METHOD FOR CHARACTERIZING A PIPELINE DEFECT

Abstract

METHOD FOR CHARACTERIZING A PIPELINE DEFECT The invention relates to a computer-implemented method for characterizing a pipeline defect. The method allows the generation of at least one label (LAB1) characteristic of a pipeline defect identified on a first image (IM1) of a pipeline. The invention also relates to a system for characterizing a pipeline defect. Figure for abstract: Fig.1

Description

Title of the invention: METHOD FOR CHARACTERIZING A DEFECT IN A PIPELINE Field of invention

[0001] The invention relates to the field of pipelines, in particular the field of pipelines in collective sanitation networks. The invention also relates to the field of structures accommodating or allowing the circulation of fluids, in particular water towers, heating or cooling networks. The invention relates to the field of maintenance of said sanitation networks. More specifically, the invention relates to the field of inspection of said pipelines, in particular the field of detection and identification of defects inside said pipelines. The invention also relates to the field of automation of detection and identification of pipeline defects. Finally, the invention relates to the field of automated recognition of defects or observations on pipeline images. State of the art

[0002] There is a need to inspect the pipes of sewerage networks to detect defects in them. Such defects are notably due to wear and fouling of these pipes.

[0003] In Europe, the ISO EN-13508-2 standard is commonly used to classify pipeline defects. This standard defines a system for coding pipeline defects in the form of a trigram. For networks and collectors, the standard includes a total of 36 different trigrams, each corresponding to a particular type of defect or a particular observation in the pipeline.

[0004] Typically, pipelines are inspected by television using remotely controlled robots. Such a robot typically has wheels and a motor to move within the pipeline, a camera to take images of the inside of the pipeline, and an umbilical cable allowing the robot to communicate with a control interface used by an operator outside the pipeline.

[0005] During the inspection, the operator controls the robot's progress. He has a screen on the interface allowing the real-time display of images from the robot's camera. Thus, when the operator detects a defect on the internal wall of the pipe, he stops the robot's movement, rotates the camera in order to position it in the best way to take a picture of the detected defect.

[0006] During this inspection, the operator then assigns a code to the fault he has detected. according to ISO EN-13508-2. Such inspections make it possible to know the condition of the pipeline network and to plan maintenance actions for them.

[0007] Such a process, although effective, nevertheless requires a lot of inspection time from the operators. In addition, there are disparities, for the same detected defect, in the code that is assigned by the operator. The results of the inspection operations are therefore not always the same depending on the operators who carry out the inspection. Finally, the training of novice inspectors is long and costly, to achieve the level of quality required in the inspection.

[0008] Certain so-called inspectable pipelines, that is to say those generally having a nominal diameter greater than 1600 millimeters, are generally inspected directly by an operator who manipulates a camera device and photographs the parts of the pipelines which concern points of interest such as defects or observations to be recorded.

[0009] In the state of the art, there are solutions whose aim is to enable automated detection of pipeline defects. Some of these solutions propose systems allowing a reconstruction of the pipeline from a video taken inside it by a robot for the purpose of detecting defects on the video. Such solutions allow a posteriori analysis of the pipeline by an operator who will be able to identify the defects present on the pipeline. Such systems do not allow real-time and automatic inspection of the pipeline.

[0010] Some of these solutions use artificial intelligence to perform this detection. Such systems do not allow for the classification of detected defects, or only allow for a superficial classification on a reduced number of types of pipeline defects. Such systems do not allow for a more specific characterization of a defect, i.e. for assigning an additional level of characterization to the type of detected defect.

[0011] There are also systems that allow for good classification of defects. These systems require exporting the data to a third-party system to process the data and perform the classification. Such systems do not allow for real-time analysis of the pipeline, as they require taking the video of the entire pipeline, before it is exported to the "cloud" in order to carry out the analysis on the video. These systems therefore pose operability problems in a field pipeline inspection context.

[0012] An aim of the invention is to remedy the aforementioned drawbacks of existing solutions and systems. Summary of the invention

[0013] For this purpose, the invention relates to a computer-implemented method of characterizing- repair of a pipeline defect which includes the steps of:

[0014] Generation of a plurality of probabilities of presence of a plurality of first classes of defect on a first image of the interior of a pipeline from a first trained labeling learning function, said first learning function implementing a machine learning model trained from a set of images each comprising a plurality of first labels characterizing at least one defect;

[0015] Comparison for each first class of defect of the value of the probability with a threshold characteristic of each first class of defect;

[0016] Selection of each first class of defect for which the estimated probability is greater than the threshold value associated with said class;

[0017] Selection for at least one first selected defect class, of at least one second specification class from among a plurality of second specification classes, on the first image, from a second specification learning function; and

[0018] Generation of at least one label characteristic of each first class of defect and of each second class of selected specification.

[0019] The method according to the invention advantageously allows the identification and characterization, on an image of a pipeline, of at least one defect in the pipeline. The method according to the invention allows the classification of this defect and its labeling automatically. The method according to the invention also allows a second level of classification to be defined from the first class of the selected defect. This additional specification allows better characterization and a level of detail and precision not available in the systems of the prior art.

[0020] The method according to the invention also allows automation and standardization in the characterization of pipeline defects. Finally, the method according to the invention makes it possible to facilitate the training of new operators by allowing a proposal for characterizing defects.

[0021] According to one aspect, the invention relates to a method in which a plurality of first labels are generated, each first label preferably being displayed on a display with its probability of presence.

[0022] According to one embodiment, the second specification learning function is trained, said second specification learning function implementing a machine learning model trained from the set of images each comprising a plurality of second labels characterizing at least one specification class.

[0023] According to one embodiment, the second specification learning function selects the second specification class based on a predefined rule allowing the association of said second specification class with the first class of defects.

[0024] According to one embodiment, the method further comprises a step of selecting a specific subclass of a second specification class, said step of selecting a specific subclass comprising the steps of: • Selection of at least one rule within a rule base based on the value of the first label, each rule defining an association between a characteristic value of said first image and a second predefined label of a specific subclass; and • Generation of data including the value of the first label and the value of the second label.

[0025] According to one embodiment, the characteristic value of the first image is its first label.

[0026] According to one embodiment, the characteristic value of the first image is generated from an image processing algorithm applied to the first image.

[0027] According to one embodiment, the method comprises a step of receiving data associated with the first image, said data being acquired by means of a user interface, said data being recorded in a memory so as to be associated with the characteristic value of the first image, said data being compared to a plurality of attribute values ​​associated with each first label, said comparison resulting in the exclusion of at least one first label from the list of first labels generated when the value of an attribute of said excluded first label is greater than or less than a given threshold.

[0028] According to one embodiment, the data associated with the first image is modeled in at least one discrete symbol in a natural language and normalized within a third label attached to the first label and / or to the second label comprising a plurality of discrete symbols in a natural language, said modeling and said normalization allowing relearning of the first trained labeling learning function and / or of the second trained specification learning function.

[0029] According to one embodiment, a plurality of first labels is generated, each first label being displayed on a graphical interface of a display, the graphical interface comprising a component for validating the presence of each first label displayed by an operator and a component for entering a new first label making it possible to correct the value of a first label displayed, significant data of the number of differences between the defects identified by the method of the invention and the defects identified by the operators on the same images being calculated.

[0030] According to one embodiment, the exceeding, by the significant data of the number of differences between the defects identified by the method of the invention and the defects identified by the operators, of a characteristic performance threshold, causes the retraining of the machine learning model of the first learning function and / or of the machine learning model of the second learning function.

[0031] According to one embodiment, the modification of the first label by an operator comprises an analysis of the number of symbols of the label having changed, an analysis of the position of the symbol in the label having changed and the type of the at least one symbol which has been modified, said analysis making it possible to generate a retraining action of at least one model of a second learning function.

[0032] According to one embodiment, the method comprises, before the generation step, the steps of: • Reception of at least one raw image of the interior of a pipeline; • Resizing the image to obtain a resized image; • Normalization processing of the hue, saturation and / or brightness of the resized image to obtain the first image.

[0033] According to one embodiment, the first trained labeling learning function and / or the second trained classification learning function is trained from a trained convolutional neural network comprising a plurality of layers, at least one of the layers of which is re-trained from the set of images each comprising a plurality of first labels characterizing at least one defect.

[0034] According to one embodiment, the pre-existing convolutional neural network used for training the first trained labeling learning function and / or for the second trained specification learning function is a network of the MobileNet, ResNet, EfficientNet, TRansformeur or DenseNet type.

[0035] According to one embodiment, the set of images comprising the plurality of first labels is increased by adding a second set of images resulting from the geometric and / or photometric transformation of at least a portion of the set of images comprising the plurality of first labels.

[0036] According to one embodiment, a loss function quantifying a deviation between predictions of the first learning function and the first labels is a function of one of the following types: • Binary cross entropy, • Weighted cross-entropy, • soft Fl-score or dice loss, • soft macro Fl-score, or • double soft macro Fl-score.

[0037] According to one embodiment, the method further comprises a display step on a user interface of each first label.

[0038] The invention also relates to a system for characterizing a pipeline defect, the system comprising: • A pipeline exploration device comprising at least one optical sensor; • A user interface configured to display in real time a data stream from the optical sensor of the pipeline exploration device,

[0039] the system being configured to implement the method according to the invention. Brief description of the figures

[0040] Other characteristics and advantages of the invention will emerge on reading the detailed description which follows, with reference to the appended figures, which illustrate:

[0041] [Fig-1]: a flowchart illustrating a method for characterizing a fault in a ca finalization according to an embodiment of the invention;

[0042] [Fig.2]: a flowchart illustrating the processing of a raw image according to a mode of realization of the invention;

[0043] [Fig.3]: a flowchart illustrating the step of selecting a rule in a database rules according to one embodiment of the invention; and

[0044] [Fig.4]: a block diagram illustrating the operation of a system according to a embodiment of the invention. Description of the invention

[0045] [Fig.l] represents a flowchart illustrating a method for characterizing a pipeline surface defect according to an embodiment of the invention.

[0046] In the following description, the term "pipe" will be used to designate any element, for example of a sanitation system, which allows fluids to be carried, more particularly water evacuated by said sanitation system. Pipes are generally closed tubular elements, which may include connections between several tubular elements. The section of the pipes is generally circular. Also included in the scope of a pipeline will be structures receiving or allowing the circulation of fluids, in particular water towers, heating or cooling networks.

[0047] A pipeline defect, or pipeline surface defect, is understood to mean any singularity of the pipeline surface, any deterioration or alteration of the surface condition of the interior of a pipeline which is observable. More precisely, these defects are characteristics visible from the interior of the pipeline. The pipeline defect may be, for example and without limitation, a crack in the interior surface of said pipeline, a connection with another pipeline, an obstructing element deposited on the internal surface of said pipeline, an alteration of the internal surface of the pipeline, corrosion of the internal surface of the pipeline.

[0048] The method according to the invention aims to characterize a detected defect in the pipeline to determine what type of defect it is.

[0049] The method according to the invention comprises a step of generating GEN1 a plurality of probability of presence PROB1 of a plurality of first classes of defects DEFI on a first image IM1 of the interior of a pipeline from a first trained labeling learning function FONC1. The first trained labeling learning function FONC1 implements a machine learning model trained from a set of images each comprising a plurality of first labels characterizing at least one defect.

[0050] This generation step GEN1 uses a first image IM1 of the interior of a pipe. Advantageously, the image of the interior of a pipe is a selected image, because it includes the presence on the pipe and in the field of the image of at least one defect. The first trained labeling learning function FONC1 analyzes the first image IM1 to generate a probability of presence PROB1 of a first defect class DEFI on the first image IML. In other words, the first learning function FONC1 determines the probability that at least one defect belonging to the first defect class is present on said first image IML. During this generation step GEN1 of the plurality of probabilities PROB1, a probability of presence is preferably generated for each existing first defect class DEFI.Thus, for each first class of fault DEFI, we have a probability of presence on the first IML image. All of these presence probabilities PROB1 form the plurality of presence probabilities. The first trained labeling learning function FONC1 is a function that gives a probability of presence of a first class of fault from the first IML image. In other words, the first learning function analyzes the first image IM1 and provides the plurality of presence probabilities PROB1. The first trained labeling learning function FONC1 implements a model resulting from training undergone by the function. The first learning function FONC1 is trained from a set of images. This set of images includes images of the interior of pipes. Each image is associated with a label characterizing a pipe defect.In other words, the first learning function FONC1 is obtained from supervised learning based on a set of labeled images.

[0051] The method according to the invention then comprises a step of comparing the value of the presence probabilities PROB1 with a threshold value which is specific to each first class of fault DEFI. In other words, for each first class of fault DEFI for which the first learning function FONC1 provides a probability of presence PR0B1, said probability is compared with the threshold associated with said first DEFI fault class. Thus, we can have different thresholds depending on the types of first DEFI singularity which make up the plurality of first DEFI fault classes.

[0052] The method according to the invention then comprises a step SEL1 of selecting first DEFI fault classes. The first DEFI fault classes selected are the first fault classes whose associated presence probability PROB1 is greater than the characteristic threshold of said DEFI fault class. In other words, the method selects a set of first DEFI fault classes for which the presence probability PROB1 is sufficiently high. This step therefore allows the establishment of a list of fault types corresponding to the first selected DEFI fault classes.

[0053] The method according to the invention then comprises a selection step SEL2 for at least one first fault class DEFI previously selected from at least one second specification class SPE2. The second specification class SPE2 corresponds to an additional level of classification of the pipeline fault with respect to the first fault class DEFI. The second specification class SPE2 is selected from a set of second specification classes SPE2 which are characteristic of the first fault class DEFI which was previously selected. In other words, the selection SEL2 of the second specification class SPE2 is made as a function of the first fault class(es) which were previously selected. The selection of the second specification class is carried out by a second specification learning function FONC2.

[0054] According to one embodiment, the second learning function of specification FONC2 implements a trained machine learning model. The machine learning model is trained from a set of images comprising a plurality of second labels which each correspond to at least one second class of specification SPE2.

[0055] The method then comprises a step GEN2 of generating at least one label LAB1 characteristic of each first DEFI defect class and each selected second SPE2 specification class. In other words, the method makes it possible to generate a label LAB1 which specifies a pair of a first DEFI defect class and a second SPE2 specification class. In this way, the method allows automatic generation, from an image of a pipeline comprising a defect, of a label making it possible to classify this defect according to two specification levels. Advantageously, the method according to the invention also makes it possible to generate several labels when several different defects are present on the first image IM1 of the interior of the pipeline. Image processing

[0056] According to one embodiment, the method according to the invention comprises a step of receiving a raw IMBR image of the interior of a pipeline. The raw IMBR image is preferably acquired by a pipeline exploration device 20 which is preferably equipped with an image capture device or optical sensor 22. We will return later to the description of these devices. [Fig.2] is a flowchart illustrating the different steps of processing the raw IMBR image.

[0057] The raw IMBR image, which corresponds to a shot of the inside of a pipeline, is therefore received during this step of the method. The raw IMBR image comprises a set of pixels which allow a representation of the inside of the pipeline. Preferably, the raw IMBR image comprises in its field of view one or more walls of the pipeline. The raw IMBR image can be taken so as to represent substantially from the front a wall of the pipeline. According to one embodiment, the raw IMBR image comprises a representation which is inclined relative to a wall of the pipeline. According to one embodiment, the raw IMBR image is an image of a portion of pipeline. The portion of pipeline on the raw IMBR image preferably comprises at least one pipeline defect or a pipeline surface defect. The raw IMBR image therefore represents at least one defect of the pipeline.According to one embodiment, the pipeline portion of the raw IMBR image comprises several defects.

[0058] According to one embodiment, the method then comprises a step of resizing RED of the raw image IMBR to obtain a resized image IMRED. The resizing step may comprise converting the raw image into a format that will be easier to process by the various learning functions implemented by the method. For example, during resizing, the pixel resolution of the image may be changed to be a standard resolution. Preferably, the standard resolution corresponds to a resolution of the images that were used for training the first trained labeling learning function FONC1. This arrangement makes it possible to increase the efficiency of the function and to increase the precision of the presence probabilities PROB1 generated by it.According to one embodiment, the standard resolution is the resolution of the images that were used for training the second trained specification learning function FONC2. This arrangement makes it possible to increase the reliability of the selection made by the second learning function FONC2. According to one embodiment, the images used for training the first learning function FONC1 and for training the second learning function FONC2 have the same definition, which is the same as the definition of the resized image IMRED. In this way, the same resized image IMRED makes it possible to obtain . optimized results with the first learning function FONC1 and with the second learning function FONC2.

[0059] According to one embodiment, the step of resizing the raw IMBR image comprises a step of truncation of the image. This step makes it possible to remove from the image one or more parts thereof, for example parts which would not be relevant for the analysis thereof by the learning functions FONC1, FONC2. For example, the truncation of the raw IMBR image allows the removal of a side band thereof. This arrangement is particularly advantageous when the channeling defect(s) are not present on a part of the image, or when the latter is not well centered. The truncation also makes it possible to modify, for example reduce, the resolution of the raw IMBR image. This arrangement also makes it possible to obtain the standard resolution.

[0060] According to one embodiment, the step of resizing the raw image IMBR comprises a step of geometric transformation of the image. For example, the geometric transformation of the image may comprise a step of rotating the raw image IMBR. This rotation of the raw image advantageously makes it possible to place the image in an optimal angular orientation for its processing by the first learning function FONC1 and by the second learning function FONC2. For example, this rotation makes it possible to replace the image so that its orientation is consistent with the orientation of the pipe in which the image is taken. Indeed, if the optics of the shooting device are oriented obliquely relative to the vertical, a rotation of the image makes it possible to obtain a resized image which is oriented along a vertical axis.

[0061] According to one embodiment, the image undergoes a TRAIT processing step for normalizing the light intensity. By TRAIT processing step for normalizing the light intensity, we mean a step during which the contrast of the image is adjusted to allow better processing subsequently by the different learning functions. According to one embodiment, the TRAIT processing step for normalizing the light intensity comprises a step for clipping high intensities. According to one embodiment, the TRAIT processing step for normalizing the light intensity comprises a step for adjusting the saturation of the image. The adjustment of the saturation may comprise a decrease in the saturation or an increase in the saturation. According to one embodiment, the TRAIT processing step for normalizing the light intensity comprises a step for increasing the light intensity.According to one embodiment, the TRAIT processing step for normalizing the light intensity comprises an image analysis step making it possible to deduce the processing to be carried out on the image to normalize it. Multi-label classification.

[0062] According to one embodiment, the first trained labeling learning function FONC1 takes as input a first image IM1 of the inside of a pipe. As seen previously, this first image IM1 can come from a raw image IMBR which is directly obtained by the imaging device. According to one embodiment, the first learning function directly processes the raw image IMBR of the pipe as input.

[0063] The first trained labeling learning function takes as input an image selected by an operator who triggers the taking of the raw IMBR image. Preferably, the first image comprises at least one singularity or pipeline defect. Thus, the first trained labeling learning function FONC1 only has to identify the defect(s) present on the image.

[0064] The first trained labeling learning function FONC1 generates a set of first presence probabilities PROB1 of first DEFI defect classes on the first IML image. In other words, the first trained labeling learning function FONC1 tests, for all the types of defects that exist in a classification, the probability of their presence on the first IML image. Thus, this first trained labeling learning function takes the first image IM1 as input and gives as output a plurality of presence probabilities PROB1, each presence probability PROB1 being associated with a specific DEFI defect class in the set of defect classes. To do this, the first trained labeling learning function implements an estimation method based on an analysis of the first IML image.

[0065] According to one embodiment, the first trained labeling learning function FONC1 comprises a neural network architecture. According to one embodiment, the first trained labeling learning function comprises a multi-layer neural network. According to one embodiment, the first trained labeling learning function FONC1 comprises a convolutional neural network. According to one embodiment, the first trained labeling learning function FONC1 comprises a Transformer-type neural network.

[0066] The first trained labeling learning function FONC1 has been trained previously to the implementation of the method. According to one embodiment, the training of the first trained labeling learning function FONC1 comprises supervised training. During the training, a training image base is used. The training image base comprises a plurality of images, which are preferably images of the interior of pipes, which comprise defects. Each image used for training comprises a label which has been preferably assigned by an operator to the singularity(ies) present on the image. The label comprises at least one piece of information on the first class(es). of DEFI fault present on the image. According to one embodiment, the training image base comprises at least 1000 labeled images of the interior of the pipeline. According to one embodiment, the training image base comprises at least 10000 labeled images of the interior of the pipeline. According to one embodiment, the training image base comprises at least 100000 labeled images of the interior of the pipeline. According to one embodiment, the training image base comprises at least 500000 labeled images of the interior of the pipeline. According to one embodiment, the training image base comprises at least 1000000 labeled images of the interior of the pipeline.

[0067] According to one embodiment, the images of the training image base are raw IMBR images which have undergone at least one RED resizing step as described previously. According to one embodiment, the raw IMBR images have additionally or alternatively undergone an image truncation step. According to one embodiment, the raw IMBR images have additionally or alternatively undergone a geometric image transformation step as described previously. According to one embodiment, the images have undergone a TRAIT processing step of normalizing the light intensity as described previously. In this way, the image may have undergone a brightness adjustment step, a contrast adjustment step, a brightness clipping step, and / or a saturation adjustment step.All these steps make it possible to obtain a set of training images of the learning function which is homogeneous and therefore allows better training of the function. According to one embodiment, the TRAIT processing step of normalizing the light intensity comprises a step of analyzing the image making it possible to deduce the processing to be carried out on the image to normalize it.

[0068] According to one embodiment, the number of images in the training image database is increased using data augmentation techniques. According to one embodiment, the training image database is increased by adding images from said database that have undergone geometric transformations. According to one embodiment, the geometric transformations include rotations of the images in the image database. Each image that has undergone a rotation is added to the training image database. According to one embodiment, the training image database is increased by performing processing on the colors present in the images. More specifically, color modification processing includes, but is not limited to, processing of the saturation of the images, the brightness of the images, or the contrast of the images. The images thus processed are reincorporated into the training image database in addition to the unprocessed base images.In this way, the number of images in the training image database is increased with images having different defects and color characteristics. different from the initial images of the base. This arrangement makes it possible to increase the effectiveness of the training of the learning function.

[0069] According to one embodiment, the first trained labeling learning function FONC1 is trained using a transfer learning method. A transfer learning method makes it possible to use a pre-trained image recognition model on third-party image recognition problems as a starting point for training the first trained labeling learning function FONC1. The pre-training third-party images of the first learning function comprise images that are not from pipelines. According to one example, the third-party images comprise images of faces, images of everyday objects and / or images of animals or living beings. The pre-trained model is re-trained using the training image base previously described.According to one embodiment, the pre-trained model is trained on recognition of images that are not images of the interior of a pipeline. According to one embodiment, the pre-trained model is a convolutional neural network. According to one embodiment, the pre-trained model is a convolutional neural network of the "MobileNet", "ResNet", "EfficientNet" or "DenseNet" type. According to one embodiment, the pre-trained model is a "Transformer" type model. According to one embodiment, the pre-trained model is a model as described previously which comprises at least one "Dense" type layer. According to one embodiment, the pre-trained model comprises several additional "Dense" type layers. According to one embodiment, the pre-trained model is re-trained by re-training only a portion of the layers of the neural network on the basis of training images.

[0070] According to one embodiment, the first trained labeling learning function FONC1 is trained using a hybrid model which assembles a convolutional architecture and a recurrent architecture in a single neural network. Such a training mode makes it possible to better exploit the correlations between the different defects which are present on the same first image IM1.

[0071] According to one embodiment, the first learning function at least one model. According to one embodiment, the first learning function uses at least two different models. According to one example, the learning function uses a first model for the generation GEN 1 of the presence probabilities PROB1 and a second model to also perform a generation GEN1 of the presence probabilities PROB1. According to one example, the probability that will be used during the comparison step COMP with the threshold is an average of the probability PROB 1 generated by the first model and the probability PROB1 generated by the second model. According to one example, the probability used during the comparison step COMP is an average weighted by the probability generated by the first model and the probability generated by the second model. For example, the weighting assigned to the PROB1 probability of the first model is 0.8 and that assigned to the PROB1 probability of the second model is 0.2. This arrangement is particularly advantageous for taking advantage of better identification by one model rather than another. According to one embodiment, the weighting of the average is only used for the generation of the PROB1 probability relating to a first predetermined DEFI default class. This arrangement makes it possible to take advantage of the identification performance of a model for a given DEFI default class. According to one embodiment, several different weightings are used for several first DEFI default classes.In this way, the weighting of a model is higher for the first DEFI fault classes for which its identification performance is better and can be lowered for the first DEFI fault classes for which the identification performance is lower.

[0072] According to one embodiment, the first trained labeling learning function FONC1 also provides location information for the defect(s) identified on the first image IM1. Such information may subsequently be provided to an operator to assist him in the pipeline inspection process. According to one embodiment, the location information is associated with the first image IM1. According to one embodiment, the location information comprises angular location information for the defect. This angular information preferably comprises angle information relative to a vertical line. The information preferably comprises an indication of the time location of the singularity. According to one embodiment, the location of the defect(s) on the image is based on an artificial intelligence explainability algorithm.Such an algorithm uses the decision process of the first trained labeling learning function FONC1 to derive information relating to the defects. For example, the angular location or time location of the defect(s) is determined by using an activation map of the first trained labeling learning function FONC1. According to this example, such a map provides the pixels of the first image IM1 on which the decision of the first learning function FONC1 is based. In this way, the defects are located on the first image IM1. The angular location or time location of the image can therefore be determined. Additionally or alternatively, the spatial location of the defect(s) is also determined. According to one embodiment, the first trained labeling learning function directly provides the location information.According to one embodiment, the first trained labeling learning function FONC1 was trained, at least partially, from images comprising labels comprising . fault location information. Thus, the first learning function, in addition to providing probabilities PROB1 of fault presence, also provides at least one fault location information. This location information includes, for example, time location information, angular location information, and / or spatial location information.

[0073] According to one embodiment, the first learning function FONC1 is adjusted using a loss function. In other words, parameters of the first learning function and / or the response of the first learning function FONC1 are adjusted using the loss function. A loss function is a function quantifying the difference between the predictions of the model and the actual observations of the data set used during training. According to one embodiment, the loss function used is a binary cross-entropy type function. According to one embodiment, the loss function used is a weighted cross-entropy type function. Advantageously, the loss function makes it possible to take into account an imbalance in the training image base. For example, a defect class may only be represented by a small number of images in the image base compared to the other defect classes.In this case, a weight can be assigned to such images in order to make them weigh more in the training of the first learning function FONCE. This arrangement makes it possible to make the learning function FONC1 capable of detecting a DEFI defect class which is underrepresented in the training image base, and which might otherwise not be detected. A weighted binary cross-entropy type loss function makes it possible to implement this arrangement. According to one embodiment, the loss function used is a soft Fl-score type function or a "dice loss" type function, generally named thus according to an English language term. According to one embodiment, the loss function used is a macro soft Fl-score type function or a double macro soft Fl-score type function. According to one embodiment, the loss function takes into account at least one identification performance indicator.According to one embodiment, the loss function makes it possible to choose one or more performance indicators and to reduce or increase the influence of said indicators on the loss function.

[0074] The comparison step COMP is performed for each first defect class DEFI. Indeed, the first learning function FONC1 generates a set of presence probabilities PROB1 each corresponding to a first defect class DEFI. Each value of the presence probability is compared with a threshold value which is specific to said first defect class DEFI in question. Each threshold value is adjusted according to the interest or not of having false detections or missed defects. For example, for a given first defect class, it may be desirable to avoid missing the type of defect associated with the first defect class CHALLENGE in question, because this singularity corresponds to a critical pipeline defect. In this case, the detection threshold is deliberately set quite low, to prevent the classifier from missing a defect of this type. Such an identification method can generate false positives. Conversely, for defects that are easy to identify, a higher, i.e. stricter, threshold is used. In this way, an operator carrying out subsequent image inspection is prevented from being overwhelmed by a large number of defects and false detections.

[0075] Advantageously, the threshold used for a given DEFI defect class is adapted to the recurrence of said DEFI defect class in the training image base of the first learning function FONCE Advantageously, the threshold relating to at least one DEFI defect class is set as a function of the performance of identifying the type of defect relating to said DEFI defect class during the supervised training of the first learning function FONCE In other words, the threshold is set for each DEFI defect class as a function of the performance of identifying said DEFI defect class during the training of the first learning function FONCE According to one embodiment, the threshold relating to at least one DEFI defect class is set as a function of feedback from users and / or operators using the method according to the invention.Such feedback allows the threshold to be adjusted based on the proportion of false positives identified by increasing the threshold when many false positives are detected. Similarly, the threshold may be lowered if for a first DEFI defect class, said defect is not identified several times. According to one embodiment, the thresholds are set at 0.4. According to this embodiment, when the probability of presence PROB1 associated with the first DEFI defect class is greater than 40%, then said defect class is selected during the selection step SEL1. According to one example, the threshold associated with a first DEFI defect class corresponding to a crack in the pipeline is lowered relative to the threshold of the other first DEFI defect classes. For example, this threshold is lowered to 0.3 or 0.2.Such a lowering of the threshold associated with the crack-related class is advantageous, because it allows to increase the identification of this DEFI defect class to be sure that it is detected.

[0076] According to one embodiment, the characteristic threshold of at least one first DEFI fault class is adaptive. For example, the threshold is adapted during successive uses of the method according to a success rate of the identification of the first DEFI fault class. The success rate can for example be determined according to user feedback. According to one embodiment, the characteristic threshold is defined according to at least one probability of presence PROB1 of a first DEFI fault class. For example, the characteristic threshold can be equal to the value of the highest probability of presence PROB1 of all the fault classes for for which a probability of presence is calculated. According to one example, a single DEFI defect class is selected from the DEFI defect classes for which a probability of presence PR0B1 is calculated. According to this example, only the DEFI defect class with the highest probability of presence PR0B1 is selected. According to one embodiment, the characteristic threshold of at least one DEFI defect class is defined from a third learning function F0NC3. According to one embodiment, the third learning function F0NC3 is trained from a set of labeled pipeline images. According to one example, the characteristic threshold of at least one DEFI defect class is adapted based on data measured on the first image. For example, the threshold can be adapted based on brightness data measured on the first image.This adaptation makes it possible to adapt the threshold according to the performance of the first learning function F0NC1 on the identification according to the measured data.

[0077] According to one embodiment, the method comprises five first classes of DEFI defects. According to one embodiment, the method comprises ten first classes of DEFI defects. According to one embodiment, the method comprises fifteen first classes of DEFI defects. Preferably, the method comprises twenty-one first classes of DEFI defects. These classes of defects correspond to types of pipeline defects as described in the European standard EN-13508.

[0078] According to one embodiment, the method relates to a set of first classes of DEFI defects which assess a pipeline defect. Among the first classes of DEFI defects available in the method, we find for example a crack in the pipeline, a rupture of the pipeline, a penetrating connection, a defective connection. We also find the visible sealing joints, the assembly displacements, the roots visible on an image. We also find the presence of deposits in the pipeline, which may be adherent or not, the presence of an earth inlet, a connection or a curvature of the collector. This list of defects observed forming the different first classes of DEFI defects is non-limiting and can be supplemented by any defect which it will appear relevant to identify to the person skilled in the art. Multiclass classification

[0079] The method according to the invention comprises a step of selection SEL2 for at least a first class of faults DEFI selected previously, of at least a second class of specification SPE2.

[0080] Each first class of defects does not necessarily include a second class of specification SPE2. Thus, the step of selecting SEL2 a second class of specification SPE2 may or may not include the selection of a second class of specification SPE2.

[0081] Each second SPE2 specification class comprises a specification of the label corresponding to the selected first DEFI defect class. In other words, the first DEFI defect class corresponds to a first level of specification of the defect, and the second specification class provides a second level of analysis concerning said defect. In this way, a second SPE2 specification class is specific to the first DEFI defect class in question. According to one embodiment, for the first defect class corresponding to the pipe connections, the available second SPE2 specification classes comprise the connection type among seven available connection types. According to one example, the available connection types are the connections of the breeches type, the cored saddle type, the chiseled saddle type, the cored direct tapping type, the chiseled direct tapping type, the unknown fitting type, and the other type.According to one embodiment, the unknown connection type is determined when the second learning function FONC2 does not detect one of the other types of connections. Advantageously, several second specification classes can be associated with the same first fault class DEFI to generate the first label LAB1. For example, in the case of the first fault class corresponding to the connection, a second specification class SPE2 relating to the location of the connection can be determined. For example, it is possible to have a second specification class SPE2 relating to the placement of the connection to the left of the pipe, a second specification class SPE2 relating to the placement of the connection to the right of the pipe. It is also possible to have a second specification class SPE2 relating to the time positioning of the connection in the pipe, for example a connection positioned at 10 o'clock or a connection positioned at 2 o'clock.

[0082] According to an example, in the first defect class DEFI corresponding to the assembly displacement, a second specification class SPE2 relates to the type of assembly displacement. For example, three second specification classes SPE2 are available from which the second specification learning function FONC2 selects the class relating to the defect present on the first image IM1. According to an example, the second specification classes SPE2 available for the assembly displacement are longitudinal displacement, off-centering and deviation.

[0083] According to an example, in the first defect class DEFI corresponding to the presence of roots, a second specification class SPE2 relates to the type of root present in the image. For example, three second specification classes SPE2 are available from which the second specification learning function FONC2 selects the class relating to the defect present in the first image IM1. These three second specification classes SPE2 are the presence of rootlets, the presence of a large root, or the presence of a complex set of roots.

[0084] The selection step SEL2 of the second specification class uses a second trained specification learning function FONC2. According to one embodiment, the second trained specification learning function FONC2 uses the first image IM1 as input. According to one embodiment, the second specification learning function FONC2 determines, from an analysis of the first image IM1, the second specification class SPE2 corresponding to the defect for which the first defect class DEFI was determined.

[0085] According to one embodiment, the second trained specification learning function FONC2 comprises a neural network architecture. According to one embodiment, the second trained specification learning function FONC2 comprises a multi-layer neural network. According to one embodiment, the second trained specification learning function FONC2 comprises a convolutional neural network. According to one embodiment, the second trained specification learning function FONC2 comprises a Transformer-type neural network.

[0086] The second trained specification learning function FONC2 has been trained previously to the implementation of the method. According to one embodiment, the training of the second trained specification learning function FONC2 comprises supervised training. During the training, a training image base of the second specification learning function SPE2 is used. The training image base of the second learning function FONC2 comprises a plurality of images, which are preferably images of the interior of pipes, which comprise labeled defects. Advantageously, each defect in the training base of the second learning function FONC2 belongs to the same first defect class DEFI. In this way, the second learning function FONC2 is trained on a basis allowing it to carry out the specification of a given defect from among the first defect classes DEFI.

[0087] According to one embodiment, the method comprises several second trained specification learning functions FONC2. Each second trained specification learning function FONC2 corresponds to a first fault class DEFI. In other words, each time the method calls upon a second trained specification learning function SPE2, the latter relates to the fault class DEFI which has been selected.

[0088] Each second trained specification learning function FONC2 is trained using a specific image base. The specific training image base comprises a set of labeled images comprising a defect of the first defect class DEFI to which the second specification class SPE2 is relative. The image labels include information specific to the first DEFI defect class and the second SPE2 specification class that is present on the images in question.

[0089] The specification label of each image used for training has preferably been assigned by an operator to the defect(s) present on the image. The specification label comprises information on the second specification class(es) SPE1 present on the image. According to one embodiment, the specific training image base comprises at least 10 labeled images of the interior of the pipeline. According to one embodiment, the specific training image base comprises at least 50 labeled images of the interior of the pipeline. According to one embodiment, the specific training image base comprises at least 100 labeled images of the interior of the pipeline. According to one embodiment, the training image base comprises at least 1000 labeled images of the interior of the pipeline.According to one embodiment, the training image base comprises at least 10,000 labeled images of the interior of pipelines.

[0090] According to one embodiment, the specific training image base of the second trained specification learning function FONC2 is the same base as the training image base of the first trained labeling learning function FONC1.

[0091] According to one embodiment, the images of the specific training image base of the second learning function FONC2 are a subset of the training image base of the first learning function FONC1. According to this embodiment, only the images of the training image base of the first learning function having as label the first defect class DEFI to which the second specification class SPE2 relates are used in the specific training image base. In this way, the second learning function of specification FONC2 is trained from images which each comprise one or more defects of the first defect class DEFI which one seeks to assign a specification class SPE2. This arrangement makes it possible to improve the training of the second learning function of specification FONC2 by a better selection of the training images.

[0092] According to one embodiment, the images of the specific training image base are raw IMBR images which have undergone at least one RED resizing step as described previously. According to one embodiment, the raw IMBR images have additionally or alternatively undergone an image truncation step. According to one embodiment, the raw IMBR images have additionally or alternatively undergone a geometric image transformation step as described previously. According to one embodiment, the images have undergone a TRAIT processing step for normalizing the light intensity as described above. In this way, the image may have undergone a brightness adjustment step, a contrast adjustment step, a brightness clipping step, and / or a saturation adjustment step. All these steps make it possible to obtain a set of training images specific to the second learning function of specification SPE2 which is homogeneous and therefore allows better training of the function. According to one embodiment, the TRAIT processing step for normalizing the light intensity comprises an image analysis step making it possible to deduce the processing to be carried out on the image to normalize it.

[0093] According to one embodiment, the number of images in the specific training image base of the second learning function FONC2 is increased using data augmentation techniques. This arrangement is particularly advantageous when the specific training image base is a subset of the training image base of the first learning function, since it makes it possible to increase a smaller number of images in the base. According to one embodiment, the specific training image base is increased by adding images from said base that have undergone geometric transformations. According to one embodiment, the geometric transformations comprise rotations of the images in the specific training image base. Each image that has undergone a rotation is added to the specific training image base.According to one embodiment, the specific training image base is augmented by performing processing on the colors present in the images. More specifically, color modification processing includes, but is not limited to, processing of image saturation, image brightness, or image contrast. The images thus processed are reincorporated into the specific training image base in addition to the unprocessed base images. In this way, the number of images in the specific training image base is increased with images having defects and color characteristics different from the initial images in the base. This arrangement makes it possible to increase the effectiveness of the training of the learning function.

[0094] According to one embodiment, the second trained specification learning function FONC2 is trained using a transfer learning method. A transfer learning method makes it possible to use a pre-trained image recognition model on image recognition problems as a starting point for training the second trained specification learning function FONC2. The pre-trained model is re-trained using the specific training image base previously described. According to one embodiment, the pre-trained model is trained on the recognition of images that are not images of the interior of pipes. According to one embodiment, the pre-trained model is a convolutional neural network. According to one embodiment, the pre-trained model is a convolutional neural network of the “MobileNet”, “ResNet”, “EfficientNet” or “DenseNet” type. According to one embodiment, the pre-trained model is a Transformer type neural network.

[0095] According to one embodiment, the pre-trained model is re-trained by re-training only a portion of the layers of the neural network based on specific training images.

[0096] According to one embodiment, the second trained specification learning function FONC2 is trained using a hybrid model which assembles a convolutional architecture and a recurrent architecture in a single neural network. Such a training mode makes it possible to better exploit the correlations between the different defects which are present on the same first image IM1.

[0097] According to one embodiment, the second specification learning function FONC2 selects the second specification class SPE2 according to a predefined rule. According to one example, the predefined rule makes it possible to associate a second specification class SPE2 with a first defect class. This association can be carried out automatically. This arrangement is particularly advantageous when a second specification class SPE2 represents a large majority of the cases associated with a first defect class DEFI. According to one example, the predefined rule associated with the second learning function FONC2 makes it possible to modify the first defect class DEFI selected previously. Business rules

[0098] According to one embodiment, the method comprises a step of selecting SEL3 a specific subclass SPE3 of the second specification class SPE2 selected. [Fig. 3] is a flowchart illustrating the selection of a rule in a REGI rule base. The rule base can be stored in a memory of the equipment used or in a remote memory accessible from a data network. The specific subclass SPE3 represents a third level of description of the pipeline fault which is identified on the first IML image. The step of selecting SEL3 the specific subclass SPE3 advantageously comprises a step of selecting SEL4 at least one rule within a REGI rule base as a function of the value of the first label LAB1. Thus, in this step a rule is selected which is relative to the first fault class DEFI and / or to the second specification class which have been identified previously.We therefore select a rule, which can correspond to a business rule which can be predefined from knowledge of one or more operators, or from the technical literature of a domain. of knowledge, which is characteristic of the nature of the type of defect identified. The rule base advantageously comprises a set of rules which are relative to the first labels LAB1. This rule base advantageously comprises at least one rule per first label LAB1. The rule defines an association between a characteristic value Vc of the first image IM1 and a predefined value defining a second label LAB2. The second label LAB2 is advantageously characteristic of the specific subclass SPE3. By characteristic value Vc of the first image IM1, we mean any value or data which is associated, derived from or deduced from the first image or one of its attributes.The characteristic value Vc of the first image can for example be a metadata associated with said first IML image The predefined value defines the second label LAB2 and is obtained by association with the characteristic value Vc of the first IML image The second label LAB2 labels the specific subclass SPE3 which is selected in this step of the method. The step of selecting SEL3 of the specific subclass then comprises a step of assigning the predefined value to the value of the second label LAB2. The step of selecting SEL3 of a specific subclass therefore allows the generation of a second label LAB2 specifying even more the type of defect identified in the first IML image.

[0099] According to one embodiment, the characteristic value Vc of the first image IM1 is its first label LAB1. According to this embodiment, the selected rule is characteristic of the first identified DEFI fault class and / or of the second specification class SPE2 which was previously identified. This arrangement is particularly advantageous, because it makes it possible to use a rule allowing the generation of the second label LAB2 which is relative to the type of fault which is identity. In addition, this arrangement also makes it possible to select a rule for generating the second label LAB2 as a function of the second specification class SPE2 which has been identified. According to a first example, a second specification class SPE2 identified by the method of the invention corresponds to a fault class of the “branch” type. A given rule is then automatically selected as a function of this specification class.In this case, the rule allows for the automatic assignment of an "open branch" value. In this case, it is understood that the value of the second specification class can be supplemented by a rule that assigns an additional characteristic to the defect that can only be of one type in a given case and / or that cannot be deduced directly from the image. Another case is to allow the representation of a predominant value of a label among a set of possible values. This representation advantageously helps an operator in the validation of the characterization of the defect in a large number of cases.

[0100] Indeed, one advantage is to give an operator an additional descriptor allowing him to more simply assess the qualification of the defect. One advantage is to promote the qualification, discrimination or separation of a defect from a list of defects presented by the classifier.

[0101] According to a second example, a second specification class SPE2 identified by the method of the invention corresponds to a defect class of the “crack” type, the rule in this case allows an “open crack” value to be automatically assigned. It is understood in this case that the value of the second specification class can be supplemented by a rule which assigns an additional characteristic to the defect. One advantage is to give an operator an additional descriptor allowing him to more simply assess the qualification of the defect or to invalidate the defect if this characteristic is not present. One advantage is to promote the qualification, discrimination or separation of a defect from a list of defects presented.

[0102] As a result, it is possible to specify how to select the second LAB2 label based on a sub-characteristic of the defect. This arrangement allows for greater finesse in the generation of the second LAB2 label.

[0103] According to one embodiment, the characteristic value Vc of the first image IM1 is generated from an image processing algorithm performed on the first image IM1. According to this embodiment, an analysis of the first image IM1 is performed to derive the characteristic value Vc. This arrangement makes it possible to define a rule for generating the second label LAB2 as a function of characteristics of the first image IM1. According to one example, the analysis of the image makes it possible to determine whether a dark or particularly dark area is located in a lower part of the first image IM1. This arrangement can make it possible to define whether a hole in the pipe is visible on the image IM1. This data extracted from the image makes it possible, thanks to the method of the invention, to assign a characteristic value to the image and therefore to the defect.As a result, the defect classification can be enriched since this characteristic value of the defect can be attributed to the second label LAB2 of the specification subclass SPE3. One interest is to improve the training of a learning function in particular by enriching the characterization of the defects.

[0104] According to another case, a brightness analysis makes it possible to generate a characteristic value of the image relative to a degree of confidence resulting from the image capture in order to characterize the context of the defect.

[0105] According to one embodiment, the method comprises a step of receiving a data item DON1 associated with the first image IM1. The data item DON1 associated with the first image IM1 is acquired by means of a user interface 30. The data item DON1 associated with the first image IM1 is therefore a data item which is defined according to an observation or a report made by a user. The data item associated with the first image IM1 comprises for example an indication following an observation made on the first image IM1 or on another image of the same pipeline by the user. According to one example, the data D0N1 associated with the first image IM1 is data providing information on the material of said pipeline. This data advantageously makes it possible to automatically rule out certain types of defects which are not compatible with a type of material of the pipeline. The data is recorded in a memory so as to be associated with the characteristic value Vc of the first image IM1. In this way, the observation entered in the user interface enters into the decision-making process for the allocation of the first labels LAB1 and second label LAB2.According to an example, when the user indicates that the pipeline does not have a coating material, the first DEFI defect class corresponding to the presence of a coating defect is not selectable during the SEL1 selection step of the first DEFI defect class.

[0106] In this example, the first data item DON1 may correspond to a characteristic value Vc of a descriptor of a defect. If the descriptor is encoded in the description of the defect, a rule may be used to exclude a defect class incompatible with the value of the descriptor. According to one example, the data item DON1 makes it possible to call upon a rule excluding a first defect class DEFI and / or a second specification class SPE2 already selected. According to this example, the first label LAB1 generated corresponding to the first defect class DEFI and / or to the second specification class SPE2 excluded may be deleted. These provisions make it possible to exclude labels and classes from the classification. According to one example, the user data item DON 1 is information on the material of the pipeline. This information makes it possible to exclude the first defect class DEFI corresponding to the coating defect.According to one example, if this class has already been selected, it is removed from the selection of first classes of defects. Advantageously, the first label LAB1 associated with this class is deleted. Advantageously, the collected data DON1 can allow retraining of the machine learning model of the first learning function and / or of the model of the second learning function. According to one embodiment, the operator can select an automatic retraining action when adding data DON1. According to one example, the operator can choose the model to be retrained.

[0107] According to one embodiment, the characteristic value Vc is entered by a user on the user interface. According to one example, the selected rule causes a message to be displayed to the user asking him to select the second label LAB2. According to one example, the rule causes the user to select a first DEFI defect class. For example, when one of the first DEFI defect classes such as breakage, crack, assembly displacement or branching defective is selected, then a message is displayed on the user interface asking him to choose between adding one of the following two first DEFI defect classes: soil visible by default or void visible by default. Such an arrangement is particularly advantageous, because it allows the user to be prompted on the presence or absence of a first DEFI defect class which can be critical in the maintenance of the pipeline. Indeed, if a void is visible, this means that a hole is present in the pipeline and that maintenance must be carried out quickly. In one example, the selected rule depends on the value of several first DEFI defect classes that have been identified. In one example, the selected rule influences the selection of the first DEFI defect class(es).According to one example, the first fault classes corresponding to the faulty branch and the penetrating branch can only be selected in the SEL1 selection step when the first DEFI fault class corresponding to the presence of a branch has previously been selected. In other words, certain first fault classes can only be selected when the first label LAB1 is characteristic of the presence of another characteristic first DEFI fault class. According to one example, the rule includes an indication not to select at least one first DEFI fault class. According to one example, a first DEFI fault class can no longer be selected when another first DEFI fault class has been selected. According to one example, the rule includes an indication not to select at least one second specification class SPE2 during the SEL2 selection step.

[0108] According to one embodiment, the data entered by the user is used as labeling data during the training of the first labeling learning function FONC1 and / or during the training of the second specification learning function FONC2. This arrangement makes it possible to include in the training data which are not included in the first label LAB1 and thus allows an increase in the effectiveness of the learning.

[0109] According to one embodiment, the method comprises an automatic labeling step. According to this example, for at least one DEFI fault class, the SEL2 selection of the second specification class is performed automatically to assign a second predetermined SPE2 specification class. According to one example, when the first DEFI fault class corresponding to the presence of a branch is selected, the second SPE2 specification class corresponding to the open branch is automatically selected. This arrangement advantageously makes it possible to report the fact that 90% of the branches are open. Advantageously, a user can confirm, for example after a visual observation of the first image IM1, that this second SPE2 specification class is justified. To do this, the observer uses the user interface.

[0110] An advantage of the invention is to present to an operator different classes of defect likely to characterize at least one defect in an image. The operator can then select a class presented on an interface to validate at least one of the classes of the defect. An advantage of the invention is to present different classes of defects in a hierarchical approach, that is to say to present heterogeneous classes with different levels of specificity. This approach allows a good compromise between a very generic approach not allowing to sufficiently help an operator during the validation of the defect class and a very specific approach likely to contain errors. The invention allows a mixed approach allowing to present to an operator different labels: DEFI, LAB1, LAB2 coming from two classifiers and a rule engine.

[0111] System for characterizing a pipeline defect

[0112] The invention also relates to a system designed to implement the method described above. The system according to the invention comprises a pipeline exploration device. According to one embodiment, the exploration device comprises at least one gripping means allowing it to be held in the hand by a user. This arrangement is advantageous when the pipeline is said to be visitable, that is to say when a user can move in it. According to one embodiment, the exploration device comprises an attachment means allowing it to be attached to the user. For example, the device comprises a strap or a clip allowing the device to be attached to a garment of the user. Such means also allow the device to be attached to a backpack of the user for example. According to one embodiment, the pipeline exploration device comprises an exploration robot 20.By exploration robot 20 is meant any type of device capable of moving in the pipeline. According to one example, the exploration robot comprises wheels allowing it to roll in the pipeline. According to one example, the exploration robot comprises at least one motor, preferably electric, allowing it to actuate the movement of said robot 20. The exploration device 20 comprises at least one optical sensor 22. The optical device 22 allows images to be taken of the interior of the pipeline, in particular of the first image IM1 and / or of the raw image IMBR.

[0113] The characterization system comprises at least one user interface 30. The user interface 30 is configured to display data sent by the exploration device 20. According to one example, the user interface is configured to display images received from an external system, for example images received from a device connected in a network, for example via the internet network, to the user interface. According to one example, the data sent by the exploration device 20 are received by an external system. The external system processes this data and can implement the method according to the invention, before sending display data to the user interface 30. This display data can include images of the pipeline. According to one embodiment, the interface is configured to display images taken by the optical sensor 22 of the exploration device 20. According to one embodiment, the user interface 30 displays in real time the images captured by the optical sensor 22. According to one embodiment, the user interface 30 is configured to allow the exploration device 20 to be controlled in real time. In this way, the user interface 30 allows the control of the exploration robot's motor and its direction. According to one embodiment, the exploration device 20 comprises a means for orienting the optical sensor 22.The orientation means allows the orientation of the optical sensor 22 in order to position at least the field of view of the sensor in order to take the best possible images. According to one embodiment, the exploration device comprises at least one lighting means. The lighting means makes it possible to illuminate the interior of the pipe to facilitate visibility when directing the exploration device 20. The lighting means also makes it possible to illuminate the shooting area to provide a first image IM1 a raw image IMBR having sufficient brightness for its analysis.

[0114] According to one embodiment, the exploration device 20 comprises an umbilical cable. The umbilical cable is a cable connecting the exploration device 20 to a base comprising the user interface or to the user interface 30, which are preferably located outside the pipeline. The cable advantageously serves to transmit data from the exploration device 20 to the base and vice versa. According to one embodiment, the cable also serves to transmit an electrical power supply to the exploration device 20. According to one embodiment, the connection between the user interface 30 and the exploration device 20 is carried out by a wireless connection, for example a radio connection. According to one embodiment, the exploration device 20 comprises at least one memory configured to record the images captured by the optical sensor 22.This arrangement makes it possible, for example, to take images of the inside of the pipeline and to save the images in the memory. In this way, the analysis can be carried out later after the image has been taken, without a connection to the user interface 30. According to one embodiment, the exploration device 20 is controlled by the user by means of the user interface 30. According to one embodiment, the control information for the exploration robot 20 is sent via the cable. According to one embodiment, the control information for the exploration robot 20 is sent via the wireless connection.

[0115] According to one embodiment, the user interface 30 is configured to allow the user to enter instructions during the method according to the invention.

[0116] We will now describe the use of the system according to the invention by the user.

[0117] When carrying out an inspection of a pipeline, the user takes images of the interior of the pipeline using the optical sensor 22 of the exploration device 20, which he preferably controls remotely using the user interface. Preferably, the user has, prior to the inspection, entered into the user interface one or more pieces of information or data relating to the inspected pipeline, such as the material of the pipeline for example. When the user identifies an area of ​​the interior of the pipeline which appears to have at least one defect, he takes an image of said portion of pipeline using the optical sensor 22. He then launches the method according to the invention to identify the defect(s) on the acquired image. According to one embodiment, the method is automatically launched following the taking of the image of the interior of the pipeline.During the process, the user may be required to enter information or make choices on the user interface 30. For example, the user may be required to confirm the first label(s) LAB1 and / or the second label(s) LAB2 that were generated during the pipeline fault characterization process.

[0118] According to one embodiment, a plurality of first LAB1 labels is generated during the method. Each first LAB1 label and / or each second LAB2 label is displayed on a graphical interface of a display of the user interface 30. The graphical interface comprises a component for validating the presence of each first LAB1 label and / or each second LAB2 label displayed. This component may for example be a dialog box or a table listing the first LAB1 labels and / or the second LAB2 labels generated. Each first LAB1 label and / or each second LAB2 label generated may be corrected or deleted by the user, preferably using an input component. This arrangement allows the user to modify or delete generated labels that are not correct. Advantageously, a report is generated which includes a list of the generated labels.According to one embodiment, modifying the value of a first displayed label LAB1 causes the machine learning model of the first learning function FONC1 and / or the machine learning model of the second learning function FONC2 to be retrained. This arrangement is particularly advantageous, as it allows the supervised learning of the learning functions to continue based on direct user feedback. More specifically, when the performance of identifying a type of defect is low or decreases, retraining of the machine learning model of the first learning function FONC1 and / or the machine learning model of the second learning function FONC2 can be performed. For example, data is calculated that is significant of the number of differences between the defects identified by . the method of the invention and the defects identified by the operators on the same images. When this data indicates too great a difference between the identifications by the method and the identifications carried out manually, then a retraining of the machine learning model of the first learning function FONC1 and / or of the machine learning model of the second learning function FONC2 is carried out. For example, the calculated data is compared to a characteristic performance threshold. According to one example, the calculated data is a percentage of correct identifications carried out by the method. According to one example, the characteristic performance threshold is 60%. According to this example, the method correctly identifies less than 60% of a given defect, compared to the identifications carried out manually by the user, then the model is retrained.

[0119] According to one embodiment, the modification of the first label LAB1 by an operator comprises an analysis of the number of symbols of the label LAB1 having changed, an analysis of the position of the symbol in the label having changed and the type of at least one symbol which has been modified, said analysis making it possible to generate a retraining action of at least one model of a second learning function. According to one example, the first label LAB1 comprises a sequence of symbols. The first symbols of the sequence of symbols define a characteristic code of the first DEFI fault class associated with said first label LAB1. The following symbol(s) define a characteristic code of the second specification class SPE2 associated with said first label LAB1.Analyzing which symbols have changed in the symbol sequence allows us to define, for example, that the correction action concerns the second learning function FONC2 relating to the said first label LAB 1 if it is the end symbols of the first label LAB1 that have been changed by the user. In the same way, if it is the start symbols of the first label LAB 1 that have been changed, the correction action concerns the first learning function FONC1.

[0120] The invention is not limited to the embodiments presented and other embodiments will become clear to those skilled in the art. Nomenclature:

[0121] 20: pipeline exploration device

[0122] 22: optical sensor

[0123] 30: user interface

[0124] 40: system for characterizing a surface singularity of a pipeline

[0125] GEN1: generation step

[0126] COMP: step of comparing the value of the probability

[0127] SEL1: step of selection of first singularity classes

[0128] SEL2: second specification class selection step

[0129] SEL3: step of selecting a specific subclass

[0130] SEL4: rule selection step

[0131] RED: raw image resizing step

[0132] TRAIT: intensity normalization processing step

[0133] PROB1: probability of presence of first class of singularity

[0134] LAB: label

[0135] DON1: data associated with the first image

[0136] FONC1: first learning function

[0137] FONC2: second learning function

[0138] CHALLENGE: first class of fault

[0139] SPE2: second specification class

[0140] SPE3: specific subclass of a second specification class

[0141] CLASS: specification classification

[0142] REGI: rule base

[0143] IM1: first image of the inside of a pipeline

[0144] IMRED: resized image

[0145] IMBR: raw image

Claims

Claims

1. A computer-implemented method for characterizing a pipeline defect, characterized in that it comprises the steps of: • Generating (GEN1) a plurality of probabilities of presence (PR0B1) of a plurality of first defect classes (DEFI) on a first image (IM1) of the interior of a pipeline from a first trained labeling learning function (F0NC1), said first learning function (F0NC1) implementing a machine learning model trained from a set of images each comprising a plurality of first labels characterizing at least one defect; • Comparing (COMP) for each first defect class (DEFI) the value of the probability with a threshold characteristic of each first defect class (DEFI); • Selecting (SEL1) each first defect class (DEFI) for which the estimated probability is greater than the threshold value associated with said class;• Selection (SEL2) for at least one first selected defect class (DEFI), of at least one second specification class (SPE2) from a plurality of second specification classes (SPE2), on the first image (IM1), from a second specification learning function (F0NC2); and • Generation (GEN2) of at least one first label (LAB1) characteristic of each first selected defect class (DEFI) and of each second selected specification class (SPE2).;

2. Method according to claim 1 characterized in that a plurality of first labels (LAB1) are generated, each first label (LAB1) preferably being displayed on a display with its probability of presence (PR0B1).

3. A method according to any preceding claim wherein the second specification learning function (F0NC2) is trained, said second specification learning function (F0NC2) implementing a trained machine learning model from the set of images each comprising a plurality of second labels characterizing at least one specification class (SPE2).

4. Method according to any one of claims 1 and 2 in which the second specification learning function (F0NC2) selects the second specification class (SPE2) according to a predefined rule making it possible to associate said second specification class with the first class of defects.

5. Method according to any one of the preceding claims which further comprises a step of selecting (SEL3) a specific subclass (SPE3) of a second specification class (SPE2), said step of selecting (SEL3) a specific subclass (SPE3) comprising the steps of: • Selecting (SEL4) at least one rule within a rule base (REGI) as a function of the value of the first label (LAB1), each rule defining an association between a characteristic value (Vc) of said first image (IM1) and a predefined second label (LAB2) of a specific subclass (SPE3); and • Generating data comprising the value of the first label (LAB1) and the value of the second label (LAB2).

6. Method according to the preceding claim characterized in that the characteristic value (Vc) of the first image (IM1) is its first label (LAB1).

7. Method according to claim 5 characterized in that the characteristic value (Vc) of the first image (IM1) is generated from an image processing algorithm applied to the first image (IM1).

8. Method according to claim 5 characterized in that it comprises a step of receiving a data item (D0N1) associated with the first image (IM1), said data item being acquired by means of a user interface (30), said data item being recorded in a memory so as to be associated with the characteristic value (Vc) of the first image (IM1), said data item (DON1) being compared with a plurality of attribute values ​​associated with each first label (LAB1), said comparison resulting in the exclusion of at least one first label (LAB1) from the list of first labels generated (LAB1) when the value of an attribute of said first excluded label is greater than or less than a given threshold.

9. Method according to the preceding claim in which the data (DON1) associated with the first image (IM1) is modeled in at least one discrete symbol in a natural language and standardized within a third label (LAB3) attached to the first label (LAB1) and / or to the second label (LAB2) comprising a plurality of discrete symbols in a natural language, said modeling and said standardization allowing relearning of the first trained labeling learning function (FONC1) and / or of the second trained specification learning function (FONC2).

10. Method according to any one of the preceding claims, characterized in that a plurality of first labels (LAB1) is generated, each first label (LAB1) being displayed on a graphical interface of a display, the graphical interface comprising a component for validating the presence of each first label (LAB1) displayed by an operator and a component for entering a new first label (LAB1) making it possible to correct the value of a first label displayed, significant data of the number of differences between the defects identified by the method of the invention and the defects identified by the operators on the same images being calculated.

11. Method according to the preceding claim in which the exceeding, by the significant data of the number of differences between the defects identified by the method of the invention and the defects identified by the operators, of a characteristic performance threshold, results in the retraining of the machine learning model of the first learning function and / or of the machine learning model of the second learning function.

12. Method according to one of claims 10 and 11 characterized in that the modification of the first label (LAB1) by an operator comprises an analysis of the number of symbols of the label (LAB1) having changed, an analysis of the position of the symbol in the label (LAB1) having changed and the type of the at least one symbol which has been modified, said analysis making it possible to generate a retraining action of at least one model of a second learning function.

13. Method according to any one of the preceding claims which comprises, before the generation step (GEN1), the steps of: • Receiving (REC) at least one raw image (IMBR) of inside a pipeline; • Resizing (RED) the image to obtain a resized image (IMRED); • Processing (TRAIT) to normalize the hue, saturation and / or brightness of the resized image (IMRED) to obtain the first image (IM1).

14. Method according to any one of the preceding claims in which the first trained labeling learning function (FONC1) and / or the second trained classification learning function (FONC2) is trained from a trained convolutional neural network comprising a plurality of layers, at least one of the layers of which is re-trained from the set of images each comprising a plurality of first labels (LAB1) characterizing at least one defect.

15. Method according to the preceding claim in which the pre-existing convolutional neural network used for training the first trained labeling learning function (FONC1) and / or for the second trained specification learning function (FONC2) is a network of the MobileNet, ResNet, EfficientNet, TRansformeur or DenseNet type.

16. Method according to one of the preceding claims in which the set of images comprising the plurality of first labels (LAB1) is increased by adding a second set of images resulting from the geometric and / or photometric transformation of at least part of the set of images comprising the plurality of first labels (LAB1).

17. A method according to any preceding claim wherein a loss function quantifying a deviation between predictions of the first learning function and the first labels is a function of one of the following types: • Binary cross-entropy, • Weighted cross-entropy, • soft Fl-score or dice loss, • macro soft Fl-score, or • double macro soft Fl-score.

18. Method according to any one of the preceding claims which further comprises a step of displaying on a user interface (30) each first label (LAB1).

19. System for characterizing a pipeline defect (40) characterized in that it comprises: • A pipeline exploration device (20) comprising at least one optical sensor (22); • A user interface (30) configured to display in real time a data stream from the optical sensor (22) of the pipeline exploration device (20), • The system (10) being configured to implement the method according to any one of the preceding claims.

Images