High-Speed Inspection of Rails by Passive Ultrasonic Monitoring

Defective rails are a major concern in the safety and operation of the railroad industry. The two leading causes of defective rails are transverse defects and detail fractures in the rail head which are primarily caused by manufacturing defects and rolling contact fatigue [1]. Although improved manufacturing techniques for rails have reduced the effects of inclusion and rail grinding has slowed the effects of rolling contact fatigue, broken rails from internal defects still remain a significant rail maintenance issue. A common technology for internal rail flaw detection is based on ultrasonic tests performed by piezoelectric transducers hosted in fluid-filled wheels, known as rolling search units (RSU) [2,3]. The downside of these RSUs is that they operate at speeds (∼30 mph) which are considerably lower than revenue speeds (∼60 mph), requiring careful scheduling to minimize traffic disruptions. Furthermore, the need for specialized inspection vehicles limits the number of passes that can be performed on the same rail, with little opportunity to exploit the advantages of test redundancy to enhance the overall inspection accuracy.

This paper presents a high-speed passive noncontact ultrasonic rail inspection technique that has the potential of operating from a train traveling at revenue speed (smart train). The initial implementation of this approach was presented in a previous paper [4]. Once fully developed, this capability would allow inspections in a manner that is “transparent” to normal traffic, without requiring scheduling disruptions, while at the same time taking full advantage of the test redundancy allowed by the multiple train passes over the same rail. The technique is termed as “passive” because it does not require an active (controlled) source to acoustically excite the rail. The approach uses pairs of noncontact capacitive ultrasonic air-coupled transducers [4] for extracting the Green's function (or transfer function) of the rail segment between two points. The rolling train wheels provide the acoustic excitation. In the output-only approach, a cross-power spectrum operation (or cross-correlation in the time domain) between two receiving locations results in the transfer function filtered by the energy spectrum of the source [5], with various normalization procedures that can be applied to eliminate the effect of the (uncontrolled) source if needed (e.g., Ref. [6]). In ultrasonic and acoustic structural monitoring, the output-only technique has been applied to several cases of ambient or natural excitations (e.g., Refs. [7–17]).

The paper presents the latest results from a prototype that uses the passive approach to inspect rails at high speeds. In this implementation, the passively reconstructed transfer function of the rail in specific frequency bands (20–40 kHz and 70–120 kHz) is further processed by a statistical outlier analysis [18–20] to enhance the sensitivity to rail discontinuities between the two receivers in a pair. High values of the statistical discordancy of a current location from a set of “baseline” locations (Damage Index, DI) indicate possible discontinuities such as welds, joints, or defects. The results shown here are from field tests conducted at Transportation Technology Center Inc. (TTCI). The prototype test performance is evaluated through receiver operating characteristic (ROC) curves computed for varying levels of threshold applied to the DI traces to determine the probability of detection (PD) and the probability of false alarms (PFA). This evaluation follows similar procedures adopted previously to assess the performance of a noncontact rail inspection system that used an active excitation source [21,22]. The ROC results are computed for varying operational parameters of the inspection, specifically the baseline distribution length, the test speed, and the redundancy afforded by multiple runs.

Reference [6] demonstrated that this operation, which uses a conventional “intra-segment” averaged cross-power spectrum at the numerator (along the lines of the Welch's periodogram technique) that is normalized by an “inter-segment” averaged auto-power spectrum successfully eliminates the uncontrolled source W(f) and the receiver noise N_A(f) and N_B(f) from the extraction of the desired transfer function G_AB(f).

Figure 2 shows a typical reconstructed transfer function in the time domain in a rail segment from a pair of air-coupled ultrasonic receivers positioned at a distance of 18 in. (46 cm). The wave packet arrival at ∼160 µs corresponds to the flexural/torsional wave excited by the train wheel and propagating in the rail between the two receivers. Considering the measured wave speed and the 20–40 kHz frequency band, the corresponding wavelengths are in the range of 7.2–14.4 cm. This transfer function will change due to scattering if a discontinuity is present in the rail between the two receiver locations. As mentioned above, these changes are best tracked by an outlier analysis that is briefly discussed in the next section. A statistical outlier analysis is used to track the changes in the reconstructed transfer function to determine if a probed rail segment has discontinuities.

Tests were performed with the passive rail inspection prototype at TTCI in Pueblo, CO The prototype was mounted on the Federal Railroad Administration 5229 Test Car connected to a locomotive car (Fig. 4). The sensing head consisted of 12 ultrasonic capacitive air-coupled transducers (CAP-2 by VN Instruments Inc.) with a central frequency of 120 kHz and arranged as shown in Fig. 5. The transducers were arranged in three groups with each group having four devices. This configuration resulted in four possible combinations of transfer functions from each group. For example, from group 1, the four transfer functions extracted are from transducer pairs 1–7, 1–8, 2–7, and 2–8. The transducers were positioned at 3 in. from the rail's top surface at an angle of 6 deg with the vertical based on Snell's Law to ensure unidirectional reception of the flexural/torsional guided waves leaking from the rail into the air. A laser system consisting of two lasers was attached on both ends of the prototype to ensure that the sensor heads were properly aligned with the top surface of the rail (Fig. 4). A high-speed camera (up to 100 frames per second), aided by a stroboscopic light source, was installed alongside the prototype to continuously capture images of the inspected rail during the test runs to build a library of the locations of joints, welds, and defects on the rail (“ground truth”). Finally, a global positioning system (GPS) receiver was mounted on the top of the test car to tag each data point from the transducers and the camera images.

Test runs were conducted at speeds of 25 mph, 33 mph, and 40 mph on the high-tonnage loop (HTL) and speeds of 60 mph, 70 mph, and 80 mph on the railroad test track (RTT). Both the HTL and RTT had numerous joints and welds. The HTL track also had three known locations of natural transverse defects in the rail, whereas no internal defects were known to be present in the RTT track.

Figure 6 shows the data-acquisition system used in the field tests. A National Instruments peripheral component interconnect (PCI) Extensions for Instrumentation (PXIe) unit running labview Real-Time was used for data recording and processing. Signals from the air-coupled transducers were recorded continuously at a 1 MHz data sampling rate. A Windows 10 Computer was connected to the PXIe unit for user monitoring and control of the system. A tachometer (transistor-transistor logic (TTL) pulse generator) marked the position of the test car with a resolution of approximately 1.6 in. (4 cm) and was linked to the PXIe unit through an encoder input box. The GPS locations were recorded at a rate of 10 Hz using a Novatel GPS receiver connected to the PXIe unit via RS-232.

The strength of the acoustic signals recorded by the transducers with respect to the noise floor determines the quality of the data. When the wheels of the locomotive do not excite the rails sufficiently, such as when the train is moving at slow speeds and the wheels are not flanging, the signal in the transducers may not contain enough strength to ensure a robust transfer function reconstruction. Four zones of the HTL track are highlighted in Fig. 7 corresponding to locations where the signal strength was higher than the sensor's electrical noise floor. The zones where signal strength was adequate were primarily concentrated in sections of curved tracks where the wheels were suspected to be flanging [23]. Conversely, the tangent sections of the track had relatively low acoustic signal strengths. Furthermore, as also shown in Fig. 7, increasing test speeds result in increased zones of good signal strength as a result of the stronger wheel excitations.

To compute the ROC curves for the test runs, the rail track is first discretized into “pristine” and “defective” portions. The terms “defect” and “feature” are used equivalently with the term “discontinuities” herein. The pristine sections of the track have no discontinuities, while the defective segments have either a joint, a weld, or a transverse defect. Since the locations of the discontinuities cannot be ascertained with absolute precision, a portion of the track (called the pristine margin) is eliminated before and after each feature to build the pristine sections of the rail, as shown in Fig. 8(a). Similarly, for the defective population, a portion of the track (called the defect margin) is included before and after each feature, as also shown in Fig. 8(b). The pristine margin and the defect margin are kept the same for a particular analysis. The ROC curves are computed for each discontinuity by scanning the track from the start of the run until the end of the run and observing the number of times the DI value from the reconstructed transfer functions crosses a given threshold within the scanning length (gauge length of the prototype of 1.5 ft). If the total count of crossings exceeds the threshold by the number of threshold crossings (NTC) in the defective segment of the track, a “positive detection” is recorded. If, instead, the number of crossings exceeds the NTC threshold in the pristine segment of the track, a “false alarm” is recorded. An NTC value of 7 was found to result in optimum performance and was used in computing all the ROC curves presented in this paper.

Figure 9 shows the ROC curves for one test run at 40 mph on the HTL track. The curves for welds are shown in Fig. 9(a), and those for joints are shown in Fig. 9(b). The curves were computed for three different defect margins (tolerance ranges): ±3 ft (0.9 m), ±5 ft (1.5 m), and ±10 ft (3 m). A defect search range of ±10 ft means that if the DI values crossed a threshold for a certain number of times (NTC) in the range of ±10 ft from the actual location of the defect, a true detection is assigned to that location. The plots also show the AUC metric that indicates the overall damage detection performance of the system for different threshold levels. A higher AUC represents better performance and visually represents a curve shifted toward the upper-left corner of the ROC graph.

The results in Fig. 9 show that as the “defect margin” increases, the curves shift toward the top-left with a corresponding increase in the AUC metric, indicating an increase in overall detection performance (higher PD and lower PFA). This is the case for both weld and joint detections. This result is expected because an increase in the “defect margin” means that the true defect could lie within a larger distance from the identified location and still be considered a positive detection. A defect margin of ±10 ft seems like a reasonable range given the uncertainties of the location index and the required phase of manual defect verification that follows a detection. Considering this margin, for example, the plots suggest that a PD of 90% for either welds or joints can be achieved at the expense of a PFA of 35%. Further increasing the PD also results in an increase in PFA. This level of PFA is probably unacceptable to an actual inspection system to be deployed in service, and improvements to the prototype are being made in preparation for future tests as discussed in the Conclusions section to address this issue.

Figures 10(a)–10(c) show the ROC curves for the welds, joints, and defects in the HTL track at three different test speeds of 25 mph, 33 mph, and 40 mph. As in the previous figure, the AUC metric is presented for each of the curves. A discontinuity margin of ±10 ft was used for these results. The results show that higher speeds generally produce improved detection performance, as expected by the stronger wheel-generated excitation of the rail. For example, from Fig. 10(c), at a test speed of 40 mph, the internal defects can be detected with a PD of 100% with a PFA as low as 17%. At a speed of 33 mph, the rate of detection drops to 63% (PD = 63%) for the same rate of false alarms (17%). If the speed is lowered to 25 mph, the rate of detection further drops to 40%. As found in the previous figure, these PFA are still too high for system deployment in service, although they represent a good first start upon which improvements can be made.

The length of the baseline distribution in the statistical outlier analysis used to compute the DI traces plays a key role in the system's detection performance. The effects of changing the length of the baseline distribution are analyzed in terms of ROC curves in Fig. 11. The figure shows the ROC curves at 80 mph on the RTT track for welds and joints for the different baseline distribution lengths of 30, 60, 120, and 240 points. At 80 mph they correspond to approximately 3.5 ft (1 m), 7 ft (2 m), 14 ft (4 m), and 28 ft (8 m) of rail segment preceding the current location (adaptive baseline). The results indicate that a reduction in baseline length generally improves the discontinuity detection performance. This can be explained by the fact that a shorter baseline “focuses” the analysis to a more local portion of the rail. With a 3.5 ft baseline length, for example, the plots show a PD of 90% for welds with a PFA of 33%, and a PD of 90% for joints with a PFA of 27%.

It was discussed in the Introduction section how the ability of a train to perform rail inspection during normal traffic operations can provide the test redundancy that would result from multiple train passes over the same segment of rail. Such redundancy is expected to improve the defect detection performance by, for example, reducing the rate of false alarms (PFA) for the same PD.

A preliminary analysis on the effect of test redundancy has been completed and is shown in Fig. 12. The figure presents two ROC curves computed for the weld discontinuities using a single run and two separate runs on the HTL track at a speed of 40 mph. When compounding the two runs, similar locations flagged in both runs are considered “true detections,” whereas different locations flagged in the two runs are discarded as “false alarms” as shown in Fig. 13. The ROC curve for the two runs (Fig. 12) shows a slight improvement compared to that for the single run. For example, for a PD of 80%, the PFA for the single run is 30% and it decreases to 27% for the two runs. It can be reasonably expected that the advantages of redundancy will increase with an increasing number of runs.

This paper presents an initial analysis of a high-speed noncontact ultrasonic rail inspection concept that utilizes the rolling train wheels as the excitation. The transfer function of the rail between two points is extracted passively from a normalized cross-correlation operation on the signals recorded by two transducers. The attraction of this idea is the possibility to install it in trains running at revenue speed, therefore with no or minimal disruption of normal train traffic operations and a great opportunity for test redundancy. Data from a prototype tested in the field at TTCI were analyzed in terms of ROC curves to evaluate its detection performance for joints, welds, and internal defects. In particular, the performance was evaluated as a function of various relevant operational parameters, including different excitation strengths, defect position margins, test speed, length of statistical baseline distribution, and multiple runs. The results generally show expected trends, such as improved performance with higher speeds, larger defect position margins, smaller baseline lengths, and multiple runs.

The rate of false alarms (PFA) computed in the tests is still too high for field implementation of this technique. Ways are being currently devised to reduce the PFA to acceptable levels. One current investigation is the use of a continuous acoustic source to augment the wheel excitations and provides signal stability also in instances of low excitation such as slow speeds and tangent track with little or no wheel flanging. Additional work is being conducted to consider additional signal metrics in the feature vector of the outlier analysis to further enhance the robustness of the discontinuity detection. Plans are in place for the next field test that will be conducted in the next several months with the improved prototype based on the lessons learned in the previous tests. An interesting exercise that could be carried out in a future paper is to process the data in the opposite direction of the running train, i.e., by comparing a “current” location with a baseline collected “ahead” of the train. This analysis should theoretically provide the same ROC results, although there may be ways to exploit the double directions for improved detection performance.

This project was funded by Federal Railroad Administration (Contract #693JJ620C000024) with Dr. Robert Wilson as the Program Manager. All test support personnel at the Transportation Technology Center are greatly acknowledged. Dr. Albert Liang, former doctoral student at the University of California San Diego (UCSD) and currently with SpaceX, California, is also greatly acknowledged for his key role in designing, assembling, and testing various versions of the test prototype. The other graduate students who helped with the field tests, namely Ranting Cui, Ali Hosseinzadeh, and Izabela Batista, also provided invaluable help. The U.S. Federal Railroad Administration (Grant No. 693JJ620C000024).

There are no conflicts of interest.

The datasets generated and supporting the findings of this article are obtainable from the corresponding author upon reasonable request.

x =

feature vector extracted from the transfer function

$x¯$ =

mean of the feature vector from the baseline distribution

N =

total number of segments in a snapshot

D_i =

number of discontinuities detected during the test run

D_p =

total number of discontinuities spuriously identified in pristine rail segments

D_t =

total number of discontinuities present in the probed rail

P_t =

total number of pristine rail segments scanned

G_AB(f) =

transfer function in frequency domain between two points A & B

G_AB(t) =

transfer function in time domain between two points A & B

N_A(f) =

additive noise at location A

N_B(f) =

additive noise at location B

O_A(f) =

response at transducer location A in frequency domain

O_A(t) =

response at transducer location A in time domain

O_B(f) =

response at transducer location B in frequency domain

O_B(t) =

response at transducer location B in time domain

W(f) =

input wheel excitation spectrum

WA(f) =

transfer function between the wheel and location A

Cov =

covariance matrix of the baseline distribution

τ_ij =

time-lag between two segments with indices i & j of a snapshot