Karst region large-diameter pile construction process monitoring method based on computer vision

Abstract

The application discloses a karst area large-diameter pile construction process monitoring method based on computer vision, which comprises image acquisition of mud, quantitative extraction of visual features, construction of a visual feature library, establishment of a classification model and discrimination of disease types and severity; by deploying high-definition industrial cameras at the inlet and outlet of the mud tank and the sedimentation tank, synchronous quantitative extraction of color features, turbidity features, bubble features and solid particle features of the mud is realized, typical disease types such as karst cave fissure leakage, hole wall collapse and mud performance degradation and their severity are effectively discriminated based on computer vision technology, the rapid response capability of construction risks is significantly improved, and the probability of accident expansion and economic losses are reduced.

Description

Technical Field

[0001] This invention belongs to the field of pile foundation construction monitoring technology, and relates to a computer vision-based method for monitoring the construction process of large-diameter piles in karst areas. Background Technology

[0002] Large-diameter piles, as a type of deep foundation with excellent bearing capacity, are widely used in large-scale projects such as large bridges, super high-rise buildings, and port terminals. However, the geological structure of karst areas is characterized by complex and variable features, mainly consisting of sinkholes, dissolution fissures, solution channels, karst buds, and soil caves. When using the mud-wall circulation drilling process for large-diameter pile construction in karst areas, these complex and variable geological conditions can easily lead to accidents such as seepage from sinkholes and fissures, collapse of the borehole wall, and deterioration of mud properties during construction. These accidents can seriously threaten construction safety, delay the construction period, increase costs, and may even cause casualties and equipment damage.

[0003] Currently, condition monitoring during the construction of large-diameter piles in karst areas mainly relies on manual inspections and traditional sensor monitoring. Manual inspections depend on operators' experience observing the mud level in the mud tank, mud properties, and slurry return at the borehole opening. This method suffers from significant drawbacks, including low efficiency, strong subjectivity, difficulty in achieving continuous real-time monitoring, and insensitivity to early, subtle anomalies, making it difficult to detect early signs of mud leakage or borehole wall instability in a timely manner. Traditional sensor monitoring methods use level gauges, flow meters, pressure gauges, turbidity meters, and other sensors for automated monitoring of macroscopic physical quantities. However, they cannot deeply analyze the complex visual characteristics of the mud's internal state, such as subtle changes in mud color, spatial differences in turbidity distribution, the morphology and origin of bubbles, and the morphology and dynamic migration characteristics of solid particles. These visual characteristics are precisely the most direct and abundant early signal sources for identifying key defects such as seepage from karst caves and fissures, borehole wall collapse and fragment detachment, and mud performance deterioration. In summary, existing monitoring methods are insufficient to achieve real-time and accurate perception of the status of the mud wall circulation hole forming process. Therefore, they cannot determine the specific type of disease in the early stage and provide necessary early warnings, and thus cannot achieve rapid construction response based on the type and severity of the disease.

[0004] Therefore, to improve the safety and efficiency of large-diameter pile construction in karst areas and effectively prevent in-hole accidents, there is an urgent need for a method that can monitor the mud slurry circulation process in real time, comprehensively, and accurately, and provide early warnings based on the monitoring results. Computer vision-based monitoring technology has advantages such as non-contact operation, rich information, and good real-time performance. Applying it to construction monitoring in this field is expected to address the shortcomings of existing monitoring methods and provide strong protection for construction safety. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes a computer vision-based monitoring method for the construction process of large-diameter piles in karst areas. By combining computer vision technology, deep learning methods, and closed-loop control strategies, it achieves real-time, accurate, and intelligent perception and early warning of key defects in the mud-wall cyclic drilling process of large-diameter piles in karst areas, which is beneficial to improving construction safety, efficiency, quality, and intelligence.

[0006] A computer vision-based monitoring method for the construction process of large-diameter piles in karst areas includes the following steps:

[0007] S1. Image acquisition of mud:

[0008] High-definition industrial cameras are deployed at the inlet and outlet of the mud tank and above the sedimentation tank to continuously collect real-time image data at a preset frame rate and transmit it to the control center in real time. The real-time image data includes initial state image data of the mud before it is injected into the hole, state image data of the mud after the circulation in the hole is completed, and state change image data of the mud during the sedimentation process. The high-definition industrial cameras are dustproof, waterproof and vibration resistant.

[0009] S2. Quantitative extraction of visual features:

[0010] After receiving the real-time image data, the control center extracts the color features, turbidity features, bubble features, and solid particulate matter features from the real-time image data, thereby combining them to form a real-time feature vector. O 1 The color features include mean hue value, variance of hue value, skewness of hue value, mean saturation value, variance of saturation value, and skewness of saturation value; the turbidity features include turbidity gradient amplitude and area ratio of turbidity gradient abrupt change regions; the bubble features include bubble density, bubble size, and bubble outline roundness; and the solid particulate matter features include particle concentration, particle concentration gradient amplitude, and particle outline fractal dimension.

[0011] S3. Construction of the visual feature library:

[0012] By constructing a sample dataset, training a feature extraction model, generating a disease feature vector set, and establishing a dynamic update mechanism, a visual feature library of mud diseases in karst areas is built and dynamically updated.

[0013] S4. Establishing a classification model:

[0014] A multi-class classification model based on Support Vector Machine (SVM) is used to establish real-time feature vectors. O 1 With disease probability vector O 2The mapping relationship; wherein, when training the SVM-based multi-classification model, 70% of the samples are randomly selected from the visual feature database of mud diseases in the karst area as the training set and the remaining 30% of the samples are selected as the validation set; after training, the real-time feature vector is input. O 1 The disease probability vector is obtained. O 2 =[ P leak , P collape , P degradation , P normal ],in, P leak The probability of leakage from fissures in a karst cave. P collape The probability of the borehole wall collapsing. P degradation For the deterioration of mud properties, P normal The probability of being in a normal state; P leak , P collape , P degradation , P normal The values ​​of are all in the range [0,1], and P leak + P collape + P degradation + P normal =1;

[0015] S5. Determination of Disease Type and Severity:

[0016] The real-time feature vector generated in step S2 based on the real-time image data O 1 The input is fed into the SVM-based multi-classification model trained in step S4, and the output is a disease probability vector. O 2 =[ P leak , P collape , P degradation , P normal Define the probability of occurrence of the most severe disease. P max =max{ P leak ,P collape , P degradationl}, and accordingly determine the most severe disease type. T max ∈{karst cave fissure leakage, borehole wall collapse, mud performance deterioration}; Based on historical defect data of large-diameter piles in karst areas and engineering safety standards, early warning thresholds for defects are set. P max ];when P max ≤[ P max When [the situation is assessed], the current construction status is determined to be low-risk, and the regular monitoring frequency is maintained; when [the situation is assessed], ... P max >[ P max When determining that there is a significant risk of disease, a severity index is calculated. I M As shown in the following formula:

[0017]

[0018] in β 1 and β 2 All are weighting coefficients, and their values ​​range from [0,1]. β 1 + β 2 =1; d P max / d t The rate of change of the probability of the target disease over time; when 0.5 ≤ I M When 1 < 1, a Level 1 response is triggered, and the control center outputs a low-frequency intermittent audible and visual warning signal to alert operators to potential risks; when 1 ≤ 1 I M When the alarm level is <2, a level 2 response is triggered, and the control center outputs an audible and visual warning signal and escalates the alarm level to prompt the operator to take intervention measures; when I M When the value is ≥2, a level 3 response is triggered, and the control center outputs a continuous high-intensity audible and visual warning signal to prompt the operator to take strong intervention measures.

[0019] Preferably, in step S1, a dimmable industrial light source is configured next to the high-definition industrial camera; the dimmable industrial light source adopts a ring or strip structure design, and monitors the ambient light intensity in real time through a built-in light sensor to dynamically adjust the light intensity and achieve adaptive brightness compensation.

[0020] Preferably, in step S2, the quantization extraction of visual features includes the following steps:

[0021] S201, Color feature extraction: Calculate the color features of each pixel in the real-time image data using the HSV color gamut analysis model. x , y hue value H ( x , y ) and saturation value S ( x , y The system calculates the color features of the real-time image data, including the mean hue value, the variance of the hue value, the skewness of the hue value, the mean saturation value, the variance of the saturation value, and the skewness of the saturation value.

[0022] S202, Extraction of turbidity features: Convert the real-time image data into a grayscale image and extract each pixel point ( x , y grayscale value M 1 ( x , y Then, the turbidity gradient amplitude is calculated using the central difference method. M 2 ( x , y The expression is as follows:

[0023]

[0024] Among them, boundary pixels are mirrored to maintain image size; when pixel ( x , y ) turbidity gradient amplitude M 2 ( x , y )>2 M 3 When, mark the pixel point ( x , y ) represents abruptly changed pixels, where M 3 Under normal construction conditions M 2 ( x , y The median of ) ; the percentage of the area of ​​the region with abrupt change in turbidity gradient M 4 As shown in the following expression:

[0025]

[0026] in, M5 This represents the total number of mutated pixels. M 6 The total number of pixels;

[0027] S203. Bubble Feature Extraction: The visual regions representing bubbles in the real-time image data are segmented and morphologically analyzed to extract bubble features, which are used to distinguish the source of bubbles and assist in disease judgment. The bubble features include bubble density, bubble size, and bubble outline roundness. The bubble density is the number of bubbles per unit area, calculated as the ratio of the number of pixels occupied by bubbles to the total number of pixels. The bubble size is the average equivalent diameter of the bubble, calculated as the average diameter of a circle with the same area as the bubble. The bubble outline roundness... M 7 As shown in the following expression:

[0028]

[0029] in M 8 The area of ​​the bubble. M 9 The perimeter of the bubble outline;

[0030] S204. Extraction of solid particulate matter characteristics: Identify and analyze suspended or deposited solid particulate matter in the mud from the real-time image data, and extract solid particulate matter characteristics reflecting their source, state, and dynamics; the solid particulate matter includes primary rock fragments, collapsed debris, and metal fragments; the solid particulate matter characteristics include particle concentration. Q 1 Particle concentration gradient amplitude Q 2 and particle profile fractal dimension Q 3 Among them, particle concentration Q 1 The percentage of pixels occupied by solid particles in a unit area image is shown in the following expression:

[0031]

[0032] in, Q 4 This represents the total number of pixels occupied by solid particles. Q 5 This represents the total number of pixels in the image.

[0033] Particle concentration gradient amplitude Q 2 Based on particle concentration Q 1 The central difference method is used for calculation;

[0034]

[0035] Fractal dimension of particle profile Q 3 The box counting method is used for calculation, with a grid size of... S The mesh covers the particle outline, and S ∈[ S min , S max For each size S Calculate the minimum number of mesh boxes required to cover the contour. Q 6 ( S If the particle profile fractal dimension is 0, then... Q 3 As shown in the following expression:

[0036]

[0037] In specific calculations, the fractal dimension of the particle profile... Q 3 Fit log using linear regression Q 6 ( S ) and log(1 / S The slope between ) is obtained.

[0038] Preferably, step S3 involves constructing and dynamically updating a visual feature database of mud sludge diseases in karst areas, specifically including the following steps:

[0039] S301. Construct a sample dataset: Collect historical image data of the mud slurry wall protection and circulation hole formation process during the construction of large-diameter piles in karst areas, covering four scenarios, including karst cave fissure leakage, hole wall collapse, mud slurry performance deterioration and normal state;

[0040] S302. Training the feature extraction model: A convolutional neural network model is used as the feature extraction model. The convolutional neural network model includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The preprocessed labeled images are used as the training set, and the model parameters are optimized through the backpropagation algorithm.

[0041] S303. Generate disease feature vector sets: Using a trained convolutional neural network model, extract features from historical image data to generate disease feature vector sets for the four scenarios. Each set of disease feature vectors is correlated with the real-time feature vectors. O 1 The dimensions are consistent;

[0042] S304. Establish a dynamic update mechanism: Adopt an incremental learning algorithm. When new construction scenarios introduce unrecorded disease features, the new samples are automatically included in the training set. The disease feature vector set is updated by fine-tuning the model parameters of the convolutional neural network model to ensure the adaptability of the feature library to complex geological conditions.

[0043] In summary, compared with existing technologies, the beneficial effects of this invention are as follows: Addressing the problems of low data accuracy, poor real-time performance, and weak disease interpretation in the monitoring technology for karst large-diameter pile drilling with mud wall protection, this invention proposes a computer vision-based monitoring method for the construction process of large-diameter piles in karst areas. This method includes mud image acquisition, quantitative extraction of visual features, construction of a visual feature library, establishment of a classification model, and determination of disease type and severity, achieving the following breakthrough improvements:

[0044] 1) Achieve comprehensive monitoring of mud condition: By deploying high-definition industrial cameras at the mud tank inlet, outlet and sedimentation tank, the color characteristics, turbidity characteristics, bubble characteristics and solid particulate matter characteristics of mud are simultaneously and quantitatively extracted. This is conducive to capturing early micro-change signals of diseases such as seepage from caves and fissures, collapse of borehole walls and deterioration of mud performance, and provides a high-information-density data foundation for disease identification.

[0045] 2) Early identification of karst diseases: Based on the visual feature library of mud diseases in karst areas, combined with the disease probability vector output by the SVM-based multi-classification model, the matching degree and severity index of the real-time feature vector and the standard feature vector in the feature library are calculated. This can provide early warning when there are subtle abnormalities in mud color, turbidity distribution, bubble morphology or particulate matter characteristics. It can also effectively distinguish typical disease types such as seepage from caves and fissures, collapse of borehole walls, and deterioration of mud performance from normal disturbances, thus improving the timeliness and accuracy of disease identification.

[0046] 3) Real-time risk warning: Based on the monitoring data, the most serious disease type and its severity are calculated and analyzed. The risk level is judged in real time and quantitatively. It can effectively realize the early warning of serious accidents such as seepage from karst caves and fissures and collapse of borehole walls, significantly improve the rapid response capability of construction risks, and reduce the probability of accident expansion and economic losses. Attached Figure Description

[0047] Figure 1 This is a flowchart of the computer vision-based monitoring method for the construction process of large-diameter piles in karst areas, as described in this invention. Detailed Implementation

[0048] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments given herein are for illustration and explanation only and are not intended to limit the present invention.

[0049] The first aspect of this application discloses as follows: Figure 1 The computer vision-based monitoring method for the construction process of large-diameter piles in karst areas, as shown, includes the following steps:

[0050] S1. Image Acquisition of Mud: High-definition industrial cameras are deployed at the mud tank inlet, mud tank outlet, and above the sedimentation tank. Real-time image data is continuously acquired at a preset frame rate and transmitted to the control center in real time. The real-time image data includes initial state image data of the mud before injection into the borehole, state image data of the mud after completing borehole circulation, and state change image data during the mud sedimentation process. The high-definition industrial cameras are dustproof, waterproof, and vibration-resistant, adaptable to the unfavorable conditions of large-diameter pile construction sites in karst areas. In specific implementation, a dimmable industrial light source is configured next to the high-definition industrial camera. The dimmable industrial light source adopts a ring or strip structure design, and uses a built-in light sensor to monitor ambient light intensity in real time, dynamically adjusting the luminous intensity to achieve adaptive brightness compensation under complex lighting conditions such as day-night cycles and changes in weather. The dimmable industrial light source can effectively suppress interference from uneven ambient light, specular reflection, equipment shadows, and water vapor refraction on image quality, ensuring the color fidelity, texture clarity, and spatiotemporal consistency of the real-time image data, providing a reliable data foundation for subsequent visual feature quantitative analysis.

[0051] S2. Quantitative Extraction of Visual Features: After receiving the real-time image data, the control center extracts the color features, turbidity features, bubble features, and solid particulate matter features of the real-time image data, thereby combining them to form a real-time feature vector. O 1 The color features include the mean hue value, variance hue value, skewness hue value, mean saturation value, variance saturation value, and skewness saturation value; the turbidity features include the turbidity gradient amplitude and the area ratio of turbidity gradient abrupt change regions; the bubble features include bubble density, bubble size, and bubble outline roundness; the solid particulate matter features include particle concentration, particle concentration gradient amplitude, and particle outline fractal dimension; in specific implementation, the following steps are included:

[0052] S201, Color Feature Extraction: Based on the color change characteristics formed by the mixture of soil particles, air bubbles, and karst seepage media in the mud, the HSV color gamut analysis model is used to calculate the color features of each pixel in the real-time image data. x , y hue value H ( x , y ) and saturation value S ( x , yThe system calculates the color features of the real-time image data, including the mean hue value, the variance of the hue value, the skewness of the hue value, the mean saturation value, the variance of the saturation value, and the skewness of the saturation value. Among these, an abnormal increase in the variance of the hue value is correlated with seepage from the fissures in the karst cave, and a sudden change in the saturation skewness is correlated with the phenomenon of rock debris being mixed into the mud due to the collapse of the borehole wall.

[0053] S202. Turbidity Feature Extraction: To achieve dynamic monitoring of the spatial distribution of suspended particle concentration, the real-time image data is converted into a grayscale image, and each pixel is extracted. x , y grayscale value M 1 ( x , y Then, the turbidity gradient amplitude is calculated using the central difference method. M 2 ( x , y The expression is as follows:

[0054] (1)

[0055] Among them, boundary pixels are mirrored to maintain image size; when pixel ( x , y ) turbidity gradient amplitude M 2 ( x , y )>2 M 3 When, mark the pixel point ( x , y ) represents abruptly changed pixels, where M 3 Under normal construction conditions M 2 ( x , y The median of ) ; the percentage of the area of ​​the region with abrupt change in turbidity gradient M 4 As shown in the following expression:

[0056] (2)

[0057] in, M 5 This represents the total number of mutated pixels. M 6 The total number of pixels; the percentage of the area of ​​the region with abrupt changes in turbidity gradient. M 4 It is related to the phenomenon of mud turbidity caused by seepage from cave fissures and collapse of borehole walls;

[0058] S203. Bubble Feature Extraction: The visual regions representing bubbles in the real-time image data are segmented and morphologically analyzed to extract bubble features, which are used to distinguish the source of bubbles and assist in disease judgment. The bubble features include bubble density, bubble size, and bubble outline roundness. The bubble density is the number of bubbles per unit area, calculated as the ratio of the number of pixels occupied by bubbles to the total number of pixels. The bubble size is the average equivalent diameter of the bubble, calculated as the average diameter of a circle with the same area as the bubble. The bubble outline roundness... M 7 As shown in the following expression:

[0059] (3)

[0060] in M 8 The area of ​​the bubble. M 9 The perimeter of the bubble outline; high-density microbubble clusters are correlated with seepage from cavern fissures, and bubbles with significantly irregular outlines are correlated with severe turbulent disturbances caused by pore wall instability;

[0061] S204. Extraction of solid particulate matter characteristics: Identify and analyze suspended or deposited solid particulate matter in the mud from the real-time image data, and extract solid particulate matter characteristics reflecting their source, state, and dynamics; the solid particulate matter includes primary rock fragments, collapsed debris, and metal fragments; the solid particulate matter characteristics include particle concentration. Q 1 Particle concentration gradient amplitude Q 2 and particle profile fractal dimension Q 3 Among them, particle concentration Q 1 The percentage of pixels occupied by solid particles in a unit area image is shown in the following expression:

[0062] (4)

[0063] in, Q 4 This represents the total number of pixels occupied by solid particles. Q 5 This represents the total number of pixels in the image.

[0064] Particle concentration gradient amplitude Q 2 Based on particle concentration Q 1 The central difference method is used for calculation;

[0065] (5)

[0066] Fractal dimension of particle profile Q 3 The box counting method is used for calculation, with a grid size of... S The mesh covers the particle outline, and S ∈[ S min , S max For each size S Calculate the minimum number of mesh boxes required to cover the contour. Q 6 ( S If the particle profile fractal dimension is 0, then... Q 3 As shown in the following expression:

[0067] (6)

[0068] In specific calculations, the fractal dimension of the particle profile... Q 3 Fit log using linear regression Q 6 ( S ) and log(1 / S The slope between ) is obtained.

[0069] S3. Construction of the Visual Feature Database: By constructing a sample dataset, training a feature extraction model, generating a disease feature vector set, and establishing a dynamic update mechanism, a visual feature database of mud diseases in karst areas is constructed and dynamically updated. This includes the following steps:

[0070] S301. Construct a sample dataset: Collect historical image data of the mud slurry wall protection and circulation hole formation process during the construction of large-diameter piles in karst areas, covering four scenarios, including karst cave fissure leakage, hole wall collapse, mud slurry performance deterioration and normal state, and the sample size of each of the four scenarios is not less than 5000 frames; preprocess the image data, including denoising, distortion correction, illumination normalization and annotation, the annotation content includes the type of disease, the bounding box of the feature area and key parameters;

[0071] S302. Training the feature extraction model: A convolutional neural network model is used as the feature extraction model. The convolutional neural network model includes an input layer, 5 convolutional layers, 3 pooling layers, 2 fully connected layers, and an output layer. The preprocessed image data is used as the training set, and the model parameters are optimized through the backpropagation algorithm to improve the accuracy of extracting deep visual features of mud color, turbidity, bubbles, and solid particles.

[0072] S303. Generate a set of feature vectors for the disease: Using a trained convolutional neural network model, extract features from historical image data to generate standard feature vector sets for the four scenarios. The feature vector set includes 2000 sets for karst cave and fissure seepage, 1800 sets for borehole wall collapse, 2200 sets for mud performance deterioration, and 2500 sets for normal state. Each set of feature vectors is correlated with the real-time feature vectors. O 1 The dimensions are consistent, including color features, turbidity features, bubble features, and solid particulate matter features;

[0073] S304. Establish a dynamic update mechanism: Adopt an incremental learning algorithm. When new construction scenarios introduce unrecorded disease features, the new samples are automatically included in the training set. The feature template is updated by fine-tuning the model parameters of the convolutional neural network model to ensure the adaptability of the feature library to complex geological conditions, and the update cycle does not exceed 72 hours.

[0074] S4. Classification Model Establishment: A multi-classification model based on SVM is adopted to establish real-time feature vectors. O 1 With disease probability vector O 2 The mapping relationship; wherein, when training the SVM-based multi-classification model, 70% of the samples are randomly selected from the visual feature database of mud diseases in the karst area as the training set and the remaining 30% of the samples are selected as the validation set; after training, the real-time feature vector is input. O 1 The disease probability vector is obtained. O 2 =[ P leak , P collape , P degradation , P normal ],in, P leak The probability of leakage from fissures in a karst cave. P collape The probability of the borehole wall collapsing. P degradation For the deterioration of mud properties, P normal The probability of being in a normal state; P leak , P collape , P degradation , P normal The values ​​of are all in the range [0,1], and Pleak + P collape + P degradation + P normal =1.

[0075] S5. Disease type and severity determination: The real-time feature vector generated in step S2 based on the real-time image data is used to determine the disease type and severity. O 1 The input is fed into the SVM-based multi-classification model trained in step S4, and the output is a disease probability vector. O 2 =[ P leak , P collape , P degradation , P normal Define the probability of occurrence of the most severe disease. P max =max{ P leak , P collape , P degradationl}, and accordingly determine the most severe disease type. T max ∈{karst cave fissure leakage, borehole wall collapse, mud performance deterioration}; Based on historical defect data of large-diameter piles in karst areas and engineering safety standards, early warning thresholds for defects are set. P max ];when P max ≤[ P max When [the situation is assessed], the current construction status is determined to be low-risk, and the regular monitoring frequency is maintained; when [the situation is assessed], ... P max >[ P max When determining that there is a significant risk of disease, a severity index is calculated. I M As shown in the following formula:

[0076] (7)

[0077] in β 1 and β 2 All are weighting coefficients, and their values ​​range from [0,1]. β 1 + β 2 =1; dP max / d t The rate of change of the probability of the target disease over time; when 0.5 ≤ I M When 1 < 1, a Level 1 response is triggered, and the control center outputs a low-frequency intermittent audible and visual warning signal to alert operators to potential risks; when 1 ≤ 1 I M When the alarm level is <2, a level 2 response is triggered, and the control center outputs an audible and visual warning signal and escalates the alarm level to prompt the operator to take intervention measures; when I M When the value is ≥2, a level 3 response is triggered, and the control center outputs a continuous high-intensity audible and visual warning signal to prompt the operator to take strong intervention measures.

[0078] In specific implementation, the most severe disease type determined in step S5 is used as the basis. T max and severity index I M This triggers a multi-level risk response control strategy, as follows:

[0079] like T max For seepage from cave fissures: when 0.5 ≤ I M When 1 < 1, a Level 1 response is triggered, outputting a low-frequency intermittent audible and visual warning signal to alert operators to potential risks; when 1 ≤ 1 I M When the speed is less than 2, a level 2 response is triggered, outputting an audible and visual warning signal and escalating the alarm level. The drilling speed is adjusted to 60% to 80% of the original speed to reduce disturbance within the hole, and the pre-discharge procedure for plugging materials is initiated. I M When ≥2, a level 3 response is triggered, outputting a continuous high-intensity audible and visual warning signal, controlling the drilling rig to stop drilling operations and raising the drill bit to a safe height of 1 to 2 meters below the borehole opening, and adding special plugging material into the mud at a volume of 5%-15% of the mud volume in the hole;

[0080] like T max For borehole wall collapse: when 0.5 ≤ I M When 1 < 1, a level 1 response is triggered, outputting a low-frequency intermittent audible and visual warning signal; when 1 ≤ 1, a level 1 response is triggered. I M When the value is less than 2, a secondary response is triggered, outputting an audible and visual warning signal and escalating the alarm level. The drilling rig's drilling speed is adjusted to 40% to 60% of its original speed to reduce disturbance within the borehole. Simultaneously, the mud circulation system is controlled to increase the mud density to a preset value. I MWhen the value is ≥2, a level 3 response is triggered, outputting a continuous high-intensity audible and visual warning signal, controlling the drilling rig to stop drilling operations and raising the drill bit to a safe height of 1 to 2 meters below the borehole opening, and starting the high-pressure grouting pump to inject high-viscosity, high-shearing-force mud into the borehole, with an injection volume of 1.2 to 1.5 times the borehole volume;

[0081] like T max Deterioration of mud properties: when 0.5 ≤ I M When 1 < 1, a level 1 response is triggered, outputting a low-frequency intermittent audible and visual warning signal; when 1 ≤ 1, a level 1 response is triggered. I M When the alarm level is <2, a secondary response is triggered, outputting an audible and visual warning signal and escalating the alarm level, while adding targeted mud treatment agent; when I M When the value is ≥2, a level 3 response is triggered, outputting a continuous high-intensity audible and visual warning signal to control the drilling rig to stop drilling operations and raise the drill bit to a safe height of 1 to 2 meters below the borehole opening, thereby discharging the original mud in the borehole and replacing it with newly prepared qualified mud.

[0082] The above describes one or more embodiments of the present invention in a relatively specific and detailed manner, but it should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this patent should be determined by the appended claims.

Claims

1. A computer vision-based monitoring method for the construction process of large-diameter piles in karst areas, characterized in that, Includes the following steps: S1. Image acquisition of mud: High-definition industrial cameras are deployed at the mud tank inlet, mud tank outlet and above the sedimentation tank to continuously acquire real-time image data and transmit it to the control center in real time; the real-time image data includes the initial state image data of mud before injection into the borehole, the state image data of mud after completing the borehole circulation, and the state change image data of mud during the sedimentation process. S2. Quantitative Extraction of Visual Features: After receiving the real-time image data, the control center extracts the color features, turbidity features, bubble features, and solid particulate matter features of the real-time image data, thereby combining them to form a real-time feature vector. O 1 ; The extraction of solid particulate matter features involves identifying and analyzing suspended or deposited solid particles within the mud from the real-time image data, extracting solid particulate matter features reflecting their source, state, and dynamics; the solid particles include primary rock fragments, collapsed debris, and metal fragments; the solid particulate matter features include particle concentration. Q 1 Particle concentration gradient amplitude Q 2 and particle profile fractal dimension Q 3 Among them, particle concentration Q 1 The percentage of pixels occupied by solid particles in a unit area image is shown in the following expression: in, Q 4 This represents the total number of pixels occupied by solid particles. Q 5 This represents the total number of pixels in the image. Particle concentration gradient amplitude Q 2 Based on particle concentration Q 1 The central difference method is used for calculation; Fractal dimension of particle profile Q 3 The box counting method is used for calculation, with a grid size of... S The mesh covers the particle outline, and S ∈[ S min , S max For each size S Calculate the minimum number of mesh boxes required to cover the contour. Q 6 ( S If the particle profile fractal dimension is 0, then... Q 3 As shown in the following expression: In specific calculations, the fractal dimension of the particle profile... Q 3 Fit log using linear regression Q 6 ( S ) and log(1 / S The slope between ) is obtained; S3. Construction of the visual feature library: By constructing a sample dataset, training a feature extraction model, generating a disease feature vector set, and establishing a dynamic update mechanism, a visual feature library of mud diseases in karst areas is constructed and dynamically updated. S4. Classification Model Establishment: A multi-classification model based on SVM is adopted to establish real-time feature vectors. O 1 With disease probability vector O 2 The mapping relationship; wherein, when training the SVM-based multi-classification model, 70% of the samples are randomly selected from the karst area mud disease visual feature database as the training set and the remaining 30% of the samples are selected as the validation set; disease probability vector O 2 =[ P leak , P collape , P degradation , P normal ],in, P leak The probability of leakage from fissures in a karst cave. P collape The probability of the borehole wall collapsing. P degradation For the deterioration of mud properties, P normal The probability of being in a normal state; P leak , P collape , P degradation , P normal The values ​​of are all in the range [0,1], and P leak + P collape + P degradation + P normal =1; S5. Disease type and severity determination: The real-time feature vector generated in step S2 based on the real-time image data is used to determine the disease type and severity. O 1 The input is fed into the SVM-based multi-classification model trained in step S4, and the output is a disease probability vector. O 2 =[ P leak , P collape , P degradation , P normal Define the probability of occurrence of the most severe disease. P max =max{ P leak , P collape , P degradationl }, and accordingly determine the most severe disease type. T max ∈{karst cave fissure leakage, borehole wall collapse, mud performance deterioration}; Based on historical defect data of large-diameter piles in karst areas and engineering safety standards, early warning thresholds for defects are set. P max ];when P max ≤[ P max When [the situation is assessed], the current construction status is determined to be low-risk, and the regular monitoring frequency is maintained; when [the situation is assessed], ... P max >[ P max When determining that there is a significant risk of disease, a severity index is calculated. I M As shown in the following formula: in β 1 and β 2 All are weighting coefficients, and their values ​​range from [0,1]. β 1 + β 2 =1; d P max / d t The rate of change of the probability of the target disease over time.

2. The method for monitoring the construction process of large-diameter piles in karst areas based on computer vision according to claim 1, characterized in that, High-definition industrial cameras are dustproof, waterproof, and vibration-resistant.

3. The method for monitoring the construction process of large-diameter piles in karst areas based on computer vision according to claim 1, characterized in that, In step S1, a dimmable industrial light source is configured next to the high-definition industrial camera; the dimmable industrial light source adopts a ring or strip structure design, and monitors the ambient light intensity in real time through a built-in light sensor to dynamically adjust the light intensity and achieve adaptive brightness compensation.

4. The computer vision-based monitoring method for the construction process of large-diameter piles in karst areas according to claim 1, characterized in that, In step S2, the color features include the mean hue value, variance hue value, skewness hue value, mean saturation value, variance saturation value, and skewness saturation value; the turbidity features include the turbidity gradient amplitude and the area ratio of turbidity gradient abrupt change regions; the bubble features include bubble density, bubble size, and bubble outline roundness; and the solid particulate matter features include particle concentration, particle concentration gradient amplitude, and particle outline fractal dimension. The color features, turbidity features, bubble features, and solid particulate matter features are extracted through the following steps: S201, Color feature extraction: Calculate the color features of each pixel in the real-time image data using the HSV color gamut analysis model. x , y hue value H ( x , y ) and saturation value S ( x , y This allows for the calculation of the mean hue value, variance hue value, skewness hue value, mean saturation value, variance saturation value, and skewness saturation value of the real-time image data. S202, Extraction of turbidity features: Convert the real-time image data into a grayscale image and extract each pixel point ( x , y grayscale value M 1 ( x , y Then, the turbidity gradient amplitude is calculated using the central difference method. M 2 ( x , y The expression is as follows: Among them, boundary pixels are mirrored to maintain image size; when pixel ( x , y ) turbidity gradient amplitude M 2 ( x , y )>2 M 3 When, mark the pixel point ( x , y ) represents abruptly changed pixels, where M 3 Under normal construction conditions M 2 ( x , y The median of ) ; the percentage of the area of ​​the region with abrupt change in turbidity gradient M 4 As shown in the following expression: in, M 5 This represents the total number of mutated pixels. M 6 The total number of pixels; S203. Bubble Feature Extraction: The visual regions representing bubbles in the real-time image data are segmented and morphologically analyzed to extract bubble features, which are used to distinguish the source of bubbles and assist in disease judgment. The bubble features include bubble density, bubble size, and bubble outline roundness. The bubble density is the number of bubbles per unit area, calculated as the ratio of the number of pixels occupied by bubbles to the total number of pixels. The bubble size is the average equivalent diameter of the bubble, calculated as the average diameter of a circle with the same area as the bubble. The bubble outline roundness... M 7 As shown in the following expression: in M 8 The area of ​​the bubble. M 9 This represents the perimeter of the bubble outline.

5. The computer vision-based monitoring method for the construction process of large-diameter piles in karst areas according to claim 1, characterized in that, In step S3, a visual feature library of mud sludge diseases in karst areas is constructed and dynamically updated, specifically including the following steps: S301. Construct a sample dataset: Collect historical image data of the mud slurry wall protection and circulation hole formation process during the construction of large-diameter piles in karst areas, covering four scenarios, including karst cave fissure leakage, hole wall collapse, mud slurry performance deterioration and normal state; S302. Training the feature extraction model: A convolutional neural network model is used as the feature extraction model. The convolutional neural network model includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The preprocessed labeled images are used as the training set, and the model parameters are optimized through the backpropagation algorithm. S303. Generate disease feature vector sets: Using a trained convolutional neural network model, extract features from historical image data to generate disease feature vector sets for the four scenarios. Each set of disease feature vectors is correlated with the real-time feature vectors. O 1 The dimensions are consistent; S304. Establish a dynamic update mechanism: Adopt an incremental learning algorithm. When new construction scenarios introduce unrecorded disease features, the new samples are automatically included in the training set. The disease feature vector set is updated by fine-tuning the model parameters of the convolutional neural network model to ensure the adaptability of the feature library to complex geological conditions.

6. The method for monitoring the construction process of large-diameter piles in karst areas based on computer vision according to claim 1, characterized in that, In step S5, when 0.5 ≤ I M When 1 < 1, a Level 1 response is triggered, and the control center outputs a low-frequency intermittent audible and visual warning signal to alert operators to potential risks; when 1 ≤ 1 I M When the alarm level is less than 2, a level 2 response is triggered, and the control center outputs an audible and visual warning signal and upgrades the alarm level to prompt the operator to take intervention measures. when I M When the value is ≥2, a level 3 response is triggered, and the control center outputs a continuous high-intensity audible and visual warning signal to prompt the operator to take strong intervention measures.

Images