BR102018011661A2 - method and system for mobile thermographic inspection of energy distribution networks - Google Patents

Abstract

the present invention relates to a method for generating preventive diagnostic reports of failures in elements of the electricity distribution network using thermographic and optical images obtained from cameras positioned on a vehicle, thermographic (ct) and optical images (co ) being captured simultaneously by means of a robot that performs the positioning of the cameras.

Description

METHOD AND SYSTEM FOR MOBILE THERMOGRAPHIC INSPECTION OF ENERGY DISTRIBUTION NETWORKS ”

Field of Invention [0001] The present invention refers to a method and system for generating preventive diagnostic reports of failures in elements of the electricity distribution network using thermographic and optical images obtained from cameras positioned on a vehicle, the images thermographic and optical images being captured simultaneously by means of a robot that performs the positioning of the cameras. Background of the Invention [0002] With technological development in the electronic area, the optical and thermographic cameras available on the market provide images with increasingly better resolutions. Obtaining such images can be used for a wide range of applications, including the monitoring of equipment and electrical components during operation.

[0003] Different technologies are currently employed in an attempt to efficiently monitor electrical elements with possible defects. Such elements may present overheating, requiring maintenance, exchange or load shifting in the network.

[0004] Patent document No. US 2017/0150069 A1, published on May 25, 2017, under the title: THERMAL IMAGING BASED MONITORING SYSTEM ”, on behalf of SEEK THERMAL, INC, refers to a system for monitoring any field of view using matrix infrared radiation sensors. After acquisition, the thermal image is

Petition 870180049151, of 06/08/2018, p. 18/52

2/23 segmented into two or more regions, identifying the temperature of these regions from the information provided by the camera manufacturer and generating an alert in case of temperature above a reference. In a different way, the present invention uses optical and thermal cameras for an intelligent system that identifies elements with irregular functioning in different points of the electricity distribution network. The method of the present invention does not use region separation and temperature analysis of those regions. In addition, the system in US 2017/0150069 A1 does not deal with a mobile system. US 2017/0150069 A1 describes a thermographic image monitoring system based on networks with multiple cameras powered by batteries. In this system described in US 2017/0150069 A1, a wireless transmission network (WiFi) transmits the images captured by thermographic cameras to a remote computer. US 2017/0150069 A1 describes the system by which the batteries will control the cameras and refers to a method to save energy from the batteries and increase the life of the batteries, in order to extend the periods between possible recharges. The method of the present invention does not aim to save the energy of the batteries that power thermographic cameras, it does not use a wireless network for transmitting thermographic images. The method of the present invention does not have an algorithm that aims to save battery power; it is not a monitoring system that transmits radio frequency images to a remote computer. The method of the present invention comprises capturing images

Petition 870180049151, of 06/08/2018, p. 19/52

3/23 thermographic and optical images from four thermal cameras and four optical cameras, which are installed on the upper roof of a motor vehicle. These images are transmitted via wires (electrical cables) to a computer inside the vehicle. This computer inside the vehicle extracts characteristics from the standard RGB color images, extracts the characteristics from the images by the Gradient Descent Minimization method, classifies the images by the Convolution Neural Network method, compares it with a vector of characteristics from a stock image ( which are numeric vectors), identifies the elements of the electrical network for which the system was previously trained and automatically produces a report that brings the optical and thermographic photos of all the elements photographed, showing heating points (hot spots) in network elements requiring maintenance. All of this information is contained in the report automatically generated by the method of the present invention. None of these features is part of the description contained in US 2017/0150069 A1.

[0005] CN 104809722 A, published on July 29, 2015, under the title: “ELECTRICAL DEVICE FAULT DIAGNOSIS METHOD BASeD ON INFRARED THERMOGRAPHY”, on behalf of STATE GRID CORP CHINA; SHANDONG ELEC POWER RES INST, refers to a method of analyzing thermal images based on Otsu thresholding, feature extraction (Zernike moment) and training by neural networks. CN 104809722 A describes a method of diagnosing faults in

Petition 870180049151, of 06/08/2018, p. 20/52

4/23 elements of the electrical network based on infrared thermography, where infrared thermography images are subjected to gray level processing and segmented using the OTSU threshold method. The characteristic parameters of the segmented infrared thermography images are extracted to serve as input parameters of a BackPropagation type neural network (abbreviated by BP) and the BP neural network is trained with extracted images. The diagnosis of thermal defect is performed on the electrical device through the trained BP neural network and a diagnostic result is generated. In the method of the present invention, both images of the same element are analyzed: the thermographic image and also the optical image in the RGB standard (without transforming into gray scales). In the method of the present invention, images in color (RGB standard) are analyzed by the computer, unlike the system proposed by the document CN 104809722 A, which works with the images in gray scales. In the system described in document CN 104809722 A, the characteristics of the grayscale images are extracted using the Zernicke Moments method. In one embodiment of the present invention, the characteristics of color images are extracted using the Gradient Descending Minimization (GDM) method, which is part of the image identification technique known in the literature as Deep Learning. In the system described by document CN 104809722 A, the classification of images is made from a Neural Backpropagation Network (BP) and the classification of images from

Petition 870180049151, of 06/08/2018, p. 21/52

5/23 according to the method of the present invention is carried out through a Convolutional Neural Network (CNN = Convolution Neural Network).

[0006] The article made available through the link www.mfap.com.br/pesquisa/arquivos/20081126150533-174.pdf, under the title: ROBOT FOR MONITORING HOT POINTS THROUGH INFRARED CAMERAS IN ELECTRIC ENERGY SUBSTATIONS ”, in The name of MV Garbelotti et al (undated document) of the company TBE - Transmissoras Brasileiras de Energia, deals with the use of a robot to monitor electricity substations using an infrared camera. The robot is controlled remotely (does not have the intelligence to position itself), having a pulse count as a guide. The images are analyzed only to identify temperatures above a pre-set limit. Unlike the present invention, the system by Garbelotti et al is very limited, no processing is done to identify elements, optical cameras are not used, reports are not generated. In the method of the present invention, the camera is dynamically aimed at the objects of interest through an Image Recognition Algorithm based on the technique known in the literature as Deep Learning: the elements are recognized through their shape through a convolutional neural network. In the system described in the document by Garbelotti et al, the entire structure is controlled by a program residing on a PC located at a distance of approximately 1km from the robot. In the method of the present invention, the control of the entire system is located inside the vehicle itself, at a distance of

Petition 870180049151, of 06/08/2018, p. 22/52

6/23 a few meters from the camera, which gives greater immunity to electromagnetic interference in relation to the system described in the article by Garbelotti et al.

[0007] US Patent No. 7,989,769 B2, issued on August 2, 2011, under the title: “IN-CABINET THERMAL MONOTORING METHOD AND SYSTEM”, on behalf of ROCKWELL AUTOMATION TECHNOLOGIES, INC, describes a system for monitoring frames of energy, with low cost sensors and with the identification of a single set of elements. A mechanical system in the frame allows scanning of the elements of the frame (static scanning). Very differently, the method of the present invention uses an intelligent system that identifies elements at different points in the distribution network.

[0008] Therefore, there is a need for a method for monitoring hot spots that is capable of capturing, in a mobile way, optical and thermal images to be processed in order to identify electrical components using pattern recognition algorithms. present on the pole and generate reports on the identified components.

Summary of the Invention [0009] The hot spot monitoring method of the present invention is capable of recognizing the types of electrical elements with and without overheating along overhead distribution lines. The identification of the elements is performed during the displacement of a land vehicle equipped with an array of optical and thermographic cameras and controlled by a robot. The system has a series of

Petition 870180049151, of 06/08/2018, p. 23/52

7/23 algorithms, among them: software for Pan-tilt control, software for shooting control and control of cameras by computer, Post tracking software, software for combining images of elements captured by different cameras, software for recognition of standards that use neural networks (Deep Learning), software to interface with maps and software for generating reports. The present invention also features a specific damping system for vibration compensation.

[0010] Among the functionalities provided by the thermographic inspection method of the present invention, the following stand out:

[1] Capture images of electrical elements along the aerial power distribution line without the presence or intervention of an instrumentalist or thermographer during the route.

[2] Carry out step [1] with the land vehicle in motion;

[3] Image processing for pattern recognition of electrical elements.

[4] Image processing to identify elements with overheating;

[5] Combining images from several cameras to generate a single report per element with images at different angles.

[6] Combination of thermographic and optical images for processing;

[7] Ability to capture and recognize electrical elements simultaneously on two sides of the track;

Petition 870180049151, of 06/08/2018, p. 24/52

8/23 [8] Ability to identify track width variation through the use of a front view camera, image processing and pan-tilt control software for correct positioning of the rear cameras from the front camera information.

[9] Ability to generate an image library of new elements for neural network feedback.

[10] Ability to compensate for temperature measurements and compare the measurement with temperature history of geolocalized electrical elements stored in a database, with the objective of recommending preventive action.

[0011] The method of the present invention then combines items 1 to 10 in order to generate reports of thermographic inspection without the need for intervention by instrumentalists and without the need to stop the vehicle.

[0012] Such objectives and advantages are achieved through a method for mobile thermographic inspection of energy distribution networks that comprises the steps of:

automatically and remotely adjust the position of the optical and thermal cameras from a medium to automatically and remotely adjust the position of the optical and thermal cameras in the vertical and rotational directions (or axes) triggered from a width of the obtained track through a front camera;

capture, with the vehicle in motion, a plurality of optical images and a plurality of thermal images to be

Petition 870180049151, of 06/08/2018, p. 25/52

9/23 starting from pairs of optical cameras and thermal cameras arranged on a support base on a vehicle;

calculate, in real time, the temperature compensation from a database with nominal current values of the electrical elements detected at each point as a function of wind speed and ambient temperature measurements;

perform pole pattern recognition using Deep Learning with the addition of a thermal image layer and perform KCF (Kernelized Correlation Filter) tracking for the post with the largest area;

perform the pre-processing of optical and thermal images;

locate points of interest through a Deep Learning network and select images associated with a defined label;

report whether a particular region or element of the pole is overheating according to its respective overheating threshold.

Brief Description of the Figures [0013] The objectives and advantages of the present invention will become clearer through the following detailed description of a preferred, but not limiting embodiment of the invention, in view of the attached figures, in which:

[0014] Figure 1 illustrates an arrangement of optical and thermal cameras installed in pairs on the roof of a vehicle;

[0015] Figure 2 illustrates an example of measuring the distance between gutters through the Hough transform;

Petition 870180049151, of 06/08/2018, p. 26/52

10/23 [0016] Figure 3 illustrates the procedure for converting the image in the Bayer format to the RGB format;

[0017] Figure 4 illustrates the detection of elements contained in the pole using the modified YOLO algorithm;

[0018] Figure 5 illustrates an example of traditional Architecture of Convolutional Neural Networks. The first stages are composed of layers of convolution and pooling and the last layers are completely connected;

[0019] Figure 6 illustrates an application of max pooling in a 4x4 image using a 2x2 filter;

[0020] Figure 7 shows a convolutional network for the RGB + T image;

[0021] Figure 8 shows a pole detection using the modified YOLO algorithm;

[0022] Figure 9 represents an image of an electric pole obtained by the optical camera, image resolution is 1024 x 643 (rescheduled from 1936 x 1216);

[0023] Figure 10 illustrates an image of an electric pole using the thermal camera;

[0024] Figure 11 illustrates a thermal image converted to the scale of the optical image, now with 1024 x 643 resolution (before it was 640 x 480);

[0025] Figure 12 illustrates a rescheduled thermal image superimposed on the optical image;

[0026] Figure 13 illustrates a Gaussian Difference of the optical image;

[0027] Figure 14 illustrates a Gaussian Difference of the thermal image;

Petition 870180049151, of 06/08/2018, p. 27/52

11/23 [0028] Figure 15 illustrates the edges of the thermal image using the initial projection matrix;

[0029] Figure 16 illustrates the superposition of the projection of the thermal image with the optical image;

[0030] Figure 17 illustrates a distance between the center of the pole and the center of the optical image;

[0031] Figure 18 shows a flowchart of the pan-tilt tracking steps.

Detailed Description of the Invention

Optical and thermal cameras [0032] As shown in figure 1, the camera system consists of a front camera (CF) and 4 pairs of optical (CO) and thermal (CT) cameras which are: right front (FD), front left (FE), right rear (TD) and left rear (TE).

[0033] According to one embodiment, optical CO cameras record images in Bayer format with a fixed value of 2.4 MBytes per image with a resolution of 1936 x 1216 pixels.

[0034] However, it is evident for a person versed in the technique, the use of other formats and resolutions.

Support Base for Optical and Thermal Cameras [0035] A mechanical base is provided to allow the fixation of CO optical and CT thermal camera arrangements, together with so-called pan-tilts, which are equipment that allow immunity to terrain irregularities and also the positioning control of the cameras in the vertical and rotational directions. The objectives achieved by the base are listed below:

Petition 870180049151, of 06/08/2018, p. 28/52

12/23

- Design of a flexible mechanical base that allows arrangements for CO optical and CT thermal cameras during the project without the need for re-fabrication;

- Obtaining a base with reduced weight;

- Base with damping preventing excessive vibration of the cameras;

- Support the installation of Pan-tilts;

- Possibility of controlling the movement of all cameras by software;

- Possibility that intelligent algorithms take control of the cameras and make it possible to reduce the number of cameras; and

- Possibility of arrangement with a smaller number of cameras.

Automatic Height Compensation of the Pan-Tilt System [0036] All cameras in this image acquisition system, both optical and thermal, are controlled by a Pantilt positioning system. It is a mechanical-robotic system with 2 degrees of freedom that allows the cameras to move in vertical and horizontal directions via software.

[0037] Pan's movement is defined in 360 degrees in the horizontal plane in relation to the ground, and the Tilt movement is defined as being 180 degrees in the vertical plane, orthogonal to the plane of the Pan movement. As the streets become narrower, there is a need to increase the height of the cameras so that the areas of interest, where the elements of the power grid are located, continue to be photographed properly.

Petition 870180049151, of 06/08/2018, p. 29/52

13/23 [0038] A front camera is then used to measure the width of the streets, in order to measure the narrowing or widening of the road to automatically correct the position of the engines of the PanTilts systems. The principle of the proposed algorithm is to apply the Hough Transform to measure the distance (fixing 2 fixed points on the image) between the gutters of the street guides, as shown in Figure 2. The edges (guides) of the streets are straight segments .

[0039] When measuring a new distance between the road guides, the system automatically corrects the Tilt position, changing the height of the cameras according to the greater or lesser proximity of the vehicle that captures the images in relation to the posts.

[0040] The following script is then applied in order to obtain the width of the road.

1. Reading the image of the vehicle's front camera; 2. Binarization of the captured image using half gives range intensity maximum image; 3. Application of the Sobel filter, in order to if highlight the regions contour of the image; 4. Application of the Hough in the picture

resulting from the Sobel filter application step.

5. Application of the Distance Transform using the main lines obtained from the previous step.

6. Variations in time measurements from step 5 will be provided for Pan-tilt control.

Temperature compensation from environmental conditions

Petition 870180049151, of 06/08/2018, p. 30/52

14/23 [0041] Before the pattern recognition application step, temperature correction must be carried out, according to GED 3485, depending on the wind and the line loading. For this, it is necessary a database containing the nominal values of the working current of the elements of each point. The measurement of wind speed and ambient temperature, at the time of measurement, must be obtained in real time.

[0042] Initially, the current temperature rise is calculated, which is the difference between the temperature the material is in and the room temperature:

Ati = Tmat - Tamb (1)

Where:

At ₁ is the temperature rise measured with the current load

T _mat is the temperature of the measurement point (material)

It is also the ambient temperature at the time the inspection was made.

[0043] The temperature rise due to loading can be calculated by:

Át2 = Ati / Loading ^ ² i 100% \ ² (Loadingi) = ^ ^ti '(loading% i) = Ati.FCC (2) [0044] For air speed, the correction is:

At2 = Ati ·

Speed

Speed2 /

0.448 = .FCVA (3)

Where:

Speed1 is the airspeed at the time of measurement

Speed ₂ is the air speed for which you want to find the temperature rise, usually for an air speed of 1 m / s.

Petition 870180049151, of 06/08/2018, p. 31/52

15/23 [0045] Once the corrected temperature variation is obtained, exemplary actions at each temperature level that should be used are illustrated in Table 1 below:

Priority Temperature range compared to simulated components under the same conditions (° C) Difference when compared to room temperature (° C) Action Recommended 4 1-3 1-10 Possible disability, search for occurrence 3 4-15 11-20 Indicates disability, repair when time permits 2 21-40 Monitor to what extent corrective measures can be carried out 1 > 15 > 40 Large discrepancy; Immediate repair

TABLE 1

1. Hot spot identification

1.1 - Pre-processing

Petition 870180049151, of 06/08/2018, p. 32/52

16/23 [0046] As illustrated in figure 3, a conversion from a Bayer format to RGB is carried out, in accordance with an embodiment of the present invention. An image in Bayer format is divided into 4x4 sub-images and, from this, a pixel in RGB format is calculated.

[0047] Additionally, the image is rescheduled to 1024x643 pixels, keeping the original image format, and a JPG compression will also be performed to reduce the storage space. With this conversion, the image will have an average size of 135.31 KBytes (18 times smaller than the original image).

[0048] Once the rear cameras receive information about the position of the pole, the pole tracking algorithm (for the optical image only) will be activated for Pan-tilt control and the regions of interest around the pole will be extracted. post that will be the entrance to the new deep learning algorithm. However, in this case, the elements that belong to the pole are identified, such as knife keys, fuse keys, transformer, etc.

[0049] Figure 4 shows an example of detecting the elements contained in the pole using the modified YOLO algorithm. This procedure is more accurate than when applied to the entire image, so the image of the pole is rescheduled to a resolution of 448 x 448, so that it is the input of the convolutional network algorithm.

1.2 - Hot spot detection [0050] The hot spot identification is based on two types of images: thermal and optical. The images

Petition 870180049151, of 06/08/2018, p. 33/52

17/23 thermal units are made up of pseudo-colors and are accompanied by data provided by the equipment manufacturer, which reveal an estimate of the temperature of the objects present. The optical image is obtained from a conventional camera, has no additional information, but allows a larger amount of pixels when compared to the thermal image.

1.2.1. Location of points of interest [0051] The initial phase consists of finding an image that has the objects of interest: posts and network elements. The premise is to classify an image by obtaining its characteristics. For this, it is first necessary to detect local characteristics and then group them into a global representation within the image. Finally, train a classifier to be able to associate a label to the image. So, if we want to classify a new image, this classifier will know which tag should be related.

[0052] The detection of posts will be carried out using a Convolutional Neuronal Network, also called deep learning or Deep Learning. Currently, a typical architecture of a Convolutional Neuronal Network is divided into a series of stages. The first stages are composed of two types of layers, the convolution layers and the pooling layers, as shown in Figure 5. The convolution layer consists of attribute maps, connected to each unit of the previous layer through a set of parameters shared among all units and has ReLUs (Rectified Linear Units), neurons with activation function defined as non-linearity in the form described

Petition 870180049151, of 06/08/2018, p. 34/52

18/23 in the following equation applied to the output of each convolutional layer:

f (x) = max (0, x) (4) [0053] Pooling layers are a form of downsampling. A typical pooling layer computes the maximum location of a given region of the attribute map, as can be seen in Figure 6. They are useful for eliminating non-maximum values, reducing the dimension of the data representation and consequently the computation necessary for the next ones. layers, besides creating an invariance to small local changes and distortions. Two or three stages of convolution, nonlinearity and pooling are stacked, followed by more layers of convolution and completely connected layers. The convolutional and pooling layers are directly inspired by classic notions of simple cells and complex cells in visual neuroscience.

[0054] The procedure performed with the two pairs of right cameras is the same performed on the two pairs of left cameras, with this it will only be explained for the right side. The pair of FD cameras will have the task of detecting the pole, being responsible for notifying the pair of right rear cameras that a pole has been detected. To perform the pole detection, a modification of the architecture of the deep neural network used in the YOLO algorithm will be used, adding the thermal image layer. For this, the thermal image with 640x480 resolution will be rescheduled to 1024x643 and thus build an RGB + T image with 1024x643 pixel resolution with 4 color layers, as shown in Figure 7.

[0055] This network reschedules the image to a square matrix of 448 x 448 x 4 resolution

Petition 870180049151, of 06/08/2018, p. 35/52

19/23 and is formed by 24 convolutional layers and, in the end, by two completely connected layers. This algorithm has a high computational cost, which is why it will be complemented with the KCF (Kernelized Correlation Filter) tracking algorithm. Once one or more posts are detected, the tracking algorithm for the post with the largest area will be triggered and Pan-tilt will follow the center of the post area. If for 5 detections (using the modified YOLO algorithm) in 10ms intervals, the pole region coincides in at least 90% of the region area with the KCF tracking algorithm, then we will confirm that the detected object will be a pole, informing then to the pair of right rear cameras, according to the detection example in Figure 8.

1.2.2. Segmentation and recognition of the electrical network elements [0056] Once the rear cameras receive information about the position of the pole, the pole tracking algorithm (for the optical image only) will be activated to control the Pan-tilt and will be extracted the regions of interest around the pole that will be the entrance to the new deep learning algorithm (identical to the one shown in Figure 8). However, in this case, the elements that belong to the pole will be identified, for example, knife keys, fuse keys, transformer, etc.

[0057] To identify the temperature of the elements in the optical image, it will be necessary to calculate the projection matrix between the optical and the thermal image. In Figure 9 the image obtained by the Dalsa Genie Nano C1920 optical camera is presented, this image was converted from the Bayer standard of this camera.

Petition 870180049151, of 06/08/2018, p. 36/52

20/23 [0058] Figure 10 shows the thermal image, which was converted from 16 bits and uses a color space to represent its temperatures (using 16 bits it is not possible to see hot regions).

[0059] As the cameras are fixed to the car (see Figure 1), the projection matrix between them is invariant to scale and rotation, so only their translation parameters will be changed.

[0060] First using the projection matrix we convert the thermal image into the optical image space, as shown in Figure 11.

[0061] Performing the superposition of the optical images and the projection of the thermal, an example of the result obtained can be seen in Figure 12.

[0062] To calculate the parameters of the homography matrix, the Gaussian differences of the optical and thermal images are calculated to obtain the edges of the image, as shown in Figures 14 and 15. The Gaussian differences work by performing two different Gaussian blurs in the image, with a different radius of defocus in each one, and subtracting them to obtain the result. This algorithm is widely used in artificial vision and is very fast when compared to other techniques for the same purpose, as they use very efficient methods to make Gaussian blur. The most important parameters are the smoothing radii. This method is the most suitable to obtain the edges in the thermal image in comparison to the traditional Canny or Sobel algorithms. An I (x, y) image is converted to L (x, y, o), using the Gaussian with standard deviation σ. This new image is produced by the convolution of a Gaussian function, G (x, y ^), with the image I (x, y).

Petition 870180049151, of 06/08/2018, p. 37/52

21/23

L (x, y, o) = G (x, y, o) * I (x, y)

Where:

G (x, y, o) = / (x ² + y ² ) / I ^ e'l M

2πσ ² (5) (6) [0063] The Gaussian difference function is defined as:

D (x, y ^) = L (x, y, l ^) - L (x, y ^) (7) where 'k' is the scale of the Gaussian difference (in the space of scales). For thermal images, (^ σ) = (5, 1.6) and for optical images, a (^ σ) = (6, 1.6) is used.

[0064] To calculate the translation, the edges of the thermal image will be converted using the homography matrix and then the translation parameters will be modified in a small neighborhood, where the largest amount of pixel of the logical operator and between the images will be the translation of the matrix .

[0065] Figure 15 shows the result of the projection of the edges of the thermal image using the initial projection matrix, now it will have a resolution of 1920 x 1200 pixels.

[0066] Figure 16 shows the conversion of the thermal image using the homography matrix with the modified translation parameters superimposed with the optical image. In this image it is already possible to obtain the temperatures of the elements.

1.3 - Pan-tilt control [0067] According to an exemplary embodiment, the Pan-tilt control is performed using the PELCO D protocol, which is made up of 7 bytes, in the following format (values in hexadecimal):

byte 1: initial synchronization, FF.

byte 2: camera address (camera 1 has address 01).

Petition 870180049151, of 06/08/2018, p. 38/52

22/23 byte 3 and 4: commandol and command2 as described below.

byte 5: date 1, speed for PAN, ranging from 00 stopped to 3F - high speed and FF - turbo.

byte 6: date 2, speed for TILT, ranging from 00 stopped to 3F - high speed.

byte 7: checksum, sum of bytes 2 to 6 and then module 100.

[0068] Bytes 3 (commandol) and 4 (command2) encode the actual desired controls. For example, when command2 has the value of 0000.0100 binary (or 04 hexa), the addressed camera must perform a PAN operation on the left; if it has a value of 0000.0010 binary, the operation will be PAN on the right; if it has a value of 1000.0000, the focus will be adjusted away.

[0069] Using the pole detection and tracking algorithms, the position of the center of the pole (C _x , C _y ) in the optical image is obtained, as shown in figure 17. The difference in pixels is obtained with reference to the center of the image (Px, Py) and if these points are close together then it is not necessary to rotate the Pan-tilt. As shown in figure 18, in case this difference is considerable, it will be necessary to rotate the Pan-tilt to the center of the image, so the center of the pole will always be close to the center of the image. The combination of the two types of images allows a refinement of the technique recognition, allowing accuracy rates greater than 0.9.

1.4 Creation of the inspection report [0070] With the previous steps it is possible to know if the region of the element contained in the pole has overheating or not, and each element has its own threshold of overheating. So if, for example,

Petition 870180049151, of 06/08/2018, p. 39/52

23/23 a hot region is detected in the rectangular region of the knife wrench, it is possible to identify whether the irregularity occurs in the upper or lower part of the element. This procedure is specified for each possible element found on the pole.

[0071] For the inspection report, a good quality image with less vibration effect will be chosen.

[0072] Although the present invention has been described in connection with a preferred embodiment, it should be understood that it is not intended to limit the invention to that particular embodiment. On the contrary, it is intended to cover all possible alternatives, modifications and equivalents within the spirit and scope of the invention, as defined by the appended claims.

Claims (9)

1. Method for mobile thermographic inspection of energy distribution networks characterized by the fact that it comprises the steps of: automatically and remotely adjust the position of at least one optical camera (CO) and at least one thermal camera (CT) from a medium to automatically and remotely adjust the position of the optical and thermal cameras in the vertical and rotational directions (or axes) driven from a width (L) of the track obtained through a front camera (CF); capture, with the vehicle in motion, a plurality of optical images and a plurality of thermal images from a pair of optical camera (CO) and thermal camera (CT); calculate, in real time, the temperature compensation from a database with nominal current values of the electrical elements detected at each point as a function of wind speed and ambient temperature measurements; perform post pattern recognition using a Convolutional neural network with the addition of a thermal image layer and perform the KCF tracking for the post with the largest area; perform the pre-processing of optical and thermal images; locate points of interest using a Convolutional neural network with the addition of a thermal image layer with synchronism of types of images established by a Gaussian differences algorithm; Petition 870180049151, of 06/08/2018, p. 41/52 2/3 report if a specific region or element of the pole has overheating according to its respective overheating threshold. 2. Method according to claim 1, characterized by the fact that the information from the front camera is a distance obtained from the Hough transform of two fixed points in the image of the front camera (CF). 3. Method, according to claim 1, characterized by the fact that the step of performing pole pattern recognition is to apply a deep learning method. 4. Method, according to claim 1, characterized by the fact that the detection of elements of the electricity distribution network is performed through the deep learning method. 5. Method, according to claim 1, characterized by the fact that the pre-processing step consists of: convert from a Bayer format to RGB; reschedule the image keeping O image format original; and compress the Image rescheduled in JPG. 6. Method, in a deal with The claim 1, characterized by the fact that the stage of locating points of interest comprises identifying elements such as Knife Switch, Fuse Switch, Oil Switch, Connectors, Feeders, Transformers, Posts, Lightning Rods. Petition 870180049151, of 06/08/2018, p. 42/52 3/3 7. Method, according to claim 1, characterized by the fact that the classification step also includes the use of non-linear classifiers. 8. Method, according to claim 1, characterized by the fact that the pair of optical camera (CO) and thermal camera (CT) is arranged on a support base fixed to a vehicle. 9. Method, according to claim 8, characterized by the fact that the stage of capturing images uses a vehicle such as an automobile, drone, robot, among others. 10. System for mobile thermographic inspection of energy distribution networks characterized by the fact that it comprises: four pairs of cameras formed by optical cameras (CO) and thermal cameras (CT) arranged on a support base on a vehicle; a central processing unit capable of performing calculations and real-time comparisons of values measured by optical (CO) and thermal (CT) cameras; a database storing parametric definitions and collected data; wherein the system is adapted to carry out the steps of the method as defined in any of claims 1 to 9.