Tunnel joint and water burst detection early warning method based on structured light unmanned aerial vehicle

Abstract

The invention discloses a tunnel joint and water burst detection early warning method based on a structured light unmanned aerial vehicle, and relates to the technical field of tunnels, and the method comprises the following steps: collecting three-dimensional point cloud data in a tunnel, and transmitting the three-dimensional point cloud data to a data display terminal through a communication tool; in combination with a convolutional neural network, surrounding rock joint density information in the tunnel is obtained, and the tunnel joint density is identified by using an AlexNet architecture; a random forest algorithm is adopted to construct a water inflow prediction model; and estimating the water inflow by using the water inflow prediction model, performing early warning judgment on the water inflow of the tunnel based on an early warning threshold value, and executing a preset early warning strategy according to a judgment result. Compared with a traditional joint density detection method, the structured light scanning technology can reduce the detection cost, data processing is relatively simple, the joint density estimation precision is greatly improved, evaluation of the permeability of rock is facilitated, and the safety of tunnel construction is guaranteed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of tunneling, in particular to a tunnel joint and water gushing detection and early warning method based on a structured light unmanned aerial vehicle. BACKGROUND

[0002] Joint density is very important for the engineering properties and stability of rock during tunnel excavation, as joints affect the strength, permeability, stability, and ease of engineering exploitation of rock. Therefore, during tunnel construction, detailed measurement and analysis of joint density are often required in order to make reasonable engineering design and engineering management. During tunnel construction, especially during excavation and support, if the joint density of rock is high, the stability of the rock mass will be reduced, and more measures need to be taken to prevent accidents caused by tunnel collapse.

[0003] Joint density has a large impact on groundwater flow. Rock with a higher joint density usually has a higher permeability, and water can flow rapidly through the joints, leading to the occurrence of water gushing in the tunnel; changes in joint density during tunnel construction can change the original hydrological conditions, leading to new water gushing problems; changes in joint density can make water gushing prediction and monitoring more complex, increasing the risk during construction.

[0004] Currently, visual and manual measurement methods are the most traditional and intuitive methods for measuring joint density. By observing the rock surface or core with the naked eye or with the aid of tools such as magnifying glasses or microscopes, the number and distribution density of joints are manually recorded and calculated. This method is simple and easy to use, and is suitable for small-scale core or rock surface analysis. In addition, commonly used methods include photography and image processing technology and laser scanning technology. Although the current water gushing early warning method in tunnel construction can effectively monitor potential water gushing risks, it also has some defects and limitations.

[0005] In summary, the existing technology has the following defects:

[0006] 1. Traditional visual and manual measurement methods require manual measurement and judgment, which have the problems of subjectivity and error.

[0007] 2. Joint density detection methods usually involve a large amount of data processing and analysis work, especially when dealing with high-resolution images or laser scanning data. This may require specialized data processing software and algorithms, making data interpretation and result extraction complex and time-consuming.

[0008] 3. It is difficult to adapt to complex construction environments. In complex terrain or surface structure conditions, joint density detection methods may encounter challenges, as these methods usually assume that the surface is flat or regular. Non-uniformity and irregularity can lead to difficulties in data analysis and an increase in errors.

[0009] 4. In some complex faults, fissure water development in geological conditions, engineering needs to detect the ability of existing technology, affect the water inflow forecast effect, tunnel construction process has a certain water inflow risk.

[0010] In view of the problems in the related art, no effective solution has been proposed so far. SUMMARY

[0011] (I) Technical problems to be solved

[0012] In view of the deficiencies in the prior art, the present application provides a tunnel joint and water inflow detection and warning method based on a structured light unmanned aerial vehicle, which has the purpose of predicting fissure water inflow and timely warning, thereby solving the problem of poor water inflow prediction effect.

[0013] (II) Technical solutions

[0014] To achieve the above purpose of predicting fissure water inflow and timely warning, the present application adopts the following specific technical solutions:

[0015] The tunnel joint and water inflow detection and warning method based on a structured light unmanned aerial vehicle comprises the following steps:

[0016] S1, based on the structured light scanner mounted on the unmanned aerial vehicle, collecting three-dimensional point cloud data inside the tunnel, and using a communication tool to transmit the three-dimensional point cloud data to a data display terminal;

[0017] S2, according to the three-dimensional point cloud data inside the tunnel, combining a convolutional neural network to obtain the surrounding rock joint density information inside the tunnel, and using an AlexNet architecture to identify the tunnel joint density to obtain joint density data;

[0018] S3, by fusing the joint density data, hydrogeological parameters and advanced geological prediction data, using a random forest algorithm to construct a water inflow prediction model;

[0019] S4, using the water inflow prediction model to estimate the water inflow size, based on the warning threshold to make a warning judgment on the water inflow of the tunnel, and executing a preset warning strategy according to the judgment result.

[0020] Preferably, the structured light scanner mounted on the unmanned aerial vehicle collects three-dimensional point cloud data inside the tunnel, and uses a communication tool to transmit the three-dimensional point cloud data to a data display terminal, comprising the following steps:

[0021] S11, based on the laser radar receiver and sensor built-in the unmanned aerial vehicle, respectively acquiring the first position information and attitude information of the unmanned aerial vehicle, and using the attitude information to calculate the second position information of the unmanned aerial vehicle;

[0022] S12, based on the first position information of the unmanned aerial vehicle and the second position information of the unmanned aerial vehicle, the first position information of the unmanned aerial vehicle and the second position information of the unmanned aerial vehicle are fused by a fusion algorithm, and the final position information of the unmanned aerial vehicle is obtained;

[0023] S13, based on the preset reference point, the running parameters of the unmanned aerial vehicle are adjusted according to the final position information of the unmanned aerial vehicle, so that the unmanned aerial vehicle moves to the reference point and scans the tunnel face working area, and obtains the three-dimensional point cloud data inside the tunnel.

[0024] Preferably, the first position information of the unmanned aerial vehicle and the second position information of the unmanned aerial vehicle are fused by a fusion algorithm, and the final position information of the unmanned aerial vehicle is obtained, including the following steps:

[0025] S121, the first position information of the unmanned aerial vehicle and the second position information of the unmanned aerial vehicle are time stamp aligned;

[0026] S122, based on the second position information of the unmanned aerial vehicle, the motion model of the unmanned aerial vehicle is constructed, and the current position of the unmanned aerial vehicle is predicted by using the motion model of the unmanned aerial vehicle, and the prediction result is obtained;

[0027] S123, based on the first position information of the unmanned aerial vehicle and the prediction result, the fingerprint observation vector winding processing is carried out, the credibility of the first position information is judged, and the first position of the unmanned aerial vehicle is processed based on the credibility judgment result, and the credible first position information is obtained;

[0028] S124, the prediction result and the credible first position information are fused and filtered by using the unscented Kalman filtering method, and the final position information of the unmanned aerial vehicle is obtained.

[0029] Preferably, the first position information of the unmanned aerial vehicle and the prediction result are processed by the fingerprint observation vector winding processing, the credibility of the first position information is judged, and the first position of the unmanned aerial vehicle is processed based on the credibility judgment result, and the credible first position information is obtained, including the following steps:

[0030] S1231, according to the maximum flight speed of the unmanned aerial vehicle, the sampling time interval and the tunnel characteristics, the displacement component boundary and the positioning deviation boundary of the unmanned aerial vehicle are calculated;

[0031] S1232, the displacement component of the prediction result of the unmanned aerial vehicle in the preset direction is calculated, and the displacement component of the first position information in the same preset direction is calculated;

[0032] S1233, compare the displacement component of the first position information with the displacement component of the prediction result and the displacement component boundary respectively, obtain a comparison result of the displacement component, and calculate a positioning deviation of the first position information from the prediction result and compare it with the positioning deviation boundary to obtain a positioning deviation boundary comparison result;

[0033] S1234, integrate the displacement component comparison result and the positioning deviation comparison result to judge the credibility of the first position information;

[0034] S1235, based on the credibility judgment result, if the first position information is credible, it is directly used, if it is not credible, the prediction result is used or combined with historical credible data for correction processing to obtain credible first position information.

[0035] Preferably, the step of using the unscented Kalman filter to fuse and filter the prediction result and the credible first position information to obtain the final position information of the unmanned aerial vehicle comprises the following steps:

[0036] S1241, initialize the unscented Kalman filter parameters, including the state vector, the covariance matrix, and the process noise and the observation noise of the unmanned aerial vehicle;

[0037] S1242, based on the motion model of the unmanned aerial vehicle, generate a Sigma point set through unscented transformation, and perform state prediction on the Sigma point set to obtain the mean value and the covariance matrix of the predicted state vector;

[0038] S1243, taking the credible first position information as an observation value, calculating the Kalman gain combined with the predicted state vector, and updating the state vector and the covariance matrix of the unmanned aerial vehicle using the Kalman gain;

[0039] S1244, extract the final position information of the unmanned aerial vehicle from the updated state vector, and output the information as the optimal estimation result of the current time position of the unmanned aerial vehicle.

[0040] Preferably, the step of obtaining the surrounding rock joint density information inside the tunnel according to the three-dimensional point cloud data inside the tunnel, combining a convolutional neural network, and identifying the tunnel joint density using an AlexNet architecture to obtain joint density data comprises the following steps:

[0041] S21, preprocessing the three-dimensional point cloud data inside the tunnel, and converting the preprocessed three-dimensional point cloud data into two-dimensional image data;

[0042] S22, performing labeling processing on the converted two-dimensional image data, marking the area representing the tunnel surrounding rock joint in the two-dimensional image and identifying the surrounding rock joint density to obtain a labeled data set, and dividing the labeled data set into a first training set, a first validation set and a first test set;

[0043] S23, a convolutional neural network model is built based on the AlexNet architecture, and the convolutional neural network model is trained, verified and tested by using the first training set, the first verification set and the first test set;

[0044] S24, the tunnel joint density of the latest tunnel internal three-dimensional point cloud data is identified by using the test completed convolutional neural network model, and the joint density data is obtained.

[0045] Preferably, the AlexNet architecture comprises a convolutional layer, a pooling layer and a fully connected layer.

[0046] The convolutional layer is used to extract joint features in the tunnel image.

[0047] The pooling layer is used to reduce the feature map dimension to reduce the calculation amount.

[0048] The fully connected layer is used to map the extracted joint features to the joint density category or value, and the ReLU activation function is used to enhance the nonlinear expression ability of the model.

[0049] Preferably, the convolutional neural network model is built based on the AlexNet architecture, and the convolutional neural network model is trained, verified and tested by using the first training set, the first verification set and the first test set, comprising the following steps:

[0050] S231, the first training set data is input into the convolutional neural network model based on the AlexNet architecture, and the convolutional neural network model is trained by using the first training set data.

[0051] S232, in each training cycle of the convolutional neural network model, the prediction result is obtained by forward propagation, and the loss value is calculated by using the loss function, and according to the loss value, the parameters of the convolutional neural network model are updated by using the back propagation algorithm.

[0052] S233, in the training process, the first verification set data is input into the current trained convolutional neural network model, and the verification result is calculated, and the hyperparameters of the convolutional neural network model are adjusted according to the verification result.

[0053] S234, the first test set data is input into the verified convolutional neural network model, the prediction result of the test set is obtained, and the performance of the verified convolutional neural network model is evaluated according to the prediction result of the test set, if the preset performance requirement is reached, the current convolutional neural network model is used as the final convolutional neural network model, otherwise, return to step S231.

[0054] Preferably, the random forest algorithm is used to build the water inflow prediction model by fusing the joint density data, the hydrogeological parameters and the advanced geological prediction data.

[0055] S31, collect the hydrogeological parameters of the tunnel, the advanced geological prediction data and the water inflow, and combine the joint density data to construct a prediction data set;

[0056] S32, data checking and standardization processing are performed on the prediction data set to obtain a standardized prediction data set;

[0057] S33, based on the standardized prediction data set, the correlation between the hydrogeological parameters, the advanced geological prediction data and the joint density data and the water inflow is calculated by using the correlation analysis method, and according to the correlation result, the parameters meeting the preset correlation threshold are selected, and are divided into a second training set, a second verification set and a second test set;

[0058] S34, an initial water inflow prediction model is constructed based on the random forest algorithm, and the second training set, the second verification set and the second test set are used to train, verify and test the initial water inflow prediction model, and a final water inflow prediction model is obtained.

[0059] Preferably, the initial water inflow prediction model based on the random forest algorithm is constructed, and the second training set, the second verification set and the second test set are used to train, verify and test the initial water inflow prediction model, and a final water inflow prediction model is obtained. The steps include:

[0060] S341, parameter initialization processing is performed on the initial water inflow prediction model constructed based on the random forest algorithm;

[0061] S342, the initial water inflow prediction model subjected to parameter initialization processing is trained by using the second training set, each decision tree in the initial water inflow prediction model is used to perform deep learning on the second training set, and the association between the data features and the tunnel water inflow phenomenon is determined by splitting and judging the data features of the second training set;

[0062] S343, in the training process of the initial water inflow prediction model, the second verification set is used to optimize the performance of the initial water inflow prediction model, and the grid search and random search algorithms are used to adjust the hyperparameters of the initial water inflow prediction model to obtain the optimal hyperparameter combination;

[0063] S344, the optimal hyperparameter combination is applied to the initial water inflow prediction model, and the initial water inflow prediction model is tested and evaluated by using the second test set to obtain a final water inflow prediction model.

[0064] (Three) beneficial effects

[0065] Compared with the prior art, the tunnel joint and water inflow detection and early warning method based on the structured light unmanned aerial vehicle has the following beneficial effects:

[0066] (1) Compared with the traditional joint density detection method, the structured light scanning technology can reduce the detection cost, the data processing is relatively simple, and the joint density estimation accuracy is greatly improved, which helps to evaluate the permeability of the rock and ensure the safety of the tunnel construction.

[0067] (2) The unmanned aerial vehicle structured light scanning method can realize the control and monitoring of the construction quality in the tunnel construction process, and ensure the accuracy and stability in the construction process.

[0068] (3) By determining the joint density information of the tunnel surrounding rock, combining the hydrogeological conditions and the advanced prediction results, using artificial intelligence technology, a random forest prediction model is constructed to realize the intelligent early warning of the tunnel water gushing, analyze the feature importance, help to understand the influence of different factors on water gushing, realize the intelligent early warning of the tunnel water gushing, and improve the accuracy of water gushing early warning.

[0069] (4) The present application provides a method combining structured light scanning and unmanned aerial vehicle for detecting the joint density of surrounding rock in tunnel construction and predicting water gushing, using an accurate identification device to realize rapid and accurate detection of the joint density of surrounding rock in tunnel construction, and combining hydrogeological conditions and advanced geological prediction results to construct a random forest model using artificial intelligence technology to predict and timely warn of the fracture water gushing, improve the efficiency and safety of construction. BRIEF DESCRIPTION OF DRAWINGS

[0070] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings described below are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0071] Figure 1 is a flowchart of the tunnel joint and water gushing detection and early warning method based on the structured light unmanned aerial vehicle according to the embodiments of the present application. DETAILED DESCRIPTION

[0072] In order to further illustrate the embodiments, the present application provides drawings, which are part of the disclosure of the present application, mainly used to illustrate the embodiments, and can explain the operating principle of the embodiments in conjunction with the related description of the specification. With reference to these contents, those skilled in the art should understand other possible embodiments and advantages of the present application. The components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0073] According to the embodiments of the present application, a tunnel joint and water gushing detection and early warning method based on a structured light unmanned aerial vehicle is provided.

[0074] It should be noted that, in order to realize the tunnel joint and water gushing detection and early warning method based on the structure light unmanned aerial vehicle, the following technologies also need to be implemented:

[0075] The positioning system, the structure light scanning technology, the data transmission and processing system, and the intelligent early warning system.

[0076] The positioning system is an inertial navigation system (INS) and a laser radar positioning system, which provides a reference point position for the identified joint.

[0077] The structure light scanning technology is a technology for three-dimensional object scanning and modeling, which projects light and observes the reflected light of the object surface to obtain the surface joint information of the object surface, and is not affected by light changes and object textures, and has the characteristics of high resolution and low power consumption. By emitting a laser beam through a three-dimensional laser scanning technology, when the laser beam is projected onto an irregular tunnel surface, the light pattern will be deformed due to the shape and characteristics of the surface. These deformations contain information about the surface geometry. At the same time, the structure light scanning technology is used to analyze the deformation of the light pattern to accurately capture the details of the tunnel surface, which is suitable for capturing complex surfaces and shallow structures. By combining the two technologies, high-precision point cloud data containing the geometric shape of the tunnel lining, joint and other disease information is generated, so that accurate joint density information is obtained.

[0078] The data transmission system is based on a real-time transmission system of an unmanned aerial vehicle, which transmits the collected data to a tunnel outside workstation, and then performs denoising, alignment, registration and fitting on the data to obtain accurate joint density information.

[0079] After the data processing result is fed back to the workstation, a water gushing prediction model based on machine learning is constructed by combining hydrogeological conditions and advanced geological prediction results. According to the threshold set by the prediction result, the possibility of tunnel water gushing and the size of water gushing are judged and timely warning is given. Visualization tools such as Matplotlib and Seaborn are used to display the comparison between the prediction result of the model and the actual water gushing amount, analyze the importance of features, and help understand the influence of different factors on water gushing.

[0080] The present application will be further described in conjunction with the drawings and specific embodiments, as shown in Figure 1 The tunnel joint and water gushing detection and early warning method based on the structure light unmanned aerial vehicle according to the embodiment of the present application includes the following steps:

[0081] S1, based on the structure light scanner mounted on the unmanned aerial vehicle, three-dimensional point cloud data inside the tunnel is collected, and the three-dimensional point cloud data is transmitted to a data display terminal by using a communication tool;

[0082] Specifically, the selection and configuration of the drone in this invention requires: a drone suitable for tunnel construction, with stable flight capabilities and the ability to perform flexible flight operations within the narrow tunnel. Secondly, the drone's flight time and load capacity must also meet actual requirements.

[0083] Among them, when processing three-dimensional point cloud data, a structured light scanning system is required. The structured light scanning system consists of two parts: a built-in 3D structured light scanner and a data processing module. The built-in 3D structured light scanner is located at the bottom of the drone and can scan the tunnel face working area 360° in all directions; the data processing module can process and analyze the data through algorithms such as filtering, alignment and matching, thereby extracting relevant information about the tunnel joint density.

[0084] As a preferred embodiment, the method of collecting three-dimensional point cloud data inside the tunnel based on a structured light scanner mounted on a drone and transmitting the three-dimensional point cloud data to a data display terminal using a communication tool includes the following steps:

[0085] S11, based on the laser radar receiver and sensor built into the drone, respectively obtain first position information and attitude information of the drone, and calculate second position information of the drone using the attitude information;

[0086] It should be noted that during tunnel construction, a LiDAR receiver installed on a drone can obtain real-time information about the drone's position in the tunnel. After receiving the signal, the LiDAR receiver uses the obtained position information to adjust the drone to various reference points.

[0087] By installing an INS sensor on the drone, the drone's attitude information can be obtained in real time, the drone's acceleration and angular velocity can be measured, and the drone's attitude angle can be calculated using the principles of kinematics and dynamics. When using the INS sensor to calculate the drone's attitude angle, the body coordinate system and the geographic coordinate system must be defined first. Next, the accelerometer and gyroscope are used to collect the acceleration and angular velocity components of the drone in the body coordinate system. For the acceleration data, the static state calibration method is used to subtract the influence of gravity acceleration. First, select an appropriate acceleration sensor, connect it to the microcontroller or data acquisition device, initialize it, and set the sampling frequency, range and other parameters. Next, place the sensor horizontally and collect data for a certain period of time to calculate the average value of gravity acceleration on each coordinate axis g. x 、g y 、g z Then, when the object moves, the acceleration data is collected in real time at a predetermined sampling frequency. x 、a y 、a z , and by formula a x′ =a x-g x , a y′ = a y -g y , a z′ = a z -g z , subtract the effect of gravity acceleration, get the real motion acceleration of the object, and perform numerical integration on the acceleration to get the velocity;

[0088] The angular velocity data is updated by means of the quaternion method, and then the quaternion is converted into Euler angles. An initial quaternion is set to represent the initial attitude of the UAV. It is usually set as a unit quaternion, that is, the real part is 1 and the imaginary parts are all 0. The unit quaternion means no rotation, that is, the UAV is in the initial reference attitude. The sampling time interval of the angular velocity data is determined. This time interval is determined by the sampling frequency of the INS sensor, which represents how long the angular velocity data is obtained every time, which will be used in the subsequent calculation of the quaternion update. The angular velocity of the UAV in the body coordinate system is obtained by means of the gyroscope sensor installed on the UAV. The angular velocity is generally represented by a three-dimensional vector, which corresponds to the rotation speed around the three axes (X, Y, Z axes) of the body coordinate system, and the unit is usually rad / s. For example, the obtained angular velocity vector ω = [ωx, ωy, ωz], wherein ωx is the angular velocity around the X axis, ωy is the angular velocity around the Y axis, and ωz is the angular velocity around the Z axis. In order to update the attitude by means of the quaternion, the obtained angular velocity vector needs to be converted into a pure imaginary quaternion Ω. The real part of the pure imaginary quaternion is 0, and the imaginary part is the three components of the angular velocity vector, that is, Ω = [0, ωx, ωy, ωz].

[0089] The derivative of the quaternion describes the rate of change of the quaternion with time. According to the kinematics equation of the quaternion, the quaternion q' is related to the current quaternion q and the angular velocity quaternion Ω, and the calculation formula is Here, represents the quaternion multiplication.

[0090] After obtaining the derivative of the quaternion, the current quaternion can be updated. The first-order Euler integration method is used to approximately calculate the quaternion q new at the next time, and the formula is q new = q + q' Δt, wherein Δt is the sampling time interval determined in the initialization. This method is simple and intuitive, but may produce certain errors when the sampling time interval is large.

[0091] In the updating process of the quaternion, the modulus of the quaternion may deviate from 1 due to calculation errors and other reasons. Only the unit modulus quaternion can accurately represent the rotation, so the updated quaternion q newNormalization. Normalization is to divide each component of the quaternion by its modulus, so that the modulus of the normalized quaternion is 1.

[0092] Convert the normalized quaternion to Euler angles, which usually include yaw, pitch and roll, which can more intuitively describe the attitude of the UAV in the geographical coordinate system. The specific conversion formula is as follows:

[0093] Pitch angle (Pitch): θ = arcsin(2(wy-zx));

[0094] Roll angle (Roll): φ = arctan2(2(wx+yz), 1-2(x 2 +y 2 ));

[0095] Yaw angle (Yaw): ψ = arctan2(2(wz+xy), 1-2(y 2 +z 2 ));

[0096] Where w, x, y, z are the real and imaginary parts of the normalized quaternion. arcsin is the inverse sine function, and arctan is the four quadrant inverse tangent function, which can correctly determine the quadrant of the angle according to the numerator and denominator.

[0097] Through the above steps, the real-time solution of the UAV attitude can be realized, and high-precision Euler angle output is provided for flight control. Finally, the displacement data is transmitted back to the control center, and at the same time, in order to reduce the accumulation of integral error, laser radar, vision sensor and other data can also be combined to compensate for errors through Kalman filtering and other algorithms. First, define the system state and measurement model, initialize the state estimation, covariance matrix and other parameters of the filter, then predict the state and covariance according to the system dynamic model, then calculate the Kalman gain to update the state and covariance according to the new measurement data, and repeatedly predict and update to fuse INS sensor, laser radar, vision sensor and other multi-sensor data, realize the optimal estimation of the UAV attitude and displacement, and reduce the error accumulation.

[0098] Finally, the motion information obtained by the sensor is used to calculate the displacement of the UAV, and the data is transmitted back to the control center for setting the next reference point.

[0099] S12, based on the first position information of the UAV and the second position information of the UAV, the first position information of the UAV and the second position information of the UAV are fused by a fusion algorithm to obtain the final position information of the UAV;

[0100] As a preferred embodiment, the first position information of the unmanned aerial vehicle and the second position information of the unmanned aerial vehicle are fused by a fusion algorithm to obtain the final position information of the unmanned aerial vehicle, including the following steps:

[0101] S121, time stamp alignment processing is performed on the first position information of the unmanned aerial vehicle and the second position information of the unmanned aerial vehicle;

[0102] It should be noted that the time stamp alignment processing is to ensure the consistency of the two kinds of position information in time, so as to perform subsequent fusion processing, including interpolation, extrapolation or adjustment of the time stamp, so that the two kinds of position information can be corresponded in time.

[0103] S122, a motion model of the unmanned aerial vehicle is constructed based on the second position information of the unmanned aerial vehicle, and the current position of the unmanned aerial vehicle is predicted by using the motion model of the unmanned aerial vehicle to obtain a prediction result;

[0104] Specifically, the motion model of the unmanned aerial vehicle is a mathematical model used to describe the motion law of the unmanned aerial vehicle in space, including the change of its position, velocity, acceleration and other state variables with time. The motion model of the unmanned aerial vehicle is constructed based on the dynamic characteristics of the unmanned aerial vehicle, flight environment and control input, etc. Through the motion model, the future position and state of the unmanned aerial vehicle can be predicted according to its historical state and control input. Among them, the motion model of the unmanned aerial vehicle is a nonlinear system, assuming that the position of the unmanned aerial vehicle in the two-dimensional plane is (x, y), the velocity is (v x ,v y ), and the acceleration is (a x ,a y ), the motion model of the unmanned aerial vehicle can be represented by the following differential equation set:

[0105]

[0106] In the formula, t represents time. In actual application, the acceleration (a x ,a y ) is determined by the control input of the unmanned aerial vehicle (such as thrust, pitch angle, roll angle, etc.) and external disturbance (such as wind, gravity, etc.).

[0107] S123, based on the first position information of the unmanned aerial vehicle and the prediction result, a fingerprint observation vector winding processing is performed, the credibility of the first position information is judged, and a credible first position information is obtained based on the credibility judgment result.

[0108] As a preferred embodiment, the first position information based on the unmanned aerial vehicle and the prediction result are subjected to fingerprint observation vector winding processing, the credibility of the first position information is judged, and the first position of the unmanned aerial vehicle is subjected to credibility processing based on the credibility judgment result to obtain credible first position information, including the following steps:

[0109] S1231, according to the maximum flight speed of the unmanned aerial vehicle, the sampling time interval and the tunnel characteristics, the displacement component boundary and the positioning deviation boundary of the unmanned aerial vehicle are calculated;

[0110] It should be noted that the maximum flight speed refers to the maximum flight speed of the unmanned aerial vehicle, which is a key factor limiting the maximum distance it can move within a certain time. The sampling time interval refers to the sampling time interval of the unmanned aerial vehicle position information, which determines the maximum distance the unmanned aerial vehicle can move between two samplings. The tunnel characteristics refer to the geometric shape, size, obstacle distribution and other characteristics of the tunnel, which may affect the flight path and positioning accuracy of the unmanned aerial vehicle. The displacement component boundary refers to the maximum and minimum displacement components (e.g. displacement in x, y, z directions) that the unmanned aerial vehicle can move within the sampling time interval. The positioning deviation boundary refers to the maximum and minimum values of the positioning deviation caused by various errors (such as sensor errors, environmental interference, etc.).

[0111] S1232, the displacement component of the prediction result of the unmanned aerial vehicle in a preset direction is calculated, and the displacement component of the first position information in the same preset direction is calculated;

[0112] S1233, the displacement component of the first position information is compared with the displacement component of the prediction result and the displacement component boundary respectively, the comparison displacement component comparison result is obtained, and the positioning deviation of the first position information and the prediction result is calculated and compared with the positioning deviation boundary, the positioning deviation boundary comparison result is obtained;

[0113] S1234, the credibility of the first position information is judged by synthesizing the displacement component comparison result and the positioning deviation comparison result;

[0114] Specifically, the credibility evaluation standard sets some evaluation standards (such as displacement component difference threshold, positioning deviation threshold, etc.), when the first position information meets these standards, it is considered that the credibility is higher; otherwise, it is considered that the credibility is lower.

[0115] S1235, based on the credibility judgment result, if the first position information is credible, it is directly used, if it is not credible, the prediction result or the historical credible data is combined for correction processing to obtain the credible first position information.

[0116] It should be noted that if the credibility of the first position information is high, it is directly used as the credible position information at the current time. If the credibility of the first position information is low, the prediction result can be used as the position information at the current time, or the first position information can be corrected by combining historical credible data (such as using Kalman filtering, particle filtering, etc. to fuse and correct data).

[0117] S124, fusing and filtering the prediction result and the credible first position information by using an unscented Kalman filtering method to obtain the final position information of the unmanned aerial vehicle.

[0118] As a preferred embodiment, the step of fusing and filtering the prediction result and the credible first position information by using an unscented Kalman filtering method to obtain the final position information of the unmanned aerial vehicle includes the following steps:

[0119] S1241, initializing unscented Kalman filtering parameters, including a state vector of the unmanned aerial vehicle, a covariance matrix, and process noise and observation noise;

[0120] It should be noted that the state vector is a vector describing the state of the unmanned aerial vehicle (such as position, velocity, acceleration, etc.). When initializing, the initial value of the state vector needs to be set according to the actual situation and initial conditions of the unmanned aerial vehicle. The covariance matrix describes the correlation between each state variable in the state vector and its uncertainty. When initializing, it is usually set as a diagonal matrix, and the elements on the diagonal line represent the initial variance (i.e. the measure of uncertainty) of the corresponding state variable. Process noise reflects the uncertainty in the motion model of the unmanned aerial vehicle, such as wind disturbance, control input error, etc. Process noise is usually represented by a covariance matrix, which describes the distribution of noise in each state variable. Observation noise reflects the uncertainty in the sensor measurement data. Similar to process noise, observation noise is also represented by a covariance matrix, which describes the distribution of noise in each observation variable (such as position, velocity, etc.).

[0121] S1242, based on the motion model of the unmanned aerial vehicle, generating a Sigma point set through unscented transformation, and performing state prediction on the Sigma point set to obtain the mean value and covariance matrix of the predicted state vector;

[0122] Specifically, unscented transformation (UT) is a method for approximating the probability distribution of a nonlinear function. The core idea is to select a set of specific sample points (i.e. Sigma points), which can capture the main characteristics (such as mean and covariance) of the input distribution, then transform these sample points through a nonlinear function, and finally use the transformed sample points to approximate the statistical characteristics of the output distribution. The specific implementation process is as follows:

[0123] Assume that the state vector of the UAV is g∈R n (n is the dimension of the state vector), the mean value of which is gˉ, and the covariance matrix is P gg For an n-dimensional state vector, 2n+1 Sigma points are usually selected, including one mean point and 2n symmetrically distributed points, to obtain a Sigma point set.

[0124] The generated Sigma point set is substituted into the motion model of the UAV to perform one-step prediction to obtain the predicted state vector of each Sigma point at the next time.

[0125] According to the predicted Sigma point set, the mean value and the covariance matrix of the predicted state vector are calculated.

[0126] S1243, the trusted first position information is taken as an observation value, the Kalman gain is calculated in combination with the predicted state vector, and the state vector and the covariance matrix of the UAV are updated by using the Kalman gain;

[0127] It should be noted that in the Kalman filtering framework, the observation model describes how to obtain the observation value from the state vector. Assume that the observation model of the UAV is: z k =h(g k ,v k );

[0128] Wherein: z k is the observation value (i.e. the trusted first position information) at time k, h(·) is an observation function, which describes the mapping relationship from the state vector to the observation value, v k is the observation noise at time k, and the covariance matrix thereof is R k .

[0129] The Kalman gain is a weight factor for balancing the contributions of the predicted state vector and the observation value in the state update. The calculation process is as follows:

[0130] For each Sigma point, the corresponding predicted observation value is calculated, and then the mean value of the predicted observation value is calculated.

[0131] According to the mean value of the predicted observation value, the observation covariance matrix and the mutual covariance matrix of the state and the observation are calculated, and the observation covariance matrix and the mutual covariance matrix of the state and the observation are multiplied to obtain the Kalman gain.

[0132] Finally, the observation value and the predicted state vector can be fused by using the Kalman gain to obtain the updated state vector and the covariance matrix.

[0133] S1244, the final position information of the UAV is extracted from the updated state vector, and the information is output as the optimal estimation result of the current time position of the UAV.

[0134] S13, based on the pre-set reference point, adjusting the operation parameters of the unmanned aerial vehicle according to the final position information of the unmanned aerial vehicle, so that the unmanned aerial vehicle moves to the reference point and scans the tunnel face working area, and obtains the three-dimensional point cloud data inside the tunnel.

[0135] It should be noted that the structured light scanner device should be selected to have high precision, high resolution and large measurement range. The structured light scanner device is used to scan the tunnel construction face. The device emits a laser beam, which is reflected back to the scanner after intersecting with the construction surface. The scanner calculates the three-dimensional coordinates of the construction surface by measuring the reflection time and angle of the laser beam. The process is as follows: the laser in the scanner generates a high-intensity laser beam, which is focused by optical devices and directed to the construction surface. The reflected light is captured by the receiver and converted into an electrical signal. By recording the time difference between emission and reception, combined with the propagation speed of laser in air, the distance between the scanner and the irradiated point on the construction surface is calculated by the formula d = c x t / 2, where c represents the propagation speed of laser in air, and t represents the time difference between the emission of laser beam from the scanner to the construction surface and the reflection back to the scanner.

[0136] The internal structure of the structured light scanner contains a precise angle measurement system, such as an encoder, which measures the scanning angles of the laser beam in horizontal and vertical directions in real time during scanning. The ground three-dimensional laser scanner makes the laser beam scan the measured area through the rotation of two synchronous mirrors, and the built-in precise clock controls the encoder to measure the transverse scanning angle observation value α and the longitudinal scanning angle observation value θ of each laser pulse synchronously. Then, according to the measured distance and angle information, combined with the scanner's own coordinate system, the three-dimensional coordinates of the points on the construction surface are calculated by trigonometric functions. In the left-handed coordinate system determined by the three-dimensional laser scanner, assuming that the position of the scanner is the origin (0, 0, 0), the distance d, the horizontal angle α and the vertical angle θ are known, and the three-dimensional coordinates (x, y, z) of a certain measured point P are obtained by the formulas x = d x sin θ x cos α, y = d x sin θ x sin α, z = d x cos θ. The scanner scans the construction surface at different positions and angles multiple times to obtain the three-dimensional coordinates of a large number of measurement points to form point cloud data.

[0137] In specific application, the laser emits a large number of beams to hit the ground, and the receiver collects the returned light and records the time and intensity of each reflection, thereby measuring the distance and reflection intensity, which is then marked by the receiver to form three-dimensional point cloud data. Then, based on the transmission controller, the collected point cloud data can be transmitted to the data display terminal through transmission lines by using communication tools such as the Internet and mobile communication networks for further analysis and processing. The data display terminal receives the point cloud data transmitted from the transmission controller and presents it on the screen.

[0138] S2, obtaining the surrounding rock joint density information inside the tunnel according to the three-dimensional point cloud data inside the tunnel in combination with the convolutional neural network, and identifying the tunnel joint density by using the AlexNet architecture to obtain the joint density data;

[0139] As a preferred embodiment, the obtaining the surrounding rock joint density information inside the tunnel according to the three-dimensional point cloud data inside the tunnel in combination with the convolutional neural network, and identifying the tunnel joint density by using the AlexNet architecture to obtain the joint density data comprises the following steps:

[0140] S21, preprocessing the three-dimensional point cloud data inside the tunnel, and converting the preprocessed three-dimensional point cloud data into two-dimensional image data;

[0141] It should be noted that the collected data is preprocessed by denoising, filtering, registration and the like to improve the quality. The three-dimensional point cloud data is converted into two-dimensional image data suitable for CNN processing. The projection method can be used to project the point cloud onto a plane to generate a depth image or an intensity image, or the point cloud data is converted into multi-channel image data, each channel representing different point cloud attributes. In order to increase the diversity of training data and improve the generalization ability of the model, data enhancement operations such as random rotation, flipping, scaling, adding noise and the like are performed on the converted image data.

[0142] S22, labeling the converted two-dimensional image data to mark the area representing the tunnel surrounding rock joint in the two-dimensional image and identifying the surrounding rock joint density to obtain a labeled data set, and dividing the labeled data set into a first training set, a first validation set and a first test set;

[0143] It should be noted that the preprocessed image data is labeled by professional geologists or technicians, and the joint area in the image is marked, and the joint area is labeled as a specific class or label as supervision information for model training. The labeled data set is divided into a training set, a validation set and a test set according to a certain proportion. The division proportion is 70% training set, 15% validation set and 15% test set. The training set is used for model training, the validation set is used for adjusting the hyperparameters of the model during training, evaluating the performance of the model, preventing overfitting, and the test set is used for final evaluation of the generalization ability and recognition accuracy of the model.

[0144] S23, building a convolutional neural network model based on the AlexNet architecture, and training, validating and testing the convolutional neural network model by using the first training set, the first validation set and the first test set;

[0145] As a preferred embodiment, the AlexNet architecture comprises a convolutional layer, a pooling layer and a fully connected layer.

[0146] The convolutional layer is configured to extract joint features in the tunnel image.

[0147] The pooling layer is configured to reduce the dimension of the feature map to reduce the amount of calculation.

[0148] The fully connected layer is configured to map the extracted joint features to the category or value of joint density, and a ReLU activation function is used to enhance the nonlinear expression ability of the model.

[0149] As a preferred embodiment, the convolutional neural network model is built based on the AlexNet architecture, and the training, verification and testing of the convolutional neural network model using the first training set, the first verification set and the first test set include the following steps:

[0150] S231, input the first training set data into the convolutional neural network model built based on the AlexNet architecture, and train the convolutional neural network model using the first training set data;

[0151] S232, in each training cycle of the convolutional neural network model, the predicted result is obtained by forward propagation, and the loss value is calculated using the loss function, and the parameters of the convolutional neural network model are updated using the back propagation algorithm according to the loss value;

[0152] S233, in the training process, input the first verification set data into the currently trained convolutional neural network model, calculate the verification result, and adjust the hyperparameters of the convolutional neural network model according to the verification result;

[0153] S234, input the first test set data into the verified convolutional neural network model to obtain the predicted result of the test set, and evaluate the performance of the verified convolutional neural network model according to the predicted result of the test set, if the performance requirement is met, the current convolutional neural network model is used as the final convolutional neural network model, otherwise, return to step S231.

[0154] S24, using the completed convolutional neural network model, the latest three-dimensional point cloud data of the tunnel interior is used to identify the tunnel joint density to obtain the joint density data.

[0155] Specifically, in the identification of tunnel joint density, the classic CNN architecture AlexNet is selected. The AlexNet architecture includes convolutional layers, pooling layers, and fully connected layers. The convolutional layers are responsible for extracting joint features in the tunnel image, the pooling layers reduce the dimension of the feature map to reduce the amount of calculation, and the fully connected layers map the extracted features to specific categories or values of joint density, while the ReLU activation function enhances the non-linear expression ability of the model. The loss function uses the cross-entropy loss function, combined with optimizers such as stochastic gradient descent (SGD) or Adam to update the model parameters to minimize the loss function. During model training, the training set data is input into the model. In each training period, the model first performs forward propagation to obtain the predicted results, then calculates the loss value according to the loss function, and updates the parameters through the back propagation algorithm. During training, the performance is monitored using the validation set, and the learning rate, batch size, and other hyperparameters are adjusted according to the validation set indicators. To optimize model performance, techniques such as L1 and L2 regularization, Dropout, and different hyperparameter combinations and fine-tuning architectures are used to prevent overfitting. The optimal model is selected after multiple experiments. In the model evaluation stage, the accuracy, recall, F1 value, mean square error, and other indicators are calculated using the test set. If the performance does not meet the requirements, the model structure or hyperparameters are adjusted and retrained for optimization.

[0156] Finally, the qualified model is applied to the actual task. The tunnel structure light scanning data collected is preprocessed and input into the model to obtain the joint region identification result. The joint density is calculated by counting the number or length of joints per unit area, providing decision-making basis for the design, construction, and management of tunnel engineering.

[0157] S3, by fusing joint density data, hydrogeological parameters and advanced geological prediction data, a random forest algorithm is used to construct a water inflow prediction model;

[0158] As a preferred embodiment, the step of constructing a water inflow prediction model by fusing joint density data, hydrogeological parameters and advanced geological prediction data, and using a random forest algorithm includes the following steps:

[0159] S31, collecting hydrogeological parameters, advanced geological prediction data and water inflow of the tunnel, and combining with joint density data to construct a prediction data set;

[0160] S32, performing data inspection and standardization processing on the prediction data set to obtain a standardized prediction data set;

[0161] S33, based on the standardized prediction data set, the correlation between the hydrogeological parameters, advanced geological prediction data and joint density data and the water inflow is calculated using the correlation analysis method, and according to the correlation results, the parameters that meet the preset correlation threshold are selected and divided into a second training set, a second validation set and a second test set;

[0162] S34, an initial water inflow prediction model is constructed based on a random forest algorithm, and the initial water inflow prediction model is trained, verified and tested by using the second training set, the second verification set and the second test set to obtain a final water inflow prediction model.

[0163] As a preferred embodiment, the step of constructing an initial water inflow prediction model based on a random forest algorithm and training, verifying and testing the initial water inflow prediction model by using the second training set, the second verification set and the second test set to obtain a final water inflow prediction model includes the following steps:

[0164] S341, parameter initialization processing is performed on the initial water inflow prediction model constructed based on the random forest algorithm;

[0165] S342, the initial water inflow prediction model after parameter initialization processing is trained by using the second training set, and each decision tree in the initial water inflow prediction model is used to perform deep learning on the second training set, and the association between data features and tunnel water inflow phenomena is determined by splitting and judging the data features of the second training set;

[0166] S343, in the training process of the initial water inflow prediction model, the performance of the initial water inflow prediction model is optimized by using the second verification set, and the hyperparameters of the initial water inflow prediction model are adjusted by using the grid search and random search algorithms to obtain an optimal hyperparameter combination;

[0167] S344, the optimal hyperparameter combination is applied to the initial water inflow prediction model, and the initial water inflow prediction model is tested and evaluated by using the second test set to obtain a final water inflow prediction model.

[0168] Specifically, based on the obtained point cloud data and joint density information, a solid and reliable basis is provided for the prediction and identification of geological disaster hazards. By determining the joint density information of the tunnel surrounding rock, combined with the hydrogeological conditions, the results of advanced geological prediction and the correlation prediction of joints and water inflow, a random forest model is constructed.

[0169] Random forest belongs to the method of ensemble learning, which is composed of multiple decision trees. Decision tree is a model based on tree structure for decision making, the internal node is attribute test, the branch is test output, and the leaf node is class or value. Random forest improves the accuracy and stability of the model by combining multiple decision trees.

[0170] The collected data is checked. For missing values, deletion, imputation, etc. can be used for processing; for outliers, correction or rejection can be performed. Standardize the data of different features to have the same scale, use z-score standardization, the formula is z=x=(x-u) / σ. Where x is the original data, μ is the mean of the data, σ is the standard deviation of the data. Select features that have important influence on tunnel water gushing, such as surrounding rock joint density, groundwater level, rock permeability, etc. and create new features by combining existing features. The preprocessed data is divided into training set, validation set and test set according to a certain proportion, 70% training set, 15% validation set, 15% test set. The training set is used for model training, the validation set is used for adjusting model hyperparameters, and the test set is used for evaluating the final performance of the model.

[0171] Then, the number of decision trees is accurately set, which directly affects the model's ability to mine the complex nonlinear relationship between data features and tunnel water gushing. Too few decision trees can easily lead to underfitting, which cannot fully reveal the data correlation and seriously damage the prediction accuracy; while too many decision trees can improve the fitting degree of training data, but will greatly increase the calculation cost, and also easily cause overfitting and reduce the model's generalization ability. At the same time, carefully determine the maximum number of features, which determines the number of features considered when each node splits. Reasonably setting this parameter can effectively enhance the model's generalization ability and make it better adapt to the feature distribution of different data sets. After completing the hyperparameter setting, the carefully prepared training set data is input into the random forest model. During the training process, each decision tree in the model learns the training set data in depth, and through continuous splitting and judging of data features, it gradually explores and clarifies the inherent relationship between data features and tunnel water gushing phenomena.

[0172] The validation set is used to optimize the model performance. Grid search and random search are used to adjust the hyperparameters. Within the pre-set range of hyperparameter values, all possible combinations of hyperparameters are exhaustively enumerated, and then the model performance is strictly evaluated on the validation set one by one to select the optimal combination of hyperparameters. Randomly select some hyperparameter combinations within the value range for evaluation. Compared with grid search, random search can significantly reduce the computational load and improve the tuning efficiency to a certain extent when the value range of hyperparameters is extensive. Through continuous and fine-tuning of hyperparameters, the model performance is continuously optimized, the error on the validation set is minimized, and the generalization ability and prediction accuracy of the model are effectively improved.

[0173] The tuned model is applied to the test set to comprehensively and deeply evaluate its prediction performance. The prediction results of the model are compared with the actual water inflow in detail and in depth. With the help of the feature importance scores provided by the random forest model, the influence of each factor on the tunnel water inflow is clearly and accurately determined. The higher the feature importance score, the more critical the factor plays in the model's prediction of the tunnel water inflow. These accurate analysis results provide valuable scientific basis for tunnel water inflow prediction and prevention work, helping engineers to develop targeted and efficient prevention measures, significantly reducing the risk of tunnel water inflow, and effectively ensuring the safety of tunnel construction and the stability of subsequent operation.

[0174] S4, using the water inflow prediction model to estimate the size of the water inflow, based on the early warning threshold to make early warning judgment on the water inflow of the tunnel, and according to the judgment result to execute the preset early warning strategy.

[0175] Specifically, based on historical data statistical analysis, the historical water inflow and related parameters are deeply mined, and statistical quantities such as mean and standard deviation are calculated. Thresholds are set according to quantiles, and are recalculated periodically with new data accumulation. According to risk classification and expert experience, the water inflow risk level is divided, and the threshold of each level is set combined with expert judgment, and is adjusted in time according to the feedback of experts on water inflow events. The change rate of water inflow is monitored, and different levels of change rate thresholds are set. The risk level and early warning level are comprehensively judged according to the absolute value of water inflow, so as to realize the accurate early warning of water inflow risk. When the threshold is reached, the control system will trigger the early warning device, and send early warning signals to the construction personnel through sound, light and other ways, so as to remind them to pay attention to potential dangers, so that decision makers can take effective measures in time to reduce the risk of geological disasters, and ensure the safe progress of tunnel engineering.

[0176] Specifically, the synergy of each part of the present application can achieve the goal of calculating the joint density of tunnel surrounding rock and conducting intelligent warning of water gushing. In the method, the combination of laser radar, Beidou and inertial navigation technology can provide high-precision and stable positioning data, providing accurate basic data for subsequent measurement work; the structured light scanning system is one of the core components of the present application, which can provide joint density identification information of surrounding rock in the tunnel construction process; the data transmission system adopts communication network transmission technology, which can ensure the rapid transmission and stability of data, and provide accurate data support for the subsequent warning system; the warning system can realize water gushing prediction through machine learning method, timely find the problem of high-risk area of tunnel water gushing, and make geological disaster risk discrimination, and send warning signal to the operator, so as to take timely measures to avoid accidents; the present application realizes the detection of joint density in tunnel construction, and realizes the intelligent warning method of water gushing prediction by fusing the joint density information with hydrogeological conditions and advanced geological prediction and using artificial intelligence technology; the method has the advantages of simple operation, accurate prediction effect, etc., and can improve the safety and efficiency of tunnel construction.

[0177] In summary, by means of the above technical solutions of the present application, compared with the traditional joint density detection method, the structured light scanning technology can reduce the detection cost, the data processing is relatively simple, the joint density estimation accuracy is greatly improved, the permeability of rock is evaluated, and the safety of tunnel construction is ensured. The unmanned aerial vehicle structured light scanning method of the present application can realize the control and monitoring of construction quality in the tunnel construction process, and ensure the accuracy and stability in the construction process. By determining the joint density information of tunnel surrounding rock, combining with the hydrogeological conditions and the advanced prediction results, and using artificial intelligence technology, the present application realizes water gushing warning through constructing a random forest prediction model, analyzes the feature importance, helps to understand the influence of different factors on water gushing, realizes intelligent warning of tunnel water gushing, and improves the accuracy of water gushing warning. The present application provides a method combining structured light scanning and unmanned aerial vehicle for detecting the joint density of surrounding rock in tunnel construction and predicting water gushing, adopts an accurate identification device to realize rapid and accurate detection of the joint density of surrounding rock in tunnel construction, and constructs a random forest model by using artificial intelligence technology to predict the fracture water gushing condition and timely warning, thereby improving the efficiency and safety of construction.

[0178] The above merely describes the preferred embodiments of the present application and should not be used to limit the present application, and any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A tunnel joint and water inrush detection and early warning method based on structured light drones is characterized by: The following steps are involved: S1. Using a structured light scanner mounted on a drone, collect 3D point cloud data from inside the tunnel and transmit the data to a data display terminal using a communication tool. S2. Based on the 3D point cloud data inside the tunnel, combined with a convolutional neural network, the surrounding rock joint density information inside the tunnel is obtained. The AlexNet architecture is used to identify the tunnel joint density and obtain the joint density data. S3. By integrating joint density data, hydrogeological parameters and advanced geological forecast data, a water inflow prediction model is constructed using the random forest algorithm; S4. Use the water inflow prediction model to estimate the amount of water inflow, make a warning judgment on the water inflow of the tunnel based on the warning threshold, and execute the preset warning strategy according to the judgment result.

2. The tunnel joint and water inrush detection and early warning method based on structured light drone according to claim 1 is characterized in that: The method of collecting three-dimensional point cloud data inside the tunnel based on a structured light scanner mounted on a drone and transmitting the three-dimensional point cloud data to a data display terminal using a communication tool includes the following steps: S11, based on the laser radar receiver and sensor built into the drone, respectively obtain first position information and attitude information of the drone, and calculate second position information of the drone using the attitude information; S12, based on the first position information of the drone and the second position information of the drone, fusing the first position information of the drone and the second position information of the drone through a fusion algorithm to obtain final position information of the drone; S13. Based on the pre-set reference point and the final position information of the UAV, the operating parameters of the UAV are adjusted so that the UAV moves to the reference point, scans the working area of ​​the tunnel face, and obtains three-dimensional point cloud data inside the tunnel.

3. The tunnel joint and water inrush detection and early warning method based on structured light drone according to claim 2 is characterized in that: The method of fusing the first position information of the drone and the second position information of the drone using a fusion algorithm to obtain the final position information of the drone includes the following steps: S121, performing timestamp alignment processing on the first location information of the drone and the second location information of the drone; S122: constructing a motion model of the drone based on the second position information of the drone, and using the motion model of the drone to predict the current position of the drone to obtain a prediction result; S123: Based on the first position information of the UAV and the prediction result, perform fingerprint observation vector twisting processing to determine the credibility of the first position information, and perform credibility processing on the first position of the UAV based on the credibility determination result to obtain credible first position information; S124. Use the unscented Kalman filter method to perform fusion filtering processing on the prediction result and the reliable first position information to obtain the final position information of the UAV.

4. The tunnel joint and water inrush detection and early warning method based on structured light drone according to claim 3 is characterized in that: The method of performing fingerprint observation vector twisting processing based on the first position information of the drone and the prediction result, judging the credibility of the first position information, and performing credibility processing on the first position of the drone based on the credibility judgment result to obtain credible first position information includes the following steps: S1231. Calculate the displacement component boundary and positioning deviation boundary of the UAV based on the maximum flight speed of the UAV, the sampling time interval, and the tunnel characteristics; S1232: Calculate the displacement component of the prediction result of the drone in a preset direction, and calculate the displacement component of the first position information in the same preset direction; S1233: Compare the displacement component of the first position information with the displacement component and the displacement component boundary of the prediction result to obtain a comparison result of the displacement components; calculate the positioning deviation between the first position information and the prediction result; and compare the comparison result with the positioning deviation boundary to obtain a positioning deviation boundary comparison result; S1234. Determine the credibility of the first position information by combining the displacement component comparison result and the positioning deviation comparison result; S1235. Based on the credibility judgment result, if the first position information is credible, it is directly adopted; if it is not credible, the prediction result is adopted or corrected in combination with historical credible data to obtain credible first position information.

5. The tunnel joint and water inrush detection and early warning method based on structured light drone according to claim 3 is characterized in that: The method of using the unscented Kalman filter method to perform fusion filtering on the prediction result and the reliable first position information to obtain the final position information of the drone includes the following steps: S1241. Initialize the unscented Kalman filter parameters, including the UAV's state vector, covariance matrix, process noise, and observation noise; S1242. Based on the UAV's motion model, generate a Sigma point set through unscented transformation, and perform state prediction on the Sigma point set to obtain the mean and covariance matrix of the predicted state vector; S1243: Using the trusted first position information as an observation value, calculating the Kalman gain in combination with the predicted state vector, and using the Kalman gain to update the state vector and covariance matrix of the UAV; S1244. Extract the final position information of the UAV from the updated state vector and output the information as the optimal estimate of the UAV's position at the current moment.

6. The tunnel joint and water inrush detection and early warning method based on structured light drone according to claim 1 is characterized in that: The method of obtaining the surrounding rock joint density information inside the tunnel based on the three-dimensional point cloud data inside the tunnel in combination with a convolutional neural network and identifying the tunnel joint density using the AlexNet architecture to obtain the joint density data includes the following steps: S21, preprocessing the three-dimensional point cloud data inside the tunnel, and converting the preprocessed three-dimensional point cloud data into two-dimensional image data; S22, annotating the converted two-dimensional image data, marking the area representing the tunnel surrounding rock joints in the two-dimensional image and identifying the density of the surrounding rock joints, obtaining an annotated data set, and dividing the annotated data set into a first training set, a first validation set, and a first test set; S23. Build a convolutional neural network model based on the AlexNet architecture, and train, verify, and test the convolutional neural network model using the first training set, the first validation set, and the first test set; S24. Using the tested convolutional neural network model, the latest three-dimensional point cloud data inside the tunnel is used to identify the tunnel joint density and obtain the joint density data.

7. The tunnel joint and water inrush detection and early warning method based on structured light drone according to claim 6 is characterized in that: The AlexNet architecture includes: convolutional layers, pooling layers and fully connected layers; Wherein, the convolution layer is used to extract joint features in the tunnel image; The pooling layer is used to reduce the dimension of the feature map to reduce the amount of calculation; The fully connected layer is used to map the extracted joint features to the category or value of the joint density, and at the same time adopts the ReLU activation function to enhance the nonlinear expression ability of the model.

8. The tunnel joint and water inrush detection and early warning method based on structured light drone according to claim 6 is characterized in that: The method of building a convolutional neural network model based on the AlexNet architecture and training, verifying, and testing the convolutional neural network model using a first training set, a first validation set, and a first test set includes the following steps: S231, inputting the first training set data into a convolutional neural network model built based on the AlexNet architecture, and training the convolutional neural network model using the first training set data; S232. In each training cycle of the convolutional neural network model, a prediction result is obtained by forward propagation, and a loss value is calculated using a loss function. According to the loss value, the parameters of the convolutional neural network model are updated using a backpropagation algorithm. S233. During the training process, input the first validation set data into the currently trained convolutional neural network model, calculate the validation results, and adjust the hyperparameters of the convolutional neural network model according to the validation results; S234. Input the first test set data into the verified convolutional neural network model to obtain the prediction result of the test set, and evaluate the performance of the verified convolutional neural network model based on the prediction result of the test set. If the preset performance requirements are met, the current convolutional neural network model is used as the final convolutional neural network model. Otherwise, return to step S231.

9. The tunnel joint and water inrush detection and early warning method based on structured light drone according to claim 1 is characterized in that: The method of constructing a water inflow prediction model by integrating joint density data, hydrogeological parameters and advanced geological prediction data using a random forest algorithm includes the following steps: S31. Collect the hydrogeological parameters, advanced geological prediction data, and water inflow of the tunnel, and combine them with the joint density data to construct a prediction data set; S32, performing data inspection and standardization processing on the prediction data set to obtain a standardized prediction data set; S33. Based on the standardized prediction data set, the correlation between the hydrogeological parameters, the advanced geological prediction data, the joint density data, and the water inflow is calculated using a correlation analysis method, and based on the correlation results, the parameters that meet the preset correlation threshold are selected and divided into a second training set, a second validation set, and a second test set; S34. Construct an initial water inflow prediction model based on the random forest algorithm, and use the second training set, the second validation set and the second test set to train, validate and test the initial water inflow prediction model to obtain the final water inflow prediction model.

10. The tunnel joint and water inrush detection and early warning method based on structured light drone according to claim 9, characterized in that: The method of constructing an initial water inflow prediction model based on the random forest algorithm, and training, verifying, and testing the initial water inflow prediction model using the second training set, the second validation set, and the second test set to obtain a final water inflow prediction model includes the following steps: S341, performing parameter initialization processing on the initial water inflow prediction model constructed based on the random forest algorithm; S342: Using the second training set to train the initial water inflow prediction model with initialized parameters, using each decision tree in the initial water inflow prediction model to conduct deep learning on the second training set, and determining the correlation between the data features and the tunnel water inflow phenomenon by splitting and judging the data features of the second training set; S343. During the training of the initial water inflow prediction model, the performance of the initial water inflow prediction model is optimized using the second validation set, and hyperparameters of the initial water inflow prediction model are adjusted using grid search and random search algorithms to obtain an optimal hyperparameter combination. S344. Apply the optimal hyperparameter combination to the initial water inflow prediction model, and use the second test set to test and evaluate the initial water inflow prediction model to obtain the final water inflow prediction model.

Images