learning model, used to generate the final detection results. The detection head introduces the iRMB inverted residual block attention mechanism. Embedded before the convolutional operations of the detection head, the iRMB inverted residual block attention mechanism weights the input feature map. By focusing the attention map on effective regions containing rock texture features and suppressing responses from irrelevant background regions, the iRMB inverted residual block attention mechanism enhances the descriptive ability of hard rock targets.

[0061] After processing by the aforementioned network structure, the lithology identification module 7 outputs multi-dimensional rock surface information through the detection head. The information output by the lithology identification module 7 includes the current rock surface lithology category, the estimated rock mass integrity score (RQD), and the fracture density level. Lithology categories include granite, sandstone, limestone, and shale. The lithology identification module 7 transmits the identification results to the rock hardness mapping module via the controller local area network bus.

[0062] The rock hardness mapping module establishes communication connections with the lithology identification module 7 and the adaptive control module, respectively. The rock hardness mapping module receives the lithology category, rock mass integrity score estimate, and fracture density level output by the lithology identification module 7, and converts the above multidimensional features into quantified uniaxial compressive strength values ​​of the rock, so that the adaptive control module can generate control commands.

[0063] The rock hardness mapping module internally stores a preset lithology-uniaxial compressive strength correspondence table. This table, constructed based on rock mechanics experimental data, covers the hardness distribution range of major rock types in coal-bearing strata. The rock hardness mapping module uses the lithology category output by the lithology identification module 7 as the index key to query the lithology-uniaxial compressive strength correspondence table to obtain an estimated hardness range for the current rock. In this embodiment, the data configuration of the lithology-uniaxial compressive strength correspondence table is as follows: for granite, the corresponding uniaxial compressive strength range is set to 100 MPa to 250 MPa; for sandstone, the corresponding uniaxial compressive strength range is set to 30 MPa to 150 MPa; and for shale, the corresponding uniaxial compressive strength range is set to 10 MPa to 50 MPa.

[0064] To obtain more accurate continuous hardness values, the rock hardness mapping module also integrates a lightweight regression model. This lightweight regression model is constructed using the LightGBM regression algorithm. It takes the lithological feature vector output from the backbone network of the improved YOLOv8n deep learning model in lithology identification module 7 as input data. By performing nonlinear mapping operations on the lithological feature vector, the lightweight regression model directly predicts and outputs continuous uniaxial compressive strength values. After training based on a rock core test database, the lightweight regression model achieves a determination coefficient of 0.91 and a mean absolute error controlled within 10 MPa, realizing refined quantification of rock hardness.

[0065] The rock hardness mapping module is equipped with an online learning and dynamic correction mechanism to improve the system's robustness under atypical geological conditions. The rock hardness mapping module receives signals from the cutting motor current sensor, the cutting head vibration spectrum sensor, and the temperature sensor at the cutting section of the tunnel boring machine 1. When the rock hardness mapping module detects abnormal feedback signals such as a sudden increase in the current value displayed by the cutting motor current sensor exceeding a preset threshold, abnormal high-frequency components appearing in the vibration frequency of the cutting head 2, or a temperature increase, it triggers a confidence verification process for the current output result of the lithology identification module 7.

[0066] In the confidence verification process, the rock hardness mapping module compares the consistency between the visual recognition results and the physical feedback signals. When the confidence level of the visual recognition is lower than a preset threshold, the rock hardness mapping module uses fuzzy decision logic to perform a weighted calculation by combining the output probability of the lithology recognition module 7 with the hardness value calculated by the physical feedback, generating a comprehensive uniaxial compressive strength value. Simultaneously, the rock hardness mapping module dynamically corrects the mapping parameters in the lithology-uniaxial compressive strength correspondence table or triggers a weight update for the lightweight regression model based on abnormal feedback signals to adapt to actual changes in the lithology and hardness of the working face.

[0067] Please see the appendix Figure 8 The adaptive control module is integrated into the main programmable logic controller (PLC) or edge computing coprocessor of the tunneling machine 1. The communication latency between the adaptive control module and the underlying actuators of the tunneling machine 1 is configured to not exceed 100 milliseconds to ensure real-time response to changes in operating conditions. The adaptive control module is responsible for dynamically generating command parameters for controlling the operation of the tunneling machine 1 based on the rock hardness value output by the rock hardness mapping module. The command parameters specifically include the rotational speed of the cutting head 2, the longitudinal feed speed of the cutting arm 3, and the upper limit of the power output of the cutting motor.

[0068] The adaptive control module has a pre-set mapping table of control parameters for different rock hardness ranges. The adaptive control module compares the received rock hardness value (i.e., uniaxial compressive strength UCS) with the preset hardness threshold to determine the current cutting condition mode and adjust the control strategy accordingly.

[0069] When the rock hardness is less than or equal to 50 MPa, the adaptive control module determines that the current working condition is within the low-hardness rock cutting range. Within this range, the rock structure is soft and easily broken, and the adaptive control module adopts a control strategy that prioritizes cutting efficiency. The adaptive control module generates control commands, setting the rotational speed of the cutting head 2 to 60 rpm, the feed speed of the cutting arm 3 to 80 mm / min, and the upper limit of the cutting power to 60% of the rated power. This parameter combination utilizes the easily broken characteristics of soft rock to achieve rapid advancement.

[0070] When the rock hardness value is greater than 50 MPa and less than or equal to 100 MPa, the adaptive control module determines that the current working condition is in the medium-to-low hardness rock cutting range. The rock in this range has a certain strength, and both cutting efficiency and cutting tooth wear must be considered. Therefore, a control strategy balancing efficiency and durability is adopted. The adaptive control module generates control commands to adjust the rotational speed of the cutting head 2 to 55 rpm, the feed speed of the cutting arm 3 to 70 mm / min, and set the upper limit of cutting power to 70% of the rated power. This configuration effectively reduces the thermal load and wear rate of the cutting teeth while maintaining a high feed speed, thus extending their service life.

[0071] When the rock hardness value is greater than 100 MPa and less than or equal to 180 MPa, the adaptive control module determines that the current working condition is in the medium-to-high hardness rock cutting range. At this point, the rock hardness increases significantly, cutting resistance increases, and mechanical impact and fatigue damage to the cutting teeth are easily triggered. Therefore, the control strategy shifts to prioritizing equipment protection while also considering efficiency. The adaptive control module generates control commands to reduce the rotational speed of the cutting head 2 to 50 rpm, reduce the feed speed of the cutting arm 3 to 60 mm / min, and strictly limit the cutting power to 80% of the rated power. By appropriately reducing the rotational and feed speeds, the single-tooth cutting thickness and dynamic load are reduced, thereby alleviating the impact stress on the cutting system, ensuring stable equipment operation, and extending the service life of the cutting head 2.

[0072] When the rock hardness exceeds 180 MPa, the system determines that it has entered the high-hardness rock cutting zone. This type of rock is extremely hard, and conventional cutting methods can easily cause cutting teeth to chip, tool failure to occur prematurely, and damage to transmission components. Therefore, the control strategy prioritizes maximizing equipment safety. The system automatically activates a high-hardness parameter set: the rotational speed is reduced to 45 rpm, the feed rate is reduced to 50 mm / min, and the power limit is strictly capped at 90% of the rated power. This parameter combination significantly reduces the cutting energy input per unit time, avoids severe impacts, effectively extends the cutting head's life, and improves operational safety.

[0073] The adaptive control module can also employ a fuzzy logic controller to achieve continuous parameter adjustment. The fuzzy logic controller uses the Mamdani inference method to establish a nonlinear mapping relationship between rock hardness values ​​and cutting parameters. However, in the preferred embodiment for practical engineering applications, the adaptive control module uses the aforementioned hierarchical lookup table method to ensure the stability and predictability of the control system.

[0074] To prevent abrupt and drastic changes in cutting parameters due to abrupt changes in lithology, which could impact the mechanical structure and hydraulic system of the tunneling machine 1, a parameter smoothing transition unit is integrated into the adaptive control module. Located between the cutting parameter decision logic and the underlying execution drive, this unit smooths the target speed command of the cutting head 2, the target feed speed command of the cutting arm 3, and the power limit command in the time domain. The parameter smoothing transition unit ensures that control commands are sent to the actuators of the tunneling machine 1 in a continuous and gradual manner, thereby avoiding instantaneous surges in the cutting motor current and overload oscillations in the mechanical transmission chain.

[0075] The parameter smoothing transition unit uses an exponentially weighted moving average algorithm to calculate the actual control output value at the current moment. The parameter smoothing transition unit operates based on a discrete-time control system, with a control period set to 100 milliseconds. Within each control period, the parameter smoothing transition unit calculates the final output command based on the output value of the previous moment and the target value of the current moment using a preset smoothing formula.

[0076] The smoothing formula is specifically expressed as follows:

[0077] ;

[0078] in, Indicates the current time The control output value; This represents the theoretical target control value obtained by querying or calculating based on the current rock hardness value; Indicates the previous moment The actual control output value; This represents the smoothing coefficient.

[0079] In this embodiment, the smoothing coefficient The value is set to 0.3. This value has been verified through dynamic simulation, ensuring the system's response speed to lithological changes while effectively filtering out parameter fluctuations caused by image recognition noise or frequent lithological fluctuations. When the lithology identification module 7 detects a sudden change in lithology, such as a sudden transition from soft rock to hard rock, A step change will occur, and after processing with the smoothing formula... It will present an exponentially approaching smooth curve, so that the rotational speed of the cutting head 2 will smoothly transition to the target value within several control cycles.

[0080] In addition to the exponential smoothing algorithm mentioned above, the parameter smoothing transition unit further introduces rate-of-change limiting logic as a secondary protection. The rate-of-change limiting logic is used to forcibly constrain the maximum change in control commands between two adjacent control cycles. The parameter smoothing transition unit calculates the currently calculated... and The absolute value of the difference.

[0081] When the rate of change of the physical quantity corresponding to the absolute value of the difference exceeds the preset safety threshold, the parameter smoothing transition unit will forcibly truncate it. The value. Specifically, if the calculated compared to If the increase exceeds the maximum allowed increment, then... Assigned value The sum of the reduction and the maximum permissible increment; if the reduction exceeds the maximum permissible reduction, then... Assigned value The difference between the amount and the maximum allowable reduction.

[0082] For the rotational speed control of the cutting head 2, the rate-of-change limiting logic restricts the maximum rate of change of rotational speed to no more than 10% of the rated rotational speed per second; for the feed speed control of the cutting arm 3, the rate-of-change limiting logic restricts the maximum rate of change of feed speed to no more than 10% of the maximum feed speed per second. Through the combined effect of the exponential smoothing formula and the rate-of-change limiting logic, the adaptive control module eliminates high-frequency spikes in the control signal, realizing smooth switching of the tunneling machine 1 between working states of different hardness rock layers.

[0083] The machine vision-based hard rock lithology identification and cutting parameter adaptive system constructs a real-time verification and closed-loop correction mechanism for visual recognition results by introducing physical sensor data. The feedback optimization module, as the execution unit of this mechanism, establishes data connections with the current sensor installed on the cutting section of the tunneling machine 1, the vibration sensor installed on the cutting arm 3, the lithology identification module 7, and the rock hardness mapping module. The feedback optimization module aims to solve the recognition deviation problem that occurs when relying solely on visual images under extreme conditions such as extremely high dust concentrations or rock surfaces covered with coal slime.

[0084] The feedback optimization module collects real-time physical load data of the tunneling machine 1 during the cutting operation. This physical load data specifically includes the effective value of the three-phase current of the cutting motor, the longitudinal vibration acceleration amplitude of the cutting head 2, and the oil temperature data of the cutting reducer. The feedback optimization module has a pre-set load and hardness inversion model based on the cutting specific energy theory. This model can inversely calculate the equivalent cutting impedance of the current cutting object, i.e., the physically inverted hardness value, based on the current cutting motor current value, the rotational speed of the cutting head 2, and the feed speed of the cutting arm 3.

[0085] The feedback optimization module executes consistency verification logic for multi-source data. It compares in real-time the predicted rock hardness value output by the rock hardness mapping module based on the visual image with the physically inverted hardness value. The feedback optimization module calculates the absolute value of the deviation between the two hardness values. When the absolute value of the deviation is less than a preset tolerance threshold, the feedback optimization module determines that the visual recognition result is accurate, and the adaptive control module continues to execute control based on the visual recognition result.

[0086] When the absolute value of the deviation exceeds a preset tolerance threshold, the feedback optimization module determines that the credibility of the visual recognition result has decreased and triggers the fusion decision process. In this process, the feedback optimization module uses a Kalman filter algorithm or a Bayesian inference method to perform weighted fusion of the visual prediction value and the physical inversion value. Given that the physical load data directly reflects the interaction force between the rock and the cutting teeth and has higher physical authenticity, the feedback optimization module assigns a higher weight coefficient to the physical inversion hardness value and a lower weight coefficient to the visual prediction hardness value in the fusion calculation, thereby generating a corrected comprehensive rock hardness value. The adaptive control module then switches to generating control commands based on the comprehensive rock hardness value to prevent improper cutting parameter settings due to visual misjudgment.

[0087] The feedback optimization module also possesses a long-term model self-evolution function. When a persistent and significant difference is detected between the visual recognition result and the physical feedback result, the feedback optimization module packages and stores the high-definition visible light image acquired by the image acquisition module 4, the original recognition result output by the lithology identification module 7, and the true hardness label corrected by physical data as difficult example sample data. The difficult example sample data is uploaded to the ground server or used for the background training process of edge computing devices via the communication interface. The difficult example sample data is used for incremental training and parameter fine-tuning of the improved YOLOv8n deep learning model in the lithology identification module 7 and the lightweight regression model in the rock hardness mapping module. Through this mechanism, the machine vision-based hard rock lithology identification and truncation parameter adaptive system can gradually adapt to the specific changes in the geological environment of a particular mine, continuously improving the feature extraction capability and hardness mapping accuracy of complex lithological features.

[0088] After startup, the machine vision-based hard rock lithology identification and cutting parameter adaptive system first executes an initialization self-test program. The edge computing device establishes a communication handshake with the main control system of tunneling machine 1 through the Controller Area Network (CAN) bus to confirm the stability of the communication link and the online status of each sensor node.

[0089] After entering normal operation mode, the image acquisition module 4 begins image acquisition. To ensure image clarity, the miniature air curtain nozzle 46 is activated before exposure by the high-resolution industrial camera 42. The miniature air curtain nozzle 46 receives high-pressure gas from the compressed air source 8 through the gas delivery pipeline 5 and forms a continuous high-speed airflow barrier on the surface of the high-strength quartz glass 44 to blow away dust and splashed water mist generated during the cutting operation. At the same time, the supplementary lighting device 45 automatically turns on based on the brightness information fed back by the ambient light sensor. The supplementary lighting device 45 adjusts the luminous intensity of the visible light illumination unit and the near-infrared illumination unit to provide a uniform and high-contrast lighting environment for the rock surface to be cut. After the optical environment meets the requirements, the high-resolution industrial camera 42 captures a high-definition visible light image of the rock surface in front of the tunneling machine 1 through the fixed focal length lens 43.

[0090] The acquired high-resolution visible light images are transmitted in real time via a highly flexible shielded cable 6 to the lithology identification module 7 deployed in the edge computing device. The lithology identification module 7 first calls the adaptive image enhancement unit to preprocess the high-resolution visible light images. The adaptive image enhancement unit executes the Retinex dehazing algorithm and the contrast-limited adaptive histogram equalization algorithm to remove the haze effect in the image and enhance the details of the rock texture. The preprocessed image is then input into the improved YOLOv8n deep learning model. The improved YOLOv8n deep learning model utilizes the GSConvV2 module and the multi-scale detail fusion module in its backbone network to extract deep semantic features and shallow texture features of the rock. The lithology identification module 7 outputs the lithology category, rock mass integrity score, and fracture density level of the current rock surface via the detection head.

[0091] The identification results output by the lithology identification module 7 are transmitted to the rock hardness mapping module. Based on a preset lithology-uniaxial compressive strength correspondence table, the rock hardness mapping module initially determines the hardness range of the current rock. Simultaneously, the rock hardness mapping module calls a lightweight regression model to perform nonlinear regression calculations on the lithology feature vector, calculating a precise quantified value for the uniaxial compressive strength.

[0092] The adaptive control module receives the quantified value of the uniaxial compressive strength and formulates a cutting control strategy accordingly. The adaptive control module determines the hardness range of the quantified uniaxial compressive strength value. If it is determined to be in the soft rock range, the adaptive control module generates control commands with high rotational speed and high feed rate to improve efficiency; if it is determined to be in the hard rock range, the adaptive control module generates control commands with low rotational speed and low feed rate to protect the cutting head 2.

[0093] Before being sent to the tunneling machine 1, the generated raw control commands must be processed by the parameter smoothing transition unit. The parameter smoothing transition unit uses an exponentially weighted moving average algorithm and a rate-of-change limiting logic to smooth the target rotational speed of the cutting head 2 and the target feed speed of the cutting arm 3. The parameter smoothing transition unit eliminates abrupt changes in the control commands, generating a continuous and stable sequence of target commands.

[0094] Throughout the system's operation, the feedback optimization module executes closed-loop verification logic in parallel. The feedback optimization module continuously collects current data from the cutting motor and vibration spectrum data from the cutting head of the tunneling machine 1, and calculates the physically inverted hardness value. The feedback optimization module compares the physically inverted hardness value with the predicted hardness value output by the rock hardness mapping module. If the deviation is within acceptable limits, the main control system of the tunneling machine 1 directly executes the smoothed target instruction sequence. If the deviation exceeds a threshold, the feedback optimization module corrects the hardness value using a Kalman filter algorithm and triggers the adaptive control module to regenerate control instructions using the corrected comprehensive hardness value. Simultaneously, it marks and stores the current abnormal data samples for subsequent online model updates and iterations. Through this process, the machine vision-based hard rock lithology identification and cutting parameter adaptive system achieves closed-loop automated operation from environmental perception and intelligent decision-making to precise execution.

Claims

1. A machine vision-based system for hard rock lithology identification and adaptive cutting parameters, characterized in that, It includes an image acquisition module (4), which is connected in sequence to a lithology identification module (7), a rock hardness mapping module, and an adaptive control module; The image acquisition module (4) is used to acquire high-definition visible light images of the rock surface in front of the working face of the tunneling machine (1); The image acquisition module (4) is connected to the lithology identification module (7) via a highly flexible shielded cable (6), and the highly flexible shielded cable (6) is laid along the internal groove of the cutting arm (3) of the tunneling machine (1); The lithology identification module (7) runs a deep learning model, and the input end of the deep learning model integrates an adaptive image enhancement unit; The adaptive image enhancement unit is configured to execute the Retinex dehazing algorithm to decompose the high-definition visible light image into illumination and reflection components, and remove the illumination component to eliminate uneven illumination and fogging effects. The adaptive image enhancement unit is also configured to execute a contrast-limited adaptive histogram equalization algorithm to divide the high-definition visible light image into multiple non-overlapping small regions and to perform histogram equalization processing on each small region independently. The deep learning model includes a backbone network, a neck network, and a detection head; The backbone network adopts the GSConvV2 module, which uses a multi-path parallel structure to perform standard convolution operations and depthwise separable convolution operations respectively, and performs concatenation and shuffling operations on the output feature maps; the backbone network also includes a multi-scale detail fusion module, which concatenates and fuses low-level features with high-level features through cross-layer connection channels. The neck network adopts a weighted bidirectional feature pyramid network, which has bidirectional cross-scale connectivity and a weighted feature fusion mechanism; The detection head introduces an iRMB inverted residual block attention mechanism, which is embedded before the convolution operation of the detection head; The lithology identification module (7) is used to process the high-definition visible light image to output lithology category, rock mass integrity score and fracture density level; The rock hardness mapping module is used to determine the rock hardness value based on the lithology category, rock mass integrity score and fracture density level output by the lithology identification module (7); The adaptive control module is used to generate cutting parameter instructions based on the rock hardness value and send them to the control system of the tunneling machine (1).

2. The machine vision-based hard rock lithology identification and adaptive cutting parameter system according to claim 1, characterized in that, The image acquisition module (4) is fixedly connected to the body of the cutting arm (3) by a rigid mounting bracket (9); The image acquisition module (4) includes a dustproof and waterproof housing (41), a self-cleaning lens assembly and a supplementary lighting device (45). The dustproof and waterproof housing (41) is located outside the cutting arm (3). The front end of the dustproof and waterproof housing (41) is provided with an optical window made of high-strength quartz glass (44). The self-cleaning lens assembly is located inside the dustproof and waterproof housing (41) and includes a high-resolution industrial camera (42), a fixed-focus lens (43), and a miniature air curtain nozzle (46). The miniature air curtain nozzle (46) is arranged in a ring and surrounds the outer periphery of the fixed-focus lens (43) to spray airflow to remove dust and water mist adhering to the surface of the high-strength quartz glass (44).

3. The machine vision-based hard rock lithology identification and adaptive cutting parameter system according to claim 2, characterized in that, The micro air curtain nozzle (46) is connected to the compressed air source (8) through the gas delivery pipeline (5); The supplementary lighting device (45) consists of multiple sets of independently controllable LED light sources, arranged in a ring array inside the front end of the dustproof and waterproof housing (41); The supplementary lighting device (45) includes a visible light illumination unit and a near-infrared illumination unit, supporting dual-mode illumination of visible light and near-infrared light; the lens surface of the fixed focal length lens (43) is coated with a wide-band anti-reflection film, and the working band of the wide-band anti-reflection film can transmit visible light and near-infrared light.

4. The machine vision-based hard rock lithology identification and adaptive cutting parameter system according to claim 1, characterized in that, The rock hardness mapping module stores a preset lithology-uniaxial compressive strength correspondence table, which is used to query and obtain the estimated value of the hardness range of the current rock according to the lithology category. The rock hardness mapping module also integrates a lightweight regression model. The lightweight regression model uses the lithological feature vector output by the lithology identification module (7) as input data and directly predicts and outputs continuous uniaxial compressive strength values ​​through nonlinear mapping operations.

5. The machine vision-based hard rock lithology identification and adaptive cutting parameter system according to claim 1, characterized in that, It also includes a feedback optimization module; The feedback optimization module is communicatively connected to the current sensor, vibration sensor and lithology identification module (7) and rock hardness mapping module respectively installed in the cutting section of the tunneling machine (1); The feedback optimization module is configured to acquire the cutting motor current data collected by the current sensor and the vibration spectrum data of the cutting head (2) collected by the vibration sensor, and use the load and hardness inversion model to deduce the physical inversion hardness value. The feedback optimization module compares the rock hardness value output by the rock hardness mapping module with the physical inversion hardness value. When the absolute value of the deviation exceeds the preset tolerance threshold, the Kalman filter algorithm is used to weight and fuse the visual prediction value and the physical inversion hardness value to generate a comprehensive rock hardness value.

6. The machine vision-based hard rock lithology identification and adaptive cutting parameter system according to claim 1, characterized in that, The adaptive control module has a pre-set control parameter mapping table for different rock hardness ranges. The adaptive control module is configured to compare the rock hardness value with a preset hardness threshold to determine the cutting mode: When the rock hardness value is determined to be in the soft rock cutting range, a parameter combination of high rotation speed and high feed rate is generated. When the rock hardness value is determined to be within the hard rock cutting range, a parameter combination of low rotation speed and low feed rate is generated.

7. The machine vision-based hard rock lithology identification and adaptive cutting parameter system according to claim 1, characterized in that, The adaptive control module integrates a parameter smoothing transition unit. The parameter smoothing transition unit is configured to use an exponentially weighted moving average algorithm to calculate the current control output value based on the output value of the previous time step, the theoretical target control value of the current time step, and a preset smoothing coefficient. The parameter smoothing transition unit also includes a rate of change limiting logic, which is used to calculate the absolute value of the difference between the currently calculated control output value and the output value at the previous moment. When the rate of change of the physical quantity corresponding to the absolute value of the difference exceeds a preset safety threshold, the value of the control output value is forcibly truncated.

8. The machine vision-based hard rock lithology identification and adaptive cutting parameter system according to claim 1, characterized in that, The lithology identification module (7) is configured in an edge computing device, which has an IP65 or higher protection rating and is fixed to the body of the tunneling machine (1) by bolts; The edge computing device is equipped with a controller local area network bus interface, and establishes a bidirectional communication connection with the main control system of the tunneling machine (1) through the controller local area network bus interface.

Images