Power inspection unmanned aerial vehicle vision / inertia robust positioning method based on temperature compensation

Abstract

The invention discloses a power inspection unmanned aerial vehicle vision / inertia robust positioning method based on temperature compensation, and the method comprises the steps: firstly constructing an IMU temperature online compensation network, and introducing a Teaching Forcing mechanism to improve the learning capability of the network; then, an IMU error model is constructed, an LSTM-based unmanned aerial vehicle IMU zero offset temperature online compensation network is trained through IMU measurement information subjected to offline calibration compensation, the temperature, the temperature change rate and zero offset true values of a gyroscope and an accelerometer, and network parameters of an unmanned aerial vehicle temperature model are obtained; in the operation process of the unmanned aerial vehicle, trained parameters are called, online compensation is carried out on the zero offset of the gyroscope and the accelerometer corresponding to IMU measurement at each moment, and the problem that the zero offset error of the unmanned aerial vehicle becomes large due to high-frequency vibration and temperature change in the field inspection process is solved; and finally, designing a self-adaptive visual factor, performing joint estimation by using IMU measurement after zero offset compensation and a visual residual error, and resolving to obtain optimized positioning information.

Description

Technical Field

[0001] The invention belongs to the technical field of machine vision and combined navigation, and mainly relates to a visual / inertial robust positioning method for an electric power inspection UAV based on temperature compensation. Background Art

[0002] my country's power industry has widely adopted drone technology for power inspection. Drones can collect a large amount of data through high-definition cameras and sensors to help discover problems and hidden dangers of power equipment and improve the reliability and safety of the power system. Since 2009, State Grid Corporation of China has officially launched the development of drone power inspection systems. By 2020, the market size of my country's power inspection drones has exceeded 8 billion, becoming a mainstream power inspection method. At present, drone power inspections have participated in more than 60% of power inspections. With the continuous development of drone technology, power inspections are moving towards a more autonomous direction. Drone power inspections will gradually get rid of manual work and achieve full automation. Compared with traditional power inspection methods, the emergence of drone technology has brought new hope to the power industry. In power inspection tasks, when the Global Navigation Satellite System (GNSS) signal is denied or there is electromagnetic interference, drones must rely on other positioning methods, such as visual and inertial visual odometry (VIO) to obtain positioning information. In the case of GNNS denial, changes in light, texture, and environmental structure in the field environment will affect the positioning accuracy of drones. Therefore, it is of great practical significance to study how to improve the positioning accuracy of UAVs in field environments.

[0003] Visual / inertial odometry is one of the effective means to achieve autonomous positioning of drones. Although the performance of visual / inertial odometry is better than that of pure visual systems and pure inertial systems, these solutions are mostly tested using open source data sets or handheld devices during the verification phase. In the application scenarios of drone field inspections, there are often many complex and harsh factors, such as GNSS denial; the high-frequency vibration of drones has a greater impact on the zero bias of MEMS-IMU, and the zero bias error is easily affected by related independent variables such as temperature, showing nonlinear changes; in low-light conditions or areas lacking texture, visual sensors may not be able to provide reliable feature points for positioning. These problems have greatly limited the application of visual inertial positioning systems, making it necessary to further improve the effective application of visual inertial positioning systems on drones. Summary of the invention

[0004] The present invention is aimed at the problems existing in the prior art, and proposes a visual / inertial robust positioning method for power inspection drones based on temperature compensation. First, an IMU temperature online compensation network is constructed, and a Teacher Forcing mechanism is introduced to improve the learning ability of the network; then an IMU error model is constructed, and the IMU measurement information, temperature and temperature change rate, and the zero bias true value of the gyroscope and accelerometer that have been compensated by offline calibration are used to train the drone IMU zero bias temperature online compensation network based on LSTM to obtain the network parameters of the drone temperature model; during the operation of the drone, the trained parameters are retrieved, and the zero bias of the gyroscope and accelerometer corresponding to the IMU measurement at each moment is compensated online to solve the problem of increased zero bias error caused by high-frequency vibration and temperature change during the field inspection of the drone; finally, the IMU measurement after zero bias compensation and the visual residual are used for joint optimization estimation to solve the optimized positioning information. The method of the present invention also includes an adaptive visual factor to overcome the problem of positioning decline caused by visual degradation during the field power inspection of the drone, which is more in line with actual needs.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is: a visual / inertial robust positioning method for power inspection UAV based on temperature compensation, comprising the following steps:

[0006] S1. Construct an IMU temperature online compensation network: The IMU temperature online compensation network takes historical IMU measurement data and temperature data as input, and introduces a TeacherForcing mechanism to correct the network prediction error; the TeacherForcing mechanism is specifically: during network training, the zero bias true value is used as the LSTM network input at a fixed time to predict the zero bias at the next time, and correct the network prediction error;

[0007] S2. Constructing a UAV IMU zero-bias temperature model based on LSTM: The steps include the following steps:

[0008] S21: construct an IMU measurement model and use the installation error parameters obtained by offline calibration to compensate the IMU measurement information;

[0009] S22: training the IMU temperature online compensation network constructed in step S1 according to the IMU measurement information, ambient temperature information, and zero bias true values ​​of the gyroscope and accelerometer obtained in step S21, and obtaining the network parameters of the UAV IMU zero bias temperature model;

[0010] S23: Based on the network parameters of the drone temperature model in step S22, a drone IMU zero-bias temperature model based on LSTM is constructed;

[0011] S3, adaptive joint optimization estimation: The LSTM-based UAV IMU zero-bias temperature model constructed in step S2 is applied to the UAV operation process, and the gyroscope and accelerometer zero bias corresponding to the IMU measurement at each moment is compensated online. The IMU measurement after zero bias compensation and the visual residual are adaptively jointly optimized and estimated to obtain the optimized positioning information.

[0012] As an improvement of the present invention, the IMU temperature online compensation network in step S1 is defined by an LSTM neural network function F and related parameters θ, specifically:

[0013]

[0014] The left side of the above equation is the network input, including the IMU measurements from time k to time k+1. The temperature T measured inside the IMU at the kth moment k , and the accelerometer zero bias b at time k a,k and gyro bias b g,k ,in, Represent the measurements of the accelerometer and gyroscope respectively;

[0015] The network output on the right side of the above equation is the IMU accelerometer and gyroscope zero bias estimate at time k+1 Use the predicted zero bias to compensate the IMU measurement and calculate the corresponding IMU state quantity

[0016] As an improvement of the present invention, in the TeacherForcing mechanism of step S1, in the zero-bias training network, the TeacherForcing ratio is set to 0.6; in the zero-bias prediction network, TeacherForcing is turned off.

[0017] As another improvement of the present invention, in step S21, the IMU measurement model constructed is specifically:

[0018]

[0019] ω s (t)=(K g +T g )ω b (t)+b g0 (t)+b Tg0 (t)+n g (t)

[0020] Among them, a s (t) and ω s (t) represent the measurement values ​​of the accelerometer and gyroscope, respectively, K a and Kg Denote the scale factors of the accelerometer and gyroscope, T a and T g are the temperatures of the accelerometer and gyroscope, respectively, and a b (t) and ω b (t) represent the true values ​​of the accelerometer and gyroscope, respectively, and b a0 (t) and b g0 (t) are the accelerometer and gyro bias errors, b Ta0 (t) and b Tg0 (t) are the accelerometer and gyro bias temperature errors, is the rotation matrix from w to b, g w =[0,0,g] T is the acceleration of the w system, n a (t) and n g (t) are the noise of accelerometer and gyroscope respectively;

[0021] During online calibration, the IMU measurement model becomes:

[0022]

[0023]

[0024] in:

[0025] b a (t) = b a0 (t)+b Ta0 (t)+δK a a(t)

[0026] b g (t) = b g0 (t)+b Tg0 (t)+δK g ω(t)

[0027] In the formula is the IMU measurement after calibration, b a (t) and b g (t) are the bias errors of the accelerometer and gyroscope, δK a and δK g is the scale factor error.

[0028] As another improvement of the present invention, in step S22, when training the gyro bias prediction network and the accelerometer bias prediction network, the training network adopts the Adam optimizer.

[0029] The gyro bias training network batch size is 128, the training round is 200, and the learning rate is 1×10 -5 , the weight decay coefficient is 1×10-5 ;

[0030] The accelerometer bias training network batch size is 64, the training round is 150, and the learning rate is 5×10 -5 , the weight decay coefficient is 1×10 -5 ;

[0031] The ratios of training set, validation set, and test set in the training data are 0.7, 0.15, and 0.15.

[0032] As another improvement of the present invention, in the step S22 of training the gyro bias prediction network and the accelerometer bias prediction network, supervised learning is used to train the network, the mean square error is used as the evaluation index, and the loss function is:

[0033]

[0034] Among them, n represents the number of training samples, b k is the true value of the zero bias at time k; is the zero-biased estimate of the neural network prediction at time k.

[0035] As another improvement of the present invention, the adaptive joint optimization function in step S3 is specifically:

[0036]

[0037] Among them, X represents the state variable to be optimized, r prior To marginalize the prior information, represents the mth IMU measurement residual in the sliding window, ω visual represents the weight of the visual reprojection residual participating in the optimization, represents the nth visual measurement residual in the sliding window, Φ(ω visual ) represents the penalty function of the adaptive visual factor, N1 and N2 represent the number of IMU measurement residuals and visual reprojection residuals in the sliding window, respectively.

[0038] As a further improvement of the present invention, the adaptive visual factor s visual The penalty function Φ(ω visual ) specifically:

[0039] s visual =u / w

[0040]

[0041] Where u represents the number of visual feature tracking and w represents the size of the sliding window.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] (1) The method of the present invention realizes intelligent autonomous positioning of UAVs in the case of GNSS denial during power inspection, which can easily cover various parts of the power system, including hard-to-reach areas, and improve the autonomy and accuracy of inspection.

[0044] (2) The LSTM-based zero-bias temperature online compensation network proposed in the present invention overcomes the problem that the zero bias of the MEMS-IMU of the UAV cannot be accurately estimated and compensated in time under high-frequency vibration and complex temperature fields, thereby improving the positioning accuracy of the UAV in the field environment.

[0045] (3) The method of the present invention also includes an adaptive visual factor, which eliminates the impact of outliers of the visual sensor on the integrated navigation system in scenes with varying illumination distance and sparse features, improves the reliability of visual features participating in back-end optimization, and further improves the robustness of the visual / inertial integrated navigation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a flow chart of the steps of the method of the present invention;

[0047] Figure 2 is a schematic diagram of a visual / inertial integrated navigation factor graph model of the present invention;

[0048] Figure 3 is a schematic diagram of the UAV platform constructed in the test example of the present invention;

[0049] Figure 4 It is a navigation trajectory result diagram after using different methods in the test example of the present invention and comparing with the method of removing the temperature compensation and adaptive visual factor modules respectively. DETAILED DESCRIPTION

[0050] The present invention will be further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0051] Example 1

[0052] A temperature-compensated visual / inertial robust positioning method for power inspection UAVs, such as Figure 1 As shown, the specific steps are as follows:

[0053] Step S1: Design an IMU temperature online compensation network and introduce the Teacher Forcing mechanism to eliminate the accumulated error in the network and improve the learning ability of the network.

[0054] The zero bias of IMU changes with time and can be regarded as a time series with before and after correlation. In order to predict the future zero bias value, the historical IMU measurement data and temperature data are used as input and predicted through the neural network model. An LSTM neural network function F and related parameters θ are defined for this prediction task:

[0055]

[0056] The left side of the above equation is the network input, including the IMU measurements from time k to time k+1 The temperature T measured by the temperature sensor integrated inside the IMU at the kth moment k , and the accelerometer zero bias b at time k a,k and gyro bias b g,k ,in, Represent the measurements of the accelerometer and gyroscope respectively. The network output on the right side of the above equation is the estimated value of the IMU accelerometer and gyroscope zero bias at time k+1 Use the predicted zero bias to compensate the IMU measurement and calculate the corresponding IMU state quantity

[0057] During the training process, relying on the data of the previous moment for a long time will cause the accumulation of errors. Therefore, the TeacherForcing mechanism is introduced. During training, the true value of the zero bias is used as the input of the LSTM network to predict the zero bias of the next moment to eliminate the error accumulation propagation and speed up the convergence. In the zero bias training network, the TeacherForcing ratio is set to 06. In the zero bias prediction network, TeacherForcing is turned off.

[0058] Step S2: During the network training process, an IMU measurement model is constructed. First, the IMU measurement information is compensated using the installation error parameters obtained by offline calibration. Then, the LSTM network is trained using the compensated IMU measurement information, ambient temperature information, and the zero bias true values ​​of the gyroscope and accelerometer to obtain the network parameters of the drone temperature model and construct a drone IMU zero bias temperature model based on LSTM.

[0059] The network model proposed in step S1 is trained. Before network training, the IMU measurement model is first constructed. At time k, the IMU measurement model is rewritten as follows:

[0060]

[0061] ω s (t)=(K g +T g )ω b (t)+b g0 (t)+bTg0 (t)+n g (t)

[0062] Among them, a s (t) and ω s (t) represent the measurement values ​​of the accelerometer and gyroscope, respectively, K a and K g Denote the scale factors of the accelerometer and gyroscope, T a and T g Represent the temperature of the accelerometer and gyroscope, a b (t) and ω b (t) represent the true values ​​of the accelerometer and gyroscope, respectively, and b a0 (t) and b g0 (t) are the accelerometer and gyro bias errors, b Ta0 (t) and b Tg0 (t) are the accelerometer and gyro bias temperature errors, is the rotation matrix from w to b, g w =[0,0,g] T is the acceleration of the w system, n a (t) and n g (t) are the noise of the accelerometer and gyroscope respectively. In the online state, the installation error of MEMS-IMU changes little, and the installation error parameters obtained by offline calibration are used for compensation. At this time, the IMU measurement model during online calibration becomes:

[0063]

[0064]

[0065] in:

[0066] b a (t) = b a0 (t)+b Ta0 (t)+δK a a(t)

[0067] b g (t) = b g0 (t)+b Tg0 (t)+δK g ω(t)

[0068] In the formula is the IMU measurement after calibration, b a (t) and b g (t) are the bias errors of the accelerometer and gyroscope, including bias error, bias temperature error, and scale factor error δK a and δK gThree items. IMU zero bias is easily affected by external factors such as carrier vibration and ambient temperature, and changes nonlinearly over time. It needs to be calibrated online and compensated in time. Therefore, the LSTM network in step S1 is used for online compensation. The LSTM network is trained through IMU measurement information, ambient temperature information, and the true value of the zero bias of the gyro and accelerometer, and the gyro zero bias prediction network and the accelerometer zero bias prediction network are trained respectively.

[0069] The Adam optimizer is used to train the gyro bias prediction network and the accelerometer bias prediction training network. To prevent overfitting, the gyro bias training network batch size is 128, the training round is 200, and the learning rate is 1×10 -5 , the weight decay coefficient is 1×10 -5 ; The accelerometer bias training network batch size is 64, the training round is 150, and the learning rate is 5×10 -5 , the weight decay coefficient is 1×10 -5 The ratios of training set, validation set, and test set in the training data are 0.7, 0.15, and 0.15. Supervised learning is used to train the network, and the mean square error is used as an evaluation indicator to evaluate the training effect. The loss function is set as follows:

[0070]

[0071] Among them, n represents the number of training samples, b k is the true value of the zero bias at time k; is the zero bias estimate of the neural network prediction at time k. Finally, the network parameters of the drone temperature model are obtained through training, and the drone IMU zero bias temperature model based on LSTM is constructed.

[0072] Step S3: During the operation of the drone, the trained drone zero bias model is called to perform online compensation for the gyro and accelerometer zero bias corresponding to the IMU measurement at each moment, and then a visual / inertial combined navigation joint optimization system is constructed, such as Figure 2 shown. Figure 2 The figure shows the schematic diagram of the visual / inertial integrated navigation factor graph model. The factors are composed of IMU, vision, and zero bias factors. The lines between the factors reflect the mutual constraints between sensors. prior is the initial zero bias of IMU, obtained by initialization of VINS-Mono; b k is the IMU zero bias obtained at the k-th frame image time; f k is the kth camera frame, x k is the IMU measurement corresponding to the kth camera frame.

[0073] In order to reduce the complexity of the optimized visual-inertial tightly coupled integrated navigation algorithm, the algorithm selectively removes some older or unsatisfactory camera frames from the sliding window and marginalizes the IMU state x corresponding to the camera frame. k The marginalization process uses the Schur complement method. The marginalized target will no longer calculate the pose and landmark points of the target corresponding to this frame, but retain the constraint relationship of this frame to other frames in the sliding window, that is, convert the associated constraints into marginalized prior information items r prior Added to the optimization problem, the new prior information will be added to the existing prior information.

[0074] Using the visual inertial bundle adjustment method, the mean of all marginalized prior information and the sum of the Mahalanobis norm of the IMU measurement residual and the visual reprojection residual are minimized to obtain the measurement residual loss function:

[0075]

[0076] Among them, X represents the state variable to be optimized, r prior To marginalize the prior information, represents the mth IMU measurement residual in the sliding window, represents the nth visual measurement residual in the sliding window, N1 and N2 represent the number of imu measurement residuals and visual reprojection residuals in the sliding window, respectively.

[0077] In the field of power inspection, visual sensors are easily affected by factors such as drastic changes in light intensity and sparse environmental features, which reduces the ability to extract and match image features. When image outliers participate in combined navigation, the positioning accuracy of the system will decrease. Therefore, by designing an adaptive visual factor s visual To measure the degradation of the visual sensor.

[0078] s visual =u / w

[0079]

[0080] Where u represents the number of visual feature tracking and w represents the size of the sliding window. The more visual features tracked in the sliding window, the more accurate the visual information. visual Reflects the weight of the visual reprojection residual participating in the optimization. Φ(ω visual ) represents the penalty function of the adaptive visual factor. Finally, according to the adaptive visual factor s visual To construct an adaptive joint optimization function:

[0081]

[0082] Use the iterative optimization method to solve the joint optimization function of the above formula. First, fix the last moment t-1 To solve the state variable X at the current time t t , then fix the current time X t Then solve the current moment Iterate sequentially to obtain the optimal state variable X and visual reprojection residual ω visual The final solution obtains the optimized positioning information of the visual inertial navigation system, eliminates the influence of outliers of the visual sensor on the integrated navigation system in scenes with varying illumination distance and sparse features, improves the reliability of visual features participating in back-end optimization, and thus improves the robustness of the visual / inertial integrated navigation system.

[0083] Test Case

[0084] In order to verify the effectiveness of the temperature-compensated visual / inertial robust positioning method for power inspection UAVs proposed in this invention, a UAV experimental platform was built. The platform is based on the PX4 autopilot, and the power equipment selects 50A four-in-one ESC and 2810-900KV brushless motor; the flight control selects CUAV V5 nano; the power supply selects 6s 5500mAh 75C aircraft model battery; the sensors select Intel D435i camera, Xsens Mti-7 MEMS-IMU, Livox mid360 laser radar, and MTF-01 optical flow module; and it is equipped with an Intel onboard computer. The built UAV platform is as follows: Figure 3 As shown in the figure. Since flying drones near power lines is dangerous and requires official reporting, we first conducted experimental tests in a complex environment simulating power inspection. We chose a patio to simulate a GNSS-denied environment and night to simulate a visual degradation scenario. The semi-open patio environment also has temperature field changes.

[0085] The IMU zero bias true value required for training is obtained by combined navigation of IMU and LiDAR. The IMU LiDAR combined navigation algorithm FAST-LIO used is a LiDAR / inertial odometer framework with high computational efficiency and robustness. FAST-LIO uses a tightly coupled iterative extended Kalman filter to fuse IMU measurements with LiDAR point clouds, compensates for motion distortion through the back-propagation process, and outputs the IMU zero bias state quantity that is continuously updated during the information fusion process. After aligning the timestamp with the IMU measurement, it is used as the IMU zero bias true value during network training. The training time of the gyro bias network and the accelerometer bias network is about 40 minutes in total.

[0086] First, the visual inertial system is jointly calibrated, and the calibration results of relevant configuration parameters in the integrated navigation are as follows.

[0087] (1) IMU random error

[0088] Table 1 IMU random error calibration results

[0089]

[0090] Due to the influence of severe vibrations generated by the drone during flight, the IMU random error calibrated under delicate conditions is not accurate under dynamic conditions. Therefore, when configuring parameters, it is necessary to adjust the values ​​according to the current drone maneuvering state and debugging experience.

[0091] (2) Camera distortion parameters

[0092] Table 2 Calibration results of camera internal distortion parameters

[0093]

[0094] (3) Camera IMU external parameters

[0095] Table 3 Camera IMU external parameter calibration results

[0096]

[0097] Using these parameters to configure the visual / inertial integrated navigation system, the visual / inertial integrated navigation method based on the VINS-MONO algorithm is compared with the complete method of the present invention, and also compared with the method of removing the temperature compensation and adaptive visual factor modules respectively. The navigation effect is as follows: Figure 4 shown.

[0098] from Figure 4 It can be seen that the method of the present invention greatly improves the positioning accuracy of the UAV visual / inertial combined navigation system. In-depth analysis shows that in this scene, the adaptive visual factor has a greater impact on the positioning robustness. When the adaptive visual factor is removed, the positioning error of the system becomes significantly larger. This is due to the poor lighting conditions in the scene and insufficient visual feature extraction. If the adaptive visual factor is not added, the visual outliers will have a greater impact on the reliability of the system. At the same time, removing the temperature compensation also reduces the positioning accuracy of the system. This is because there are temperature changes in the scene. If the effect of temperature on zero bias is not considered, the positioning accuracy of the system will decrease.

[0099] The specific absolute position error of the navigation trajectory and the percentage improvement of navigation position accuracy compared with the VINS-Mono method are shown in Table 4 below:

[0100] Table 4. Absolute position error of navigation trajectory of data set

[0101]

[0102]

[0103] From the analysis in Table 4, it can be seen that the absolute trajectory error of the method of the present invention with the addition of adaptive visual factor and temperature compensation is reduced by 54.1% compared with VINS-Mono. From the quantified results, it can also be seen that the method of the present invention has a good effect on improving the positioning accuracy.

[0104] It should be noted that the above content only illustrates the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. For ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications all fall within the protection scope of the claims of the present invention.

Claims

1. A temperature-compensated visual / inertial robust positioning method for power inspection drones, characterized in that: The steps include: S1. Construct an IMU temperature online compensation network: The IMU temperature online compensation network takes historical IMU measurement data and temperature data as input, and introduces a TeacherForcing mechanism to correct the network prediction error; the TeacherForcing mechanism is specifically: during network training, the zero bias true value is used as the LSTM network input at a fixed time to predict the zero bias at the next time, and correct the network prediction error; S2. Constructing a UAV IMU zero-bias temperature model based on LSTM: The steps include the following steps: S21: construct an IMU measurement model and use the installation error parameters obtained by offline calibration to compensate the IMU measurement information; S22: training the IMU temperature online compensation network constructed in step S1 according to the IMU measurement information, ambient temperature information, and zero bias true values ​​of the gyroscope and accelerometer obtained in step S21, and obtaining the network parameters of the UAV IMU zero bias temperature model; S23: Based on the network parameters of the drone temperature model in step S22, a drone IMU zero-bias temperature model based on LSTM is constructed; S3, adaptive joint optimization estimation: The LSTM-based UAV IMU zero-bias temperature model constructed in step S2 is applied to the UAV operation process, and the gyroscope and accelerometer zero bias corresponding to the IMU measurement at each moment is compensated online. The IMU measurement after zero bias compensation and the visual residual are adaptively jointly optimized and estimated to obtain the optimized positioning information.

2. The temperature-compensated visual / inertial robust positioning method for electric power inspection drones according to claim 1, characterized in that: The IMU temperature online compensation network in step S1 is defined by an LSTM neural network function F and related parameters θ, specifically: The left side of the above equation is the network input, including the IMU measurements from time k to time k+1. The temperature T measured inside the IMU at the kth moment k , and the accelerometer zero bias b at time k a,k and gyro bias b g,k ,in, Represent the measurements of the accelerometer and gyroscope respectively; The network output on the right side of the above equation is the IMU accelerometer and gyroscope zero bias estimate at time k+1 Use the predicted zero bias to compensate the IMU measurement and calculate the corresponding IMU state quantity 3. The temperature-compensated visual / inertial robust positioning method for electric power inspection drones according to claim 2, characterized in that: In the TeacherForcing mechanism of step S1, in the zero-bias training network, the TeacherForcing ratio is set to 0.6; in the zero-bias prediction network, TeacherForcing is turned off.

4. The temperature-compensated visual / inertial robust positioning method for electric power inspection drones according to claim 1, characterized in that: In step S21, the IMU measurement model constructed is specifically: ω s (t)=(K g +T g )ω b (t)+b g0 (t)+b Tg0 (t)+n g (t) Among them, a s (t) and ω s (t) represent the measurement values ​​of the accelerometer and gyroscope, respectively, K a and K g Denote the scale factors of the accelerometer and gyroscope, T a and T g Represent the temperature of the accelerometer and gyroscope, a b (t) and ω b (t) represent the true values ​​of the accelerometer and gyroscope, respectively, and b a0 (t) and b g0 (t) are the accelerometer and gyro bias errors, b Ta0 (t) and b Tg0 (t) are the accelerometer and gyro bias temperature errors, is the rotation matrix from w to b, g w =[0,0,g] T is the acceleration of the w system, n a (t) and n g (t) are the noise of accelerometer and gyroscope respectively; During online calibration, the IMU measurement model becomes: in: b a (t)=b a0 (t)+b Ta0 (t)+δK a a(t) b g (t)=b g0 (t)+b Tg0 (t)+δK g ω(t) In the formula is the IMU measurement after calibration, b a (t) and b g (t) are the bias errors of the accelerometer and gyroscope, δK a and δK g is the scale factor error.

5. The temperature-compensated visual / inertial robust positioning method for electric power inspection drones according to claim 1, characterized in that: In step S22, when training the gyro bias prediction network and the accelerometer bias prediction network, the training network uses the Adam optimizer. The gyro bias training network batch size is 128, the training round is 200, and the learning rate is 1×10 -5 , the weight decay coefficient is 1×10 -5 ; The accelerometer bias training network batch size is 64, the training round is 150, and the learning rate is 5×10 -5 , the weight decay coefficient is 1×10 -5 ; The ratios of training set, validation set, and test set in the training data are 0.7, 0.15, and 0.

15.

6. The temperature-compensated visual / inertial robust positioning method for electric power inspection drones according to claim 5, characterized in that: In the step S22, in which the gyro bias prediction network and the accelerometer bias prediction network are trained, supervised learning is used to train the network, and the mean square error is used as the evaluation index. The loss function is: Among them, n represents the number of training samples, b k is the true value of the zero bias at time k; is the zero-biased estimate of the neural network prediction at time k.

7. The temperature-compensated visual / inertial robust positioning method for electric power inspection drones according to claim 1, characterized in that: The adaptive joint optimization function in step S3 is specifically: Among them, X represents the state variable to be optimized, r prior To marginalize the prior information, represents the mth IMU measurement residual in the sliding window, ω visual represents the weight of the visual reprojection residual participating in the optimization, represents the nth visual measurement residual in the sliding window, Φ(ω visual ) represents the penalty function of the adaptive visual factor, N1 and N2 represent the number of IMU measurement residuals and visual reprojection residuals in the sliding window, respectively.

8. The temperature-compensated visual / inertial robust positioning method for electric power inspection drones according to claim 7, characterized in that: The adaptive visual factor s visual The penalty function Φ(ω visual ) Specifically: s visual =u / w Where u represents the number of visual feature tracking and w represents the size of the sliding window.

Images