John Rudlin, Amin Muhammed and Charles Schneider
Paper published in Insight Vol,48, No.6, pp348-351, June 2006.
Periodic inspection of railway axles is carried out to avoid axle failures. Analytical methods for determining the periodicity of inspection require knowledge of load conditions, materials properties and inspection reliability. The Wheel set Integrated Design and Maintenance (WIDEM) Project is being carried out to acquire some of this data and to develop methodologies for axle design.
An example is given in this paper of the measurement of inspection reliability from a set of 11 cracked axles. The method estimates the POD of the in-situ inspection technique from the response vs. size method, the size in this case being measured with Alternating Current Potential Drop (ACPD), phased array ultrasonics and time of flight diffraction. The results show that useful information is generated by such methods and a further experiment is planned on a different range of axles to obtain wider applicability of the data obtained.
Design of rail axles is generally on the basis of rules that have been laid down by experience and this limits the possibilities when, for example, new materials with different properties become available. Therefore some scientific information backing up and enhancing the rules may enable the development of new designs or design philosophies. The WIDEM Project started in January 2005. The objective of the project is to provide information that will assist in the future design of axles. TWI is involved primarily with attempting to measure NDT performance and the methodology of inspection periodicity within this project.
Fig.1. Basic axle geometry
Designs of rail axles vary depending on their application ( Fig.1). Except for simple designs, most axles will have wheel seats, but some will have additional seats for driving or braking. The design of the seats and the relative diameters of the seat and the main axle body determine the position of the cracking in these areas.
The inspection issues
Axles tend to crack either in mid-span or under or close to the wheel seats ( Fig.2). The crack morphology can also vary. Fig.2 is from a UK axle approximately mid-span showing a typical crack group. Fig.3 shows a crack formed on a brake seat. Fig.4 shows a crack formed close to a seat.
Fig.2. Crack positions and ultrasonic axle test geometries
Fig.3. Example of crack group on UK axle
Fig.4. Crack under brake seat
In the UK various inspection methods have been tried or developed for this particular application. Surface inspection methods (particularly MPI and electromagnetic) have been introduced for accessible areas since the Rickerscote accident in 1996. 
However, where the crack initiates from an inaccessible surface the inspection is by ultrasonics. The methods adopted are generically known as the high angle scan (applied from the axle body), the near end scan and the far end scan (applied from the axle ends). These scans are shown in Fig.5
Fig.5. Crack close to seat
Hollow axles are also used, and the ultrasonic test used in this case is an angled beam scan from a rotating probe in the bore. This inspection is mechanized.
The primary difficulty and skill required for the inspections is discriminating between geometrical echoes and crack signals.
The inspection performance of these techniques (and from this the inspection periodicity) has been estimated previously.  This reference shows a 90% POD of the near end and high angle scans for flaws of around 3mm depth, and a 90% POD at around 12mm depth for the far end scan. Because the far end scan is not very sensitive, this leads to a situation where the inspection, which requires removal and replacement of the bearing cover, has to be applied all too frequently. A question has arisen as to whether this technique is effective in preventing axle failures because of its low sensitivity. However, no experimental data exists to more accurately estimate the sensitivity and therefore the inspection continues to be deployed. Part of the WIDEM project is to get better data for the inspection sensitivity.
Determination of probability of detection
The generic methods and statistical processes of determining POD have been described earlier.  An experimental approach requires a set of samples with known, preferably natural, fatigue cracks. These are then inspected and either hit/miss or response vs. size methods can be used to estimate POD.
Experiment carried out at LBF (Darmstadt, Germany)
A set of cracked axles from a metro system has been held at LBF. An opportunity arose to estimate the inspection sensitivity of the procedure used to inspect the axles in service (essentially a mid axle inspection technique applied from the axle end).
The work involved consisted of:
- Inspecting the cracks with ACPD, TOFD and phased array UT to estimate crack depth;
- Inspecting the axles with the specified procedure (and two similar procedures) and recording the signal amplitude from each crack;
- Estimating the POD using response vs. size method.
The samples at LBF are from an urban Metro system. There are 13 samples, 11 of which are cracked (the uncracked samples were not used).
The general geometry and crack location is given in Fig.6.
Fig.6. Axle geometry and crack location (axles at LBF)
The cracks were initially detected with magnetic particle inspection (MPI). An example of these cracks has already been given in Fig.4. These indications were used to measure crack length.
The ACPD (alternating current potential drop) method was used to measure crack depth on all the cracks. This gives an estimate of crack depth at individual points on the crack and a profile can be built up. An example of the type ofdata is given in Fig.7.
Fig.7. Typical ACPD profile of crack depth
Phased array pulsed echo ultrasonics was used as a check on the ACPD results for larger cracks. Smaller cracks could not be visualised with this technique applied from the same side as the crack due to the beam width of the available probe. This probe was also not large enough to provide a focussed beam when applied to the other side of the axle. An example of the results from this technique is given in Fig.8.
Fig.8. Phased array image of crack
Time-of-Flight-Diffraction (TOFD) ultrasonics were also attempted. This was difficult to apply because the geometry of the component did not allow ideal positioning of the probe (equidistant from the crack at the surface). The shallow areas of the crack were hidden, but some indications were obtained. Fig.9
shows an example of one of these.
Fig.9. TOFD image of crack
The crack dimensions finally obtained are given in Table 1.
|Length (mm)||Depth (mm)|
The standard procedure for inspecting these axles in-situ specifies a 5MHz zero degree compression probe. The amplitude of the ultrasonic reflection is set up on a 3mm deep slot on an axle to 80% of full screen height.
In addition to this procedure a zero degree 2MHz compression wave probe was also used.
The reporting threshold (-8.5dB from the slot signal) was noted. The signals from each crack were optimised and recorded as a level above the reporting level.
Analysis and results
The response vs. size method was used to analyse the data. The data points are plotted on log/log scale of response versus size ( Fig.10). Using maximum likelihood estimation, two parallel regression lines are plotted through the two sets of data respectively, together with the corresponding 10% and 90% confidence limits. 50% of each population is therefore above the appropriate regression line and 50% below. Thus, the point at which the regression line crosses the reporting threshold corresponds to 50% POD. The POD curve can be derived from the confidence levels that correspond to the various points along the reporting threshold line (e.g. 90% POD occurs when the 90% confidence limit crosses the reporting threshold line and so on).
Fig.11 gives the resulting PODs for the 0 degree probes.
Fig.10. Response vs. size analysis for 0 degree compression wave probes
The results from Fig.11 show that the 90% POD is at about 3mm for the 5MHz probe and about 4mm for the 2MHz probe. Since this method is detecting cracks at a relatively short range (less than the far end scan in full size axles) although using a method similar to the far end scan, this result appears to be slightly optimistic compared with the Benyon and Watson estimates.
Fig.11. Predicted POD curves for 0 degree beams
However there are some limitations to the experiment that may have led to this result and these need to be investigated further. The measurements were made by TWI personnel with no specific time limit to carry out the inspection rather than the normal personnel used for the inspection. Also the brake drum would be in place when the in-situ experiment was carried out. The result has been obtained by comparison with NDT data (mostly ACPD for the smaller cracks)and this is of course subject to inaccuracy. Further work is also needed to establish the repeatability of the information obtained.
The difference between the 5MHz and 2MHz probes are also in line with expectations. The 2MHz frequency has a wavelength of around 3mm so the response to flaws of this order of size is likely to vary considerably. It is therefore likely that some flaws will be missed, and the POD affected.
Inspection periodicity - possible alternative methodologies
The deterministic method of establishing periodicity is to make the period such that the maximum expected crack growth rate at the known loads will enable a crack to be detected before failure (usually giving two opportunities). Of course there is a great deal of uncertainty in both loads and material properties. This uncertainty is being addressed by a series of projects with which WIDEM is collaborating (see below).
An alternative method is to work backwards from a probability of failure. On average there are less than two axle failures per year on UK trains, but there are no figures available for how many potential failures have been saved byNDT (axles are simply scrapped if cracks are detected). Using reliability methods the uncertainty in the various parameters (including POD) can be included in probability distributions and therefore a different calculation is carried out to obtain the variation of failure probability in service. Then inspection can be set at appropriate intervals to limit failure probability to acceptable levels.
TWI is planning (with other partners) to compare the different methods and their outcome. Such calculations will enable the effectiveness of an inspection method to be directly related to an expected number of failures.
TWI has now obtained a collection of 19 axles. Some of these were withdrawn from service because cracks were detected. Others were withdrawn from service for other reasons and have subsequently had cracks induced in them. Pooling these specimens with a collection of axles owned by Applied Inspection, further tests will be carried out to estimate POD. These tests will include some features to address human factors. Further work on the Darmstadt axles will also be carried out, and a series of trials on hollow axles are also planned.
The WIDEM project
The WIDEM Project (Wheelset Integrated Design and Maintenance) is a part funded EU project managed by Lucchini (Italy). The other partners are Politecnico di Milano (Italy), LBF (Germany), D2S (Belgium), UNIFE (Belgium), Alstom(France), TWI (UK), Microsystems (Italy), MTB (Sweden) and VUZ (Czech Republic).
The industry advisory group for the project includes UK Railway Safety and Standards Board (RSSB) and Trenitalia. The project is exchanging information with the UKAxle Project (funded by RSSB and led by AEA Technology - Rail) andthe DeuFrako project (funded by Deutsch Bahn (DB) and SNCF).
The authors would like to thank the EU and RSSB for funding the work, and Roy Archer of Applied Inspection for sharing his experiences of axle inspection.
- Health and Safety Executive (HSE) 'Railway accident at Rickerscote'. November 1996 ISBN: 071761171X. HMSO.
- Benyon J.A. and Watson A.S. 'The use of Monte Carlo Analysis to increase inspection intervals' Proc. International Wheelset Conference, Rome 2001.
- Schneider C.R.A. & Rudlin J.R. 'Review of statistical methods used in quantifying NDT reliability'. Insight 46 (2) February 2004.