IMPROVING THE DETECTION OF INTERNAL RAIL CRACKS BY USING RADON TRANSFORM OF BSCAN IMAGE

Transcription

1 IMPROVING THE DETECTION OF INTERNAL RAIL CRACKS BY USING RADON TRANSFORM OF BSCAN IMAGE Dr. Patrice AKNIN INRETS, French National Institute for Transport and Safety Research 2 Av. du Gnl. Malleret-Joinville, Arcueil Cedex, France aknin@inrets.fr Hervé CYGAN SNCF IEM-RM2 17 rue d Amsterdam, Paris, France KEYWORDS : transversal crack, ultrasonic, radon transform, BSCAN, detection, signal to noise ratio ABSTRACT Ultrasonic inspection systems generally use ASCAN mode to detect rail crack. A flight time window is open where a flaw echo is validated when it occurs with a sufficient level. The maintenance operators well know the difficulties to adjust the threshold and the position of the time window : if the threshold is too high, the system misses cracks, if it is too low the system generates false alarms ; if the time window is too close to the origin -corresponding to the rail surface-, noises disturb the detector, if it is far from the origin, the sub-surface cracks are not detected. The only solution to increase the detection robustness is to validate it only in the case of 2 or 3 successive threshold overshoots. This article introduces a new decision space that uses the Radon transform of the BSCAN image, a transform often used in straight line recognition. This approach takes benefit of the spatial coherence of the crack. The real-time implementation is possible because the computation time can be drastically reduced : only a particular angle of the Radon transform is computed, directly related to the direction of the incident ultrasonic beam. Tests on simulated and real BSCAN signatures demonstrate the good behaviour of the new detector compared to the usual one. INTRODUCTION Most of the railway networks, use onboard devices to detect internal cracks into the rail head. Specific train inspects the two rails with one or several ultrasonic probes, up to 50 km/h (Cygan & al, 2003). The following figures show the SNCF inspection train and support structure. Figure 1. SNCF ultrasonic inspection train Figure 2. Ultrasonic probes support Three probes observe each rail : one vertical probe for longitudinal cracks and two inclined probes for transversal cracks with privileged angle (cf. figure 3). Ultrasonic pulses are triggered at constant space distance, driving by the running speed of the train. 1/6 Railway Engineering. London 2004

2 water 3 probes direction of motion horizontal crack transverse crack x Figure 3. Mounting of inclined and vertical probes In order to improve the acoustic energy transmission, water is continuously injected between the probe support and the rail during the inspection. A-SCAN DETECTION PROCEDURE OF TRANSVERSE CRACKS The usual detection procedure of transverse cracks works on the ASCAN defect signature. In such a space, the reflected amplitude of ultrasonic (US) echo is drawn, function of time of flight. At each longitudinal position x of the probe along the rail, a curve is plotted. The beginning of the ASCAN curve corresponds to the surface echo, where the acoustic pulse reaches the rail head. If any defect occurs, the ASCAN amplitude remains very low (cf. figure 4), excepted near the surface. If a transverse crack intercepts the ultrasonic wave cone, this induces echo. The echo time is directly proportional to the defect depth, and the echo amplitude is linked to the equivalent surface of the defect. When the probe is moving along the defect, the echo is coming closer to 0 and the amplitude is firstly growing then decreasing (cf. figure 5). Ascan amplitude Ascan amplitude surface echo surface echo Detector window echoes of an internal crack h a t 1 t 2 time of flight Figure 4. ASCAN signal in no defect case time of flight Figure 5. Set of ASCAN signals on transverse crack 2/6 Railway Engineering. London 2004

3 The detection is usually checked (case 1 in the formula (1)) when the ASCAN signal overtakes a detector window (cf. Figures 4, 5). > 1 detect ascan = Max [ ASCAN] t1 <t<t 2 h a (1) < 0 3 parameters have to be tuned : the threshold h a, the initial time t 1 and final time t 2 of the window. The tuning of these three parameters is not so easy. If the threshold is too high, little cracks are ignored, if it is too low, the background noise could induce false alarms. If the window is timing localized too close to zero, the unavoidable surface echo could induce false alarm. If it is far from zero, some cracks will be missed. The only solution to increase the robustness of the detector is to validate the defect only in the case of successive threshold overshoots. Even in this case, the [false alarm]/[missing detection] compromise is difficult to reduce. THE B-SCAN VISUALIZATION SPACE Ultrasonic device of new generation presents new facilities. Their storage capabilities in conjunction with real time signal processing possibilities allow to design original detector in a different space, the BSCAN space. BSCAN representation is built by ASCANs aggregation (Krautkramer 1990). As shown in figure 6, the BSCAN is often presented in false color mode (color is dependant of the US echo amplitude), where each column is a single ASCAN signal obtained with one position x of the probe, and each line codes a time of flight. Figure 6. BSCAN mode versus ASCAN mode for transverse crack. As shown at figure 6, the BSCAN mode highlights the spatial coherence of the defect that is visualized as a segment of straight line. Not only the ultrasonic echo amplitude grows and decreases, but the maximum position describes a particular path. The orientation of the segment is directly linked to the space sampling dx and the direction of ultrasonic beam in the target. The new detector will take advantage of this coherence. RADON TRANSFORM, A SPECIFIC TOOL Among all the image processing tools, Radon Transform has a central status for line detection (Deans 1981) (Duda 1972). It calculates the perpendicular image projection on a turning axle. The RT(α,ρ) point sums all the pixel that belongs to the Δ line (cf. figure 7). 3/6 Railway Engineering. London 2004

4 RT(α,ρ)= BSCAN(P) with P=(x,y) and Δ: y = ρ sinα 1 tanα x (2) P Δ Obviously, the RT(α,ρ) becomes maximum when α=α 0 is the complement of the defect angle : α 0 =θ 0-90 THE NEW DETECTION PROCEDURE Figure 7. Radon transform principle Taking advantage of the Radon transform properties, the new detector is based on the calculation of RT(α 0,0) that is thresholded instead of equation 1. In this case the Δ line becomes Δ 0 that is parallel to the crack into the BSCAN space. detect bscan = RT(α 0,0) > 1 < 0 h b (3) In this way, the detector sums several data along the specific direction Δ 0. The conjunction of these data pointed out the defect (case 1 in the formula (3)). The robustness of such a kind detector lies in its capability to take into account successive ASCAN and then to average the noise effect. Concerning the time consuming, we can note that only one point of the Radon transform (RT(α 0,0)) is required. For a specific probe and a constant space sampling, α 0 is unique and can be pre-calculated. TESTS AND COMPARISON OF ASCAN AND BSCAN MODES In order to compare the two detection procedures, the figures 8 to 11 give an example of very light transverse crack with maximum echo about 25% of surface echo. Moreover, the signature has been deliberately noised (Signal to Noise Ratio SNR=-20dB). This defect cannot be detected in ASCAN mode because the signal does not reach the threshold h a. As seen in figure 9, if the threshold is moving down, the noise could well trigger false alarms. The detection of this defect is impossible in ASCAN mode. On the other hand, the spatial coherence of the defect in BSCAN mode allows to obtain a sufficient contrast as shown in figure 11. 4/6 Railway Engineering. London 2004

5 detectascan SNR=-20dB ha=threshold ha=threshold Figure 8. Noised transversal crack ASCAN Figure 9. Detection signal in ASCAN mode detectbscan hb=threshold Figure 10. Same noised transverse crack (SNR=-20dB) in BSCAN mode. contrast Figure 11. Detection signal in BSCAN mode A complete study about the influence of signal to noise ratio (SNR) on the detector contrast has been carried out. The contrasts have been calculated on the typical crack shown at figure 12 without noise. The maximum echo reaches 25% of the surface echo. The standard deviation of the signal is established after suppressing the surface echo. Figure 12. Typical light crack without noise 5/6 Figure 13. Example of noise realization Railway Engineering. London 2004

6 The injected noise sums a homogeneous white noise with variable standard deviation and a horizontal structured noise in order to model the real electronic noise and the weak natural echos due to the probe imperfections and 3D geometry of the rail (cf. figure 13). db BSCAN contrast ASCAN contrast SNR (db) Figure 14. ASCAN and BSCAN detection contrasts, function of SNR The BSCAN contrast is always 30dB over than the ASCAN one. Moreover the BSCAN detector remains usable until SNR=30dB while ASCAN detection is impossible under SNR= 20dB. So the BSCAN detection is really more effective than the ASCAN one. CONCLUSION This article introduces a new decision space for the inner defect detection by means of ultrasonic inspections. It uses the Radon transform of the BSCAN signature, a transform often used in straight line recognition. This approach takes benefit of the spatial coherence of the crack in BSCAN mode. The realtime implementation is easy and some tests on simulated and real BSCAN signatures demonstrate the good behaviour of the new detector compared to the usual one. AKNOWLEDGEMENTS These works are supported by SNCF Company, Infrastructure department. Especially, we thank Louis Girardi and Anne-Marie Riollet for their collaboration. REFERENCES H. Cygan, P. Aknin, P. Simard, L. Girardi (2003) B-Scan ultrasonic image analysis for internal rail defect detection. WCRR, Edinburgh S.R. Deans (1981) Hough transform from the Radon transform. IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 3, no. 2, p R.O. Duda, P.E. Hart. (1972) Use of the Hough Transformation to Detect Lines and Curves in Pictures. Comm.ACM., vol. 15, p J. Krautkramer, H. Krautkramer (1990) Ultrasonic Testing of materials. Springer-Verlag 6/6 Railway Engineering. London 2004