A surface displacement measurement method based on UAV monitoring blind area data completion

Abstract

The present invention relates to the field of automated engineering measurement technology, and in particular to a surface displacement measurement method based on the completion of drone monitoring blind area data. It includes the steps of controlling a drone to scan the slope to be measured, the step of monitoring blind area identification, and the step of predicting blind areas. The present invention takes into account the situation where there is an obstruction above the slope, and can automatically identify and correct the data that is measured incorrectly due to the obstruction, so that the application range of drone slope monitoring is wider and has a higher market value. At the same time, a drone state prediction model is also designed. According to the current state of the drone, the state of the drone at the next moment can be predicted, and then the predicted drone state is compared with the actual state of the drone to obtain the gap between the predicted control and the actual control, that is, the trajectory correction value; therefore, each time the drone is controlled, it is corrected based on the last deviation, which can significantly increase the tracking ability and stability of the drone, and help improve the monitoring effect of the laser radar.

Description

Technical Field

[0001] The present invention relates to the field of automated engineering measurement technology, and in particular to a surface displacement measurement method based on unmanned aerial vehicle monitoring blind area data completion. Background Art

[0002] Slope surface displacement measurement has always been an important link in the field of engineering measurement technology. Traditional slope settlement observations mostly use ground measuring instruments or ground sensors. The disadvantage is that the measuring instruments or sensors are fixedly installed on the slope, which requires preliminary construction and is time-consuming and labor-intensive. Moreover, each measuring instrument or sensor can only monitor a small section of the slope. If the slope is long, the initial investment cost is too high and it is not convenient to remove it.

[0003] Based on the above shortcomings, in recent years the industry has begun to try to use drones equipped with lidar to achieve slope monitoring, which can just make up for the shortcomings of ground measurement technology and has the advantages of no need for preliminary construction, long monitoring distance, and easy transplantation.

[0004] For example, the invention patent entitled "UAV-based high slope monitoring method" disclosed in CN115930898A adopts a monitoring method using UAVs in combination with lasers, including route planning, data collection, three-dimensional modeling, data processing and other steps, which is a relatively common monitoring method. Its shortcomings are: 1. It does not take into account the situation of slope vegetation obstruction, and has no adaptability to slopes with more complex ground conditions; 2. When monitoring slopes, there are high requirements for the accuracy of collected data, and the route tracking ability and shooting stability of UAVs will have a serious impact on slope monitoring. The patent does not disclose a special control method for UAVs suitable for laser measurement, which will lead to poor monitoring results. Summary of the invention

[0005] The present invention discloses a surface displacement measurement method based on unmanned aerial vehicle monitoring blind area data completion, which is characterized by the following specific method:

[0006] According to the slope to be measured, the measurement route is pre-planned;

[0007] Control the drone to fly according to the preset measurement route, and use the laser radar on the drone to scan the slope to be measured;

[0008] The slope data measured by the laser radar is divided into a number of window data, and each window data is input into the diagnosis model to determine whether there is a monitoring blind area;

[0009] If there is a monitoring blind spot, obtain the blind spot range and mark the corresponding window as a blind spot window;

[0010] Retrieve the data adjacent to the blind area, combine the data within the blind area, and build a complete data set;

[0011] According to the completed data set, predict the predicted data of the blind area range.

[0012] The advantage of this embodiment is that it takes into account the situation where there is obstruction above the slope. The erroneous data measured due to the obstruction can be automatically identified and corrected, so that the application range of drone slope monitoring is wider and it has a higher market value.

[0013] Furthermore, the UAV is a quad-rotor UAV, and the dynamic model is as follows:

[0014]

[0015] In the formula, They represent the acceleration of the quadrotor drone, They represent the acceleration of the roll angle, pitch angle and yaw angle of the quadrotor drone, are the speeds of the roll angle, pitch angle and yaw angle of the quadrotor drone, I xx ,I yy ,I zz They represent the moment of inertia of the quadrotor drone around the three coordinate axes of the coordinate system, m is the mass of the quadrotor drone, and g is the acceleration due to gravity;

[0016] in:

[0017]

[0018]

[0019] Where U L is the total tension, τ x , τ y , τ z are the rotational moments in three directions around the body axis, ω i is the rotation speed of the four propellers, l is the distance from the center of the propeller to the center of gravity of the quadrotor drone, C T is the thrust coefficient, C M is the counter torque coefficient.

[0020] The advantage of this embodiment is that the stability and maneuverability of the quad-rotor drone meet the measurement requirements of the drone carrying a laser radar, and the cost is low, making it the optimal drone model.

[0021] Further, the specific method of controlling the drone is as follows:

[0022] The trajectory correction value X at the next moment c (k+1) Correct the expected trajectory X at the next moment ref (k+1), get the new state X at the next moment e(k+1);

[0023] The new state quantity X at the next moment e (k+1), solve the optimal control quantity U(k) at the current moment through optimization;

[0024] The optimal control quantity U(k) at the current moment is input into the UAV state prediction model to obtain the predicted state X of the UAV at the next moment pre (k+1);

[0025] Control the quadrotor drone with the current optimal control value U(k), and record the actual state Y(k+1) of the quadrotor drone at the next moment;

[0026] According to the next moment drone predicted state X pre (k+1) and the actual state of the drone at the next moment Y(k+1), and then calculate the trajectory correction value X at the next moment c (k+2).

[0027] In this embodiment, a drone state prediction model is designed. The drone state at the next moment can be predicted based on the drone's current state. The predicted drone state is then compared with the drone's actual state to obtain the gap between the predicted control and the actual control, that is, the trajectory correction value. Therefore, each time the drone is controlled, a correction is made based on the previous deviation, which can significantly increase the drone's tracking ability and stability, and help improve the monitoring effect of the lidar.

[0028] Furthermore, the current trajectory correction value X c (k), the expected trajectory X at the next moment ref (k+1), new state quantity X e (k+1), predicted state X pre (k+1) and the actual state Y(k+1) both include drone state attributes, and the drone state attributes include:

[0029] Drone position coordinates x, y, z, drone speed Drone acceleration

[0030] Drone attitude angle θ, ψ, attitude angular velocity Attitude angular acceleration

[0031] Furthermore, the optimal control quantity U(k) at the current moment is optimized and solved. The specific method is as follows:

[0032] ω 1 ,ω 2 ,ω 3 ,ω 4Determine the constraint conditions as optimization variables;

[0033] According to the initialized optimization variable U 0 (k) and the actual state Y(k) of the UAV at the current moment, calculate the predicted state attributes corresponding to the next moment

[0034] Calculate the fitness of this configuration optimization variable;

[0035] Change the optimization variable configuration within the constraint conditions, and recalculate the fitness corresponding to the changed optimization variable configuration;

[0036] Repeat changing the optimization variable configuration and calculating the corresponding fitness until the fitness is less than the preset target or the number of iterations is completed;

[0037] Select ω corresponding to the minimum fitness 1 、ω 2 、ω 3 、ω 4 as the optimal control quantity U(k) at the current moment;

[0038] Calculate the predicted state attributes corresponding to the next moment The specific method is as follows:

[0039] If the trajectory correction value |X c (k + 1)| < th, where th is the threshold, then input the optimization variable U(k) and the actual state Y(k) of the UAV at the current moment into the UAV dynamics model to calculate the predicted state attributes corresponding to the next moment Its fitness function formula is:

[0040]

[0041] If the trajectory correction value |X c (k + 1)| ≥ th, then input the optimization variable U(k) and the actual state Y(k) of the UAV at the current moment into the UAV state prediction model to calculate the predicted state attributes corresponding to a moment 1), that is Its fitness function formula is:

[0042]

[0043] The advantage of this embodiment is that by using the fitness function to optimize the rotational speeds of the four rotors, the optimal solution of the quadrotor can be obtained. The constraint conditions can exclude combinations of conditional places, improving the safety of UAV control. When the trajectory correction value is large, the UAV control accuracy is poor. At this time, using the UAV state prediction model to construct the fitness function can make the result output by the UAV dynamic model quickly approach the new state quantity X at the next momente (k+1), the new state quantity X at the next moment e (k+1) will also keep getting closer to the actual state Y(k+1) in each iteration. Therefore, through the dual correction of the drone state prediction model and the trajectory correction value in one iteration, the input result of the optimal control quantity U(k) can quickly approach the actual state Y(k+1). In addition, in one iteration, the drone state prediction model is used once when constructing the fitness function, and the drone predicted state X pre (k+1) can be used once to increase the training data of the drone state prediction model, speed up the acquisition of training data and the fitting speed of the model. When the trajectory correction value is large, the drone state prediction model is no longer used, and only the new state quantity X at the next moment is used in one direction. e (k+1) is used to bridge the error and avoid fitness value fluctuation caused by over-adjustment.

[0044] Furthermore, the UAV state prediction model is a BP neural network model. The actual state Y(k) of the UAV at the current moment and the optimal control quantity U(k) at the current moment are input into the trained BP neural network model, and the BP neural network model outputs the predicted state X of the UAV at the next moment. pre (k+1).

[0045] Furthermore, the trajectory correction value calculation formula is:

[0046] X c (k+1)=X pre (k)-Y(k)

[0047] The new state quantity calculation formula at the next moment is:

[0048] X e (k+1)=X ref (k+1)+X c (k+1)

[0049] Furthermore, the diagnostic model is a convolutional neural network model, and the specific method for determining whether there is a monitoring blind spot is as follows:

[0050] Calculate the average value of all windows and convert each value of each window into RGB value according to the size;

[0051] Input several RBG values ​​of each window into the convolutional neural network model;

[0052] The convolutional neural network model outputs whether there is a monitoring blind spot and the area of ​​the monitoring blind spot.

[0053] The advantage of this embodiment is that the height values ​​measured by the lidar are converted into RGB values, that is, image data. A mature convolutional neural network model can be used to identify abnormal color shapes in the image. For example, if a small area of ​​woods appears, the height value of the woods area is higher than the value outside the area as a whole. This can highlight the characteristics of the obstruction not only in color, but also in shape, thereby improving recognition accuracy.

[0054] Furthermore, the supplemented data set includes a data area adjacent to a preset area at the edge of the blind area, and a blank area after deleting the data in the blind area;

[0055] The prediction data of the prediction blind area range is as follows:

[0056] Make the blank blind area into a number of rows and columns;

[0057] Retrieve the two most adjacent values ​​on the left and right of each row of blind area, and fill each blank data grid in the blind area by interpolation. The filled content is named row prediction data.

[0058] Retrieve the two most adjacent values ​​in each row of the blind area, and fill each blank data grid in the blind area with interpolation. The filled content is named column prediction data.

[0059] According to the prediction weight, the row prediction data and the column prediction data are weighted to obtain the blind area prediction data.

[0060] Furthermore, the prediction weight acquisition method is as follows:

[0061] Initialize the weights of all blank data grids in the blind area;

[0062] The weight of each blank data grid in the blind area is adjusted individually to construct a weight matrix;

[0063] The absolute value of the difference between the weighted row and column predicted data of the weight matrix and the real data in the blind area is used as the evaluation index;

[0064] The weight matrix corresponding to the evaluation index with the smallest absolute value is selected as the prediction weight.

[0065] The advantage of this embodiment is that slope changes generally have a gradual characteristic, so the interpolation method can be used to predict the change of blind spot height values, and the height values ​​​​obtained from the fusion of row and column directions will not lack the characteristics of height value changes, and the weights can be used to allocate the consideration ratio of row features and column features according to different situations, which can further improve the accuracy of blind spot height prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] The accompanying drawings of the present invention are as follows.

[0067] Figure 1It is a schematic diagram of the process of the present invention.

[0068] Figure 2 This is a schematic diagram of the UAV control process.

[0069] Figure 3 Schematic diagram of the prediction data for predicting the blind area range. DETAILED DESCRIPTION

[0070] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0071] A surface displacement measurement method based on UAV monitoring blind area data completion, such as Figure 1 As shown, the specific method is as follows:

[0072] S1. Preset and plan the measurement route according to the slope to be measured.

[0073] In step S1, the route is measured according to the preset plan, and the expected trajectory of each node on the route can be calculated to facilitate the calculation in the subsequent steps.

[0074] S2. Control the UAV to fly according to the preset measurement route, and use the laser radar on the UAV to scan the slope to be measured.

[0075] In step S2, the drone is controlled to fly according to the preset measurement route. Figure 2 As shown, the specific steps are:

[0076] S21, the trajectory correction value X at the next moment c (k+1) Correct the expected trajectory X at the next moment ref (k+1), get the new state X at the next moment e (k+1).

[0077] S22, the new state quantity X at the next moment e (k+1), and solve the optimal control quantity U(k) at the current moment through optimization.

[0078] Specifically, in step S22, the optimal control amount U(k) at the current moment is optimized and solved, and the specific method is as follows:

[0079] S221, with ω 1 ,ω 2 ,ω 3 ,ω 4 As optimization variables, determine the constraints.

[0080] S222, according to the initialization optimization variable U 0 (k) and the actual state of the drone at the current moment Y(k), calculate the predicted state attribute corresponding to the next moment

[0081] S223. Calculate the fitness of the configuration optimization variable.

[0082] Specifically, calculate the predicted state attributes corresponding to the next moment The specific method is as follows:

[0083] S2231. If the trajectory correction value |X c (k + 1)| < th, where th is a threshold, then input the optimization variable U(k) and the actual state Y(k) of the UAV at the current moment into the UAV dynamics model to calculate the predicted state attributes corresponding to the next moment Its fitness function formula is:

[0084]

[0085] S2232. If the trajectory correction value |X c (k + 1)| ≥ th, then input the optimization variable U(k) and the actual state Y(k) of the UAV at the current moment into the UAV state prediction model to calculate the predicted state attributes corresponding to a moment That is Its fitness function formula is:

[0086] J = |X pre (k + 1) - X ref (k + 1)|

[0087] S224. Change the optimization variable configuration within the constraint conditions and recalculate the fitness corresponding to the changed optimization variable configuration;

[0088] S225. Repeat changing the optimization variable configuration and calculating the corresponding fitness until the fitness is less than the preset target or the number of iterations is completed.

[0089] S226. Select ω 1 , ω 2 , ω 3 , ω 4 corresponding to the minimum fitness as the optimal control quantity U(k) at the current moment.

[0090] S23. Input the optimal control quantity U(k) at the current moment into the UAV state prediction model to obtain the predicted state X of the UAV at the next moment pre (k + 1).

[0091] S24. Control the quadrotor UAV with the current optimal control quantity U(k) and record the actual state Y(k + 1) of the quadrotor UAV at the next moment.

[0092] S25. According to the predicted state X of the UAV at the next moment pre(k+1) and the actual state of the drone at the next moment Y(k+1), and then calculate the trajectory correction value X at the next moment c (k+2).

[0093] In step S21 to step S25, the current trajectory correction value X c (k), the expected trajectory X at the next moment ref (k+1), new state quantity X e (k+1), predicted state X pre (k+1) and the actual state Y(k+1) both include drone state attributes, which include: drone position coordinates x, y, z, drone speed Drone acceleration Drone attitude angle θ, ψ, attitude angular velocity Attitude angular acceleration After determining the above 18 variables, the status of the drone can be fully determined.

[0094] The same drone state prediction model is used in step S2232 and step S23. Specifically, the drone state prediction model is a BP neural network model. The actual state Y(k) of the drone at the current moment and the optimal control quantity U(k) at the current moment are input into the trained BP neural network model, and the BP neural network model outputs the predicted state X of the drone at the next moment. pre (k+1).

[0095] In step S2231, the UAV dynamics model is used to calculate the predicted state attributes corresponding to the next moment. 1), the UAV dynamics model is as follows:

[0096] The UAV is a quad-rotor UAV, and its dynamic model is as follows:

[0097]

[0098] In the formula, They represent the acceleration of the quadrotor drone, They represent the acceleration of the roll angle, pitch angle and yaw angle of the quadrotor drone, are the speeds of the roll angle, pitch angle and yaw angle of the quadrotor drone, I xx ,I yy ,I zz They represent the moment of inertia of the quadrotor drone around the three coordinate axes of the coordinate system, m is the mass of the quadrotor drone, and g is the acceleration due to gravity;

[0099] in:

[0100]

[0101]

[0102] Where U L is the total tension, τ x , τ y , τ z are the rotational moments in three directions around the body axis, ω i is the rotation speed of the four propellers, l is the distance from the center of the propeller to the center of gravity of the quadrotor drone, C T is the thrust coefficient, C M is the counter torque coefficient.

[0103] S3. Divide the slope data measured by the laser radar into a number of window data, and input each window data into the diagnosis model to determine whether there is a monitoring blind area.

[0104] S4. If there is a monitoring blind spot, obtain the blind spot range and mark the corresponding window as a blind spot window.

[0105] S5. Retrieve the data adjacent to the blind area, and combine the data within the blind area to construct a complete data set.

[0106] S6. Based on the completed data set, predict the prediction data of the blind area range, such as Figure 3 shown.

[0107] In step S6, the prediction weight acquisition method is as follows:

[0108] S61, initializing the weights of all blind area blank data grids.

[0109] S62. Individually adjust the weight of each blank data grid in the blind area to construct a weight matrix.

[0110] S63, taking the absolute value of the difference between the weighted row and column prediction data of the weight matrix and the real data of the blind area as the evaluation index.

[0111] S64. Select the weight matrix corresponding to the evaluation index with the smallest absolute value as the prediction weight.

[0112] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A surface displacement measurement method based on UAV monitoring blind area data completion, characterized in that: The specific method is as follows: According to the slope to be measured, the measurement route is pre-planned; Control the drone to fly according to the preset measurement route, and use the laser radar on the drone to scan the slope to be measured; The slope data measured by the laser radar is divided into a number of window data, and each window data is input into the diagnosis model to determine whether there is a monitoring blind area; If there is a monitoring blind spot, obtain the blind spot range and mark the corresponding window as a blind spot window; Retrieve the data adjacent to the blind area, combine the data within the blind area, and build a complete data set; According to the completed data set, predict the prediction data of the blind area range; The diagnostic model is a convolutional neural network model, which determines whether there is a monitoring blind spot. The specific method is as follows: Calculate the average value of all windows and convert each value of each window into RGB value according to the size; Input several RBG values ​​of each window into the convolutional neural network model; The convolutional neural network model outputs whether there is a monitoring blind spot and the area of ​​the monitoring blind spot; The completed data set includes the adjacent data area of ​​the preset area at the edge of the blind area, and the blank area after deleting the data in the blind area; The prediction data of the prediction blind area range is as follows: Make the blank blind area into a number of rows and columns; Retrieve the two most adjacent values ​​on the left and right of each row of blind area, and fill each blank data grid in the blind area by interpolation. The filled content is named row prediction data. Retrieve the two most adjacent values ​​in each row of the blind area, and fill each blank data grid in the blind area with interpolation. The filled content is named column prediction data. According to the prediction weight, the row prediction data and the column prediction data are weighted to obtain the blind area prediction data.

2. The surface displacement measurement method based on the UAV monitoring blind area data completion as claimed in claim 1 is characterized in that: The UAV is a quad-rotor UAV, and its dynamic model is as follows: In the formula, They represent the acceleration of the quadrotor drone, They represent the acceleration of the roll angle, pitch angle and yaw angle of the quadrotor drone, are the speeds of the roll angle, pitch angle and yaw angle of the quadrotor drone, I xx ,I yy ,I zz They represent the moment of inertia of the quadrotor drone around the three coordinate axes of the coordinate system, m is the mass of the quadrotor drone, and g is the acceleration due to gravity; in: Where U L is the total tension, τ x , τ y , τ z are the rotational moments in three directions around the body axis, ω i is the rotation speed of the four propellers, l is the distance from the center of the propeller to the center of gravity of the quadrotor drone, C T is the thrust coefficient, C M is the counter torque coefficient.

3. The surface displacement measurement method based on the UAV monitoring blind area data completion as claimed in claim 1 is characterized in that: The specific method of controlling the drone is as follows: The trajectory correction value X at the next moment c (k+1) Correct the expected trajectory X at the next moment ref (k+1), get the new state X at the next moment e (k+1); The new state quantity X at the next moment e (k+1), solve the optimal control quantity U(k) at the current moment through optimization; The optimal control quantity U(k) at the current moment is input into the UAV state prediction model to obtain the predicted state X of the UAV at the next moment pre (k+1); Control the quadrotor drone with the current optimal control value U(k), and record the actual state Y(k+1) of the quadrotor drone at the next moment; According to the next moment drone predicted state X pre (k+1) and the actual state of the drone at the next moment Y(k+1), and then calculate the trajectory correction value X at the next moment c (k+2).

4. The surface displacement measurement method based on the UAV monitoring blind area data completion as claimed in claim 3 is characterized in that: The current trajectory correction value X c (k), the expected trajectory X at the next moment ref (k+1), new state quantity X e (k+1), predicted state X pre (k+1) and the actual state Y(k+1) both include drone state attributes, and the drone state attributes include: Drone position coordinates x, y, z, drone speed Drone acceleration Drone attitude angle θ, ψ, attitude angular velocity Attitude angular acceleration 5. The surface displacement measurement method based on the UAV monitoring blind area data completion as claimed in claim 4 is characterized in that: Optimize and solve the optimal control quantity U(k) at the current moment. The specific method is as follows: Take ω1, ω2, ω3, and ω4 as optimization variables and determine the constraints; According to the initial optimization variable U0(k) and the actual state of the drone Y(k) at the current moment, calculate the predicted state attribute corresponding to the next moment Calculate the fitness of the configuration optimization variable; Change the optimization variable configuration within the constraints and recalculate the fitness corresponding to the changed optimization variable configuration; Repeatedly change the optimization variable configuration and calculate the corresponding fitness until the fitness is less than the preset target or the number of iterations is completed; When the fitness is the smallest, the corresponding ω1, ω2, ω3, and ω4 are selected as the optimal control quantity U(k) at the current moment; Calculate the predicted state attributes corresponding to the next moment The specific method is as follows: If the trajectory correction value |X c (k + 1)| < th, where th is the threshold, then the optimization variable U(k) and the actual state Y(k) of the UAV at the current moment are input into the UAV dynamics model to calculate the predicted state attributes corresponding to the next moment The formula for its fitness function is: If the trajectory correction value at the next moment |X c (k+1)|≥th, then the optimization variable U(k) and the actual state of the drone at the current moment Y(k) are input into the drone state prediction model to calculate the predicted state attribute corresponding to a moment. Right now The fitness function formula is: J=|X pre (k+1)-X ref (k+1)|。 6. The surface displacement measurement method based on the UAV monitoring blind area data completion as claimed in claim 4 is characterized in that: The UAV state prediction model is a BP neural network model. The actual state Y(k) of the UAV at the current moment and the optimal control quantity U(k) at the current moment are input into the trained BP neural network model, and the BP neural network model outputs the predicted state X of the UAV at the next moment. pre (k+1).

7. The surface displacement measurement method based on the UAV monitoring blind area data completion as claimed in claim 4 is characterized in that: The trajectory correction value calculation formula is: X c (k+1)=X pre (k)-Y(k) The new state quantity calculation formula at the next moment is: X e (k+1)=X ref (k+1)+X c (k+1)。 8. The surface displacement measurement method based on the UAV monitoring blind area data completion as claimed in claim 1, characterized in that: The prediction weight acquisition method is as follows: Initialize the weights of all blank data grids in the blind area; The weight of each blank data grid in the blind area is adjusted individually to construct a weight matrix; The absolute value of the difference between the weighted row and column predicted data of the weight matrix and the real data in the blind area is used as the evaluation index; The weight matrix corresponding to the evaluation index with the smallest absolute value is selected as the prediction weight.

Images