MOVING TRAIN

Field of the Invention The present invention relates to detection methods of wheel flatten defects on a moving train, especially relates to methods using acoustic emission sensors.

Background of the Invention At a global scale and in the last few decades, rail networks are gradually getting busier with rolling stock travelling at higher speeds and carrying heavier axle loads. Therefore, in-service inspection and maintenance have become a major issue for the safe, efficient and reliable operation of the rail network (Papaelias, Roberts, & Davis, A review on non-destructive evaluation of rails: state-of-the-art and future development, 2008). The transport systems are one of the most important elements that contribute to the economic strategy and to enhance the market equilibrium of countries. More high-speed lines are constantly being designed and constructed to help stimulate mobility and economic growth across the European continent (Internationan Railway Statistics, UIC, 2008). In order to maintain economic growth, mobility and cooperation within Europe, it is imperative that the rail industry meets the operational, socioeconomic and environmental demands and expectations as described in the latest White Paper for Transport published by the European Commission in 2011 (EUR-Lex Access to European Union Law, 2010). The immediate key challenges faced by the rail industry are the improvement in the safety of the railway systems, the development of new railways to accommodate the continued growth in demand, innovative developments to maximise the capacity

of the existing infrastructure and contributing to a more sustainable railway, in both environmental and financial terms, by delivering efficiencies and exploiting technological innovation (Suarez, Chover, Rodriguez, & Gonzalez, 2011). The Railway industry is an interdisciplinary industry composed of a complex engineering system with numerous critical components of the railway infrastructure together with the rolling stock. Deterioration of the structural integrity of these critical railway components can result in emergency maintenance being required or even potential accidents, causing severe disruption on normal services, delays and unnecessary costs (Papaelias, Amini, Huang, Vallely, Dias, & Kerkyras, 2014). The aim of this project is for the rail industry to significantly reduce delays, to increase network capacity, to improve efficiency and to minimize accidents. Traffic density on rail networks has been continuously increasing in recent years with a new high speed line being built to carry more passengers with heavier axle loads. The effective and efficient application of advanced structural integrity inspection systems for railways can help infrastructure managers and rolling stock operators to control the condition of critical assets efficiently and cost effectively (Council Directive 96/48/EC of 23 July 1996 on the interoperability of the trans-European high-speed rail system, 1996).

The patent document no. US6951132 discloses a system and method for determining at least one parameter related to a train traversing on a railway track is provided. The system comprises a sensor coupled to a detection location and configured for sensing acoustic signals at the detection location on the railway track and a processor coupled to the sensor and configured for analyzing a temporal progression of a frequency spectrum corresponding to the acoustic signals.

The patent document no. US4129276 discloses a method and apparatus for detecting the presence of flat wheels on railroad cars, comprising an electro- acoustic transducer located on the track wayside so as to pick up the vibrations generated by a passing train. If a flat wheel is present it will generate a periodic clanging sound at a frequency proportional to train speed and wheel diameter. The invention capitalizes particularly on the measurement of train speed to control the response of an adaptive filter so as to enhance the periodic clanging frequency with respect to the background noise, thereby to improve the signal-to-noise ratio; the enhanced signal is further auto-correlated for ten wheel revolutions and if a periodic signal is present in the narrow frequency band of interest, a large periodic autocorrelation output will result and, as a consequence, any wheel flat will be readily detected and will act to trigger an alarm to alert the train crew of the condition.

Summary of the invention

The object of the invention is to realize a method for detection of wheel flatten defect on a moving train which is operable by using only one acoustic emission sensor and also operable by using plurality of sensors if desired. Detailed Description of the Invention

A method realized to fulfil the objective of the present invention is illustrated in the accompanying figures, in which: Figure 1 is the flowchart of the method.

Figure 2 is the measured signal by acoustic sensors.

The steps illustrated in the figures are individually numbered where the numbers refer to the following:

100. Method

101. Acquiring data during a train passes on rail, by at least one acoustic emission sensor mounted on at least one predetermined location on the rail,

102. Modeling a noisy periodic impulsive signal that contains data of flattened surfaces hit the rail and data of the external noise, using the data acquired by at least one sensor, 103. Dividing measured signal into non-overlapping blocks which are sized according to wheel diameter and train speed, with a predetermined sampling period,

104. Detecting occurrence time of peaks in the signal and signal power at that time by calculating Root Mean Square (RMS) power at each block with a sliding window,

105. Determining periodically repeated peaks by calculating deviations of the occurrence times and the signal powers,

106. Calculating a defect score curve which is a ratio of number of peaks whose time and power differences are within a predetermined limit defined by parameters of deviations in time and signal power, to the total number of peaks,

107. Calculating a likelihood ratio which measures ratio of total number of points in defect score curve that are greater than a predetermined threshold to number of the points in defect score curve that are less than the threshold,

108. Checking if the likelihood ratio is greater than a defect decision parameter, corresponding desired detection rate and false alarm rate, which is calculated by experimental data,

109. Detecting wheel flatten defect if the likelihood ratio is greater than the defect decision parameter,

110. Determining that the wheel doesn't have a flattening on its surface if the likelihood ratio is less than the defect decision parameter.

When a train that has a wheel with flattening surface travels on the rail, the Acoustic Emission sensor measures a noisy periodic impulsive signal x(t)), where the signal period depends on the train speed {9) and the wheel diameter (D). In that case, it is desired to detect the event, where an unknown impulsive signal periodically repeats with a specified period of πΌ/9: Since the unknown signal p(t) is impulsive, the location and signal power of the impulsive signals are determined with computationally efficient Root-Mean-Square (RMS) signal power calculation method and identify the peaks. Then the time delays and power differences between the consecutive peaks is defined to generate a "defect score curve" that represents the joint time delay and power deviation of the detected peaks. Then, a decision is made about the wheel flatten defect by comparing the measured defect score curve with the threshold curve, which is obtained in the experimental data. The details of each step are explained in the following sections.

When the flattening part of the wheel, hits the rail, it creates an acoustic vibration on the rail and a resulting signal is measured by the Acoustic Emission sensor(s) located on the rail (101). This phenomenon repeats at every turn of the wheel and a measured signal x(t)) can be modeled for a total number of K turns as

K— 1

x{t) = ^e~ ° ^:C ( - ^r k ) + n(t) ,

where p(t) and n(t) are the impulsive signal generated at each hit and the noise, respectively; a is the attenuation coefficient and ¾ is the time that the flattening surface of the wheel hits the rail at the k't turn and <%k ^{= β} · _? where d corresponds to the distance between the first impact point to the position of the acoustic emission sensor. Note that the characteristics of the signal p(t) may change with the type and the size of the flattening surface. Since the aim is to detect a wheel flatten defect with any type and size, p(t) is considered as an unknown parameter as well as the noise n(t). On the other hand, the parameters a and T are independent of the type and the size of the flattening surface. The attenuation parameter a determines how the signal power degrades with the distance between the measurement point and the point, where the flattening surface hits the rail and depends on the rail characteristics. The parameter ¾ determines the k'th hit time of the flattening surface and when the train travels it directly relates with the train speed 9 and wheel diameter D as

In this case, the measured signal x(t)) becomes a noisy periodic impulsive signal (102). The goal is to decide whether there is a wheel flatten defect or not for a given signal x(t)) measured when the train travels. It is handled as a binary classification problem, where observing the periodic impulsive signal is considered as "Defect" and observing noise-only signal is considered as "Normal".

One of the steps in the proposed defect detection method is to identify the peaks in the measured signal x(t) that corresponds to the hit events, i.e., the event, where the flattening surface of the wheel hits the rail. There are many methods for the peak identification that can be used for this application. Since the acoustic emission sensor(s) measures all the signal harmonics, derivative search based peak detection methods may result too much peaks, which cannot be handled effectively. Moreover, these methods are computationally inefficient for long sequence signals. Therefore, a simpler and computationally more efficient peak identification method is proposed that is based on searching the maximum signal power within predefined window. In step 103, the measured signal is divided into non-overlapping blocks Xb(t) for b = 0, 1, . . . , B - 1 as where T is the sampling period, Γ and B are the block size and the total number of blocks, respectively. Since, identifying the peaks in the signal that are generated when the flattening surface of the wheel hits the rail is desired, the block size Γ is selected by considering the wheel diameter D and the train speed 9- as

where l^ l is the nearest integer greater than or equal to x E R. In this case, it is

guaranteed that when there is a wheel flatten defect, the impulsive signal p(t) is observed at each block. Then, the occurrence time of the peaks in the signal and the signal power at that time can be measured by calculating the Root-Mean- Square (RMS) power at each block with a sliding window as

where W and F are the window size and the offset between the successive windows in samples, respectively (104). At each block, the RMS values are calculated for M point, which is related with the window size W, offset size F and the block size B as

M

F

Note that since the signal p(t) is impulsive as shown in Figure 6, the point in r _¾ [m] with the maximum RMS power corresponds to the b'th peak in the measured signal x(t). Thus, the occurrence time ¾ and the power σ _¾ of the b'th peak can be defined as where

Note that since the RMS power values of the signal in time domain is used for the peak identification, ^m * corresponds to a set of indexes that are used for RMS calculation. Therefore, the occurrence time of the peaks can be estimated with an ambiguity, i.e., the exact location of the peak time domain can be any point within the window size W. More clearly, W- ^■ > where is the actual occurrence time of the b'th peak. Without loss of generality, all the occurrence time of the peaks are selected as the first point of the window in

To increase the accuracy of ¾, the window size W should be decreased. On the other hand, small W may result selecting the impulsive noise with high instantaneous power and small duration as a peak instead of longer duration impulsive signal, which is the one desired to be located and decreases the estimation accuracy of . Therefore, the window size W should be selected by considering the characteristics of the impulsive signal p(t) and the noise signal measured on the rail.

After identifying the peaks in terms of the occurrence time and the signal power, the deviations of the occurrence time and the signal power are calculated to determine the peaks, which are periodically repeated, i.e., the peaks generated when the flattening surface of the wheel hits the rail (105). To that end, the time deviation and the power deviation ^A ¾ of the b'th peak are defined for b = 0, 1, . . . , B -2 as Δ;. — ~ i — i tr ^. i - m _¾ ) FT 4· Γ:Γ,

Note that when the peaks are observed due to the flattening surface of the wheel, which we want to detect, the locations and the values of the maximum RMS value at consecutive blocks are similar, i.e., *'*Η 5 and ^{,! '} " ^ ¹ . In that case, the time and power deviations can be written as ^;¾; ^ and ^;¾; i. On the other hand, when the peaks are generated by the noise or the arbitrary events such as the vibration of gravels in the railway, both the time and power deviations can be vary arbitrarily. Considering these, a step (106) is proposed for identifying the type of the event that generates the peaks by defining defect score curve Ω as

where a and β parameters specify the deviations in time and power, respectively;

andf(x, y) E (0, 1} is the selection function, which is defined as

The defect score curve 0 < Ω(α, β) < 1 measures the ratio of the number of case, where the time and power differences between two consecutive peaks are within the limits defined by parameters a and β over the total number of peaks. When there is a flattening surface on the wheel, it generates periodic peaks due to the observed impulsive signals and the time and power differences between consecutive peaks are small, which results large Ω. On the other hand, when the flattening surface is not exist on the wheel, the detected peaks are most probably due to the noise or any arbitrary events that is not periodic and in this case Ω is small as compared with the previous case for the same parameters a and β. Therefore, a likelihood test condition to identify the case is proposed (107), where there exist a flattening surface on the wheel, i.e., by using the defect score curve Ω(α, β) for various a and β values as

The likelihood ratio L measures the ratio of the total number of points that are greater than the threshold Ψ(α, β), to the total number of points that are less than the threshold over the defined set of parameters A. Here, 0 < Π < 1 is a user defined parameter that controls the desired detection and false alarm rates

_ .a

and the sets * ^' v and ^: - ^f define the regions, where the threshold Ψ(α, β) is determined according to experimental data. Then, the test condition can be defined for Wheel Flatten Defect detection as se

where R = 1 corresponds to the case, where Wheel Flatten Defect is detected.

The critical point in the proposed decision mechanism is to select the threshold

Ψ(α, β) and the search region defined by ?■ and P properly. In the following, a training method for estimating the threshold and identifying the search region for a given labeled data, i.e., Defect and Normal by considering the desired detection rate and false alarm rate is defined.

In a preferred embodiment of the invention to estimate the threshold parameter for wheel flatten defect detection, a search based supervised training method is proposed that is controlled by the desired detection and false alarm rates. The method estimates the optimum threshold values or optimum threshold curve in terms of required minimum detection rate and maximum false alarm rate at each selected time and power deviation parameters a and β independently and identifies the valid regions for these parameters. To that end first, each measured data is labeled as either "Defect" or "Normal". A defect score curve Ω(α, β) is calculated for each data for a given parameter sets a and β. Let ) denotes the defect score for the z'th signal in the data set ¾%that contains Wheel Flatten

Defect and — " denotes the defect score for the /th signal in the data set ^« that do not contain wheel flatten defect, i.e., "Normal". Then, at each parameter set (a, β), we search for the optimum threshold value that satisfies the desired minimum detection rate d and maximum false alarm rate as

u. / ^j 1¾«, β, o} } / {Pf { , 0 _t ), ) > 0, where are detection rate and false alarm rate of the scores il^u . -'j ^ ajjjj β» ) ^fc: respectively for the threshold ψ and defined as

The condition in that the optimum threshold value is defined above guarantees that the estimated threshold value Ψ(α, β) satisfies the user defined detection and false rate limits at each parameter set (α, β) and the minimization operation with respect to Φ that select the threshold value as the maximum distant point to the boundaries of both sets, i.e., Defect and Normal. On the other hand, it is not guaranteed that the condition is satisfied for all parameter sets (α, β). In that case, the threshold values are set in these parameter sets (α, β) as invalid and are not used for the wheel flatten defect detection. The valid parameter sets i¾ P) can be defined that are used for the Wheel Flatten Defect detection in (13) as

In a preferred embodiment of the invention, the size of the wheel defect can be calculated according to power of the peak determined in the step "Determining periodically repeated peaks by calculating deviations of the occurrence times and the signal powers (105)". In a preferred embodiment of the invention a trigger system is used for activating the acoustic emission sensors. As the transmission speed of sound on the rails is very high, acoustic emission sensors may start to generate signals even if the train is very far away from the location that the data acquiring desired to be started. To prevent acquiring unnecessary data, a trigger system that comprises; - a trigger unit for detecting the passage of the train wheel, located on a predetermined location on the rail,

- a communication unit,

- an activation unit that activates the at least one acoustic emission sensor after receiving the data of the wheel passage from trigger unit via a communication unit is used.

The trigger unit may be an axle counter, an acoustic probe, a guard sensor, a laser sensor or an infrared sensor.

Within the scope of these basic concepts, it is possible to develop a wide variety of embodiments of the inventive "contrast stretching method (100)". The invention cannot be limited to the examples described herein; it is essentially according to the claims.