Method for aiding the detection of migrating bodies within a fuel assembly

Abstract

The invention relates to a method for assisting in the detection of migrating objects within a fuel assembly of a nuclear power plant, and more particularly on the debris grid of the lower end of said assembly. In this method, at least one camera is directed towards said assembly, and the image stream recorded by said camera is directed to a human-machine interface comprising at least one screen enabling a first operator to view said video image stream. The method is characterized in that it comprises a first step of detecting said migrating objects using an image recognition algorithm, and at a minimum, a second step of alerting said operator if said algorithm has detected the potential presence of at least one migrating object. Figure 3 (for the abstract)

Description

Title of the invention: Method for assisting in the detection of migrating bodies within a fuel assembly

[0001] GENERAL TECHNICAL DOMAIN

[0002] The present invention relates to the field of detecting migrating bodies at the base of fuel assemblies in nuclear power plants.

[0003] It is situated more specifically in the context of underwater television inspections of fuel assemblies.

[0004] CONTEXT OF THE INVENTION AND PRIORITY OF THE TECHNOLOGY

[0005] In the attached [Fig.1] is represented a fuel assembly 1 among the plurality of assemblies constituting the core 2 of a nuclear reactor.

[0006] It comprises two upper end caps 10 and lower end caps 11 between which are the structure of the assembly (guide tubes 12 and grids 13), as well as fuel rods 14.

[0007] The fuel assembly 1 rests on the lower plate of the core 2 by the lower end 11.

[0008] When setting up such assemblies in the core, it is necessary that their positioning in their respective dedicated location 20 gives them a stable hold, in a vertical position with reference to the X-X' axis of the largest dimension of the assembly 1.

[0009] The fuel assembly rests on the lower core plate 2 on four support points, which constitute feet 15a and 15b visible in the attached [Fig.2].

[0010] The two feet 15b are perfectly flat, while the other two are provided with a bore allowing good positioning of the tip 11 when the assembly 1 is put in place in the center.

[0011] The stability of the assembly is obtained when the lower end 11 of the fuel assembly 1 rests on these four feet 15a and 15b, without any foreign body or any asperity interposed between these support points 15a and 15b and the bottom of the core 2.

[0012] It should be noted that in [Fig.2] the anti-debris grid 16 of the lower end 11 is also visible, allowing foreign bodies to be filtered in order to prevent them from rising into the assembly, between the fuel rods.

[0013] During the storage of fuel assemblies 1, foreign bodies, also called migrating bodies CM, may become lodged and attached under the feet 15a and 15b of the lower end fitting 11. These bodies must be removed before the assemblies are introduced into the reactor. Otherwise, they could com- promise stability.

[0014] Document FR-A-3118270 in the name of the present applicant is described a cleaning installation which makes it possible to get rid of migrant bodies which may have lodged under the feet of the lower tip.

[0015] In the context of the present invention, the terms "foreign body" or "migrating body" CM shall be understood interchangeably as any object of any material. They may have various origins commonly found in any industrial site and may be endogenous or exogenous to it.

[0016] By way of non-limitation, this may include debris from welding or machining operations. It may also include small components such as, for example, screws, bolts, washers, springs, metal chips, etc., which may migrate and become fixed on the debris grid 16.

[0017] Fig. 2 illustrates an example of a migrant body CM stopped by the anti-debris grid 16. These CM migrant bodies come to lodge themselves under or on the anti-debris grid 16 where they are blocked depending on the case.

[0018] In a pressurized water reactor, the nuclear fuel that produces heat is contained in zirconium alloy tubes known as fuel rods. These fuel rods are grouped into bundles of 264 rods, held together by a frame closed by end caps. The entire assembly, frame plus rods, forms a "fuel assembly." The core of a reactor in a nuclear power plant consists of 157, 193, or 205 (241 in an EPR) of these fuel assemblies, particularly, but not exclusively, in the case of French nuclear power plants such as those owned by the present applicant.

[0019] The lower end 11 of a fuel assembly 1 rests on the lower plate of the core 2, which is perforated with holes to allow the circulation of cooling water in the vertical direction, from bottom to top. This cooling water, as it circulates in the primary circuit, may become laden with debris, namely migrating bodies CM, as explained above.

[0020] During a nuclear power plant shutdown in which the reactor core is unloaded, the assemblies are extracted for underwater inspection from all angles using cameras to search for various defects, and then returned to the core for the next cycle or placed in a cooling pool. During these video inspections (VIIs), the debris screen 16 is carefully inspected for any potential migrating objects CM.

[0021] This inspection is necessary because a migrating body present on the grid 16 can have several impacts if the assembly 1 is reloaded as is into the core 2.

[0022] Indeed, by clogging part of the filter formed by the anti-debris grid 16, it alters the circulation of the cooling water, can induce vibrations / turbulence bubbling and causing a "fretting" phenomenon, i.e. friction of the pencils 14 on the skeleton of the assembly 1, thus causing premature wear of the latter.

[0023] Furthermore, if a migrating body CM is still present on the anti-debris grid 16 during reloading, it may possibly, during the reloading operations, pass through the anti-debris grid and end up between pencils 14.

[0024] If it is a metallic body, there is a risk of puncturing the jacket of the pencils, so that the core is no longer clean.

[0025] Furthermore, during reloading and in particular when the lower tip 11 of the assembly 1 arrives on the lower core plate 2, a still-present migrating body can apply a deformation to this assembly.

[0026] Potential impacts on the pencils 14 may damage them, or even cause a rupture of the cladding which constitutes the first of the three safety barriers which prevents the dispersion of radioactive products contained in the fuel.

[0027] Furthermore, the possible deformation of assembly 1 due to a migrating body CM during reloading may make future extraction of this assembly from core 2 difficult.

[0028] It is therefore very important, during the television inspections (abbreviated ITV) of the anti-debris grid 16, that no CM migrating body escapes inspection by the operators dedicated to this task.

[0029] However, such an inspection mission is difficult because it requires high vigilance from the operators over a long period. It causes significant visual and attentional fatigue, as it involves very repetitive, monotonous actions, during which the events to be detected are very rare (on the order of one CM migrating body per fifty assemblies inspected, which represents approximately 2.5 hours of video to view).

[0030] The consequence is that CM migrant bodies are sometimes not detected by the operator, despite the fact that in the current protocol, each video is inspected individually by several operators, successively, according to the organization specific to the control center.

[0031] This current television inspection technique is unassisted, so the video feed returned by the inspection camera is raw, with a human operator responsible for detecting CM migrant bodies from the video feed.

[0032] An example of an inspection organization is described below.

[0033] Initially, a first operator is tasked with a "live" inspection during the acquisition of images using a camera at the bottom of the pool. The operator brings the fuel assembly 1 above the camera, then zooms in on the legs 15 and the debris grid 16, which he inspects in detail (between 2 and 3 minutes per as- assembly).

[0034] He notes, in the form of a report, the suspicions of CM migrant bodies and transmits the raw video (without the information on detected migrant bodies) to a second operator who re-examines the entire video in search of CM migrant bodies and also generates his report.

[0035] Finally, the third operator re-examines the video, assisted by the two reports of the two previous operators, and makes a final decision (presence or not of a migrant body to be extracted).

[0036] In practice, there is no specific training on this subject, so operators learn through apprenticeship and experience from real-life cases.

[0037] The present invention aims to improve the reliability of this operation by compensating for operator lapses in vigilance, through assistance in the detection of these migrating bodies. PRESENTATION OF THE INVENTION

[0038] Thus, the present invention relates to a method for assisting in the detection of migrating bodies within a fuel assembly of a nuclear power plant and more particularly on the debris grid of the lower end of said assembly, during which at least one camera is piloted towards said assembly and the stream of images recorded by said at least one camera is directed towards a human-machine interface which includes at least one screen enabling a first operator to view said stream of video images, characterized in that it includes a first step of detection, using an image recognition algorithm, of said migrating bodies, as well as at least a second step of alerting said operator if said algorithm has detected the potential presence of at least one migrating body.

[0039] Thanks to the invention, the operator is provided, in a way, with a "virtual colleague" who can indicate the migrant bodies he detects. It is understood that this facilitates the operator's task, prevents a lapse in vigilance, and helps ensure that no migrant body escapes surveillance.

[0040] According to other advantageous and non-limiting features of this process, taken alone or according to a technically compatible combination of at least two of them:

[0041] - said second alert stage includes the triggering of an audible alert and / or a visual alert on said screen, of the location where said migrant body is potentially placed and / or the automatic generation of an inspection report;

[0042] - when said algorithm has detected the potential presence of at least one body migrant, we store the corresponding video file as well as the detection file(s) performed by said algorithm and we send them to a second operator with a temporal indication, in said video file, of the instant(s) at which one (or more) migrant body(ies) was / were potentially detected, this second operator validating the presence or lack of presence of at least one migrant body;

[0043] - said detection files are transmitted to a third operator who validates, definitively, the presence or lack of presence of at least one migrant body;

[0044] - it includes a machine learning step for detecting migrating bodies when the second operator, respectively the third operator, has validated the absence of a migrant body;

[0045] - an image recognition software of the "network of" type is used convolutional neurons”, which is trained on a learning base consisting of video archives of migrating bodies, which base is potentially enriched each time said recognition software makes a detection error such as a “false positive” or a “false negative”, this error having been validated by the second and / or third operator;

[0046] - said time indication has the form of at least one bar displayed on the line of Time progression of the video.

[0047] Throughout this application, including the claims, the terms "false positive" and "false negative" respectively mean the detection by the software of something that is not a migrating body, and the failure to detect a real migrating body, not seen by the software, even though it is present in the video. DESCRIPTION OF THE FIGURES

[0048] Other features and advantages of the invention will become apparent from the description which will now be given, with reference to the attached drawings, which represent, by way of example but not limitation, different possible embodiments.

[0049] On these drawings:

[0050] [Fig. 1] is a perspective view of a combustible assembly in accordance with the prior art and discussed above;

[0051] [Fig.2] is a simplified perspective and low-angle view of the lower face of the lower end of the assembly of [Fig.1];

[0052] [Fig.3] is a flowchart intended to illustrate the implementation of the process according to the invention;

[0053] [Fig.4] is a schematic front view of a human-machine interface screen used for implementing the invention, and more particularly intended to illustrate how the information visible on this screen is presented. DETAILED DESCRIPTION OF THE INVENTION

[0054] The present invention relates essentially to a method for detecting migrating bodies CM on the anti-debris grid 16 of a lower tip 11 of a combustible assembly 1, which is implemented by a program which processes data from television inspection images (hereinafter ITV).

[0055] Overall, this method makes it possible to analyze in real time and / or in delayed time, the video stream from a camera which is directed towards the fuel assembly 1 in order to detect migrating bodies CM and present them to the operators in charge in different ways.

[0056] The method can be applied to each of the inspection steps. By the expression "a direction-controlled camera", it is understood that said camera is fixed and that it zooms and performs rotations directed towards the assembly.

[0057] In the case of real-time use by the operator manipulating the camera, the method generates an "alert," notably a visual one on the video stream using a detection algorithm, to signal that it has detected "something" at "a certain location," so that the operator pays attention to it and may decide to examine that area more specifically. It is also possible to generate another type of alert, such as an audible alert.

[0058] It can also allow the automatic generation of an inspection report.

[0059] At the end of this step, we then have the raw inspection video, a file which gathers the detections made by the algorithm, and the operator's inspection report.

[0060] This corresponds to step El shown in the attached [Fig.3].

[0061] In the case of delayed uses, for example for an immediate second stage (as soon as the first operator has completed their inspection) of independent inspection by a second operator, as is the case in several nuclear power plants, and for the final stage of reviewing the videos to resolve doubts and decide whether or not to intervene to remove the CM, the video file of the inspection, along with the detection file generated in the previous stage, is provided and used by the method to present the same assistance as to the first operator, but also indicating, in the video's time progress bar, the times at which migrating bodies are suspected by the program of being present. This corresponds to stage E2 of [Fig. 3].

[0062] Finally, a third and final operator, responsible for a final review based on the reports of the first two operators, and with the objective of deciding on the actions to be taken, may also benefit from the algorithm's detections in its review, in particular from the functionality that allows direct access to the relevant times of the video where the migrant body is seen. This corresponds to step E3 of [Fig.3].

[0063] The "doubt clearance" by this last operator is on the critical path of the unit shutdown. The critical path is the entry into a reverse schedule counted from a previously fixed date of a future event - for example, reactor criticality - such that any delay prior to this event will lead to its shift in time, with penalizing consequences for operation (production failure, etc.).

[0064] It follows from the above that the present method thus makes it possible to considerably reduce the time required for video analysis. Indeed, a video image capture for an assembly 1 lasts on average 3 minutes (more precisely between 2 and 5 minutes), to be multiplied by 157 to 205 assemblies (approximately 8 hours for a slice outage).

[0065] Furthermore, the present method allows operators to focus directly on the sections of the video where CM migrant bodies are suspected, without having to go through the entire video to search for them. ALGORITHM

[0066] The method according to the invention is based on an image recognition and CM migrant body detection algorithm which can be made up of a version of "Yolov3" (convolutional neural network type detection algorithm), which is trained on a specific learning base made up of archives of migrant body videos.

[0067] The learning base is created iteratively and following a specific methodology to minimize the human labeling effort.

[0068] Thus, for example, the database can be enriched each time the absence of at least one CM migrating body has been validated by the second and / or third operator (in other words, when the operator has validated that it was a "false positive"), and each time a CM has not been automatically detected. In this way, the algorithm learns, in a sense, from its errors. However, it is also possible to enrich the database each time the presence of a CM migrating body is validated.

[0069] Certain modifications can be made to the algorithm to take advantage of the specificities of video images of fuel assemblies compared to natural images, in particular with regard to the geometric transformations of the data augmentation which includes top-bottom symmetries, left-right symmetries, isotropic scaling changes (contraction / stretching) and rotations of 90°, 180° and 270°.

[0070] Yolov3 basic data augmentation uses only left / right mirroring, anisotropic scaling changes (contractions / stretchings) and no rotations.

[0071] The neural network is supplied with an image of dimension 416x416 pixels2, and if the image is of a different size, it must be resized.

[0072] The output is a list of detections.

[0073] Advantageously, each detection is characterized by a position (in the form of the four coordinates of an "axis aligned" bounding box enclosing the detection) and a confidence level (real number between 0.10 and 1.00) which reflects how the algorithm estimates the reliability of the detection.

[0074] Generally, the higher this number is, the higher the probability that it is a true positive (i.e. a genuine detection).

[0075] VIDEO PLAYBACK TOOL INTEGRATING OR INTERFACING WITH THE DETECTION ALGORITHM

[0076] The human-machine interface proposed to operators to assist them in the task of searching for CMs during ITVs can take the form of a video player in which the proposals of the detection algorithm are presented.

[0077] This video player can, for example, have three different possible operating modes in its link with the detection algorithm (this choice being configurable via a configuration file):

[0078] A. The detection algorithm can be embedded in the video player. In this case, the detection neural network is provided to the tool in the form of a file in "ONNX" format (operated via "ML.Net", i.e. the "framework" (in French "cadre") of "open source machine learning" (in French "apprentissage automatique en source ouverte") of the Microsoft Company), and the tool's configuration file references this file.

[0079] In this configuration, the video playback tool "feeds" the neural network with successive images of the video (or the stream from the camera), resizing them to 416x416, and retrieves the detections to display them.

[0080] B. The detection algorithm can be remote and run in / on a server.

[0081] In this case, the tool's configuration file references the server to which it should connect. In this configuration, the video playback tool sends successive images to the server and retrieves the detections for display.

[0082] C. The tool can be used in simple replay of detections already made previously, in which case it reads as input the "detection file" in text format which lists all the detections associated with the video in order to display them.

[0083] The association between the video and the associated detection file to be opened is automatic and is done by name. When a video is opened, if a text file with the same name as the video (but in ".txt" format) is present, then the tool offers the choice between using this existing file or performing a new detection session.

[0084] The algorithm's detections can be presented to the user in two ways, in two places, for example:

[0085] 1. On the image of the video of grid 16 of fuel assembly 1: a square surrounds Each detection, accompanied by the mention "migrant body" CM and an audible alert signals the detection;

[0086] 2. On the video's time progress bar, each detection is marked by a vertical bar, at the location of the time progress bar which corresponds to the moment of detection in the video, in order to easily locate the detections in order to be able to go there, in particular the moments / intervals in which the detections are concentrated.

[0087] In the aforementioned modes A and B, when the detection algorithm has finished processing a video (or a camera session), it offers to archive the list of detections via a text file, named identically to the video, and placed in the same location (same directory). If such a file already exists, the old one is saved and indexed in a subdirectory "backup".

[0088] Figure 4 shows how suspicions of CM migrant bodies can be presented to the operators, on a computer screen.

[0089] A number of interface elements, visible in [Fig.1], facilitate the search and analysis of migrant bodies.

[0090] These interface elements are as follows:

[0091] a. “play / pause” button (in French “lecture / pause”).

[0092] b. Buttons allowing the user to advance in the video or rewind by a specified duration. The duration in question is indicated in the configuration file.

[0093] c. Time indicator in the video (the current time and the total time are displayed when replaying a video, but only the current time is displayed in direct mode connected to the camera).

[0094] d. Time progress bar with, to its right, a speed multiplier factor allowing the video to be switched to fast forward.

[0095] e. Time position indicator in the video, which can be manipulated.

[0096] f. Detection groups (one detection = one vertical bar, a group of contiguous bars = one detection group).

[0097] g. An isolated detection (typical aspect of a "false positive").

[0098] h. Filter that allows filtering the positions to be displayed according to their confidence level.

[0099] i. Navigation from detection group to detection group.

[0100] j. Video display area.

[0101] k. Indication of detection of migrating body CM by the method, in the form of an "encompassing rectangle" + mention "CM" + level of confidence of the detection.

[0102] 1. Access to settings (to fill in the configuration file).

[0103] m. Capture request (for addition to the ITV report).

[0104] n. Opening of the ITV report.

[0105] o. Saving detections after editing by the user (who can add / delete detections).

[0106] p. Opening either a file for review, or the camera feed.

[0107] The concept of "detection group", mentioned above, is explained below.

[0108] In the video, the moments when a migrant body is visible on the screen are grouped together generally in blocks (for example the CM is visible from t=10.2 s to t=15.04 s then not at all, then from t= 121 s to t= 125.3 s).

[0109] In the time progress bar d, each detection is associated with a vertical bar whose height corresponds to the confidence index of the detection.

[0110] A “detection group” therefore appears as a set of such contiguous bars, and thus as a single object.

[0111] The left and right arrows i allow you to navigate directly from detection group to detection group.

[0112] In the configuration file, two parameters related to this concept are advantageously specified:

[0113] 1. The minimum number of detections for a "detection group" to be considered » (this is a way to filter out isolated detections, which are mostly false positives);

[0114] 2. A tolerance threshold for non-detections, in the form of a number of images contiguous "without detection" tolerable (which is two by default, meaning that we should only change groups if we have three or more contiguous images without detection).

[0115] A group of detections is therefore a "same detection" (i.e., a priori of the same object), but one that lasts over time. Navigating from one to another allows for very rapid visualization, for a video, of the different "candidate migrant bodies".

[0116] The advantage of displaying vertical detection bars in the time progress bar is that this visual representation takes on very different aspects in the presence of migrating bodies (contiguous groups of detections with a medium level of high confidence) or in the presence of false positives (isolated detections most often with a low level of confidence).

[0117] With practice, simply viewing the time progress bar equipped with these bars gives a good idea of ​​whether the video contains a migrating body or whether the displayed detections will likely be false positives.

[0118] Detection filtering allows the display of detections whose associated confidence level is lower than the filter value to be suppressed. The level can thus be adapted to each video independently. It allows the elimination of false positives and the visualization, in real time, of how this modifies the appearance of the progress bar. temporal d equipped with vertical bars associated with detections f and g.

[0119] Generating a report, via the m and n buttons, allows the operator to take screenshots n at different key moments, then generate a report m. This report lists the "valid" detections according to the operator (moments in the video + screenshot).

[0120] It may be transmitted to assist the final stage of doubt removal and may be used for archiving.

[0121] OTHER FEATURES YES ALLOW FOR INCREASING THE EFFICIENCY OF THE PROCESS

[0122] The video can be fast-forwarded. In this mode, the video slows down and returns to normal speed in areas where the algorithm has made detections.

[0123] To ensure real-time on machines that do not have sufficient power, the tool does not send all the images of the video (or camera stream) into the detection algorithm but only a certain ratio, namely the one that allows real-time to be maintained.

[0124] In the case where a video file (and not a camera stream) is being processed, it is possible to move within the video during processing, and the processing will jump directly to the current moment.

[0125] If this is done several times, the video will contain parts that have already been processed and parts that have not yet been processed, at different points in the video. To show this state to the user, the time progress bar is colored green (or another color of choice) for the moments that have been processed by the algorithm, and remains white (or another color) for the moments that still need to be processed.

[0126] The progress of the green zone is visualized in real time (in the same way as online videos, which show the user the areas of the video that have been loaded and those that have not yet been loaded).

[0127] Two different modes for resizing the video to the 416x416 format that the network input takes can be offered.

[0128] These videos are in rectangular format in landscape mode (width>height), but some have vertical black bars on the left and right, the central area of ​​interest being almost square.

[0129] The user therefore has the choice, in order to reduce the image to 416x416, of either reducing the video so that only the central square is extracted, which is then reduced to 416x416, or of fitting the entire video into a 416x416 square (the length is reduced to 416, and the width is reduced by the same factor). This choice between the two reduction methods is specified in the configuration file via the parameters button 1.

[0130] Finally, for the accelerated analysis of an archived campaign or a folder containing numerous videos, it is possible to offer a "batch" processing which generates For each video, in addition to the detection files, there is a "signature" image which should allow for quick and priority viewing of videos containing migrant bodies.

[0131] The principle is as follows.

[0132] For each video, once its algorithmic processing is complete, the video tool positions itself at the moment of best detection (in terms of confidence level) of the largest continuous detection group (when there are detection groups), and then takes a screenshot. This screenshot is named like the video, with a 6-digit counter as a prefix, in the format 000000_, which counts the total number of detections in that video.

[0133] All the screenshots thus generated are placed in the same directory. Then, to view the result, simply sort these image files in reverse alphabetical order (in the "Windows" file explorer (registered trademark)), and then browse them in slideshow mode.

[0134] The videos with the most detections will appear first in this sequence. One second of viewing by the operator is sufficient per screenshot, because this screenshot shows both the suspected migrant body framed and the appearance of the time progress bar equipped with detection indicators.

[0135] This image most often allows us to capture a moment when the CM is present (if it is a video that presents a real CM), and in the case where the captured moment does not show a CM, the appearance of the time progress bar gives a very good idea of ​​the presence of CM in the video.

[0136] By applying this latter feature, it was possible, in less than one minute of human processing (viewing these images in reverse alphabetical order), to detect two migrant bodies (on two different videos) that had not been detected by any of the three operators during a previous ITV campaign. PERFORMANCE CHARACTERIZATION:

[0137] The cases of non-detection are explainable. These are new types of CM, on which the algorithm had not been trained.

[0138] It is therefore sufficient to retrain the algorithm with these inputs to improve performance. Thus, the detection rate for the types of CM already seen by the algorithm approaches 100%.

[0139] False positives are very few (some entire videos do not show a single detection, over three minutes of video, i.e. 4500 images).

[0140] The average false positive rate “per image”, measured on a reference test set, is on the order of 2%. This means that in one minute of video (1500 images), there will be 30 false “spot” detections (i.e., flashing for one-twentieth of a second). In a complete test campaign, more than half of the videos of 3 minutes contain fewer than 15 false detections (less than 0.3% false positives). Furthermore, when a false positive does appear in a video, it is usually very brief. Generally, it appears as an isolated flash lasting 1 / 25th of a second and does not attract attention.

[0141] However, there are several ways to further reduce these false positives. The first is filtering out isolated detections (at a point in the assembly where there are no or too few other detections in the video).

[0142] We can also eliminate "typical" false positives which appear because of new elements which look like CMs but are not (such as locally very lightened pencils, or certain water reflections) by retraining them. IMPLEMENTATION VARIATIONS

[0143] IMAGE MODE:

[0144] It is possible to propose a mode in which images can also be processed in addition to videos.

[0145] In this mode, one can either open a single image for immediate analysis (or drag and drop the image from the "Windows" Explorer (registered trademark) to the tool window), or specify a directory containing several images.

[0146] In the case of a single image, the time navigation tools and the time progress bar are no longer of interest.

[0147] In the case of an image directory, these elements are recycled. The navigation arrows b allow movement from image to image, as in a slideshow, and the navigation arrows i, which allowed movement from detection group to detection group, here allow movement from one image containing a detection to another image containing a detection.

[0148] The display of the vertical bars f and g illustrating the detections is retained and the filter h is also functional. TRACKING THE SCREWS AS REFERENCE POINTS:

[0149] The lower end 11 of the combustible assembly 1 generally has twenty-four very characteristic screws which are placed in the same location, regardless of the supplier of the assembly.

[0150] These screws constitute easily detectable reference points with the existing algorithm (in fact, it is sufficient to train the algorithm to detect the screws in addition to the CMs). Since their respective gaps and positions are known, the screws allow, at a very low algorithmic cost, for the estimation of the position of the fuel assembly 1 (precise evaluation of the position / orientation, at each instant, of the fuel assembly relative to the camera).

[0151] The tracking of the screws thus allows the creation of a reference frame relative to the assembly 1 in which it becomes possible to perform mapping and metrology of the CM.

[0152] We can thus give:

[0153] 1. An estimate of the dimension of the CM.

[0154] 2. An estimate of the coverage of the anti-debris assembly grid 16, during riTV (by estimating the areas viewed), because on some archive videos, we see that some areas are dealt with very quickly or even forgotten.

[0155] 3. Determine the start and end of the analysis of the debris grid 16 in a video of ITV (which sometimes includes other sections than just the debris grid examination phase).

[0156] 4. Know the precise position of each detection on the anti-assembly grid debris 16.

[0157] The aforementioned functionality 2 in turn leads to several interesting possibilities:

[0158] 1. A possible almost complete automation which associates a view with a video schematic of the assembly grid on which is shown, where applicable, the location of the CM(s) detected a large number of times (parameter to be specified), as well as an image of each CM.

[0159] 2. The elimination of false positives of the reflection type or related to a perspective effect (which will, at a certain time, be seen in a place on the grid but which will no longer be seen when the same place on the grid is looked at from another angle of view, subsequently).

[0160] For this, it is sufficient to give, for each detection, its "visibility rate" when the corresponding area is on the screen (a CM should have such a rate very high, and a false positive a low rate).

[0161] 3. The counting of "different" detections (in the sense of their location on the grid), and the ability to differentiate them, via a different color for each different "object" in their representation, both on the video and in the time progress bar, as well as via "ticks" in the interface allowing to hide or show each detection.

[0162] 4. Have a reference position for the CM withdrawal operation.

[0163] 5. Have a reference to the CM's position for research / confrontation on previous ITVs of the same assembly (to check if the CM was already there). AUTOMATION OF THE ITV:

[0164] In an even more forward-looking version of the tool, an artificial intelligence (based on the same archive videos) could learn to manipulate the camera itself to search for migrant bodies by capturing the videos itself and deciding, for example, to come and examine its suspicions of CM more closely. APPLICATIONS / USER BENEFITS:

[0165] The method according to the invention can be implemented in each of the 3 steps of CM detection.

[0166] Thus, the tool can be offered to each of the three operators or only to a subset. Thus, for example, it could be given not only to the operator who performs the live inspection, but only to the "reviewers", so as not to risk a decrease in vigilance paradoxically due to the assistance).

[0167] In a version where the three steps are implemented via the method according to the invention, the assistance can be used from the first step during the initial image capture, i.e. "live" (therefore either with a machine powerful enough to use mode A described above, or the use of mode B described above).

[0168] The operator can automatically generate their report (including any screenshots they deem relevant) using the appropriate function. They can then provide the generated video and detection file as input for the second step, for independent analysis (meaning they do not provide the report).

[0169] During this second independent analysis step, the tool can be used in mode C (no detection via the algorithm, but simply a review of the previously generated detection file), and the operator can also automatically generate their report. They will also have the option to edit the detection file and delete detections by selecting them individually on the screen or by selecting a time range in the timeline.

[0170] He can also add them manually, via the manual add function (he pauses the video and draws the square on the screen).

[0171] Finally, the video, the detection file possibly enriched by editing it during the second step, and the two generated reports are given to the operator in charge of the final review and the dispelling of doubt in case of suspicion of CM.

[0172] The resolution of doubt in the event of a suspected CM is on the critical path to the block shutdown, so any time saved at this stage is a direct monetary gain.

[0173] A combustible assembly is very expensive and failure to detect a migrating body can damage it and render it unusable.

[0174] A migrating body may, possibly through various means, cause a perforation of the sheath of a pencil and cause a loss of the first safety barrier, with diffusion of fission products in the primary circuit.

[0175] On archived videos from different sites, migrant bodies missed by the method according to the current state of the art involving three operators could be detected by the process according to the invention which directly indicated at which moments in the video to watch, which, on the one hand, increases the reliability of the search for CM and, on the other hand, greatly limits the viewing time of the operators, particularly for the last operator in its role of clearing up doubt.

[0176] The solution according to the invention makes it possible to partially overcome the disadvantages associated with human fatigue in the face of such a task, which inevitably leads to forgetfulness.

Claims

Demands

1. A method for assisting in the detection of migrating bodies (MBs) within a fuel assembly (1) of a nuclear power plant and more particularly on the debris grid (16) of the lower end (11) of said assembly (1), during which at least one camera is piloted towards said assembly (1) and the image stream recorded by said at least one camera is directed to a human-machine interface which includes at least one screen enabling a first operator to view said video image stream, characterized in that it includes a first step of detection, using an image recognition algorithm, of said migrating bodies (MBs), as well as at least a second step of alerting said operator if said algorithm has detected the potential presence of at least one migrating body (MB).

2. A method according to claim 1, characterized in that said second alert step includes triggering an audible alert and / or a visual alert on said screen, from the location where said migrating body (MB) is potentially located and / or automatically generating an inspection report.

3. A method according to claim 2, characterized in that, when said algorithm has detected the potential presence of at least one migrating body (MB), the corresponding video file and the detection file(s) performed by said algorithm are stored and sent to a second operator with a time indication in said video file of the time(s) at which one (or more) migrating body(ies) (MB) was / were potentially detected, this second operator validating the presence or lack of presence of at least one migrating body (MB).

4. Method according to claim 3, characterized in that said detection files are transmitted to a third operator who definitively validates the presence or lack of presence of at least one migrating body (MB).

5. A method according to any one of claims 3 or 4, characterized in that it includes a machine learning step for detecting migrating bodies (MBs) when the second operator, respectively the third operator, has validated the lack of presence of a migrating body (MB).

6. A method according to claims 3 to 5, characterized in that one does use of image recognition software of the "convolutional neural network" type, which is trained on a learning base consisting of archives of videos of migrant bodies (MB), which base is potentially enriched each time said recognition software makes a detection error such as a "false positive" or a "false negative", this error having been validated by the second and / or third operator.

7. Method according to claim 3, characterized in that said time indication has the form of at least one bar displayed on the time progress line of the video.

Images