Data and model dual-driven unmanned aerial vehicle wing strain real-time monitoring method

Abstract

The invention discloses a data and model dual-drive unmanned aerial vehicle wing strain real-time monitoring method for real-time monitoring of unmanned aerial vehicle wing strain. Firstly, a physical model of an unmanned aerial vehicle wing is established, and a strain response rule under a load is analyzed; secondly, training actual measurement data by using a machine learning method, and constructing a data-driven high-fidelity strain prediction model; secondly, a numerical simulation method is used for obtaining calculation data, a low-fidelity data model is obtained, the model is combined with a data driving model, and a multi-fidelity strain prediction model is constructed; and finally, sensor data are accessed to the prediction model, and the prediction model is visualized, so that the wing strain of the unmanned aerial vehicle is monitored in real time, and a powerful guarantee is provided for safe flight and task execution of the unmanned aerial vehicle.

Description

Technical Field

[0001] The present invention relates to the field of real-time structural health monitoring, and particularly to a method for real-time monitoring of the strain of a drone wing driven by both data and models. Background Art

[0002] With the increasing application of drones, it has become very important to monitor their health status. Conventional detection methods are not only time-consuming and laborious, but also costly. There is a lack of monitoring methods for drone wings, and the use of strain monitoring methods can significantly reduce costs. In addition, both physical models and data-driven methods have single limitations. How to overcome the limitations of single methods, improve the accuracy of monitoring, and expand the prediction geometric space is a key problem that needs to be urgently solved in current real-time monitoring. By realizing the real-time monitoring of the strain of the drone wing, it provides a strong guarantee for the safe flight and mission execution of the drone. Summary of the Invention

[0003] The purpose of the present invention is to provide a method for real-time monitoring of the strain of a drone wing driven by both data and models to solve the problems that both physical models and data-driven methods have single limitations, resulting in the inability to improve the monitoring accuracy and the limitation of the prediction geometric space.

[0004] For the real-time monitoring of the strain of the drone wing, the present invention provides a method for real-time monitoring of the strain of a drone wing driven by both data and models.

[0005] To achieve the above purpose, the present invention adopts the following technical solutions:

[0006] A method for real-time monitoring of the strain of a drone wing driven by both data and models, comprising:

[0007] Step 1: Data acquisition, collecting strain data through strain gauge sensors and data acquisition devices pasted on the wing surface;

[0008] Step 2: Using machine learning methods to train the data collected in Step 1 to construct a data-driven high-fidelity strain prediction model;

[0009] Step 3: Using numerical simulation methods to obtain calculation data and obtain a low-fidelity data model driven by a physical model;

[0010] Step 4: Combining the data-driven model in Step 2 with the physical model in Step 3 to construct a multi-fidelity strain prediction model;

[0011] Step 5: Given an error evaluation method, judging the accuracy of the predicted value of the strain prediction model in Step 4;

[0012] Step 6: Compare the prediction results of the strain prediction model in Step 4 for each acquisition point with the experimental results obtained from the test, and judge the magnitude of the relative error between the two according to the error evaluation method in Step 5. If the relative error is greater than 10%, repeat Steps 2 to 6; if the relative errors are all less than 10%, proceed to the next step;

[0013] Step 7: Connect the sensor data to the strain prediction model in Step 3, and visualize the prediction model to achieve real-time monitoring of the strain of the UAV wing;

[0014] Further, in Steps 2 and 4, the machine learning method used is the Bootstrap aggregating (Bagging) algorithm.

[0015] Further, in Step 3, the numerical simulation method used is the finite element method, and the software used is the ABAQUS simulation software.

[0016] Further, in Step 5, the relative error δ is used as a method to evaluate the accuracy of the predicted value.

[0017] Preferably, in Step 7, the visualization of the prediction model is to display the prediction results in the form of a strain nephogram, and as the sensor data is continuously input, the strain nephogram changes to form an animation.

[0018] The present invention has the following beneficial effects compared with the prior art:

[0019] The present invention discloses a method for real-time monitoring of the strain of a UAV wing driven by both data and a model. First, a physical model of the UAV wing is established to analyze its strain response law under load; then, a machine learning method is used to train the actual measurement data to construct a data-driven high-fidelity strain prediction model; secondly, a numerical simulation method is used to obtain calculation data to obtain a low-fidelity data model, and the model is combined with the data-driven model to construct a multi-fidelity strain prediction model; finally, the sensor data is connected to the prediction model, and the prediction model is visualized to achieve real-time monitoring of the strain of the UAV wing, providing a strong guarantee for the safe flight and mission execution of the UAV; the present invention adopts a strain monitoring means, greatly reducing the cost and enabling real-time visualization; the present invention overcomes the single limitations of physical models and data-driven methods, having the beneficial effects of improving the accuracy of monitoring and expanding the geometric space of prediction. Description of the Drawings

[0020] Figure 1 It is the overall flowchart of the method for real-time monitoring of the strain of a UAV wing driven by both data and a model of the present invention.

[0021] Figure 2Schematic diagram of the installation position of strain gauges on the wing in the embodiment of the present invention.

[0022] Figure 3 Flowchart of the algorithm in the embodiment of the present invention.

[0023] Figure 4 Comparison of different training results and the average value of the test set in the embodiment of the present invention.

[0024] Figure 5 Relative error of the predicted values before and after interpolation in the embodiment of the present invention.

[0025] Figure 6 Average value of the predicted results and the test set in the embodiment of the present invention.

[0026] Figure 7 Relative error of the predicted value in the embodiment of the present invention.

[0027] Figure 8 Strain nephogram corresponding to some continuously collected real-time data in the embodiment of the present invention. Detailed implementation manners

[0028] To enable those skilled in the art to better understand the real-time monitoring method of the present invention, the present invention will be further described in detail below with specific embodiments in conjunction with the accompanying drawings, but it is not limited to a specific instance only, and this method can be applied to similar situations.

[0029] Please refer to Figures 1 to 8 , the present invention provides a real-time monitoring method for the strain of an unmanned aerial vehicle (UAV) wing driven by both data and model, including the following steps:

[0030] Step 1: Install high-precision strain gauge sensors on the wing surface to collect strain data in real time. Since the wing structure of the UAV is mainly affected by bending moment, shear force and torque during flight, for the convenience of calculation and subsequent research, only the lift force on the wing is considered, and the force on the wing is appropriately simplified to form a cantilever beam structure model. A total of 18 strain gauges are arranged on the wing, among which 15 are used as test points and 3 are used as verification points. During the experiment, in order to avoid accidental data, the wing is repeatedly loaded and unloaded, and 50 tests are carried out in total. The strain data after the wing stabilizes is taken as the measurement data. For the convenience of subsequent calculation, the strain data of all parts in this embodiment are taken as positive numbers, because the negative sign only represents compression and does not represent the magnitude. As Figure 2It is a schematic diagram of the installation position of strain gauges on the wing. The collected data is normalized to scale the data to a unified range, facilitating subsequent model training and prediction. The wing projection is discretized two-dimensionally. In this two-dimensional space, the coordinates are used as the input of the model, and the strain is used as the response value. The strain values of 15 test points are used as feature variables and imported into the two-dimensional space. The coordinates of the 15 test points are Pt, and each test point corresponds to a channel, denoted as Yt. The coordinates of the 3 verification points are Pv, corresponding to the channel Yv. The specific values are as follows:

[0031]

[0032] Y t = [1 2 3 4 6 7 8 10 11 13 14 15 16 17 18] (2)

[0033]

[0034] Y v = [5 9 12] (4)

[0035] Step 2: Use the Bagging algorithm for strain prediction. The algorithm flow chart is as Figure 3 shown. The model training process for the regression problem is as follows: (1) Divide the dataset and use the Bagging algorithm to construct an ensemble model; (2) Randomly draw multiple subsample sets from the training dataset with replacement; (3) Train a decision tree model for each subsample set; (4) Average the prediction results of all decision tree models to obtain the final prediction value, that is:

[0036]

[0037] In the formula, Γ is the final prediction result, m is the number of base models, and h i (x) is the prediction result of the base model.

[0038] The obtained 50 groups of strain data of 15 measurement points are divided into a test set and a training set. The training set is used for model training, and the test set is used for model verification. The Bagging algorithm is used for training. The number of test points is less relative to the two-dimensional space of the entire wing. Two models are considered. One is to train only on the test points; the other is to first interpolate the two-dimensional space of the entire wing based on the strain data of 15 test points to obtain the data of the entire wing plane, and then train to obtain a prediction model. Both of these compare the predicted values at the test points after training with the average value of the test set to evaluate the accuracy of the model. As Figure 4 is the comparison of the two training results and the average value of the test set, Figure 5It is the relative error of the predicted values before interpolation (PVAI) and the relative error of the predicted values after interpolation (PVAI).

[0039] Step 3: Use simulation software to conduct numerical simulation on the wing test, verify the accuracy of the numerical simulation through test data, and obtain the strain data of the entire wing.

[0040] Step 4: Index the simulation results according to the two-dimensional spatial coordinates of the wing, combine the simulation result data and the data obtained from the test to form the original dataset. Divide the original dataset into a training set and a test set at a ratio of 4:1, and use the Bagging algorithm to train the training set to obtain a multi-fidelity strain prediction model. The predicted values of 18 channels and the average value of the test set are as Figure 6 shown, and the relative error of the predicted values is as Figure 7 shown.

[0041] Step 5: Use the relative error δ as a method to evaluate the accuracy of the predicted values. The calculation formula of δ is as follows:

[0042]

[0043] In the formula, is the predicted value, and y is the measured value.

[0044] Step 6: Compare the prediction results of the strain prediction model in Step 4 for each acquisition point with the experimental results obtained from the test, and judge the relative error between the two according to the error evaluation method in Step 5. If the relative error is greater than 10%, repeat Steps 2 to 6; if the relative errors are all less than 10%, proceed to the next step. Table 1 shows the comparison between the prediction results and the test results of the prediction model. In the table, the predicted value is PV, the test value is TV, and the relative error is RE. The following numbers represent the groups, meaning that three comparisons have been made. The specific data are shown in the table:

[0045]

[0046] Step 7: The prediction model in Step 4 can respond in a timely manner to the strain data collected by the strain collector and update the strain nephogram in real time. The specific implementation process is as follows:

[0047] First, by inputting the spatial coordinates (x, y) of the corresponding acquisition point, combining with the grid data of the model, calculate the minimum distance between the input coordinate point and the grid point, so as to locate the closest grid point and display the corresponding strain nephogram.

[0048] Secondly, to ensure the real-time visualization of the results, a method of dynamically updating the graph is adopted. After each new data is collected, the system clears the current display content on the graph interface and draws a new strain nephogram.

[0049] Finally, a warning mechanism is added. If the input strain value exceeds the safety threshold (such as setting 1200 με), that is, the load on the wing is too large, the warning mechanism is triggered, and the user is reminded of the possible damage risk of the wing through sound and pop-up windows.

[0050] Based on the real-time collected data and this algorithm model, the real-time monitoring of the wing strain can be realized, and the possible damages can be adjusted in time. As Figure 8 Shown is the strain nephogram corresponding to some continuously collected real-time data.

[0051] The above are only the specific implementation steps of the present invention. The protection scope of the present invention is not limited to the use of the above methods and the implementation of specific examples. All technical solutions under the idea of the present invention belong to the protection scope of the present invention. It should be noted that for those of ordinary skill in the art, several improvements and refinements made without departing from the principle of the present invention should be regarded as within the protection scope of the present invention.

[0052] The parts not elaborated in detail in the present invention belong to the well-known technologies in the art.

Claims

1. A real-time monitoring method for UAV wing strain driven by both data and model, characterized in that: The following steps are involved: Step 1: Perform load test on the wing, install high-precision strain gauge sensors on the wing surface, and collect strain data in real time; Step 2: Use machine learning methods to train the data collected in step 1 to build a data-driven high-fidelity strain prediction model; Step 3: Use numerical simulation methods to obtain calculation data and obtain a low-fidelity data model driven by the physical model; Step 4: Combine the data-driven model in step 2 with the physical model in step 3 to build a multi-fidelity strain prediction model; Step 5: Given an error evaluation method, determine the accuracy of the predicted value of the strain prediction model in step 4; Step 6: Compare the prediction results of the strain prediction model described in step 4 for each acquisition point with the experimental results obtained by the test, and determine the relative error between the two according to the error evaluation method described in step 5. If the relative error is greater than 10%, repeat steps 2 to 6; if the relative errors are all less than 10%, proceed to the next step; Step 7: Connect the sensor data to the strain prediction model in step 3 and visualize the prediction model to achieve real-time monitoring of the UAV wing strain.

2. The method for real-time monitoring of UAV wing strain driven by data and model according to claim 1 is characterized in that: In step 2 and step 4, the machine learning method used is a guided clustering algorithm.

3. The real-time monitoring method for UAV wing strain driven by data and model according to claim 1 is characterized in that: In step three, the numerical simulation method used is the finite element method, and the software used is ABAQUS simulation software.

4. The method for real-time monitoring of UAV wing strain driven by data and model according to claim 1 is characterized in that: In step five, the relative error δ is used as a method to evaluate the accuracy of the predicted value.

5. The method for real-time monitoring of UAV wing strain driven by data and model according to claim 1 is characterized in that: In the step seven, the prediction model visualization is specifically to display the prediction result in the form of a strain cloud map. As the sensor data is continuously input, the strain cloud map changes accordingly to form an animation.

Images