Parachute-based unmanned aerial vehicle inspection anti-falling method for complex power grid environment, medium and equipment

Abstract

The invention discloses a complex power grid environment-oriented unmanned aerial vehicle inspection anti-falling method based on a parachute, a medium and equipment, and the method comprises the steps: inputting a circuit diagram into a first evaluation model to obtain a circuit collision score a, and inputting a background image into a second evaluation model to obtain a road collision score b; a risk score R = kT / M and T = (q * a + (1-q) * b) are obtained based on the line collision score and the road surface collision score, k is a correction factor, q is a weight factor, and M is the operation safety level of the unmanned aerial vehicle; and controlling a parachute opening strategy based on an actor-commentator network. In order to deal with a plurality of risk factors, including environmental parameters, aircraft attitude parameters, aircraft falling speed parameters and risk scores, which may occur in a complex power grid environment, the intelligent control technology is adopted to realize autonomous control, and analysis and judgment are performed according to current variable data; a corresponding parachute opening strategy is provided through autonomous perception and intelligent control, so that safe operation of the unmanned aerial vehicle is ensured in a complex power grid environment, and the method has good practicability.

Description

Technical Field

[0001] The present invention belongs to the technical field of unmanned aerial vehicle inspection, and specifically relates to a parachute-based unmanned aerial vehicle inspection anti-fall method, medium and equipment for complex power grid environments. Background Art

[0002] Autonomous inspection of transmission lines by drones based on trajectory planning still has the risk of excessive deviation in the inspection process of existing transmission lines and substations (converter stations), such as errors caused by fluctuations in positioning accuracy and errors in trajectory planning, which may cause the drone to collide with the tower (explode the aircraft). Drone anti-fall parachute technology refers to installing a special parachute system for the drone so that it can land safely in an emergency. When a drone fails in the air, the system will automatically cut off the power supply of the drone, open the parachute system, and let the aircraft slowly descend from the sky. The parachute uses air resistance to slow down and reduce the speed when falling from a high altitude to achieve a safe landing. This safety system significantly reduces the possibility of harm to people or property on the ground, and reduces the risk of damage to the drone itself, effectively reducing the loss of the drone and the harm to people on the ground, improving the reliability and safety of autonomous inspection of transmission lines by drones, and facilitating the widespread application of autonomous inspection by drones.

[0003] The UAV anti-fall parachute system based on artificial intelligence technology will also be widely used. It can analyze and control the flight status of the UAV more intelligently, so that the UAV can fly more safely and stably under abnormal conditions. If the UAV is in a dangerous state, the parachute system will issue a command to start the parachute deployment system, so that the parachute can be deployed quickly and adjusted in time. However, there are many variables that affect the fall, and the existing model control algorithm is often unable to adapt to the complex power grid environment. Summary of the invention

[0004] The purpose of the present invention is to provide a parachute-based UAV inspection and anti-fall method, medium and equipment for complex power grid environments, aiming to solve the above-mentioned problems.

[0005] The present invention is mainly achieved through the following technical solutions:

[0006] The parachute-based UAV inspection and anti-falling method for complex power grid environments includes the following steps:

[0007] Step S1: The drone collects images of the area to be landed, and processes the images to obtain a route map and a background map;

[0008] Step S2: Input the route map into the first evaluation model to obtain the route collision score a, and input the background map into the second evaluation model to obtain the road collision score b; then obtain the risk score R=kT / M based on the route collision score and the road collision score, T=(q×a+(1-q)×b), where k is the correction factor, q is the weight factor, and M is the safety level of the UAV operation;

[0009] Step S3: Control the umbrella opening strategy based on the actor-critic network;

[0010] Collect the experience data of parachute opening of each UAV inspection and obtain the state set S = {s1, s2, ..., s n} and the umbrella opening strategy set A = {a1, a2, ..., a n}, where state s n Including environmental parameters, aircraft attitude parameters, aircraft falling speed parameters and risk scores; then, based on empirical data (s t , a t ,s t+1 , r t ), and using the actor-critic algorithm to train an actor network and a critic network, the actor network is used to control the parachute opening action, and the critic network is used to comment on the parachute opening action; the trained actor network is used to control the parachute opening action;

[0011] Where: s t is the real-time status of the UAV at time t;

[0012] s t+1 is the real-time status of the UAV at time t+1;

[0013] a t The parachute opening strategy of the UAV in the time period from t to t+1;

[0014] r t For drones by s t Change to s t+1 And implement a t Reward.

[0015] In order to better implement the present invention, further, in step S1, the image is segmented based on an image segmentation algorithm to obtain a route map and an initial background map, and then the initial background map is input into an image completion model to obtain a complete background map. Preferably, the image completion model is obtained based on full convolutional neural network training.

[0016] In order to better implement the present invention, further, in step S2, the training of the first evaluation model and the training of the second evaluation model respectively include the following steps:

[0017] Step A1: Collect route pictures or landscape background pictures to form a training sample set, and score and annotate the pictures from different dimensions. The annotation information is {S1, S2, …, S k}, where S k is the score of the kth dimension;

[0018] Step A2: Use a deep neural network to train the first evaluation model or the second evaluation model to evaluate the scores of the image under several designed dimensions; input the route image or the terrain background image, output the scores under several dimensions, and train a multi-classification deep learning model so that F(Image)={S1, S2, ..., S k};

[0019] Step A3: Take the weighted sum of the scores under each dimension Σw k S k , and the final score is used as the line collision score or road collision score, where w k The weights used to score each dimension.

[0020] In order to better implement the present invention, further, in the step A1, the dimensions of the first evaluation model include the number of towers and line density, spacing and space proportion, and the dimensions of the second evaluation model include geological type and geomorphic gullies.

[0021] In order to better implement the present invention, further, in step A3, for the first evaluation model, if the line density score is greater than the set threshold, the line collision score is η×Σw k S k , where η is the intensive risk correction coefficient.

[0022] A computer-readable storage medium stores a computer program, which implements the above method when executed by a processor.

[0023] An electronic device comprises a memory and a processor; the memory stores a computer program; the processor is used to execute the computer program in the memory to implement the above method.

[0024] The beneficial effects of the present invention are as follows:

[0025] In order to cope with multiple risk factors that may appear in a complex power grid environment, such as strong winds, thunderstorms and other natural environmental factors, the present invention first detects the space to be dropped and the ground based on the pictures taken by the drone to perform risk analysis and calculate the risk score, taking into account both the power grid and the geological conditions; then combines the actor-critic network to control the parachute opening strategy, which takes into account environmental parameters, aircraft attitude parameters, aircraft falling speed parameters and risk scores, and obtains the optimal solution for the parachute opening strategy. The present invention adopts intelligent control technology to achieve autonomous control, analyzes and judges according to the current variable data, and proposes a corresponding parachute opening strategy through autonomous perception and intelligent control, thereby ensuring the safe operation of the drone in a complex power grid environment, and has good practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a flow chart of the parachute-based UAV inspection and fall prevention method for complex power grid environments of the present invention;

[0027] Figure 2 It is a principle block diagram of the present invention. DETAILED DESCRIPTION

[0028] Embodiment 1:

[0029] A parachute-based UAV inspection and anti-fall method for complex power grid environments, such as Figure 1 and Figure 2 As shown, the following steps are included:

[0030] Step S1: The drone collects images of the area to be landed, and processes the images to obtain a route map and a background map;

[0031] First, the collected image is subjected to noise removal and image enhancement processing:

[0032] Noise removal: In falling scenes, noise often appears in the image due to vibration, lighting changes and other factors during image acquisition, affecting subsequent segmentation and detection results. Gaussian filtering and median filtering can be used to eliminate these noises.

[0033] Image enhancement: It is to purposefully emphasize the overall or local characteristics of an image, such as improving the color, brightness, and contrast of an image, making an originally unclear image clear or emphasizing certain interesting features, expanding the differences between the features of different objects in the image, suppressing uninteresting features, and improving the visual effect of the image. Specifically, it includes:

[0034] (1) Histogram equalization: By adjusting the contrast of the image, the sign area is made more prominent and the performance of the segmentation model is enhanced. In this process, the histogram equalization technique is used to make the grayscale distribution of the image more uniform, thereby highlighting the outline and details of the road sign.

[0035] (2) Brightness and contrast adjustment: Automatically adjust the brightness and contrast of the image according to the lighting conditions of the shooting environment to ensure that the image remains consistent under different lighting conditions.

[0036] Based on the image segmentation algorithm, the image is segmented into a route map and an initial background map, and then the initial background map is input into the image completion model to obtain the complete background map. The image completion model is trained based on a fully convolutional neural network. The fully convolutional neural network uses convolutional layers in all layers of the network, which can accept input images of any size and produce outputs of corresponding sizes. In order to train the image completion network to maintain consistency, global and local context discriminators are used. The discriminators are trained to distinguish between real images and completed images. The global discriminator looks at the entire image to assess whether it is coherent as a whole, while the local discriminator only looks at a small area centered on the completed area to ensure the consistency of the generated patches locally. Then, the image completion network is trained to deceive these two context discriminator networks, requiring it to generate images that are indistinguishable from real images in terms of overall coherence and details.

[0037] Step S2: Input the route map into the first evaluation model to obtain the route collision score a, and input the background map into the second evaluation model to obtain the road collision score b; then obtain the risk score R=kT / M based on the route collision score and the road collision score, T=(q×a+(1-q)×b), where k is the correction factor, q is the weight factor, and M is the safety level of the UAV operation;

[0038] Specifically, the training of the first evaluation model includes the following steps:

[0039] 1. Collect line pictures and form a training sample set, and score and annotate the pictures based on the number of towers, line density, spacing, and space ratio. The annotated information is the score S of the dimension;

[0040] 2. Use a deep neural network to train the first evaluation model or the second evaluation model to evaluate the scores of the image under several designed dimensions; input the route image, output the scores under several dimensions, and train a multi-classification deep learning model so that F(Image) = {S1, S2, ..., S k};

[0041] 3. Take the weighted sum of the scores under each dimension Σw k S k , and the final score is used as the line collision score, where w k Preferably, if the line density score is greater than a set threshold, the line collision score is η×Σw k S k , where η is the intensive risk correction coefficient.

[0042] Specifically, the training of the second evaluation model includes the following steps:

[0043] (1) Collect geomorphic background images and form a training sample set. Score and annotate the images based on the dimensions of geological type and geomorphic gully. The annotation information is the dimension score S.

[0044] (2) Using a deep neural network to train a second evaluation model to evaluate the scores of the image under several designed dimensions; inputting a route image, outputting the scores under several dimensions, and training a multi-classification deep learning model so that F(Image) = {S1, S2, ..., S k};

[0045] (3) Take the weighted sum of the scores under each dimension Σw k S k , and the final score is used as the road collision score, where w k The weights used to score each dimension.

[0046] Step S3: Control the umbrella opening strategy based on the actor-critic network;

[0047] Collect the experience data of parachute opening of each UAV inspection and obtain the state set S = {s1, s2, ..., s n} and the umbrella opening strategy set A = {a1, a2, ..., a n}, where state s n Including environmental parameters, aircraft attitude parameters, aircraft falling speed parameters and risk scores; then, based on empirical data (s t , a t ,s t+1 , r t ), and using the actor-critic algorithm to train an actor network and a critic network, the actor network is used to control the parachute opening action, and the critic network is used to comment on the parachute opening action; the trained actor network is used to control the parachute opening action;

[0048] Where: s t is the real-time status of the UAV at time t;

[0049] s t+1 is the real-time status of the UAV at time t+1;

[0050] a t The parachute opening strategy of the UAV in the time period from t to t+1;

[0051] r t For drones by s t Change to s t+1 And implement a tReward.

[0052] The actor-critic algorithm combines the policy-based algorithm with the value-based algorithm, looking at both the strategies of different actions corresponding to each state and the value of each state-action. The value of each state-action is the expected cumulative benefit from this state-action to the end point. The actor-critic algorithm combines algorithms based on value functions and policy gradients, including actor networks and critic networks, and uses actor networks and critic networks for joint learning. Among them, the actor network is used to learn good strategies and action selection, and the critic network is used to evaluate the input state value and provide reinforcement signals. The critic network parameters are updated using a value function-based algorithm, and the mean square error is used as the loss function. After going through multiple sampling trajectories, the output value gradually approaches the true value.

[0053] In order to cope with multiple risk factors that may appear in a complex power grid environment, such as strong winds, thunderstorms and other natural environmental factors, the present invention first detects the space to be dropped and the ground based on the pictures taken by the drone to perform risk analysis and calculate the risk score, taking into account both the power grid and the geological conditions; then combines the actor-critic network to control the parachute opening strategy, which takes into account environmental parameters, aircraft attitude parameters, aircraft falling speed parameters and risk scores, and obtains the optimal solution for the parachute opening strategy. The present invention adopts intelligent control technology to achieve autonomous control, analyzes and judges according to the current variable data, and proposes a corresponding parachute opening strategy through autonomous perception and intelligent control, thereby ensuring the safe operation of the drone in a complex power grid environment, and has good practicality.

[0054] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification or equivalent change made to the above embodiment based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A parachute-based UAV inspection and anti-falling method for complex power grid environments, characterized in that: The following steps are involved: Step S1: The drone collects images of the area to be landed, and processes the images to obtain a route map and a background map; Step S2: Input the route map into the first evaluation model to obtain the route collision score a, and input the background map into the second evaluation model to obtain the road collision score b; then obtain the risk score R=kT / M based on the route collision score and the road collision score, T=(q×a+(1-q)×b), where k is the correction factor, q is the weight factor, and M is the safety level of the UAV operation; Step S3: Control the umbrella opening strategy based on the actor-critic network; Collect the experience data of parachute opening of each UAV inspection and obtain the state set S = {s1, s2, ..., s n } and the umbrella opening strategy set A = {a1, a2, ..., a n }, where state s n Including environmental parameters, aircraft attitude parameters, aircraft falling speed parameters and risk scores; then, based on empirical data (s t , a t ,s t+1 , r t ), and using the actor-critic algorithm to train an actor network and a critic network, the actor network is used to control the parachute opening action, and the critic network is used to comment on the parachute opening action; the trained actor network is used to control the parachute opening action; Where: s t is the real-time status of the UAV at time t; s t+1 is the real-time status of the UAV at time t+1; a t The parachute opening strategy of the UAV in the time period from t to t+1; r t For drones by s t Change to s t+1 And implement a t Reward.

2. The parachute-based UAV inspection and anti-falling method for complex power grid environments according to claim 1 is characterized in that: In the step S1, the image is segmented based on an image segmentation algorithm to obtain a route map and an initial background map, and then the initial background map is input into an image completion model to obtain a complete background map.

3. The parachute-based UAV inspection and anti-falling method for complex power grid environments according to claim 2 is characterized in that: The image completion model is obtained based on full convolutional neural network training.

4. The parachute-based UAV inspection and anti-falling method for complex power grid environments according to claim 2 is characterized in that: In step S2, the training of the first evaluation model and the training of the second evaluation model respectively include the following steps: Step A1: Collect route pictures or landscape background pictures to form a training sample set, and score and annotate the pictures from different dimensions. The annotation information is {S1, S2, …, S k }, where S k is the score of the kth dimension; Step A2: Use a deep neural network to train the first evaluation model or the second evaluation model to evaluate the scores of the image under several designed dimensions; input the route image or the terrain background image, output the scores under several dimensions, and train a multi-classification deep learning model so that F(Image)={S1, S2, ..., S k }; Step A3: Take the weighted sum of the scores under each dimension Σw k S k , and the final score is used as the line collision score or road collision score, where w k The weights used to score each dimension.

5. The parachute-based UAV inspection and anti-falling method for complex power grid environments according to claim 4 is characterized in that: In step A1, the dimensions of the first evaluation model include the number of towers and line density, spacing and space proportion, and the dimensions of the second evaluation model include geological type and geomorphic gullies.

6. The parachute-based UAV inspection and anti-falling method for complex power grid environments according to claim 5 is characterized in that: In step A3, for the first evaluation model, if the line density score is greater than the set threshold, the line collision score is η×Σw k S k , where η is the intensive risk correction coefficient.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method described in any one of claims 1 to 6 is implemented.

8. An electronic device, characterized in that: The invention comprises a memory and a processor; the memory stores a computer program; the processor is used to execute the computer program in the memory to implement the method according to any one of claims 1 to 6.

Images