Method and system for the early detection of leaks in collector tubes of a thermoelectric solar plant

Abstract

The present invention relates to a method and system for the early detection of leaks in collector tubes of a thermoelectric solar plant that includes a preliminary phase of placing a series of video cameras (6), distributed throughout the plant to obtain a general floor plan or a floor plan of a specific area (7) of the plant. In the operational phase, the images captured by each camera, together with a temporal dimension of said images, are processed in a leak detection module equipped with a deep learning algorithm, which processes a temporal sequence of images to determine the existence in said plant of both local leaks and leaks affecting a large area of the plant. The invention further discloses a method for geolocating each leak observed.

Description

[0001] DESCRIPTION

[0002] Method and system for early detection of leaks in the collector tubes of a solar thermal power plant

[0003] Technical sector

[0004] The field of this invention is the physical testing of static or dynamic equilibrium of machines or structures; these tests are intended to investigate the fluid tightness of certain structures.

[0005] The overall objective of the system is to detect, online, potential leaks in ball joints, as well as other potential failures in parts containing heat transfer fluid (HTF) in a solar thermal power plant (CSP), using images from cameras installed throughout the plant. Furthermore, these alarms are geolocated, allowing the plant to be located on a map. The system stores these alarms in a database, which can be used for viewing and confirmation.

[0006] The system is composed of different modules that together allow leak detection and location within the solar thermal power plant (CSP).

[0007] The leak detection method is based on at least one working module that is equipped with an AI algorithm (specifically, a Deep Learning algorithm) capable of detecting leaks in real time.

[0008] State of the art

[0009] In recent years, a large-scale deployment of parabolic trough solar thermal installations has been underway, leading to the development of various scientific studies on the issue of HTF (Heat Transfer Fluid) leaks. This aims to prevent or detect them as quickly as possible, minimizing environmental damage and potentially impacting plant operation, with the resulting financial costs. Some prototypes have also been developed and even commercialized; however, all of this remains in an initial phase, leaving a long path to improvement, something this project aims to contribute to.Before analyzing the current state of the art regarding HTF leak detection, it is worth first recalling the origin of the problem: A parabolic trough solar thermal plant consists of a set of mirrors, with this geometric shape, that follow the sun's path as much as possible throughout the day, to maximize the flow of incident solar energy. Following the rules of optics, this energy is reflected by the mirror and concentrated at the focus of the device, which contains a tube filled with a fluid (HTF) with the appropriate thermal properties, which is heated by the incidence of solar energy. This fluid transmits the captured thermal energy, which is subsequently converted into electrical energy through conventional mechanisms available at the plant.It should be noted that this process is a completely closed system, so there should be no leakage or contact of HTF fluid with the outside world of any kind. However, the pipes and pipe connections can deteriorate to the point where vapors or even liquid drips, i.e., HTF leaks. These spills, both in their gaseous and especially in their liquid form, are highly polluting, particularly when they cool and fall to the ground. Furthermore, when they escape from the pipe at high temperature and pressure and suddenly expand, they vaporize, significantly lowering their autoignition temperature. This sometimes causes fires when they come into contact with hot surfaces or when they pass through the beam of sunlight concentrated by the parabolic mirrors.

[0010] Some of the studies and solutions that have been proposed so far to reduce the serious problem of HTF leaks within the operation and maintenance process of CSP, consist of motorized robots with thermal cameras, which according to an experimental study carried out by Ibarguren et al. (2013), consists of a motorized robotic device that carries a thermal camera along the collector plant of the solar plant, taking thermal images of the collectors, pipes and joints between them. Or, in a system based on the previous design, which performs a complete analysis with certain improvements related to the capacity to manipulate the robots, the analysis time to detect a leak, or the optimization in terms of the number of robots that allow leaks to be detected in a time of about 15 min, etc.

[0011] The company: SST-SYSTEM (Smart Sensor Technology) has announced its FENIXST HTF leak detection system, which it has implemented at the La Dehesa CSP plant (Badajoz). It consists of a network of 1,000 wireless sensors installed in ball joints to detect HTF leaks. The system is connected to the plant's SCADA system, allowing specific valves to be opened or closed, thereby preventing liquid spills.

[0012] There are also notable references in the patent literature; for example, CN112131993(A) describes a system and method for visual gas leak detection based on image recognition technology, wherein the system comprises a detection module provided with a gas-sensitive color-changing material, a storage device for gas to be detected, image acquisition devices, and a processor, which is used to identify the color change of the gas-sensitive color-changing material.

[0013] EP3270134(A1) describes a gas leak location estimation device including: an information processor that acquires image information in a plurality of frames from an infrared camera, which is capable of executing: block discrimination processing for dividing an image area of ​​each of the frames into a plurality of blocks and discriminating whether each of those blocks is a gas area; counting processing comprising counting a number of times that each of the blocks is discriminated as a gas area, over the frames in chronological order; and leak location estimation processing for setting among the blocks a block whose counted value obtained in the counting processing is equal to or greater than a predetermined value as an estimated gas leak location.

[0014] CN111696320(A) relates to an intelligent visual gas leak early warning system, comprising a detection area, a data collection and processing module, and a central control room, wherein the data collection and processing module is responsible for real-time storage of monitoring videos and leak detection and positioning, and the central control room is responsible for displaying and early warning of organic gas leaks. The organic gas leak detection positioning module adopted by this invention belongs to a hybrid deep learning model.

[0015] Therefore, there is still a need to develop alternative technologies that, using image recognition systems, can distinguish between serious and minor leaks and also allow for their precise location, allowing for prompt intervention by the maintenance services of the solar thermal power plant. Explanation of the invention

[0016] In order to achieve the proposed objectives, mentioned in the previous section, the invention proposes a method for early detection of leaks in a solar thermoelectric plant, which has the characteristics of claim 1.

[0017] This method for early detection of leaks in the collector tubes of a solar thermal power plant involves the following phases:

[0018] A preliminary phase involves the placement of a series of video cameras, distributed throughout the solar thermal power plant, capable of moving at a specific angle and zooming, in order to obtain either a general view of the plant or a specific area of ​​the plant; among these specific areas, at least those most susceptible to leaks are selected, such as the ball-joints of the absorption tube or the plant's piping system.

[0019] An operational phase in which the images collected by each camera, together with a temporal dimension of the same, are sent to a video recorder that receives all the video streams from all the cameras, which, by means of at least one leak detection module, equipped with a deep learning algorithm, processes a temporal sequence of images to determine the existence of leaks in the plant.

[0020] The leak detection module employs two deep learning algorithms: one for detecting serious leaks, which respectively affect a large area of ​​the plant, and another for local leaks, considered less serious. The deep learning algorithm that determines the existence of serious leaks, which affect a large area of ​​the plant, is trained with image sequences captured by at least one of the cameras, which performs a general focus on the entire plant or a large area; while the deep learning algorithm that determines the existence of minor leaks is trained with image sequences captured by multiple cameras, which focus on specific areas or objects of interest, as they are more susceptible to leaks.To focus its search, the deep learning algorithm that determines the existence of soft leaks also uses a deep learning algorithm to detect ball-joints as a preliminary step, since these are the points where leaks most commonly occur. This detection algorithm is trained with images captured by various cameras focused on specific areas or objects of interest, and returns the area(s) in the image where each ball joint or group of balls is located. The algorithm that determines the existence of soft leaks only processes information related to those areas in the image.

[0021] The cameras that focus on the general shot and the specific area may be the same, in which case they change focus, switching from a general shot to a specific shot and vice versa, in a specific sequence or time period. The cameras that focus on the general shot and the specific area may also be different cameras, permanently focused on a general shot or a specific shot.

[0022] According to an important feature of the invention, along with the camera images, a series of labels relating to the location where each image was captured and / or the object of interest are transmitted to the leak detection module. This is done in order to geolocate the alarms, so that they can be clearly located within a map of the plant. In addition to the camera location, labels are transmitted with the orientation, zoom level, and calibration of each camera, along with the timestamp of each sequence, in order to determine the position of the leak using projective geometry techniques.

[0023] As indicated above, the leak detection method of the invention is based on at least one operating module, equipped with a Deep Learning algorithm. The training model is provided with sequences of RGB images captured by cameras (tensors -matrices- of pixels, with their 3 color channels and channels for the different time instants of the temporal volume considered), as well as the areas in said images where the different hinges are located and an identifying label for each hinge. The training model comprises a neural network with a set of spatiotemporal convolutional layers (trainable filters), followed by activation and upscaling layers; the network output consists of a classifier head composed of fully connected layers of neurons.The training algorithm relies on gradient descent to determine the network weights based on the provided examples, minimizing the cross-entropy between the prediction and the actual labels.

[0024] Another objective of the present invention is to provide a system for detecting, online, potential leaks in ball-joints and other points of a solar thermal power plant (CSP), as well as other possible faults (broken mirrors, vacuum leaks in tubes, etc.) using images from cameras strategically located throughout the plant. This system allows these alarms to be geolocated so that they can be clearly located within a map of the plant. Furthermore, the proposed system stores these alarms in a database, which can be used for viewing and confirming them.

[0025] This early detection system for leaks in heat collector tubes of solar thermal power plants is composed of different modules that together allow leak detection and their location within the CSP plant. It meets the characteristics of claim 7.

[0026] Description of the drawings

[0027] To complement the description being made and in order to facilitate the understanding of the characteristics of the invention, a set of drawings is attached to this specification in which, for illustrative and non-limiting purposes, the following has been represented:

[0028] Fig. 1 shows a functional block diagram detailing each of the modules that form part of the system of the invention, which will be described below in the following paragraphs of this document.

[0029] Carrying out the invention

[0030] As can be seen in Fig. 1, the system of the invention comprises:

[0031] - A set of dome-type cameras (6), mounted on staffs, that are installed on the plant on which the system is intended to be deployed.

[0032] - A sensor interface (1), which allows data to be captured from the different cameras and prepared for use by the rest of the system modules. It also implements the application programming interface (API) for controlling these cameras, allowing them to be activated and / or oriented as desired.

[0033] - A leak detection module (2) through images, which makes up the work module for operating the system, equipped with an AI algorithm (specifically Deep Learning) capable of detecting leaks in real time.

[0034] - A leak location module (3), which determines, given a possible leak detection in the image, where in the solar plant the leak is located. To do this, it uses the known geometric information of the camera (its position, orientation and zoom) and its calibration.

[0035] - An orchestrator module (4) responsible for coordinating the previous modules to carry out the detection operation. This operation consists of a systematic scan of the plant according to a predefined pattern. To do this, the orchestrator module (4) sends commands to the cameras to position themselves in a predefined direction and zoom to observe a general or specific area (7), a ball joint or a set of ball joints. It then orders the detection for that area. If something is detected, the corresponding alarm is generated. It will then move to the next area and repeat the cycle.

[0036] - A data storage module (5), responsible for storing the results of the leak detection system, which can be integrated into the CSP operation and maintenance systems. Alternatively, this information could be presented in an ad hoc application exclusively for leak detection.

[0037] The arrangement of the cameras (6) is carried out taking into account their spatial resolution, considering parameters such as position, angle with respect to the horizontal, focal length and resolution, so that using the resolution and focal length parameters of the selected camera, the system is capable of having a distance of at least 18 pixels between the ball joints visualized by the camera in the most unfavorable situation (180 meters away). In a preferred embodiment, the specifications of the cameras are as follows:

[0038] DOME Type

[0039] IP network communication

[0040] Waterproof for outdoor use. IP67 resistance rating.

[0041] - Max. 25 / 30fps @ 4K

[0042] - 40x optical zoom

[0043] Focal Length 5.6mm to 223mm

[0044] - Maximum aperture F1.4 to F4.8 - Angle of view 1-1:63, 9 e to 2 e ;v:37,3 e a1 ,1 e ;D:71 ,2 e to 2.3 e

[0045] - ONVIF Protocol Compatibility

[0046] The sensor interface module (1), which allows the capture of data from the different cameras and prepares them for use by the rest of the system modules, also includes a camera interface module, which implements an interface with the IP cameras directly and / or a LAN video recorder installed in the rooms of the control building, a video recorder that receives all the video streams from all the cameras.

[0047] This interface has a video input and image output that includes the necessary information in each case for the operation of the rest of the modules: a JSON header or similar, with a capture timestamp, a camera ID, its position and orientation, the zoom level and the camera calibration are factors included in a preferred implementation.

[0048] This sensor interface (1) also includes a camera control module, which allows the cameras to be commanded to move in a desired orientation and apply a specific zoom.

[0049] The image leak detection module (2) constitutes the working module for operating the system; it is equipped with an AI (Deep Learning) algorithm capable of detecting leaks in real time, since this type of algorithm uses logical structures that are more similar to the organization of the mammalian nervous system, having layers of processing units (artificial neurons) that specialize in detecting certain characteristics existing in the perceived objects, artificial vision being one of the areas where Deep Learning provides a considerable improvement compared to more traditional algorithms.

[0050] The type of data provided to the model for training are sequences of RGB images captured by the cameras (tensors - matrices - of pixels, with their 3 color channels and channels for the different time instants of the temporal volume considered), as well as the labels with the desired outputs. These labels correspond to the bounding box (location of the image) of the object of interest (ball joints) and the failure condition or not (leak or no leak, with its severity level (mild, severe). This neural network is composed of a set of convolutional layers (trainable filters), followed by activation and upscaling layers; the network output consists of a classifier head composed of fully connected layers of neurons. The training algorithm is based on gradient descent to determine the network weights based on the examples provided, minimizing the cross-entropy between the prediction and the actual labels.

[0051] This detection module (2) constitutes the main element of the system and includes an interface in which images of the plant under study are provided as inputs and an information report with the positive / negative leak and its location in the image is provided as outputs.

[0052] The leak location module (3) determines, given a possible leak detection in the image, where in the solar field it is located. To do this, it uses the known geometric information from the camera (its position, orientation, and zoom) and its calibration; and as output, it provides the most probable position of the leak in the plant, indicating an uncertainty zone.

[0053] The orchestrator module (4) is responsible for coordinating the previous modules to carry out the detection operation. This operation consists of carrying out a systematic scan of the plant according to a predefined pattern. To do this, the module sends commands to the cameras (6) to position themselves in a predefined direction and zoom to observe a general or specific area (7) (a ball joint or set of ball joints). It then orders the detection for that area. If something is detected, the corresponding alarm is generated. It will then move to the next area and repeat the cycle. Sequentially with this systematic scan, or at the same time as it, the orchestrator module will carry out a detection operation on the entire plant, or on a large area of ​​​​it that allows serious leaks in the installation to be detected.

[0054] The system also includes a data storage module (5) responsible for storing the results of the leak detection system so that it can be integrated into the CSP's operation and maintenance systems. Alternatively, said information could be presented in an ad hoc application exclusively for leak detection. It is hereby stated that the materials, shape, size and arrangement of the elements described may be modified, provided that this does not imply an alteration of the essential characteristics of the invention claimed below:

Claims

CLAIMS 1.- Method for early detection of leaks in the collector tubes of a solar thermal power plant, which includes: - a preliminary phase of positioning a series of video cameras (6), distributed throughout the plant, to obtain a general view and / or a specific area (7) of the plant; and, - an operational phase in which the images collected by each camera, together with a time dimension thereof, are sent to a video recorder that receives all the video streams from all the cameras and feeds them into a leak detection module equipped with a deep learning algorithm, which processes a time sequence of images to determine the existence of leaks in the plant; in the operational phase, the leak detection module employs two deep learning algorithms: - a serious leak detection algorithm, affecting a large area of ​​the plant, which is trained with image sequences captured by at least one of the cameras using a general approach; and - an algorithm for detecting local leaks, considered to be of lesser severity; which is trained with sequences of images captured by several cameras that focus on specific areas (7); where, together with the images from the cameras (6), a series of labels relative to the location where each image was captured and / or the object of interest are transmitted to the leak detection module, in order to clearly geolocate the alarms; in this method, a systematic sweep of the plant is carried out during which each of the specific areas (7) are sequentially covered, proceeding to carry out the detection process for a specific area (7) and generating the corresponding alarm if something abnormal is detected, before moving on to the next; and, sequentially with this systematic sweep, or at the same time as it, a detection operation is carried out over a large area or the entire plant. 2.- Method according to claim 1, in which in its operational phase data relating to leak detection results are stored. 3.- Method, according to any of the previous claims, in which for each camera, in addition to its location, labels are transmitted with the orientation, zoom level and calibration, together with the time stamp of each sequence, in order to determine the position of the leak using projective geometry techniques.

4. Method according to any of the preceding claims, wherein in order to focus its search, the deep learning algorithm that determines the existence of slight leaks also uses, as a prior stage, a deep learning algorithm for detecting the position of the ball joints or the points where leaks most commonly occur; this algorithm is trained with images captured by various cameras, which focus on specific areas or objects of interest, and returns the area or areas of the image where each ball joint or group of ball joints or objects of interest is located. 5.- Method, according to claim 4, in which the algorithm that determines the existence of light leaks only processes the information referring to the areas of the image in which each ball joint or group of ball joints or objects of interest are located.

6. A method according to any of the preceding claims, wherein the cameras that focus generally and on specific areas are either the same, in which case they change focus, shifting from a general shot to a specific shot and vice versa, in a specific sequence or established period of time; or they are different cameras, which remain focused on a general or local shot. 7.- Early detection system for leaks in the collector tubes of a solar thermal power plant, comprising a set of cameras (6) installed in the plant on which the system is deployed, a sensor interface (1), which allows the capture of data from the different cameras and prepares them for use by the rest of the system modules, which implements a control API for said cameras, in order to activate and / or orient them at will, - a leak detection module (2) through images, which forms the work module for operating the system, equipped with an algorithm generated through deep training, capable of detecting leaks in real time, - a leak location module (3), which determines, given a possible detection of a leak in the image, where in the solar plant said leak is located, - an orchestrator module (4) in charge of coordinating the previous modules to carry out the detection operation by performing a systematic scan of the plant according to a predefined pattern, sending commands to the cameras to position themselves in a predefined direction and zoom to observe a specific area. (7) and the leak detection module (2) to proceed with the detection of leaks in each specific area (7) sequentially, generating the corresponding alarm if it detects something abnormal; sequentially with said systematic sweep it commands the detection operation over a wide area or the entire plant, which allows serious leaks to be detected in the installation. 8.- System, according to claim 7, comprising: a data storage module (5), responsible for storing the results of the leak detection system, capable of being integrated into the plant's operation and maintenance systems. 9.- System according to any of claims 7-8, wherein the sensor interface (1), which allows the capture of data from the different cameras and prepares them for use by the rest of the system modules, comprises: a camera interface module, with direct connection via IP and / or through a video recorder that receives all the video streams from all the cameras and - a camera control module, which allows you to command the cameras to move in a desired orientation and apply a specific zoom. 10.- System according to claim 9, wherein the sensor interface (1), as output, in addition to the images, includes: a header, a time stamp, a camera ID, the position and orientation of the same, the zoom level and the camera calibration. 11.- System according to any of claims 7-10, wherein the image leak detection module (2), as input, receives data from RGB image sequences captured by the cameras (6), with the labels provided by the sensor interface (1), and as output provides a positive / negative leak report and its location in the image. 12.- System according to any of claims 7-11, wherein the leak location module (3) that, given a possible detection of a leak in the image, determines where in the plant it is located, has as input the detection of a leak, data from the camera (6) in which it is located, providing as output the most probable position of the leak in the plant.

Images