METHOD FOR NON-DESTRUCTIVE TESTING OF A MULTIPLE OF ELEMENTS

Description

technical field

[0001] The invention relates to the field of non-destructive testing of a plurality of elements, in order to determine whether an element is sound or unsound. More specifically, the invention applies to the testing of similar elements using measurement signals, each measurement signal corresponding to a different, specific element. Technological background

[0002] Non-destructive testing methods rely on detecting anomalies in measurement signals that may indicate a defect in a faulty component. These measurement signals can be of various types, and include, for example, signals resulting from ultrasonic acquisitions, or electrical responses to stimuli such as eddy current generation, radio wave detection, or even images or videos.

[0003] In these measurement signals, the search for anomalies can be challenging. Indeed, numerous factors can alter the measurement signals, such as the complex geometry of the inspected element or its lack of cleanliness, and the presence of an anomaly can be hidden among these alterations. Furthermore, if detection must be performed by a human operator analyzing the measurement signals, the inspection can be very lengthy and tedious, and its quality depends on the operator's skill. These problems can become particularly problematic in fields where the number of elements to be inspected is large and / or where it is essential to be able to detect all defects for safety reasons.

[0004] Various methods have been proposed for detecting anomalies that may correspond to defects in measurement signals. Most of these methods rely on comparing derived signal values ​​with reference values. However, each method is then specific to a particular application, and the sensitivity of the method depends heavily on the accuracy of the derived value calculation and the definition of the reference used. These methods therefore generally require extensive development and may not be suitable for detecting unanticipated defects or for inspecting components other than those for which they were designed.

[0005] For example, patent application FR3029288 proposes an inspection method for bottom penetrations in tanks; for each time sample, statistical values ​​such as the mean or standard deviation of the signal values ​​at that instant are calculated on the P available signals. These statistics are then used to construct a temporally homogenized signal, with a constant mean of zero and a constant standard deviation of 1 for each time sample. While applying this method to bottom penetrations yields satisfactory results, it is not as effective when applied to the inspection of other components such as bulkhead screws. This is because defects do not systematically generate amplitude increases, but often variations in the shape of the signals.On the other hand, the signals are not exactly synchronized with each other; local dilations / contractions / delays may appear.

[0006] Van Vaerenbergh et al. "Pattern Localization in Time Series Through Signal-To-Model Alignment in Latent Space" February 2018, DOI:10.1109 / ICASSP.2018.8461890, Conference: ICASSP 2018 - IEEE International Conference on Acoustics, Speech and Signal Processing, Calgary, Alberta, Canada" describes a method for localizing a predefined pattern sequence in a non-destructive test time series, based on machine learning to increase similarity between a synthesized time series and a real time series.

[0007] Hendrickx K et al, "A general anomaly detection framework for fleet-based condition monitoring of machines", Mechanical Systems and Signal Processing, Elsevier, Amsterdam, vol. 139, January 11, 2020, describes a method for detecting machine anomalies for monitoring a fleet of similar machines. Pairs of machines are formed, and measurement signals, each corresponding to a machine, are compared and statistically analyzed, particularly with the consideration of dynamic time warping (DTW). Presentation of the invention

[0008] The invention therefore aims to propose a non-destructive testing method for a plurality of elements based on measurement signals, which is simple to implement, applicable to the testing of a large number of similar elements, and which allows for the reliable detection of defects.

[0009] A non-destructive testing method is proposed for a plurality of similar elements based on measurement signals, each measurement signal corresponding to a different, specific element, comprising the following steps: obtaining a sequence of at least one signal portion from each measurement signal, for each pair associating two signal portions having the same position in their respective sequence, a distance between these signal portions of said pair is calculated, said distance being calculated using a dynamic time deformation, for each signal portion, determination of at least one statistical indicator from the distances of pairs involving said signal portion, the statistical indicator being associated with the element of said signal portion, for each element, comparison of at least one statistical indicator of said element with statistical indicators of other elements obtained for signal portions having the same position in their respective sequence, and depending on the comparison, determination of a healthy or unhealthy character of the element.

[0010] In this process, the plurality of elements comprises P elements, P being a natural number greater than 10, with each signal portion being involved in P-1 pairs.

[0011] Dynamic time distortion is applied here in an original way, since a statistical indicator based on couple distances is then used to detect the signal portions that stand out from the others.

[0012] This process is advantageously complemented by the following characteristics, taken alone or in any technically possible combination thereof: two consecutive signal portions in a sequence obtained from the same measurement signal partially overlap; a statistical indicator is a measure of central tendency, such as the mean or median; a statistical indicator is a measure of dispersion that measures the variability of distances, such as the standard deviation or variance; at least two statistical indicators are determined, including a measure of central tendency and a measure of dispersion; the comparison of at least one indicator of said element with indicators of other elements includes determining a difference between said statistical indicator and a closer other statistical indicator of another element; at least two statistical indicators are determined.and the comparison of the pair of indicators of one element with pairs of indicators of other elements includes a statistical modeling of a probability density of the values ​​taken by the pair of two statistical indicators of said element, each signal portion corresponding to a measurement distance at least equivalent to a typical size of defects sought.

[0013] The invention also relates to a computer program product comprising program code instructions for executing the steps of the process 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. Presentation of the figures

[0014] Other features, purposes and advantages of the invention will become apparent from the following description, which is purely illustrative and not limiting, and which should be read in conjunction with the accompanying drawings on which: there figure 1 is a diagram schematically showing the steps for implementing a process according to a possible embodiment of the invention; the figure 2 shows an example of alignment by dynamic time warping between two signal portions of healthy elements; the figure 3 shows an example of alignment by dynamic time warping between a signal portion of a healthy element and a signal portion of an unhealthy element; the figure 4 shows an example of a pair distance histogram for a healthy element; the figure 5 shows an example of a pair distance histogram for an unhealthy element; the figure 6 shows an example of the distribution of values ​​for two statistical indicators. Detailed description

[0015] The invention relates to the non-destructive testing of a plurality of similar components, a majority of which are considered to be in good condition. A component is defined as one with a defect, for example, a mechanical defect (crack, corrosion, etc.), that potentially renders it unfit for its function. For example, a fastener such as a screw is no longer considered sound if cracks are detected in its body. The components can be of various types, provided that measurement signals are available, each corresponding to a specific component. While the method can be implemented with a limited number of components, such as at least 10, it is more reliable and cost-effective to implement it on a larger number. Therefore, the plurality of components preferably comprises at least 50, and even more preferably at least 100.

[0016] Measurement signals can be of various types, as they are likely to exhibit variations related to the condition of an inspected component, and in particular, variations depending on the presence of mechanical defects. For example, measurement signals can result from ultrasonic acquisitions, or electrical responses to stimuli such as eddy current generation, radio wave detection, or even images or videos. As a non-limiting example, the method is described in an application to the inspection of mechanical fasteners such as partition screws, using measurement signals derived from ultrasonic acquisitions. Even in this application, the measurement signals could differ.It should be noted that the process may include the acquisition of measurement signals by means of a probe or sensor near each of the elements, for example successively from one element to another.

[0017] The process aims to determine whether each item is sound or unsound. It is then possible to continue the process with a more thorough inspection of the items considered unsound.

[0018] The method is implemented on a set of measurement signals, each corresponding to a different, defined element. For example, each measurement signal can be acquired specifically for a given element, typically by placing a probe near the element to be inspected: the measurement signal is then naturally associated with a particular element. In our example, an ultrasonic probe is placed against a partition screw, and then emits and receives ultrasonic waves from this position against the partition screw. It is also possible that the initial acquisition includes measurement signals from several elements. This is the case, for example, when a probe is moved along a path where it successively encounters several elements to be inspected. Segmenting the acquisition result allows the recovery of measurement signals, each corresponding to a different, defined element.

[0019] Preferably, the measurement signals are of similar sizes, in temporal and / or spatial terms, which facilitates their common manipulation in the rest of the process.

[0020] A sequence of at least one signal portion is obtained from each measurement signal (step S1). Preferably, several signal portions forming a sequence are extracted from each measurement signal. If the measurement signal originates from the propagation of a wave, for example, an ultrasonic wave, or from the movement of a probe, each signal portion is a time window corresponding to a distance. Thus, each signal portion can correspond to a measurement distance, which is chosen to be equivalent to at least one typical size of the defects being sought, and more advantageously to approximately twice a typical size of the defects being sought. Typically, the defects being sought are on the order of millimeters (cracks, fissures, etc.), so each signal portion corresponds to a measurement distance between 1 mm and 4 mm. In the following example, each signal portion corresponds to a measurement distance that is a depth of approximately 2 mm.

[0021] Preferably, at least two signal segments are extracted from the same measurement signal, and preferably at least four signal segments are extracted from the same measurement signal. The number of signal segments, however, obviously depends on the size of the measurement signal and the size of the defects being sought. It should be noted that if the measurement signal is sufficiently small, it may only be necessary to extract one signal segment, and in this case, the sequence of signal segments may consist of only one segment. Preferably, the signal segments from two different elements, occupying the same position in their respective sequences, are of comparable size, for example, in terms of measurement depth. Thus, the extraction of signal segments can also be used to ensure that comparable signal segments cover similar areas of the inspected elements, and therefore that these signal segments from different elements are comparable.

[0022] Therefore, even when measurement signals are small enough that only a single signal portion is extracted, the extraction can at least serve to ensure that signal portions from different measurement signals are similar (not necessarily identical) in terms of the physical representativeness of the measurement (e.g., a measurement distance or penetration depth). If the measurement signals are already similar to each other, obtaining the sequence of at least one signal portion can simply consist of making the measurement signal available.

[0023] Preferably, two successive signal portions within a sequence of the same measurement signal partially overlap. In other words, some data from the measurement signal is found in both of these successive signal portions, or part of one signal portion is also found in another signal portion. The overlap with another signal portion is preferably between 0.1% and 25% of a signal portion, and preferably between 0.5% and 10% of a signal portion.

[0024] The signal segments form a sequence in that they are ordered relative to each other: there is a first signal segment, followed by a second signal segment, followed by a third signal segment, and so on. We can denote by ai,n the nth signal segment of the i-th measurement signal. It should be noted that not all the segments obtained from a measurement signal necessarily belong to the sequence of measurement segments, but that two signal segments in the sequence maintain the same order as in the measurement signal: the first signal segment is in the measurement signal before the second signal segment, the third signal segment is in the measurement signal before the second signal segment, and so on.

[0025] Next, for each pair associating two signal segments with the same position in their respective sequences, a distance between these signal segments of said pair is calculated: let dtw n (ai,n , aj,n ) be the distance between the nth segment of the ith measurement signal and the nth segment of the jth measurement signal. As mentioned above, the segments obtained from a measurement signal are not necessarily part of the sequence of measurement segments, and for example, the sequence of measurement segments may only begin from a certain extracted segment, the preceding signal segments then not being used in the comparison and therefore not being part of the sequence of measurement segments. Similarly, it is possible that the signal segments in the sequence are not immediately consecutive in the measurement signal.However, it is constant that the order of the signal portions is respected: the (n+1)th portion of the ith signal is compared with a measurement portion of the jth measurement signal which is, in the measurement signal, after the portion of the jth measurement signal to which the nth portion of the ith measurement signal is compared.

[0026] Since each measurement signal is associated with a particular element, we can also define this distance dtw n (ai,n , aj,n ) as the distance between the nth portion of the ith element and the nth portion of the jth element. Distance is understood here as a criterion of similarity or, more accurately, dissimilarity (the greater the distance, the greater the dissimilarity). This is, in particular, the distance considered in dynamic temporal deformation.

[0027] For this purpose, the distance is calculated using dynamic time warping (DTW), a method already employed in other applications to measure the similarity between two sequences that may vary. Generally, DTW is a method that seeks an optimal match between two time series, subject to certain restrictions. The time series are warped by a nonlinear transformation of the time variable to determine a measure of their similarity, independent of certain nonlinear transformations of time.

[0028] The DTW defines a distance that allows for a more relevant measurement of the similarity between two signal segments, s1 and s2, than by comparing signals point by point, as is the case, for example, with Euclidean distance. With DTW, a point in s1 is associated with one or more points in s2 (and vice versa), based on the minimization of a cost function. This makes the measure particularly suitable for time series analysis because it is robust to dilations, compressions, and time shifts between the two signal segments. figure 2This shows an example of alignment by dynamic time warping between two signal segments of healthy elements. The pairings are illustrated by the lines connecting the points on the curves corresponding to the two signal segments. Although most of the lines are generally vertical, approaching a Euclidean distance between synchronous points, some pairing lines are strongly inclined, showing the influence of the transformations undergone which, thanks to DTW, do not prevent highlighting a high degree of similarity.

[0029] There figure 3 shows an example of alignment by dynamic time warping between a signal portion of a healthy element and a signal portion of an unhealthy element. Compared to the figure 2We observe greater inclinations in the matching lines, and therefore greater lengths, i.e., greater matching distances, indicating a greater distortion required to match the signal segments. Furthermore, we observe that certain points in one signal segment are matched with points in the other signal located at various points within that other signal. For example, several peaks in the signal segment at the top of the figure 3 are paired with the same point at the end of the signal portion at the bottom of the figure 3 Even after alignment, the resulting final distance, a measure of dissimilarity obtained by calculating the differences between the amplitudes of the points after matching, will remain significant, since not all of these peaks can be aligned with the same point in the other portion of the signal.

[0030] However, it is difficult to make assumptions about the quality of the DTW alignment between two signal segments from non-healthy elements. It is possible that two signal segments from non-healthy elements may exhibit similarities, and in this case, the DTW measurement may be of the same order of magnitude as the DTW between two signal segments from two healthy elements. It is also possible that the alignment between two signal segments from non-healthy elements may be very complex, and in this case, the DTW measurement may be of the same order of magnitude as the DTW between a signal segment from a healthy element and a signal segment from a non-healthy element. The preliminary step of dividing the signal into a series of signal segments allows for local and comparable DTW values. Indeed, if the DTW measurement were performed on the entire measurement signal, the impact of a deformation at a given instant would be too subtle to detect.

[0031] The DTW is applied to each pair associating two signal segments having the same position in their respective sequences, so that if the plurality of elements comprises P elements, where P is a natural number greater than 10, each signal segment is involved in P-1 pairs, resulting in P-1 distances involving that signal segment and therefore a particular element. Preferably, the distances are calculated for several signal segments, and preferably for all sequences of signal segments.

[0032] This approach can be translated using a P x P matrix Mn, where P is the number of elements and therefore measurement signals. The components Mni,j of the matrix are: Mni,j = dtwn(ai,n, aj,n), where dtwn(ai,n, aj,n) is the distance between the nth portion of the ith element and the nth portion of the jth element, denoted dtwn(ai, aj) for simplicity: M n = dtw n a 1 a 1 dtw n a 1 a 2 ⋯ dtw n a 1 a P dtw n a 2 a 1 dtw n a 2 a 2 … dtw n a 2 a P ⋮ ⋮ ⋱ ⋮ dtw n a N a 1 dtw n a P a 2 … dtw n a P a P

[0033] This matrix Mn is symmetric, with zero diagonal, and groups all possible pair distances for the same position in the sequences obtained from the measurement signals. Note that matrix notation is not essential, but it allows for the organization of the distances. Thus, column j of the matrix Mn groups the distance dtwn(aj, aj) between aj, the nth signal portion of element j, and itself, which is equal to 0, and the set of distances dtwn(ak, aj) between aj, the nth signal portion of element j, and the P-1 other nth portions ak of the other elements, k≠j.

[0034] However, having these torque distances does not allow for an immediate conclusion regarding the soundness or unsoundness of an element, since there are no reference points against which to compare them. It is therefore preferable to adopt a statistical approach aimed at identifying elements with signal portions whose characteristics deviate from others. For example, in the case of partition screw inspection, the figure 4 shows an example of a pair distance histogram for a healthy element, while the figure 5This shows an example of a pair distance histogram for an unhealthy element. The x-axis represents the pair distances for this element, and the y-axis represents the number of distances in each distance value class. Of course, the values ​​shown are only indicative, as they depend on many parameters such as the amplitudes of the measurement signals. Note that the pair distance histogram for an unhealthy element is generally shifted and expanded compared to the pair distance histogram for a healthy element. Therefore, a statistical approach can be used to identify such deviations.

[0035] Therefore, for each signal portion, at least one statistical indicator is determined from the pair distances involving said signal portion (step S03), the statistical indicator being associated with the element of said signal portion. Using matrix notation, the pair distances involving an element j, and thus involving the signal portion aj, correspond to column j of the matrix M n.

[0036] Preferably, a statistical measure is either a measure of central tendency, such as the mean or median, or a measure of dispersion that measures the variability of distances, such as the standard deviation or variance. Preferably, at least two statistical measures are determined, comprising a measure of central tendency and a measure of dispersion. Typically, the statistical measures are the mean µ and the standard deviation σ. Note that the statistical measure(s) can be determined directly from the pair distances, or from a histogram such as the one illustrated in the Figures 4 and 5 .

[0037] For each element, at least one indicator of that element is then compared with indicators of other elements (step S04). Preferably, at least two statistical indicators are determined, and the comparison is performed using pairs of indicators, with one pair of indicators from one element being compared with pairs of indicators from other elements. Based on the comparison, the healthy or unhealthy nature of an element is determined (step S05), preferably on the basis of a deviation of the statistical indicators relative to the others.

[0038] The comparison aims to identify elements for which the statistical indicator(s) deviate from those of the healthy elements, assumed to be in the majority. Several approaches can be adopted. Comparing at least one indicator of a given element with indicators of other elements may involve determining the difference between that statistical indicator and a closer statistical indicator of another element, and even between that statistical indicator and several closer statistical indicators of other elements. For example, the "K nearest neighbors" method consists of calculating, for each element, the Euclidean distance to its K-th nearest neighbor. A threshold is then chosen for this distance to determine whether the element is potentially unhealthy. Other variations can be used, such as the local outlier factor method.

[0039] There figure 6 This shows an example of a two-dimensional spatial representation of {mean, standard deviation} pairs for distances between partition screws. Means are shown on the y-axis and standard deviations on the x-axis. By applying the "K nearest neighbors" method, elements whose statistical indicators deviated from the majority were identified and appear in black on the figure 6 , while the majority of elements whose statistical indicators were close to each other appear in white, and are therefore considered healthy.

[0040] First, we observe that some healthy elements have means and standard deviations higher than those of unhealthy elements, and that some unhealthy elements have means significantly higher than those of healthy elements, particularly exceeding 70. Thus, a simple comparison of values ​​with a fixed threshold would allow us to detect some unhealthy elements, but would also determine the health of many unhealthy elements. Such an approach is therefore not entirely satisfactory. On the contrary, identifying elements whose indicators deviate from those of the majority of others allows us to detect potentially unhealthy elements even when these elements have means and standard deviations lower than the means and standard deviations of healthy elements.

[0041] Alternatively, comparing a pair of indicators for one element with pairs of indicators for other elements involves statistical modeling of the probability density functions of the values ​​taken by the pair of two statistical indicators for that element. Parametric methods (Gaussian, mixture of Gaussian) or non-parametric methods, such as kernel methods, can be used, for example.

[0042] It should be noted that there is at least one statistical indicator per signal portion, so the comparison can involve several statistical indicators from multiple signal portions of the same element. For example, it is then possible to determine an element as unhealthy if it appears that the statistical indicator differs from the others in only one position in the sequence, or conversely, to require that the statistical indicator differ from the others in several positions in the sequence for the element to be considered healthy. In order not to miss any potentially defective elements, it is preferable to consider as unhealthy an element whose statistical indicator(s) differ from the others in at least one position in the sequence of signal portions.

[0043] Once the soundness of each element has been determined, a thorough inspection of the unsound elements can be carried out, or physical intervention can be performed directly on them, such as their replacement or disposal. In this regard, it should be noted that the process is configured to detect all unsound elements, and that it is less damaging to consider a sound element unsound than to risk deeming an unsound element sound.

[0044] The invention is not limited to the embodiment described and shown in the accompanying figures. Modifications remain possible without departing from the scope of protection of the invention, which is defined by the following claims.

Claims

1. A method for non-destructive testing of a plurality of components based on measurement signals, each of the measurement signals corresponding to a predetermined different component, comprising the following steps: - obtaining (S01) a sequence of at least one signal portion based on each measurement signal, - for each pair associating two signal portions having one and the same position in their respective sequence, a distance (S02) between these signal portions of said pair is computed, said distance being computed using dynamic time warping, - for each signal portion, determining (S03) at least one statistical indicator based on the distances of pairs involving said signal portion, the statistical indicator being associated with the component of said signal portion, - for each component, comparing (S04) at least one statistical indicator of said component to statistical indicators of other components, which are obtained for signal portions having one and the same position in their respective sequence, and as a function of the comparison, determining (S05) the soundness or unsoundness of the component, wherein the plurality of components comprises P components, P being a natural integer greater than 10, and each signal portion is involved in P-1 pairs.

2. The method as claimed in claim 1, wherein two signal portions that follow one another in a sequence obtained from one and the same measurement signal partially overlap.

3. The method as claimed in any one of the preceding claims, wherein a statistical indicator is a central tendency indicator, such as the mean or the median.

4. The method as claimed in any one of the preceding claims, wherein a statistical indicator is a dispersion indicator measuring the variability of the distances, such as the standard deviation or the variance.

5. The method as claimed in any one of the preceding claims, wherein at least two statistical indicators are determined, comprising a central tendency indicator and a dispersion indicator.

6. The method as claimed in any one of the preceding claims, wherein the comparison of at least one indicator of said component to indicators of other components comprises the determination of a difference between said statistical indicator and another nearer statistical indicator of another component.

7. The method as claimed in any one of claims 1 to 6, wherein at least two statistical indicators are determined, and the comparison of the indicator pair of one component to indicator pairs of other components comprises a statistical modelling of a probability density of the values taken by the pair of two statistical indicators of said component.

8. The method as claimed in any one of the preceding claims, wherein each signal portion corresponds to a measurement distance at least equivalent to a typical size of sought defects.

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