METHOD FOR FAULT DETECTION IN A TUBE HEAT EXCHANGER

Description

Technical field

[0001] The present invention relates generally to the field of maintenance of a tube heat exchanger. More specifically, the invention applies to the search for defects in the tubes of the heat exchanger from signals from a measuring probe passing inside each tube. Technological background

[0002] A tube heat exchanger such as a steam generator is generally composed of a bundle of tubes in which the hot fluid circulates, and around which circulates the fluid to be heated. For example, in the case of a steam generator in a PWR-type nuclear power plant, the steam generators are heat exchangers that use the energy from the primary circuit resulting from the nuclear reaction to transform the water in the secondary circuit into steam that will power the turbine and thus produce electricity.

[0003] The steam generator brings the secondary fluid from a liquid water state to a vapor state just at the saturation limit, using the heat of the primary water. This circulates in tubes around which the secondary water circulates. The steam generator outlet is the highest point in temperature and pressure of the secondary circuit.

[0004] The exchange surface, physically separating the two circuits, is thus made up of a tubular bundle, composed of 3500 to 5600 tubes, depending on the model, in which circulates the primary water brought to high temperature (eg 320°C) and high pressure (eg 155 bars).

[0005] To ensure the proper functioning of the steam generator, it is necessary to periodically inspect the tubes. Since they are not directly accessible to an operator, a measuring probe is used which runs inside the tubes to obtain measurement signals, the analysis of which should allow the detection and characterization of any tube defects.

[0006] A typical example of a measuring probe is an eddy current probe, also known as a SAX probe. Eddy currents appear in a conductive material when the magnetic flux in its vicinity is varied. An eddy current probe is thus passed through a steam generator tube and a measurement signal is measured, which is a function of the environment in which the probe is located, from which information about anomalies in the tube can be extracted. A variation in magnetic induction, typically by a coil in which an alternating current flows, generates eddy currents, the induced variation of the magnetic field being detected. Typically, the voltage difference generated by the variation in the coil's impedance is measured.

[0007] Once the measurement signals resulting from the passage of a measuring probe through the steam generator tubes have been obtained, they still need to be analyzed to detect any anomalies that would be representative of the presence of a defect in a tube. Usually, this analysis is carried out by an operator who reviews the measurement signals. This is an extremely time-consuming method, particularly due to the large number of tubes. This approach also presents a significant risk of error, relying on the experience and skill of a fallible human operator.

[0008] Furthermore, the simplest approach is to detect a local amplitude overshoot, by comparing it to a threshold. While this approach works in areas of a tube where it is not in contact with other elements, it is ineffective when other elements are in the vicinity of the tubes. In addition, tubes may have variations in geometry that can make it difficult to interpret amplitude variations.

[0009] There figure 1shows an example of a measuring probe 1 running through a tube 2 of a steam generator to acquire a measuring signal. The tubes 2 of the steam generator are held by spacer plates 4 generally arranged perpendicular to the tubes passing through them. In order to allow the vaporizing water to pass through, passages 6 of tubes 2 of these spacer plates 4 are foliated, that is to say that their shape has lobes around the tubes 2. As the water passes from the liquid state to the vapor state, it deposits all the materials it contained. If the deposits of material are made in the lobes, they reduce the free section of the passage 6: this is clogging, which is therefore the progressive blocking, by deposits, of the holes or passages 6 intended for the passage of the water / steam mixture.

[0010] A tube 2 of a steam generator generally comprises a straight portion 2a of the hot branch, where the hot heat transfer fluid arrives in the tube, followed by a curved portion 2b, designated as the bend, and finally a straight portion 2c of the cold branch, from where the cooled heat transfer fluid leaves. Anti-vibration bars 8 are then present at the level of the curved portion 2b.

[0011] Thus, the presence of the spacer plates 4, of anti-vibration bars 8, or of a clogging deposit in the passages 6 of the spacer plates 4 influences the measurement signal, and a possible defect can no longer be detected simply by looking for a variation in the amplitude of the measurement signal. One solution is to compare the measurement signal in a particular zone with a reference signal corresponding to the measurement signal in this zone in the absence of a defect. For example, in the case of the passage of a spacer plate 4, the measurement signal can be compared to a previously learned reference signal which corresponds to the passage of a spacer plate 4 for a tube 2 without a defect.

[0012] There figure 2thus shows an example of variations in the amplitude of a component of a measurement signal acquired by an eddy current probe 1 along a tube 2 of a steam generator, as a function of time and therefore of the path of the probe 1 along the tube 2. A first part 10a corresponds to the rectilinear portion 2a of the hot branch, where the hot heat transfer fluid arrives in the tube. We note the presence of sudden variations in amplitude linked to the crossings of the upper edge of the tube plate, the flow distribution plate and then (from the third) the spacer plates 4, at roughly regular intervals. Then comes a second part 10b corresponding to the curved portion 2b of the tube. Here again sudden variations in amplitude appear, but this time linked to the crossings of the anti-vibration plates 8 by the measuring probe 1.Compared to the first part 10a, the period of sudden amplitude variations is shorter and less regular. Finally, a third part 10c corresponds to a rectilinear portion 2c of the cold branch, from which the cooled heat transfer fluid leaves. Here we find the sudden amplitude variations linked to the crossings of the spacer plates 4, then of the flow distribution plate and of the upper edge of the tube plate encountered in the first part 10a.

[0013] Theoretically, the reaction of the measuring probe 1 to the passage of a known element can be anticipated, and it would then be possible to use the corresponding reference signal. However, the reference signal may not be representative of the environment of the measuring probe 1, and therefore may not allow defects to be detected. As mentioned above, it happens for example that a passage 6 of tube 2 of a spacer plate 4 is clogged, with a deposit of unknown size and composition, which makes the comparison with a reference signal ineffective for detecting a defect of the tube 2. In addition, the movement of the measuring probe 1 in the tube 2 may present local variations in speed, in particular in the curved zone of the tube since the measuring probe 1 no longer moves in a straight line. The position of external elements such as the anti-vibration bars 8 may also not be known precisely.This can result in desynchronization of the measurement signals, even though it is necessary to know the correspondence between the measurement signal and the position of the measurement probe 1 in order to be able to use a possible reference signal at the right place in the measurement signal.

[0014] Therefore, there is a need for a method for detecting defects in heat exchanger tubes that can automatically identify tubes potentially containing a defect.

[0015] Non-destructive testing of heat generator tubes including synchronization of each measurement signal with a reference time series by applying dynamic time warping, DTW, to the measurement signal and the reference time series is known from VAN VAERENBERGH STEVEN ET AL: "Pattern Localization in Time Series Through Signal-To-Model Alignment in Latent Space", 2018 IEEE INTERNATIONAL CONFERENCE ON ACOUSTICS, SPEECH AND SIGNAL PROCESSING (ICASSP), IEEE, (2018-04-15), pages 2711-2715, DOI: 10.1109 / ICASSP.2018.8461890. Presentation of the invention

[0016] A method of maintaining a tube heat exchanger is proposed comprising: obtaining measurement signals resulting from the passage of a measurement probe in tubes of the heat exchanger, determining a reference time series corresponding to an average, at each instant, of the measurement signals, synchronizing each measurement signal with the reference time series by applying a dynamic time warping, DTW, on said measurement signal and the reference time series, searching for a possible anomaly by measuring a possible local deviation of a measurement signal synchronized with respect to the other measurement signals synchronized with the reference time series. Thanks to the proposed method, it is possible to automatically identify tubes potentially containing a defect. This allows the large majority of signals to be excluded from analysis, which clearly correspond to healthy tube areas. Operators can thus focus only on the few areas where the signal presents abnormal characteristics. Thus, the proposed method advantageously allows for a significant reduction in analysis time and a decrease in the risk of diagnostic errors.

[0017] This process is advantageously supplemented by the following characteristics, taken alone or in any technically possible combination thereof: each measurement signal is a multidimensional time series comprising P components, P being a natural number greater than or equal to 2, and the reference time series comprising P components; the DTW is applied multi-dimensionally on the P components simultaneously; the search for a possible anomaly is carried out component by component; the method comprises a selection, from among the measurement signals, of measurement signals resulting from the passage of a measurement probe in tubes belonging to a subset of tubes grouped according to the similarity of their physical characteristics; the search for a possible anomaly by measuring a possible local deviation of a measurement signal relative to the other measurement signals comprises the determination of a distance between a value of the measurement signal and the corresponding values of the other measurement signals; a median filter is applied to the distances determined during the search for anomalies;the measuring probe is a multi-frequency eddy current probe. ;

[0018] The invention also relates to a computer program product comprising program code instructions for executing the steps of the method according to the invention, when said program is executed on a computer. The computer program product may take the form of a non-volatile medium on which the instructions are stored, which may be any type of memory: hard disk, SSD, flash memory, etc. Presentation of figures

[0019] Other characteristics, aims and advantages of the invention will emerge from the following description, which is purely illustrative and non-limiting, and which must be read in conjunction with the appended drawings in which: there figure 1 schematically shows an example of a measuring probe running through a tube of a steam generator to acquire measurement signals, the figure 2shows an example of a graphical representation of a component of a measurement signal acquired by an eddy current probe along a tube of a steam generator, the figure 3 is a diagram showing steps of implementing a method according to a possible embodiment of the invention, the figure 4 shows an example of an extract of a first component of a measurement signal corresponding to the passage of a spacer plate by the measurement probe, without synchronization, the Figure 5 shows the example of the figure 4 after synchronization by DTW with the reference time series, the figure 6 shows an example of an extract of a second component of a measurement signal corresponding to the passage of a spacer plate by the measurement probe, without synchronization, the figure 7 shows an example of the figure 6 after synchronization by DTW with the reference time series. Detailed description

[0020] In reference to the Figure 1 and to the Figure 3 , a maintenance method will be described comprising the detection of a possible anomaly of a measurement signal which could be caused by a defect in the tube 2 of the heat exchanger. A defect is typically a crack in the wall of the tube 2, but can also be another alteration of the structure of the tube 2 capable of locally varying the measurement signal.

[0021] A first step (S01) consists of obtaining measurement signals resulting from the passage of a measurement probe 1 in tubes 2 of the heat exchanger. The measurement signals can be of various natures, since the measurement signals are likely to present variations linked to the presence of a defect in a tube 2. For example, the measurement signals can be electrical responses to stimulations such as the generation of eddy currents, as mentioned above, or can result from ultrasonic acquisition by means of an ultrasonic probe, or even result from the detection of radio waves. Preferably, the measurement signals are multidimensional, comprising several components each corresponding to an acquisition channel of the measurement probe 1. Preferably, each measurement signal is a multidimensional time series comprising P components, P being a natural integer greater than or equal to 2, and preferably greater than or equal to 4.

[0022] It is thus possible to provide a preliminary step (step S00) of passing the measuring probe 1 through the tubes 2 to acquire the measuring signals which are then collected, but the measuring signals can be obtained otherwise, and it is in particular possible to obtain measuring signals stored on a data medium, for example from acquisitions in the past.

[0023] As an example, and in a preferred embodiment used for the remainder of the description, the measurement signals come from a multi-frequency eddy current probe. A measurement signal then takes the form of a complex impedance, with real and imaginary components according to several frequencies coming from as many acquisition channels. For example, there may be four frequencies, resulting in a measurement signal with eight components. The frequencies are for example between 100 kHz and 600 kHz. Figure 2shows an example of amplitude variations of a component of a measurement signal acquired by an eddy current probe along a tube 2 of a steam generator, as a function of time and therefore of the path of the measurement probe 1 along the tube 2.

[0024] Due to the large number of tubes 2 present in a heat exchanger, a large number of measurement signals can thus be obtained. Typically, at least several hundred measurement signals are available. However, as these are acquired independently of each other, and the speed of movement of the measurement probe 1 can vary depending on the tubes 2, the measurement signals are not synchronized with each other. For example, the passage of the same spacer plate will be present at different times in two measurement signals. It is also possible that the tubes 2 have variable lengths. It is therefore necessary to synchronize the measurement signals, by matching the main amplitude variations encountered on all the measurement signals.

[0025] Previously, it is possible to resample the time series of the measurement signals to ensure that they all have the same number of time steps. Such resampling is known to those skilled in the art, and involves, for example, interpolation or the addition of intercalary values (typically zeros) followed by low-pass filtering.

[0026] After resampling, the measurement signals appear as Q time series x 1 , x 2 , ...., x Q of size N, i.e. for any integer i between 1 and Q, the time series xi is a function is a function 1 , … , N → ℝ P , with P the number of components of the measurement signal.

[0027] A reference time series is then determined (step S02). This reference time series corresponds to the arithmetic mean, at each instant, of the measurement signals: i.e. x: 1 , … , N → ℝ P the reference time series defined, for any element t ∈{1,...,N}, by: x ¯ t = 1 Q ∑ i = 1 Q x i t .

[0028] A synchronization (step S03) of each measurement signal with the reference time series is then carried out by applying a dynamic time warping, or DTW, on said measurement signal and the reference time series. For any i∈{1,...,Q}, the DTW provides a synchronization function fi :{1,..., N}→{1,..., N} which minimizes, under certain constraints (fi must in particular be increasing), the sum of the following deviations: ∑ t = 1 N x i ∘ f i t − x ¯ t where the double bars denote the Euclidean norm.

[0029] It should be noted that since each measurement signal is synchronized with the reference time series, by transitivity, the measurement signals become synchronized with each other. Furthermore, since the reference series is constructed from all the measurement signals, synchronization does not rely on the arbitrary designation of one of the measurement signals as the reference or by the choice of an external signal.

[0030] Typically, DTW implements a distance calculation between series, which requires applying the DTW to each pair of series, i.e. Q(Q-1) / 2 DTW calculations when Q series are processed. By determining and using the reference series, the number of DTW calculations is reduced to Q, with the DTW being applied to each series relative to the reference series. However, the use of DTW is very computationally intensive, and therefore time and resource intensive. This reduction in computation time makes it possible to group a large number of series and improve the quality of detection. In fact, preferably the number of measurement signals taken into account is greater than 100, preferably greater than 500, and even more preferably greater than 1000.

[0031] As mentioned above, the measurement signals resulting from the acquisition by eddy current measuring probes 1 are multidimensional. More precisely, the existence of four complex-valued channels means that they are multidimensional. ℂ 4 . The complex values are separated into real and imaginary parts in order to obtain a measurement signal with values in ℝ 8 . In the following, we will denote XFA, YFA, XF1, YF1, XF2, YF2, XF3, YF3 each of the 8 values making up the measurement signal, the XF values being derived from the real parts while the YF values are derived from the imaginary parts, the last letter A designating an acquisition in absolute mode while the presence of a number indicates an acquisition in differential mode. Of course, this is only a limiting example, and the measurement signal may be different, and the number P of components may be other than 8. Preferably, however, P is strictly greater than 3.

[0032] Preferably, the DTW is applied multi-dimensionally on the P components simultaneously of the measurement signal: the DTW is applied to the vectors in ℝ P grouping the components of the measurement signal. A measurement signal with values in ℝ P will undergo the time deformation minimizing the sum of the Euclidean distances between its points (in ℝ P ) and the corresponding points of the reference series. To be retained, a deformation must therefore be advantageous in all components of the measurement signal and not just one. Since external elements such as spacer plates 4 are visible on all components of a measurement signal, applying the DTW to measurement signals with values in ℝ P rather than separately to each component allows it to be stabilized and avoids synchronization errors (such as for example the synchronization of an aberrant peak in a component with a spacer plate edge 4 visible on the reference signal).

[0033] There Figure 4 shows an example of an extract of a first component, in this case YF2, of a measurement signal corresponding to the crossing of a spacer plate 4 of the hot branch by the measurement probe 1, before synchronization. The Figure 5 shows the example of the Figure 4 after synchronization by DTW with the reference time series, which is the average of the measurement signals. The figure 6 shows an example of an extract of a second component, in this case XF3, of a measurement signal corresponding to the crossing of a spacer plate 4 of the hot branch by the measurement probe 1, before synchronization. The Figure 7 shows an example of the Figure 6after synchronization by DTW with the reference time series, which is the average of the measurement signals. The units are arbitrary and have no importance, only the deviations from the other measurement signals count.

[0034] It can be seen that synchronization allows signals of similar shape to be brought together, and certain signals stand out from the set thus formed: the next step of the method is precisely to detect the measurement signals which stand out. The next step is thus a search (step S04) for a possible anomaly by measuring a possible local deviation of a measurement signal compared to the other measurement signals. Preferably, the search for a possible anomaly is carried out component by component, and not on a vector in ℝ P .

[0035] The search for a possible anomaly can, for example, be carried out using the local anomaly factor, better known by the acronym LOF for "Local Outlier Factor". LOF is based on a concept of local density, where the density is given by the k nearest neighbors, whose distance is used to estimate the density. By comparing the local density of a component of a measurement signal with the local densities of its neighbors, regions of similar density can be identified, and parts of a component of a measurement signal that have a lower density than its neighbors can then be considered anomalies.

[0036] LOF requires synchronous time series for its implementation, as are the measurement signals after DTW. Indeed, as soon as a pattern in a time series does not occur at the same time as a similar pattern in several other time series, the local anomaly factor of the time series increases. LOF is applied separately to each component, and therefore does not rely on the Euclidean distance between the vectors in ℝ P grouping the components of the measurement signal. For an anomaly to be detected, it is sufficient for it to be present on one of the components of the measurement signal.

[0037] The proposed approach makes it possible to highlight anomalies that would not have been detected or, conversely, to consider measurement signals with atypical values as being free of anomalies. For example, with reference to the Figure 4, the method makes it possible to highlight anomalies in the measurement signals corresponding to curves 40, 41, 42 or 43. On the other hand, despite their apparent deviation before synchronization, the measurement signals corresponding to curves 45 or 46 will ultimately not be considered as presenting anomalies. DTW makes it possible to compensate for possible time shifts between the measurement signals, which are not anomalies. Thus, on the Figure 5 , the measurement signal which corresponded to curve 45 of the Figure 4 is resynchronized with the others and no longer exhibits atypical behavior. On the other hand, the curves deviating from the majority behavior 50, 51, 52, 53, 54 are still highlighted.

[0038] In reference to the Figure 6, the method makes it possible to highlight anomalies in the measurement signals corresponding to curves 60, 61. On the other hand, despite their apparent deviation before synchronization, the measurement signals corresponding to curves 65 or 66 will not ultimately be considered as presenting anomalies. DTW makes it possible, by recalibrating the signals, to highlight atypical behaviors which did not appear to be initially visible. Thus, on the Figure 7 , the measurement signal corresponding to curve 70 presents an anomaly not initially visible, while others remain highlighted, such as for example the measurement signals of curves 60 and 61, which now correspond to curves 71 and 72, and whose deviation from the majority behavior remains visible. Other curves 75, 76, although sometimes presenting deviations, are not considered as anomalies by the method.

[0039] Other anomaly search methods can be used, especially since the measurement signals are, after DTW, synchronized with each other, which makes it possible to use other methods that need to be applied to synchronized signals, such as kernel estimators or isolation forests.

[0040] The search for a possible anomaly, for example by LOF, can be carried out on the entirety of a component of the measurement signals. Preferably, however, the search for a possible anomaly section by section, that is to say on temporal subparts of the component of the measurement signals. Thus, in the event of detection of an anomaly, information on the location of the fault in the tube 2 is directly accessible via the identification of the section in which the anomaly was detected. It is even possible to apply the detection method (LOF or other) at each instant. The method is then applied separately to each acquisition channel and at each instant, and therefore relates to scalars (elements of R). A method such as LOF remains based on the Euclidean distance, which is reduced here to the absolute value of the difference between values.

[0041] Preferably, a median filter can be applied to the result of the detection method such as the LOF by instant with a window size of the length of a typical defect (for example 2 mm), so as to regularize the identification of outliers. In this variant, only consecutive outliers - whose median LOF exceeds a certain threshold - of size greater than the length of a typical defect are retained. Thus, the instant-by-instant application of the LOF makes it possible both to identify whether the signal is defective but also to obtain the location and length of the defect in the signal.

[0042] The tubes 2 of a heat exchanger may have geometrical particularities which can be exploited, and in particular in the case of a steam generator the presence of a curved part 2b - the bend - framed by two rectilinear parts 2a, 2c - the hot branch and the cold branch. However, the size of the curved part 2b is not the same for all the tubes 2: the further the tube 2 is from the central axis of the steam generator, the more the length of the curved part 2b increases. At the same time, the tubes 2 no longer encounter the same number of anti-vibration bars 8. However, these variations in the profile of the tubes 2 are known a priori, and it is possible to group the tubes 2 having similar geometrical characteristics to improve the detection of defects.

[0043] Thus, the method may comprise the selection, from among the measurement signals - resulting from the passage of a measurement probe 1 in tubes 2 - belonging to a subset of tubes 2 grouped according to the similarity of their physical characteristics. The remainder of the method is then applied to the measurement signals thus selected.

[0044] Once an anomaly is detected in a measurement signal, the corresponding tube 2 can be considered to have a fault. It is then possible to plan a maintenance operation (step S05). For example, it is possible to carry out an additional, more in-depth inspection of this tube 2, or even to plan and carry out a repair or replacement operation for this tube if this operation is technically feasible. Defect detection can also validate the restart of the heat exchanger or, on the contrary, lead to the definitive shutdown of the heat exchanger.

[0045] The invention is not limited to the embodiment described and shown in the attached figures. Modifications remain possible, without departing from the scope of protection of the invention, which is defined in the claims.

Claims

1. A method for maintaining a tube heat exchanger, comprising: - obtaining (S01) measurement signals resulting from the passage of a measurement probe (1) in the tubes (2) of the heat exchanger, - determining (S02) a reference time series corresponding to an average, at each instant, of the measurement signals, - synchronising (S03) each measurement signal with the reference time series by applying a dynamic time warping, DTW, to said measurement signal and to the reference time series, - searching (S04) for a potential anomaly by measuring a potential local deviation of a synchronised measurement signal with respect to the other measurement signals synchronised with the reference time series.

2. The method according to claim 1, wherein each measurement signal is a multidimensional time series comprising P components, P being a natural integer greater than or equal to 2, and the reference time series comprises P components.

3. The method according to the preceding claim, wherein the DTW is applied multidimensionally on the P components simultaneously.

4. The method according to one of the two preceding claims, wherein the search for a potential anomaly is carried out component by component.

5. The method according to one of the preceding claims, comprising a selection, among the measurement signals, of measurement signals resulting from the passage of a measurement probe (1) in the tubes (2) belonging to a subset of tubes (2) grouped according to the similarity of their physical properties.

6. The method according to any one of the preceding claims, wherein the search for a potential anomaly by measuring a potential local deviation of a measurement signal with respect to other measurement signals comprises determining a distance between a value of the measurement signal and the corresponding values of the other measurement signals.

7. The method according to the preceding claim, wherein a median filter is applied to the distances determined during the search for anomalies.

8. The method according to any one of the preceding claims, wherein the measurement probe is a multifrequency eddy-current probe.

9. A computer program product, comprising program code instructions for executing the steps of the method according to any one of the preceding claims, when said program is executed on a computer.