Abnormal warning method and system for inspection robot

Abstract

The present application proposes an abnormal warning method for an inspection robot, comprising the following steps: collecting pictures of the inspection robot containing foreign objects or reference objects on the inspection route; making a first convolutional neural network, substituting straight line segment labels and non-straight line segment labels into the first convolutional neural network for training to obtain a first recognition model; making a second convolutional neural network, substituting the first label and the second label into the second convolutional neural network for training to obtain a second recognition model; installing the first recognition model and the second recognition model on the inspection robot, using the second recognition model to identify the non-straight line end, and judging whether the non-straight line segment is a foreign object. If so, the inspection robot stops operating and sends a warning message to the staff. Training the two models separately can improve the accuracy of individual model recognition and improve the recognition rate of foreign objects by the inspection robot during recognition.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment inspection, and in particular to an abnormality early warning method and system for an inspection robot. Background Art

[0002] With the rapid development of society and economy, the demand for electricity for residents and industry is rising. The safety status of transmission lines will directly affect the stable operation of the power grid and the economic development of the country. As a new, efficient and intelligent online inspection equipment, inspection robots carrying high-definition depth cameras are gradually replacing traditional manual inspection methods to improve the efficiency and accuracy of online inspections. However, when robots inspect high-voltage lines, they often encounter interference from various obstacles such as greenhouse films, discarded plastic bags, and dry straw. These foreign objects are generally small in size and difficult to be directly observed by manual inspections. Once entangled in the transmission line, it will block the rolling of the robot's walking wheels, affecting the normal inspection work, or even cause a short circuit in the line, causing a large area of ​​power outages in the surrounding area, causing great economic losses. In the existing technology, there are robots that can identify foreign objects on the inspection line, but due to the presence of circuit equipment such as shock-absorbing hammers on the inspection line, the robot is prone to misidentification during identification, thereby affecting the efficiency of the inspection. Summary of the invention

[0003] In view of the above-mentioned defects, the purpose of the present invention is to propose an abnormal warning method and system for an inspection robot, so as to improve the recognition accuracy of the robot and reduce the occurrence of misjudgment of foreign objects.

[0004] To achieve this purpose, the present invention adopts the following technical solution: an abnormal warning method for an inspection robot, comprising the following steps:

[0005] Step S1: collecting pictures of the inspection robot containing foreign objects or reference objects on the inspection route, separating and extracting the route, foreign objects and reference objects in the pictures, and obtaining a parameter data set containing the route, foreign objects and reference objects;

[0006] Step S2: Create a first convolutional neural network, and use a labeling tool to label the lines, foreign objects, and reference objects in the parameter data set, label the lines as straight line segment labels, label the foreign objects and reference objects as non-straight line segment labels, substitute the straight line segment labels and the non-straight line segment labels into the first convolutional neural network for training to obtain a first recognition model;

[0007] Step S3: making a second convolutional neural network, and using a labeling tool to label the foreign objects and reference objects in the parameter data set, labeling the foreign objects as a first label, labeling the reference objects as a second label, substituting the first label and the second label into the second convolutional neural network for training to obtain a second recognition model;

[0008] Step S4: Store the first recognition model and the second recognition model in the hard disk of the inspection robot, and set the recognition distance of the inspection robot. When the inspection robot uses the first recognition model to identify a non-straight segment within the recognition distance, the second recognition model is used to identify the non-straight end to determine whether the non-straight segment is a foreign object. If so, the inspection robot stops operating and sends a warning message to the staff.

[0009] Preferably, the inspection robot uses a binocular depth camera to acquire images.

[0010] Preferably, before setting the recognition distance of the inspection robot in step S4, the following steps need to be performed:

[0011] Calibrate the binocular depth camera to obtain the intrinsic parameters, extrinsic parameters and homography matrix of the two cameras, and use the binocular depth camera to shoot and obtain the original image;

[0012] The original image is corrected according to the calibration result and the intrinsic and extrinsic parameters, and the two corrected images are located in the same plane and parallel to each other;

[0013] Perform pixel matching on the two corrected images to obtain matching results;

[0014] The depth of each pixel is calculated based on the matching results to obtain a depth map.

[0015] Preferably, before performing step S2, the following steps need to be performed on the parameter data set:

[0016] Eliminate abnormal data with obstacle ratio less than the threshold, and perform uniform normalization on the parameter data set;

[0017] The processed parameter data set is divided into a training set and a validation set in proportion.

[0018] Preferably, step S4 comprises the following steps:

[0019] Step S41: when the inspection robot identifies a non-straight segment within the identification distance, it captures a frame of the video stream to obtain a picture containing the non-straight segment, and stores the captured picture containing the non-straight segment in the hard disk of the robot;

[0020] Step S42: obtaining the current network bandwidth of the inspection robot, and uploading the image containing the non-straight line segment in the hard disk to the cloud when the network bandwidth reaches a threshold;

[0021] Step S43: the cloud calls the first recognition model and the second recognition model to recognize the image containing the non-straight line segments, and updates the first recognition model and the second recognition model.

[0022] An abnormal warning system for an inspection robot includes: a collection module, a training module, an installation module and a warning module;

[0023] The acquisition module includes a training set acquisition module, which is used to collect pictures of the inspection robot containing foreign objects or reference objects on the inspection route, separate and extract the route, foreign objects and reference objects in the picture, and obtain a parameter data set containing the route, foreign objects and reference objects;

[0024] The training module includes a first training module and a second training module, the first training module includes a first convolutional neural network, the first training module is used to use a labeling tool to label the lines, foreign objects and reference objects in the parameter data set, mark the lines as straight line segment labels, mark the foreign objects and reference objects as non-straight line segment labels, substitute the straight line segment labels and the non-straight line segment labels into the first convolutional neural network for training to obtain a first recognition model;

[0025] The second training module includes a second convolutional neural network, and the second training module is used to use a labeling tool to label the foreign matter and the reference object in the parameter data set, label the foreign matter as a first label, label the reference object as a second label, and substitute the first label and the second label into the second convolutional neural network for training to obtain a second recognition model;

[0026] The installation module is used to store the first recognition model and the second recognition model in the hard disk of the inspection robot;

[0027] The early warning module is used to call the first recognition model and the second recognition model to identify foreign objects on the inspection route. If foreign objects are found, the inspection robot stops operating and sends an early warning message to the staff.

[0028] Preferably, the abnormal warning system of the inspection robot uses a binocular depth camera to acquire images.

[0029] Preferably, it also includes an image depth acquisition module, which is used to calibrate the binocular depth camera, obtain the intrinsic parameters, extrinsic parameters and homography matrix of the two cameras, and use the binocular depth camera to shoot to obtain the original image;

[0030] The original image is corrected according to the calibration result and the intrinsic and extrinsic parameters, and the two corrected images are located in the same plane and parallel to each other;

[0031] Perform pixel matching on the two corrected images to obtain matching results;

[0032] The depth of each pixel is calculated based on the matching results to obtain a depth map.

[0033] Preferably, it also includes a data processing module, which is used to remove abnormal data with an obstacle ratio less than a threshold value and perform unified normalization processing on the parameter data set;

[0034] The processed parameter data set is divided into a training set and a validation set in proportion.

[0035] Preferably, the acquisition module further comprises an update data acquisition module, and the update data acquisition module is used for capturing frames of the video stream when the inspection robot identifies a non-straight segment within the recognition distance, obtaining a picture containing the non-straight segment, and storing the captured picture containing the non-straight segment in the hard disk of the robot;

[0036] The network bandwidth of the current inspection robot is obtained, and when the network bandwidth reaches a threshold, the image containing the non-straight line segment in the hard disk is uploaded to the cloud;

[0037] The cloud calls the first recognition model and the second recognition model to recognize the picture containing non-straight line segments, and updates the first recognition model and the second recognition model.

[0038] The beneficial effects of this technical solution are: 1. Training two models separately can improve the accuracy of individual model recognition and improve the recognition rate of foreign objects by the inspection robot.

[0039] 2. The inspection robot does not directly call two recognition models at the same time. When inspecting the line, the inspection robot will use the first recognition model first for recognition. Only when the first recognition model recognizes a non-straight line segment will the second recognition model be called. This usage scheme can reasonably use the limited computing memory inside the inspection robot and improve recognition efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a flowchart of the abnormal warning method of the inspection robot;

[0041] Figure 2 It is a structural diagram of the abnormal warning system of the inspection robot;

[0042] Figure 3 It is a schematic diagram of a binocular depth camera acquiring an image. DETAILED DESCRIPTION

[0043] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0044] In the description of the present invention, it is to be understood that the terms “center”, “longitudinal”, “lateral”, “length”, “width”, “thickness”, “up”, “down”, “left”, “right”, “vertical”, “horizontal”, “top”, “bottom”, “inside”, “outside”, “axial”, “radial”, “circumferential”, etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present invention.

[0045] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "plurality" means two or more.

[0046] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a direct connection, or an indirect connection through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0047] like Figures 1 to 3 As shown, a method for abnormal warning of an inspection robot includes the following steps:

[0048] Step S1: collecting pictures of the inspection robot containing foreign objects or reference objects on the inspection route, separating and extracting the route, foreign objects and reference objects in the pictures, and obtaining a parameter data set containing the route, foreign objects and reference objects;

[0049] Step S2: Create a first convolutional neural network, and use a labeling tool to label the lines, foreign objects, and reference objects in the parameter data set, label the lines as straight line segment labels, label the foreign objects and reference objects as non-straight line segment labels, substitute the straight line segment labels and the non-straight line segment labels into the first convolutional neural network for training to obtain a first recognition model;

[0050] Step S3: making a second convolutional neural network, and using a labeling tool to label the foreign objects and reference objects in the parameter data set, labeling the foreign objects as a first label, labeling the reference objects as a second label, substituting the first label and the second label into the second convolutional neural network for training to obtain a second recognition model;

[0051] Step S4: Store the first recognition model and the second recognition model in the hard disk of the inspection robot, and set the recognition distance of the inspection robot. When the inspection robot uses the first recognition model to identify a non-straight segment within the recognition distance, the second recognition model is used to identify the non-straight end to determine whether the non-straight segment is a foreign object. If so, the inspection robot stops operating and sends a warning message to the staff.

[0052] In the existing technology, most recognition models only recognize straight segments and non-straight segments, which makes it easy to recognize circuit equipment such as anti-vibration hammers as foreign objects in actual operation. For this reason, the present application is provided with two recognition models, namely the first recognition model and the second recognition model. The first recognition model is used to recognize straight segments and non-straight segments, that is, to recognize lines and objects on the inspection lines. Since the first convolutional neural network has a simple task of recognizing straight lines and non-straight lines during the training process, the recognition accuracy of the first recognition model obtained after training will be very high in actual work. The second recognition model is used to recognize foreign objects and reference objects. The reference objects can be circuit equipment such as anti-vibration hammers and insulators, and the foreign objects can be discarded plastic bags, greenhouse films, dry straw, etc. In the training of the second convolutional neural network, a large amount of straight line segment label data is eliminated, so the training parameter amount of the second recognition model is greatly reduced, so as to concentrate on training to distinguish foreign objects from reference objects, and finally greatly improve the recognition efficiency and accuracy of foreign objects and reference objects.

[0053] Before the use of the inspection robot, the first recognition model and the second recognition model should be uploaded to the inspection robot, and the inspection scene should be arranged in the laboratory to place the inspection robot in the inspection scene for testing and training to determine whether the first recognition model and the second recognition model can play a recognition role. In addition, due to the limitations of the working environment of the inspection robot, its communication is unstable, and the first recognition model and the second recognition model cannot be called through the cloud. The inspection robot in this application moves on the inspection line, and in order to ensure the normal inspection of the inspection robot, the weight of the inspection robot must be strictly controlled, so it is impossible to install too many computing devices on the inspection robot. Limited by its computing power, the inspection robot will not directly call two recognition models at the same time. When patrolling the line, the inspection robot will give priority to using the first recognition model for recognition. Only when the first recognition model recognizes a non-straight segment will the second recognition model be called. This usage scheme can reasonably use the limited computing memory inside the inspection robot to improve recognition efficiency.

[0054] In addition, the present application will also adjust the inspection robot's recognition distance before the inspection robot is used. Since the inspection robot acquires images in real time through the camera during actual use, and then uses the first recognition model to recognize the acquired images, when the inspection robot's recognition distance is too long, the camera is likely to intercept information about the sky, poles, trees, hillsides, and houses, which can easily lead to misjudgment, etc. When the inspection robot's recognition distance is too short, the distance between the inspection robot and foreign objects during recognition will become shorter, reducing the inspection robot's reaction time. Preferably, the recognition distance is 3-5m, and when adjusting the recognition distance, the camera's recognition center should be aligned with the line at the same time to reduce the intake of unnecessary objects, reduce the inspection robot's recognition volume, and improve the inspection robot's recognition efficiency and accuracy.

[0055] In one embodiment of the present application, the first convolutional neural network and the second convolutional neural network use the improved yolo_v4_tiny algorithm, batch size batch_size = 128, image width and height of the input network width = height = 256, momentum momentum = 1.9, and learning rate learning_rate = 0.00275. According to the set batch_size and epoch, the number of training iterations is more than 30,000 times, and the loss function no longer decreases at about 11,100 iterations. When the loss value reaches 0.0975, the training stops at the default number of iterations 33,300, and the first recognition model and the second recognition model converge to obtain the first recognition model and the second recognition model with high precision recognition.

[0056] Preferably, the inspection robot uses a binocular depth camera to acquire images.

[0057] Since the line patrol robot is used for outdoor line patrol work, it is easily affected by the angle, intensity and shadow of light outdoors. The binocular depth camera can change the shooting angle through multiple adjustments to obtain the best central area of ​​the image to overcome the above problems. In addition, when training the first recognition model and the second recognition model, the consideration of the angle, intensity and shadow can be abandoned, thereby improving the accuracy of the line patrol robot in identifying foreign objects in actual use. If a monocular camera is used in one embodiment, the influence of the changing angle and intensity of light needs to be substituted into the training process when training the first recognition model and the second recognition model, which will undoubtedly increase the difficulty of training and the accuracy of the model.

[0058] Preferably, before setting the recognition distance of the inspection robot in step S4, the following steps need to be performed:

[0059] Calibrate the binocular depth camera to obtain the intrinsic parameters, extrinsic parameters and homography matrix of the two cameras, and use the binocular depth camera to shoot and obtain the original image;

[0060] The original image is corrected according to the calibration result and the intrinsic and extrinsic parameters, and the two corrected images are located in the same plane and parallel to each other;

[0061] Perform pixel matching on the two corrected images to obtain matching results;

[0062] The depth of each pixel is calculated based on the matching results to obtain a depth map.

[0063] The principle of the depth camera based on binocular stereo vision is different from that of the depth camera based on TOF and structured light. It is similar to human eyes and does not actively project light to the outside. It completely relies on the parallax of the two cameras, that is, the two pictures (grayscale or RGB color pictures) taken to calculate the depth. The schematic diagram of adjusting the depth camera angle to the optimal sampling angle of 40-80 degrees and obtaining the depth image is as follows Figure 3 As shown. The three points on the lower line at different distances will be projected at the same position by the lower camera, so the monocular camera alone cannot distinguish whether the image is the far point or the near point. However, the projections of these three points on the upper camera are located at three different positions, so by combining the information of the two cameras, it is possible to determine which point is being found. If a monocular camera is used in one embodiment, an additional sensor needs to be installed to detect the identification distance.

[0064] Preferably, before performing step S2, the following steps need to be performed on the parameter data set:

[0065] Eliminate abnormal data with obstacle ratio less than the threshold, and perform uniform normalization on the parameter data set;

[0066] The processed parameter data set is divided into a training set and a validation set in proportion.

[0067] Abnormal data with a small obstacle ratio can be birds, insects, etc. Such data will automatically leave during the actual inspection process of the inspection robot and will not affect the inspection of the inspection robot. After removing these abnormal data, the number of training non-straight segment labels can be greatly reduced, improving the recognition efficiency between non-straight segments and straight segments. After processing the abnormal data, the parameter data set is uniformly normalized, which can speed up the convergence of the first recognition model and the second recognition model.

[0068] In addition, since the data before training were collected by manually controlled inspection robots and then manually labeled, the number of training sets was not particularly large. In order to make the training set account for a larger proportion and the final training more sufficient, the division ratio of the training set and the validation set can be selected from 9:1 to 8:1, and multiple models will be made when making the first recognition model and the second recognition model. The data in the training set is used for the model to learn the relationship between input and output, and the validation set is configured to estimate the training level of the model. The best model can be selected based on the performance on the validation set.

[0069] Preferably, step S4 comprises the following steps:

[0070] Step S41: when the inspection robot identifies a non-straight segment within the identification distance, it captures a frame of the video stream to obtain a picture containing the non-straight segment, and stores the captured picture containing the non-straight segment in the hard disk of the robot;

[0071] Step S42: obtaining the current network bandwidth of the inspection robot, and uploading the image containing the non-straight line segment in the hard disk to the cloud when the network bandwidth reaches a threshold;

[0072] Step S43: the cloud calls the first recognition model and the second recognition model to recognize the image containing the non-straight line segments, and updates the first recognition model and the second recognition model.

[0073] Since the non-straight line segment labels are manually marked during training, and the number of training is limited, the inspection robot may encounter new obstacles (non-straight line segments) during the actual inspection process. Since there is no training data record of this type in the recognition model, the inspection robot cannot determine whether the new obstacle is a reference object or a foreign object. Therefore, the inspection robot will capture the video stream frame at this time to obtain the picture containing the non-straight line segment. In addition, due to the network environment of the inspection robot, the robot cannot upload the picture containing the non-straight line segment to the cloud in time to update the first recognition model and the second recognition model, so the picture containing the non-straight line segment can only be saved in the hard disk; and uploaded to the cloud again after there is a certain bandwidth. The staff can download the picture in the cloud, determine whether the inspection robot can continue to move forward, and label the non-straight line segment of the picture and substitute it into the first convolutional neural network and the second convolutional neural network for training and updating. This application takes into account the update of the model in the later operation, which improves the practicality and recognition accuracy of the inspection robot.

[0074] An abnormal warning system for an inspection robot includes: a collection module, a training module, an installation module and a warning module;

[0075] The acquisition module includes a training set acquisition module, which is used to collect pictures of the inspection robot containing foreign objects or reference objects on the inspection route, separate and extract the route, foreign objects and reference objects in the picture, and obtain a parameter data set containing the route, foreign objects and reference objects;

[0076] The training module includes a first training module and a second training module, the first training module includes a first convolutional neural network, the first training module is used to use a labeling tool to label the lines, foreign objects and reference objects in the parameter data set, mark the lines as straight line segment labels, mark the foreign objects and reference objects as non-straight line segment labels, substitute the straight line segment labels and the non-straight line segment labels into the first convolutional neural network for training to obtain a first recognition model;

[0077] The second training module includes a second convolutional neural network, and the second training module is used to use a labeling tool to label the foreign matter and the reference object in the parameter data set, label the foreign matter as a first label, label the reference object as a second label, and substitute the first label and the second label into the second convolutional neural network for training to obtain a second recognition model;

[0078] The installation module is used to store the first recognition model and the second recognition model in the hard disk of the inspection robot;

[0079] The early warning module is used to call the first recognition model and the second recognition model to identify foreign objects on the inspection route. If foreign objects are found, the inspection robot stops operating and sends an early warning message to the staff.

[0080] Preferably, the abnormal warning system of the inspection robot uses a binocular depth camera to acquire images.

[0081] Preferably, it also includes an image depth acquisition module, which is used to calibrate the binocular depth camera, obtain the intrinsic parameters, extrinsic parameters and homography matrix of the two cameras, and use the binocular depth camera to shoot to obtain the original image;

[0082] The original image is corrected according to the calibration result and the intrinsic and extrinsic parameters, and the two corrected images are located in the same plane and parallel to each other;

[0083] Perform pixel matching on the two corrected images to obtain matching results;

[0084] The depth of each pixel is calculated based on the matching results to obtain a depth map.

[0085] Preferably, it also includes a data processing module, which is used to remove abnormal data with an obstacle ratio less than a threshold value and perform unified normalization processing on the parameter data set;

[0086] The processed parameter data set is divided into a training set and a validation set in proportion.

[0087] Preferably, the acquisition module further comprises an update data acquisition module, and the update data acquisition module is used for capturing frames of the video stream when the inspection robot identifies a non-straight segment within the recognition distance, obtaining a picture containing the non-straight segment, and storing the captured picture containing the non-straight segment in the hard disk of the robot;

[0088] The network bandwidth of the current inspection robot is obtained, and when the network bandwidth reaches a threshold, the image containing the non-straight line segment in the hard disk is uploaded to the cloud;

[0089] The cloud calls the first recognition model and the second recognition model to recognize the picture containing non-straight line segments, and updates the first recognition model and the second recognition model.

[0090] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "illustrative embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0091] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. An abnormal warning method for an inspection robot, characterized in that: The steps include: Step S1: collecting pictures of the inspection robot containing foreign objects or reference objects on the inspection route, separating and extracting the route, foreign objects and reference objects in the pictures, and obtaining a parameter data set containing the route, foreign objects and reference objects; Step S2: Create a first convolutional neural network, and use a labeling tool to label the lines, foreign objects, and reference objects in the parameter data set, label the lines as straight line segment labels, label the foreign objects and reference objects as non-straight line segment labels, substitute the straight line segment labels and the non-straight line segment labels into the first convolutional neural network for training to obtain a first recognition model; Step S3: making a second convolutional neural network, and using a labeling tool to label the foreign objects and reference objects in the parameter data set, labeling the foreign objects as a first label, labeling the reference objects as a second label, substituting the first label and the second label into the second convolutional neural network for training to obtain a second recognition model; Step S4: Store the first recognition model and the second recognition model in the hard disk of the inspection robot, and set the recognition distance of the inspection robot. When the inspection robot uses the first recognition model to identify a non-straight segment within the recognition distance, the second recognition model is used to identify the non-straight end to determine whether the non-straight segment is a foreign object. If so, the inspection robot stops operating and sends a warning message to the staff.

2. The abnormal warning method of the inspection robot according to claim 1 is characterized in that: The inspection robot uses a binocular depth camera to acquire images.

3. The abnormal warning method of the inspection robot according to claim 2 is characterized in that: Before setting the recognition distance of the inspection robot in step S4, the following steps need to be performed: Calibrate the binocular depth camera to obtain the intrinsic parameters, extrinsic parameters and homography matrix of the two cameras, and use the binocular depth camera to shoot and obtain the original image; The original image is corrected according to the calibration result and the intrinsic and extrinsic parameters, and the two corrected images are located in the same plane and parallel to each other; Perform pixel matching on the two corrected images to obtain matching results; The depth of each pixel is calculated based on the matching results to obtain a depth map.

4. The abnormal warning method of the inspection robot according to claim 1 is characterized in that: Before proceeding to step S2, the following steps need to be performed on the parameter data set: Eliminate abnormal data with obstacle ratio less than the threshold, and perform uniform normalization on the parameter data set; The processed parameter data set is divided into a training set and a validation set in proportion.

5. The abnormal warning method of the inspection robot according to claim 1 is characterized in that: The step S4 comprises the following steps: Step S41: when the inspection robot identifies a non-straight segment within the identification distance, it captures a frame of the video stream to obtain a picture containing the non-straight segment, and stores the captured picture containing the non-straight segment in the hard disk of the robot; Step S42: obtaining the current network bandwidth of the inspection robot, and uploading the image containing the non-straight line segment in the hard disk to the cloud when the network bandwidth reaches a threshold; Step S43: the cloud calls the first recognition model and the second recognition model to recognize the image containing the non-straight line segments, and updates the first recognition model and the second recognition model.

6. An abnormal warning system for an inspection robot, using an abnormal warning method for an inspection robot according to any one of claims 1 to 5, characterized in that: include: Acquisition module, training module, installation module and early warning module; The acquisition module includes a training set acquisition module, which is used to collect pictures of the inspection robot containing foreign objects or reference objects on the inspection route, separate and extract the route, foreign objects and reference objects in the picture, and obtain a parameter data set containing the route, foreign objects and reference objects; The training module includes a first training module and a second training module, the first training module includes a first convolutional neural network, the first training module is used to use a labeling tool to label the lines, foreign objects and reference objects in the parameter data set, mark the lines as straight line segment labels, mark the foreign objects and reference objects as non-straight line segment labels, substitute the straight line segment labels and the non-straight line segment labels into the first convolutional neural network for training to obtain a first recognition model; The second training module includes a second convolutional neural network, and the second training module is used to use a labeling tool to label the foreign matter and the reference object in the parameter data set, label the foreign matter as a first label, label the reference object as a second label, and substitute the first label and the second label into the second convolutional neural network for training to obtain a second recognition model; The installation module is used to store the first recognition model and the second recognition model in the hard disk of the inspection robot; The early warning module is used to call the first recognition model and the second recognition model to identify foreign objects on the inspection route. If foreign objects are found, the inspection robot stops operating and sends an early warning message to the staff.

7. The abnormal warning system for an inspection robot according to claim 6, characterized in that: The abnormal warning system of the inspection robot uses a binocular depth camera to acquire images.

8. The abnormal warning system for an inspection robot according to claim 7, characterized in that: It also includes an image depth acquisition module, which is used to calibrate the binocular depth camera, obtain the intrinsic parameters, extrinsic parameters and homography matrix of the two cameras, and use the binocular depth camera to shoot to obtain the original image; The original image is corrected according to the calibration result and the intrinsic and extrinsic parameters, and the two corrected images are located in the same plane and parallel to each other; Perform pixel matching on the two corrected images to obtain matching results; The depth of each pixel is calculated based on the matching results to obtain a depth map.

9. The abnormal warning system for an inspection robot according to claim 6, characterized in that: It also includes a data processing module, which is used to remove abnormal data with an obstacle ratio less than a threshold value and perform unified normalization processing on the parameter data set; The processed parameter data set is divided into a training set and a validation set in proportion.

10. The abnormal warning system for an inspection robot according to claim 6, characterized in that: The acquisition module further includes an update data acquisition module, which is used to capture frames of the video stream when the inspection robot identifies a non-straight segment within the identification distance, obtain a picture containing the non-straight segment, and store the captured picture containing the non-straight segment in the hard disk of the robot; The network bandwidth of the current inspection robot is obtained, and when the network bandwidth reaches a threshold, the image containing the non-straight line segment in the hard disk is uploaded to the cloud; The cloud calls the first recognition model and the second recognition model to recognize the picture containing the non-straight line segment, and updates the first recognition model and the second recognition model.

Images