SYSTEM FOR IDENTIFYING FAULTS IN A WASTEWATER PIPE

Description

Technical Field

[0001] This paper describes a pipe fault identification system that facilitates the determination of the types of faults present in such a pipe. It also covers a method for training a statistical learning model for this purpose, as well as a method for identifying faults in a pipe.

[0002] Such an identification system can be used in particular for the maintenance of a network of non-visitable pipes, typically wastewater, in order to assess the health of the pipes, improve the accuracy of fault identification and facilitate the drafting of an inspection report. Previous technique

[0003] Underground pipeline networks, particularly wastewater networks, are subject to regular inspections to identify any defects present in the pipeline in order to plan possible maintenance of all or part of the network.

[0004] Since most of the pipes in these networks are inaccessible—that is, too small for an operator to enter—it is common practice to use a robot equipped with a camera to inspect the pipes for defects. An operator then controls the robot's movement and monitors the video feed transmitted by the robot's camera to identify any defects in real time. When a defect is detected, the operator is responsible for photographing the affected section of pipe and determining the type of defect from among the classifications defined by the NF-13508-2 standard. As defects are discovered, the operator compiles an inspection report that includes photographs of the defects along with their classification.

[0005] However, this standard comprises a large number of defect categories (around thirty) and an even greater number of subcategories (several hundred). The classification process is therefore particularly demanding and requires operators with significant skills and expertise. It is also tedious, as the operator must refer to the standard and its coding for each defect encountered. Furthermore, this manual identification process is susceptible to operator subjectivity and potential cognitive biases.

[0006] Attempts to automatically identify defects have been proposed. However, to date, these attempts are primarily academic and do not reflect real-world applications. In particular, most of the solutions proposed so far focus on characterizing a specific type of defect, such as cracks, but are unable to identify other types of defects, such as the presence of deposits or roots. The documents Guo W and AI, "Automated defect detection for sewer pipeline inspection and condition assessment," US 2021 / 082098 and US 2014 / 168408, describe examples of methods for detecting and identifying defects.

[0007] While some more general solutions exist, they are limited to a small number of defect types, rarely more than a few, and therefore do not offer sufficient accuracy to meet the level of precision required by these inspection reports. Furthermore, most solutions proposed to date are limited to classifying a single defect and are unable, in particular, to simultaneously classify several defects when multiple defects are present in the same photograph, which frequently occurs in practice, as some defects can contribute to the appearance of a second defect of a different type.

[0008] Furthermore, most existing solutions propose using special cameras, such as fisheye lenses, to create a 3D reconstruction of the pipeline, including its defects. However, these solutions are not preferred because they require investment in specific equipment and training operators in a completely different way of working. They also generate significant image distortions.

[0009] There is therefore a real need for a pipe fault identification system that is free, at least in part, from the drawbacks inherent in the aforementioned known solutions. Description of the invention

[0010] The present invention relates to a fault identification system in a pipeline according to claim 1.

[0011] Thanks to such a statistical learning model, it is possible to determine the type of defects present in photographs or at least to obtain a sufficiently accurate estimate of the probability of each defect's presence. This provides the operator with decision support for identifying each defect, for example, in the form of a suggested classification. The creation of inspection reports is thus facilitated. The accuracy and rigor of the inspection reports are also improved.

[0012] In particular, since the statistical learning model is trained to recognize a large number of different defect types, compliance with the NF-13508-2 standard, when required, is easier to achieve without significant additional effort from the operator. Furthermore, because the statistical learning model can simultaneously determine the type of several defects present in the same photograph, the operator is not required to spend time trying to isolate each defect in a separate photograph: the inspection is thus faster and the inspection report shorter, as a single photograph can describe several defects.

[0013] Furthermore, such an identification system is based on the conventional inspection process as it is currently carried out, using a robot to scan the pipeline for defects, taking photographs of the sections of pipeline exhibiting these defects, and determining the type of defect based on these photographs. The transition from the conventional process to this process assisted by the statistical learning model is therefore straightforward: in particular, existing robots can be used, and additional operator training is limited to manipulating the results of the statistical learning model.

[0014] In particular, as in conventional inspections, the determination is based on individual photographs, not on a video or a time and / or space-based series of photographs of the same defect. This allows, in particular, for the same operational conditions as past inspections to be replicated, and therefore for the operator to leverage the extensive photographic database of past inspections during the training of the statistical learning model, thus ensuring satisfactory reliability of the statistical learning model from its very first use.

[0015] In some embodiments, the pipeline is a non-visible pipeline, having a nominal diameter of less than 1600 mm, preferably between 200 and 500 mm.

[0016] In some embodiments, the pipe is a wastewater pipe.

[0017] In some embodiments, the pre-established list includes at least 15 types of defects, preferably at least 20 types of defects, and even more preferably at least 30 types of defects. The more different types of defects the statistical learning model is able to identify, the more the operator's residual workload is reduced, particularly when it is necessary to comply with the standard's coding requirements.

[0018] In some embodiments, all types of defects in the pre-established list are part of the types of defects described by the NF-13508-2 standard.

[0019] In certain embodiments, the pre-established list includes at least the following types of defects: crack (BAB); surface degradation (BAF); defective connection (BAH); visible soil (BAO); deposits and adherent deposits (BBC+BBB); rupture / collapse (BAC); penetrating branch (BAG); pipe displacement (BAJ); roots (BBA); and connection (BCA). The trigrams in parentheses correspond to the codes of standard NF-13508-2.

[0020] In some embodiments, each photograph contains at least one defect. This simplifies the task of developing a statistical learning model, which can assume that at least one defect is present. The statistical learning model can then focus on determining the type of defects, thereby improving its accuracy in this task. In particular, the statistical learning model can be trained for this purpose on a database containing exclusively photographs with at least one defect.

[0021] In some embodiments, at least some photographs include several defects.

[0022] The robot is configured to be operated by a human operator. However, in other aspects not covered by the claims, the robot could be configured to traverse the pipeline autonomously.

[0023] In some embodiments, the robot's camera is orientable. Preferably, it is oriented by the operator. In the context of autonomous robot control not covered by the claims, it could be oriented autonomously.

[0024] In some embodiments, the robot's camera is an RGB camera. It is therefore a standard camera such as those already found on this type of robot.

[0025] In some embodiments, the robot's camera is equipped with a lens having a focal length of 20 mm or greater, preferably 25 mm or greater, and even more preferably 30 mm or greater. It is therefore a conventional camera such as those already found on this type of robot. In particular, the robot's camera is not equipped with a fisheye lens. In this way, the photographs taken by the camera do not exhibit significant distortion, which facilitates the recognition process for a potential operator. Furthermore, image recognition by the statistical learning model is avoided if it has been trained on distortion-free photographs.

[0026] The photographs are triggered manually by an operator.

[0027] In some embodiments, the processing system includes an automatic fault detection module configured to detect at least certain types of faults based on the video stream from the robot's camera. Such an automatic fault detection module is particularly useful for autonomous robot operation: the pipeline navigation, fault detection, camera orientation, and acquisition of relevant photographs can all be performed autonomously, without operator intervention. In this case, this automatic fault detection module is preferably separate from the statistical learning model.

[0028] The processing system is configured to receive the photographs taken by the robot without delay and transmit them immediately to the statistical learning model for defect identification. In other words, defect identification occurs in real time, as defects are detected in the pipeline.

[0029] In other aspects not covered by the claims, the determination can be made asynchronously, based on all the photographs recorded during the journey of the section.

[0030] In some embodiments, the identification system further includes a photographic database comprising at least 1,000, preferably at least 100,000, and even more preferably at least 1,000,000 photographs of pipe sections exhibiting at least one defect, each photograph being associated with the type of each defect visible in the photograph. The larger this database, the more advanced the training of the statistical learning model will be, and therefore the more accurate it will be. Moreover, as mentioned above, it is preferable to include only photographs exhibiting at least one defect in order to specialize the statistical learning model in classifying defects rather than simply detecting them.

[0031] In some embodiments, the statistical learning model was trained on said database.

[0032] In some embodiments, the statistical learning model is a multi-label classifier. Unlike a multi-class classifier which provides multiple but mutually exclusive classes, a multi-label classifier provides multiple but non-mutually exclusive classes: in other words, based on a given photograph, such a multi-label classifier is able to determine the type of several defects belonging to different classes.

[0033] In some embodiments, the statistical learning model includes at least one neural network, preferably of the deep convolutional type.

[0034] In some embodiments, the statistical learning model includes a hybrid neural network, combining convolutional and recurrent architectures. Such a hybrid architecture makes it possible, in particular, to take advantage of correlations that may exist between certain types of defects, as some defects may promote the appearance of a second defect of a different type.

[0035] In some embodiments, the statistical learning model includes several classifiers trained on different datasets. This makes it possible to combine the predictions of these models to improve the performance and robustness of the predictions.

[0036] In some embodiments, the learning model is configured to determine a probability of presence for each of the types of defects in the pre-established list.

[0037] In some embodiments, the learning model is configured to determine that a particular type of defect is present if it determines that the probability of that defect exceeds a predetermined threshold. Setting this threshold allows for a compromise between suggesting a large number of defects, at the risk of generating a significant number of false positives that the operator must correct manually, thus reducing inspection speed, and suggesting a smaller number of defects, with a lower probability of false positives but a higher risk of false negatives, thereby compromising inspection quality.

[0038] In some embodiments, each type of defect is associated with a custom threshold. Indeed, it may be advantageous to offer the operator a higher proportion of false positives for certain types of defects that are frequently overlooked by operators or that are critical to operations, and a higher proportion of false negatives for other types of defects that the operator can easily detect and correct independently.

[0039] According to the invention, the identification system comprises a human-machine interface with at least one screen. The processing system is configured to display the current photograph on the screen, along with a proposed classification for each of the defects recognizable in the photograph. The identification system thus provides decision support to the operator, who retains control over the final classification. This accelerates the inspection process while ensuring a very high level of reliability, as any errors in the statistical learning model can be corrected by the operator.

[0040] In some embodiments, the processing system is configured to also display the position of detected defects as an overlay on the photograph. Implementing an algorithm to visualize activation maps provides an indication of the defect's location within the image by highlighting areas likely to contain them.

[0041] In some embodiments, the processing system is configured to receive confirmation or correction of the proposed defect, entered by an operator via the human-machine interface, and to record the final classification for each defect in the photograph. The correction may include correcting the proposed type for a given defect and / or adding an additional defect not recognized by the statistical learning model, along with its classification as determined by the operator.

[0042] In some embodiments, the processing system is configured to transmit the photograph and classification of each of the defects in the photograph to report writing software.

[0043] In some embodiments, the processing system is configured to add each new photograph to the database, along with the final defect classification. This addition is preferably performed after validation or correction by the operator. This allows the database to be enriched with each new inspection.

[0044] In some embodiments, the processing system is configured to evaluate the statistical learning model based on confirmations or corrections entered by the operator.

[0045] In some embodiments, the processing system is configured to adjust the parameters of the statistical learning model based on confirmations or corrections entered by the operator. This adjustment can also take into account the model's prediction scores. The statistical learning model thus continues to learn thanks to the corrections entered by the operator, thereby improving its accuracy for future inspections. Such an adjustment can be performed at regular intervals or when excessive drift is detected. Preferably, such an adjustment can be made outside of the inspection phases, i.e., after the robot has completed its path through the pipe.

[0046] In some embodiments, the processing system takes the form of a central processing unit. However, the processing system could also be at least partially delocalized: for example, the database and / or the statistical learning model could be hosted on a remote server, for example within a cloud computing environment.

[0047] In some embodiments, the identification system is mounted on a vehicle, for example a utility vehicle such as a van. The identification system can therefore be brought very close to the pipeline to be inspected in order to locate defects, classify them, and generate the report on-site, as the robot travels along the pipeline.

[0048] This presentation also relates to a method for training (not claimed) a statistical learning model for identifying defects in a pipeline, comprising the construction of a master database of photographs comprising at least 1,000, preferably at least 100,000, preferably even more at least 1,000,000, photographs of sections of pipeline exhibiting at least one defect, each photograph being associated with the type of each of the defects visible in the photograph, and the training of the statistical learning model on the master database thus constructed.

[0049] Such a statistical learning model can notably be used in any of the embodiments of the identification system presented above.

[0050] In some embodiments, the statistical learning model is a multi-label classifier.

[0051] In some embodiments, the statistical learning model includes at least one neural network, preferably of the deep convolutional type.

[0052] In some embodiments, the statistical learning model includes a hybrid neural network, combining convolutional and recurrent architectures.

[0053] In some embodiments, the statistical learning model includes several classifiers trained on different datasets. This makes it possible to combine the predictions of these models to improve the performance and robustness of the predictions.

[0054] In some embodiments, the learning process includes a data augmentation step in the main database, during which certain photographs are synthesized and / or duplicated by applying at least one geometric and / or photometric transformation while preserving the type of each visible defect in the photograph. The applied transformations may include, but are not limited to, rotations, cropping, and modifications to contrast, brightness, and / or saturation. On the one hand, such data augmentation allows for a significant increase in the number of photographs and their diversity, thereby deepening the training of the statistical learning model, particularly with regard to certain rare defects that are only sparsely represented in the initial database.On the other hand, by training the statistical learning model to recognize such flaws despite such modifications, we increase the robustness of the statistical learning model.

[0055] In some embodiments, the main database also includes, for each photograph, information about the pipe material. This additional information provides a complementary parameter to the statistical learning model, facilitating its recognition of defects based on the pipe material. In particular, certain types of defects appear more frequently in pipes made of a specific material.

[0056] In some embodiments, the main database also includes, for each photograph, information regarding the location of each defect. This information may include the defect's clockwise position, that is, its angular position along the circumference of the pipeline. Such additional information provides a supplementary parameter to the statistical learning model, facilitating its recognition of defects based on their location within the pipeline. In particular, certain types of defects appear more frequently in specific areas of the pipeline.

[0057] The present invention also relates to a method for identifying defects in a pipeline according to claim 11.

[0058] The advantages of this fault identification method stem from the advantages described above for the fault identification system. Furthermore, this fault identification method may exhibit all or some of the additional characteristics described above regarding the fault identification system.

[0059] In some embodiments, the statistical learning model has been trained using a training method according to any of the preceding embodiments.

[0060] The aforementioned features and advantages, as well as others, will become apparent upon reading the detailed description that follows, along with examples of implementations of the identification system and the proposed learning and identification processes. This detailed description refers to the attached drawings. Brief description of the drawings

[0061] The attached drawings are schematic and are primarily intended to illustrate the principles of the presentation.

[0062] In these drawings, from one figure to another, identical elements (or parts of elements) are identified by the same reference symbols. [ Fig. 1 ] There figure 1 is a schematic view of an example identification system. Fig. 2 ] There figure 2 represents this identification system in use at an inspection site. Description of the implementation methods

[0063] To make the explanation more concrete, an example of a fault identification system in a pipeline, along with the associated training and identification methods, is described in detail below, with reference to the accompanying drawings. It should be noted that the invention is not limited to these examples.

[0064] There figure 1 represents an example of an identification system 1. This identification system includes a processing system, in the form of a central unit 20 housed in a vehicle 3, and a robot 40, which can be stored and transported in the vehicle 3. The vehicle 3 further includes a control station 5 at which an operator 6 can be seated.

[0065] The central unit 20 includes a database 21, a statistical learning model 22, a human-machine interface 23 and editing software 24. It also includes all the usual components useful for its operation, including a processor, memory and a storage disk.

[0066] Database 21 contains a very large number, exceeding 1,000,000, of photographs of pipe sections taken from inside the pipe. In all these photographs, at least one defect is visible, and in at least some photographs, several defects are visible in the same photograph. For each photograph in database 21, the list of defects visible in that photograph is associated with the photograph in question in database 21: more precisely, for each defect, the type of defect according to standard NF-13508-2 is specified.

[0067] This standard codifies the various defects that can be encountered in a pipeline and classifies them into categories and subcategories. Within each category, each type of defect is coded using a specific trigram. Here are some examples of defect types with their associated trigram in parentheses: crack (BAB); surface degradation (BAF); defective connection (BAH); visible soil (BAO); deposits (BBC); adherent deposits (BBB); rupture / collapse (BAC); penetrating branch (BAG); pipe displacement (BAJ); roots (BBA); and connection (BCA).

[0068] The database thus includes at least the trigram for each defect, corresponding to its category as defined by standard NF-13508-2, sometimes with the additional indication of its subcategory. Furthermore, certain related categories may sometimes be grouped into a common class: for example, depots and member depots may be grouped into a class bearing the label "BBC+BBB".

[0069] In practice, database 21 is initially created from past inspection reports: thus, the type of all defects recorded in database 21 was determined by an operator during a past inspection or, at least, was checked and validated by an operator.

[0070] In addition, database 21 can associate with each defect information concerning the material of the pipe in which this defect was found as well as information concerning its location within the pipe, for example its clock position along the circumference of the pipe.

[0071] In this example, database 21 is available locally, but it is naturally understood that it could just as easily be hosted in a computer cloud.

[0072] The statistical learning model 22 is a deep convolutional neural network forming a multi-label classifier. As symbolized by arrow 11, it was trained on the dataset 21 to be able to determine the type of defect(s) present in a photograph of a section of pipe. In this example, the statistical learning model 22 was thus trained to distinguish ten different types of defects and to calculate, from a given photograph, the probability of the presence of each of the ten types of defects known to the statistical learning model 22.

[0073] More specifically, in this example, the multi-label classifier was built from standard architectures by retraining a number of layers with images from database 21.

[0074] Naturally, this training could be carried out upstream, in a dedicated computer system, separate from the central unit 20, before including the statistical learning model 22 resulting from this training in the processing system 20.

[0075] The statistical learning model 22 is also equipped with a thresholding function, with a threshold that can be customized for each type of defect, allowing the statistical learning model 22 to conclude whether or not each type of defect is present by checking if the probability of the presence of that type of defect is greater than the corresponding threshold.

[0076] In this example, the statistical learning model 22 is available locally, but it is naturally understood that it could just as easily be hosted in a computer cloud and queried remotely by the processing system 20 via a network connection.

[0077] The human-machine interface 23 includes a screen and an input device such as a keyboard and / or mouse. It allows the central processing unit 20 to transmit information to the operator 6, primarily in the form of images displayed on the screen, and to collect information entered by the operator 6 using the input device. The human-machine interface 23 also includes a control device for the robot 40. It may also include other elements such as a speaker and / or a microphone.

[0078] The 24 editing software is designed specifically for writing inspection reports. It may also include means for printing and / or electronically transmitting documents.

[0079] The robot 40 includes a locomotion device 41 and a steerable camera 42. The locomotion device 41 here includes wheels driven by a motor, but it could be of a completely different type; it could, for example, include tracks.

[0080] Camera 42 is an RGB (red-green-blue) camera, that is, a standard camera working in the visible light range; it also has a standard lens, which is not a fisheye type and therefore has negligible distortion.

[0081] The robot 40 also includes all the usual components necessary for its operation, including a battery and a communication device. In particular, communication with the central unit 20 can be wired, using a cable 43, or wireless, for example via Wi-Fi or radio waves.

[0082] An example of a fault identification procedure using such a fault identification system 1 will now be described with reference to figures 1 et 2 In particular, the figure 2 This illustrates the identification system 1 in use at an inspection site 90. The target of this inspection is a wastewater pipe 91 buried under the roadway 92. The pipe 91, having a nominal diameter typically between 200 and 500 mm, extends between a first manhole 93a and a second manhole 93b. This pipe has several defects 94, for example cracks 94a, surface degradation 94b, or a deposit 94c.

[0083] During the inspection, vehicle 3, housing the identification system 1, is parked near the first manhole 93a, and robot 40 is inserted into pipe 91 through the first manhole 93a. An operator 6 then takes their place at the control station 5 and begins to operate robot 40 using the human-machine interface 23: operator 6 is thus able to control the robot 40's locomotion device 41 and camera 42, and receives the video feed from camera 42 on a screen. This control system is represented by arrow 12, which corresponds to cable 43.

[0084] Operator 6 then gradually advances the robot 40 into the pipe 91 and monitors the video feed from camera 42, searching for defects. When they believe they have spotted a defect 94, operator 6 can adjust the camera's position to better visualize it and then takes a photograph of it. When two separate defects 94 are close to each other, operator 6 can take a single photograph of both.

[0085] As symbolized by arrow 13, this photograph is then passed to the statistical learning model 22. Based on this photograph, the statistical learning model 22 determines the probability of each type of defect known to the model in question. The statistical learning model 22 then applies its thresholding function to establish a proposed classification for each of the 94 defects recognized by the model.

[0086] As symbolized by arrow 14, this classification proposal is transmitted to operator 6 via the human-machine interface 23. Operator 6 can then check this proposal and, as appropriate, confirm or correct it, thus resulting in a final classification for each of the defects 94 in the photograph in question.

[0087] As symbolized by arrow 15, the photograph in question and the final classification of the defects 94 present on this photograph are transmitted to the editing software 24 in order to write, in a semi-automatic manner, an inspection report, the operator 6 being able to complete and / or correct the latter at any time using the human-machine interface 23.

[0088] Furthermore, as symbolized by arrow 16, the photograph in question is added to database 21, along with the final classification of the 94 defects present in that photograph. If, at this stage, the classification proposed by the statistical learning model 22 proved inaccurate and was therefore corrected by operator 6, the central processing unit 20 records this information in memory so that the statistical learning model can be evaluated and its drift estimated. At regular intervals, for example, once every two months, the parameters of the statistical learning model 22 can be adjusted to correct this drift.

[0089] Operator 6 can then proceed to search for a new defect 94 in pipe 91. The steps of the identification process are repeated until robot 40 has finished traversing pipe 91. Operator 6 can then finalize the inspection report, retrieve robot 40 and store it back in vehicle 3.

[0090] Although the present invention has been described with reference to specific embodiments, it is evident that modifications and changes can be made to these examples without departing from the general scope of the invention as defined by the claims. In particular, individual features of the various embodiments illustrated / mentioned can be combined in additional embodiments. Therefore, the description and drawings should be considered in an illustrative rather than restrictive sense.

[0091] It is also evident that all the characteristics described with reference to a method are applicable, alone or in combination, to a device, and conversely, all the characteristics described with reference to a device are applicable, alone or in combination, to a method. To make the explanation more concrete, an example of a device is described in detail below, with reference to the accompanying drawings. It should be noted that the invention is not limited to this example.

Claims

1. A system for identifying a fault in a pipeline, comprising: a robot (40) equipped with a locomotion device (41) and a camera (42), configured so as to be controlled by an operator, to travel along the pipeline (91) and to take photographs of portions of the pipeline (91) suspected of having one or more faults (94) on the basis of a manual triggering by the operator having located the fault(s), a processing system (20), equipped with a statistical learning model (22) configured to determine at the time, i.e. as and when the operator locates a fault on a portion of the pipeline and a photograph of said portion of pipeline is taken, on the basis of such a photograph called the current photograph, the type of each fault (94) recognised on the current photograph from a pre-established list of at least 10 fault types, and a human-machine interface (23) comprising at least one screen, the processing system (20) being configured to display the current photograph on the screen accompanied by a classification proposal for each of the faults (94) recognisable in the photograph, wherein the processing system (20) is configured to receive a confirmation or correction of the proposal, entered by an operator (6) via the human-machine interface (23), and to record the final classification for each of the faults (94) in the photograph.

2. The identification system according to claim 1, wherein the robot (40) is configured to travel along an inaccessible wastewater pipeline having a nominal diameter of less than 1600 mm, preferably between 200 and 500 mm.

3. The identification system according to claim 1 or 2, wherein the processing system is configured to receive, as input, photographs each comprising at least one fault (94), at least some photographs comprising a plurality of faults (94).

4. The identification system according to any one of claims 1 to 3, wherein the robot (40) is configured to be controlled by an operator (6), wherein the camera (42) of the robot (40) can be oriented by the operator (6), and wherein the robot (40) is configured so that the photographs are triggered manually by the operator (6).

5. The identification system according to any one of claims 1 to 4, wherein the processing system (20) is configured to receive, without delay, the photographs taken by the robot (40) and to transmit them, without delay, to the statistical learning model (22) for determination of the types of faults.

6. The identification system according to any one of claims 1 to 5, further comprising a database (21) of photographs comprising at least 1000, preferably at least 100,000, more preferably at least 1,000,000, photographs of pipeline portions having at least one fault, each photograph being associated with the type of each of the faults visible in the photograph, and wherein the statistical learning model (22) has been trained on said database.

7. The identification system according to any one of claims 1 to 6, wherein the statistical learning model (22) is a multi-label classifier comprising at least one neural network, preferably of the deep convolutional type.

8. The identification system according to any one of claims 1 to 7, wherein the learning model (22) is configured to determine a probability of presence for each of the fault types in the pre-established list, wherein the learning model (22) is configured to determine that a particular fault type is present if it determines that the probability of presence of that fault type is greater than a predetermined threshold, each fault type being associated with an individualised threshold.

9. The identification system according to any one of claims 1 to 8, wherein the processing system (20) is configured to additionally display, superimposed on the photograph, the position of the recognised faults (94).

10. The identification system according to any one of claims 1 to 9, wherein the processing system (20) is configured to add each new photograph to the database (21), in association with the final classification of faults (94), and wherein the processing system (20) is configured to evaluate the statistical learning model (22) and / or adjust the parameter setting of the statistical learning model (22) as a function of the confirmations or corrections entered by the operator (6).

11. A method of identifying a fault in a pipeline, comprising the following steps: providing a statistical learning model (22), configured to determine, on the basis of a photograph of a portion of pipeline, the type of each fault recognisable in the photograph from a pre-established list of at least 10 fault types; travelling along the pipeline (91) with the aid of a robot (40) equipped with a locomotion device (41) and a camera (42) and which is controlled by an operator, taking photographs of portions of the pipeline (91) suspected of having one or more faults (94) on the basis of manual triggering by the operator having located the fault(s), transmitting each photograph to the statistical learning model (22) and determination by the latter of the type of each fault (94) recognisable in each photograph, the fault being determined at the time, i.e. as and when the operator locates a fault in a portion of the pipeline and a photograph of said portion of pipeline is taken, said photograph being referred to as the current photograph, displaying on a screen of a human-machine interface (23) a classification proposal for each of the faults (94) recognisable in the current photograph, entering by an operator, via the human-machine interface (23), a confirmation or correction of the proposal; and recording the final classification for each of the faults (94) in the photograph.