You are on page 1of 8

Ultrasonics 53 (2013) 12511258

Contents lists available at SciVerse ScienceDirect

journal homepage:

Ultrasonic monitoring of erosion/corrosion thinning rates in industrial piping systems

Farhang Honarvar a,b,, Farzaneh Salehi a, Vahid Safavi b, Arman Mokhtari a, Anthony N. Sinclair b
a b

NDE Lab, Faculty of Mechanical Engineering, KN Toosi University of Technology, 17 Pardis St., Mollasadra Ave., Tehran, Iran Mechanical and Industrial Engineering, University of Toronto, 5 Kings College Road, Toronto, Ontario, Canada M5S 3G8

a r t i c l e

i n f o

a b s t r a c t
Monitoring pipe wall erosion/corrosion thinning rates is an important issue in petrochemical and power generation industries. In this paper, two signal processing techniques are utilized for estimating the thinning rate based on ultrasonic pipe wall thickness data collected over a short period of time. The rst is a combination of cross-correlation and polynomial curve tting and the second is a model-based estimation (MBE) scheme. These techniques are applied to data collected from an accelerated thinning rate apparatus and both show that they are capable of estimating the thinning rates quickly in short time periods with good accuracy. In laboratory applications, thinning rates as low as 10 lm/year were measured within 15 days with an uncertainty of 1.5 lm/year by both techniques. Although the MBE technique can yield marginally better accuracy, the greater stability and computational speed of the cross-correlation technique make it the preferred choice for industrial use. 2013 Elsevier B.V. All rights reserved.

Article history: Received 12 April 2012 Received in revised form 22 November 2012 Accepted 12 March 2013 Available online 28 March 2013 Keywords: Corrosion measurement Cross-correlation Model-based estimation Ultrasound

1. Introduction Many industrial piping systems and pressure vessels are subject to corrosion and/or erosion which result in gradual loss of wall thickness and can limit their useful life. This has three major implications: (1) direct cost of component replacement, (2) associated production loss during the replacement work, and (3) safety and environmental risks of catastrophic failures. Traditional nondestructive testing techniques such as ultrasonic thickness gauging, eddy current, and magnetic ux leakage have been widely used for ofine estimation of the remaining wall thickness [1,2]. Such measurements are performed during periodic maintenance shutdowns, typically once every 12 years; the response time of such techniques is therefore very slow. There are also intrusive methods used for long-term monitoring of pipeline corrosion, whereby a sensing device is inserted into the pipe and is subjected to the corroding environment. Such techniques use a variety of probe types including: coupons, electrical resistance, linear polarization, and galvanic probes for which the data collection is performed ofine and the process is labor intensive [1]. These techniques result in an estimate of the average material removal rate over a long time period, but yield no information

Corresponding author at: NDE Lab, Faculty of Mechanical Engineering, KN Toosi

University of Technology, 17 Pardis St., Mollasadra Ave., Tehran, Iran. Tel.: +98 21 88674747; fax: +98 21 88674748. E-mail address: (F. Honarvar). 0041-624X/$ - see front matter 2013 Elsevier B.V. All rights reserved.

Abbreviations: DSP, digital signal processing; MBE, model-based estimation.

on short-term process-related perturbations of the thinning rate. This shortcoming has highlighted the need for a greater emphasis on on-line monitoring of corrosion/erosion as a continuous process; this may enable optimization of a plants processes to bring an erosion or corrosion problem under control before the damage becomes extreme [3]. A variety of intrusive techniques have been attempted with the goal of on-line measurement of wall thinning rates. Electrical resistance (ER) probes require almost a week of data collection to measure a loss rate of 150 lm/year (with an uncertainty of 10% to 20%) in a corrosive environment [4]. There are other variations of this technique that make use of the high magnetic permeability of steel [5], but the permeability of a steel electrode is also sensitive to other parameters such as stress, shock loading, magnetic eld strength, and temperature. The eld signature method (FSM) [6] uses a network of sensing pins distributed over the area to be monitored, where the distance between pins is 23 times the wall thickness. The electrical eld measured by the pins is compared to a reference eld to infer the extent of corrosion. The system is overly sensitive to the state of stress in long runs of pipes, and the measurement of corrosion rates on the order of 100 lm/ year with an uncertainty of 15% would need approximately four months of data collection. Another category of on-line corrosion monitoring instruments is based on electrochemistry. These methods include linear polarization resistance (LPR), potential monitoring, electrical impedance spectroscopy (EIS), and electrochemical noise (EN) monitoring [1]. The interpretation of data for all electrochemical methods requires


F. Honarvar et al. / Ultrasonics 53 (2013) 12511258

careful calibration of the system. Moreover, the corrosive medium needs to be conductive and therefore, these methods are not suitable for pipes carrying hydrocarbons and gasses. Clearly such a system would not be able to detect any erosion. While the on-line methods discussed above are intrusive, nonintrusive on-line methods have also been investigated. Fiber optic sensors have been used for on-line thickness monitoring on pipelines [7]. They measure the strain induced in the ber caused by the increase in hoop stress of the pressurized pipe as the pipe wall thickness reduces due to corrosion. This technique can only be used on pipes which are under pressure. Another non-intrusive technique is based on on-line ultrasonic thickness measurement; it is the implementation and optimization of this technique that is the focus of this study. Our objective is to develop an online, non-intrusive ultrasonic system for quick and accurate estimation of the pipe wall thinning rate. The system should be able to estimate the thinning rate within 10%, on a pipe with periodic variations in temperature, in a time period during which the amount of wall thinning is of the order of 1 lm. The proposed system will be based largely on existing hardware, using improved methods of data analysis. One potential area of application of this system is on nuclear reactor feeder tubes, where corrosion/ erosion rates of up to 150 lm/year have been recorded [8,9]. In the power generation industry, moisture separator reheater (MSR) drains, heater drains, extraction steam piping, heater vents, condensate and feedwater piping are all subject to ow accelerated corrosion. Corrosion inhibitors are deposited on the inside surfaces of pipes in order to reduce their thinning rate. To evaluate promptly the success of such corrosion inhibition measures, one needs to determine the thinning rate before and after this type of surface treatment. Conversely, it is also important to assess quickly the potential for erosion/corrosion damage to a plants piping system following an unintended event that releases a corrosive agent or particulates into a system. In this paper we use two different signal processing schemes to analyze ultrasonic pulse-echo thickness measurements of a pipe wall. The rst technique is a combination of cross-correlation (CC) analysis with polynomial curve tting and the second is a model-based estimation (MBE) scheme. To evaluate these two approaches, we apply them to experimental data obtained from an accelerated corrosion test apparatus. 2. Background Fig. 1 shows the working principle of a typical ultrasonic thickness gage. An ultrasonic pulse, generated by an ultrasonic transducer, is sent into the specimen through a delay line. The pulse reverberates in the specimen and generates a series of regularly spaced reections (backwall echoes). The reected waveform is picked up by the same transducer. The wall thickness can then be found by multiplying half the echo spacing (in ls) in sound velocity. The primary source of error in the estimate of pipe wall thickness is the measurement of delay time between the front wall (FW) and backwall (BW) echoes. This in turn depends on a number of factors, including: (1) The jitter performance of the A/D clock which determines the uniformity of the time intervals between waveform data points. (2) Signal bandwidth and sampling frequency. (3) A/D bit resolution. (4) Signal noise. Negative effects originating from the above factors could be minimized through the use of more advanced timing circuits,

Fig. 1. Working principle of a typical ultrasonic thickness gage.

higher sampling rates, and higher bit count in the A/D converter. Random electrical noise can be minimized by time-averaging of consecutively collected waveforms. There remains, however, one inherent source of error, i.e. deterministic electrical and/or material noise that cannot be overcome through the conventional routes specied above. It may originate from signal reections at electrical connections, internal ultrasonic reections within the transducer or coupling layer, or reections from grain boundaries. This noise is particularly signicant for applications that involve high temperatures and harsh environments due to signicant reduction of the performance of the ultrasonic transducer. Even a relatively small amount of noise can have a large inuence on estimates of thinning rates based on data collected from just a few days of measurements. This noise does not disappear by time averaging as it is synchronized with the pulse repetition of the pulser. As the wall thickness slowly decreases with time, the backwall echo(es) move within the (stationary) background noise and undergo constructive or destructive interference depending on their mutual phases. This produces very small distortions and time shifts in the backwall echoes and hence subjects the delay time measurements to a very small, variable error, whose extent would depend on the signal-to-noise ratio (SNR). The time shift is normally sufciently small that it cannot be detected by casual inspection of the data, but can still inuence the short-term estimate of erosion rate. The thinning rate estimated over a time period Dt is:


Ds cL 2Dt

where Ds is the round trip travel time of the wave in the pipe wall and cL is the speed of longitudinal waves in the pipe material. The following key assumptions have been made in this study with regard to the ultrasonic echo signal received by a transducer permanently mounted on the outer wall of the pipe: (1) All random electrical noise in echoes received by a transducer can be removed through signal averaging. (2) The total system response function can be approximated by a superposition of a low-amplitude deterministic background noise with a series of equally spaced larger echoes, originating from the delay line and backwall reections. (3) The deterministic background noise is invariant; only the large backwall echoes move slowly to the left along the time axis as the material corrodes/erodes. This assumption is based on the premise that a transducers performance will not change in any given time period in which the thinning rate is to be measured (typically a few days).

F. Honarvar et al. / Ultrasonics 53 (2013) 12511258


One implication of assumption #2 is that at any given snapshot in time, the spatial variation of corrosion rate within the ultrasonic beam should be small, and the wall surfaces are smooth. This applies to uniform material removal from the surface of a pipe due to erosion/corrosion as observed in power plants. However, this type of thinning rate estimation is not expected to work for corrosion phenomena such as pitting, where an individual pit would be smaller than the area of the ultrasonic beam. Assumption #3 is expected to hold for small amounts of corrosion on the order of a few microns, during which the overall change to the low-amplitude background can be ignored. This is because the shifts in backwall echoes are very small and consequently the whole signal does not change noticeably. 3. Cross-correlation technique 3.1. Approach One of the two investigated techniques that can be used for nding the time delay between two echoes is cross-correlation, which is a measure of similarity of two waveforms as a function of a time-lag applied to one of them. For two waveforms f(t) and g(t), cross-correlation is dened as [10]:
Fig. 3. Cross-correlation curve, note the 180 degree phase difference between the two pulses.

f g t

in Fig. 3, includes two pulses; if the rst one is centered at time t = 0, the second one would be at a location corresponding to the time delay of the two echoes. 3.2. Polynomial curve tting Due to the differences in acoustic impedances of the two interfaces, the FW and BW echoes show a phase difference of 180. This results in a similar phase difference for the two pulses appearing in the cross-correlation curve in Fig. 3. To improve the time resolution of measurements, we t polynomial curves to the peak/dip of the two pulses (designated by arrows in Fig. 3). The order of these polynomial curves and number of points used for tting the curves can be found by statistical methods. We used the reduced Chisquared value as a measure of the goodness of the t [15]. Reduced Chi-squared value is dened as:

f sg t sds

where f(s) denotes the complex conjugate of f(s). In discrete form, cross-correlation function can be written as [11]:

f g k

1k 1 NX fn g nk N n0

k 0 ; 1; 2; . . . ; N 1

where fn and gn are two length-N signal blocks. According to Eq. (3), the best achievable resolution for delay time for discrete signals is equal to the sampling period. This resolution can often be improved by curve tting and interpolation of the cross-correlation function [1214]. The entire measured waveform including the FW and BW echoes is taken as f(t), and a windowed portion of f(t) containing the BW echo is taken as g(t) in Eq. (3), as demonstrated in Fig. 2. This window should include the second echo completely but its limits are arbitrary. The resulting cross-correlation curve, which is shown

v2 red

n X 1 1 ^ i 2 g i g n k 1 i 1 r 2 i

Fig. 2. Typical waveform collected during pipe wall thinning rate measurements, SNR = 25 dB; transducer central frequency = 4 MHz; sampling frequency = 200 MHz.

^i is the estimated value from the where g(t) is the measured value, g polynomial, and r2 i is the variance of observations, n is the number of points, and k is the order of the polynomial. For each echo, the Chi-squared value is calculated for polynomials with orders ranging from 2 to 10 and number of points ranging from 5 to 15. Based on the data collected in our experiments, the Response Surface Method [16] was used to nd the optimal polynomial order and its corresponding number of points. Fig. 4 shows a typical output of the Response Surface Method in which the polynomial orders 4, 8, and 10 are shown in Fig. 4ac, respectively. The number of points used for tting the polynomial is different for each curve as indicated on each gure. In most instances, the 8th order polynomial with 13 points was the optimal curve. The transducer central frequency was 4 MHz, with a sampling frequency of 200 MHz. The polynomial curve which best ts the experimental data has an order of eight with thirteen data points. However, several other combinations of polynomial order and number of data points yielded comparable results. After tting the polynomial curve to the data points surrounding the maximum or minimum point of a pulse, the location of the maximum/minimum point of the polynomial curve was found by setting the derivative of the polynomial function to zero. (Note that the BW echo is inverted with respect to the FW echo). This gives the time location of the maximum point of the polynomial curve with high accuracy.


F. Honarvar et al. / Ultrasonics 53 (2013) 12511258

The shape of the wavelet h(t) is typically estimated from a lownoise at reector such as the backwall of a very smooth plate. However, for the purpose of online thinning rate measurements, the transducer cannot be periodically removed from the pipe for this purpose. In high-temperature applications, transducer aging over time and changing plant operating conditions render the upfront collection of a reference echo ineffective. As a result, the precise prole of h(t) is unknown in our case. In the MBE algorithm, a Gaussian prole is used for modeling each ultrasonic echo. Measurement of the time-delay between modeled echoes provides a pretty accurate estimation of the true time-delay between real echoes. A Gaussian pulse is expressed as a function of ve parameters h = (a; s; fc ; u; b), as follows [18]:

x0 h; t beats cos2pfc t s u

where a is the bandwidth factor in Hz , s is arrival time in s, fc is center frequency in Hz, u is phase in radians, and b is amplitude. The three parameters (a, fc, u) dene the shape of the reference wavelet and hence are shared among all echoes received from a test specimen (assuming negligible dispersion). The other two parameters b and s represent the amplitude and time of arrival of each echo, which are different for the various echoes. Our objective is to use pulse-echo measurements acquired over a specied number of days to estimate the average erosion/corrosion rate over that time period, such as 5 or 15 days. Each FW and BW echo of every waveform is slightly distorted by the background noise in a slightly different manner. We can therefore use all of the waveforms collected over the measurement period to nd a single set of best values of a, fc, and u via the MBE algorithm. This process would remove the effect of the background noise to a large extent (xb = 0 and consequently it can be assumed that x0 (t) is a convolution of one discrete spike of x(t) with h(t). In the next step, the two remaining parameters s and b are optimized for the FW and BW echoes. Several variations of the MBE algorithm were assessed. The MBE algorithm that showed good stability in minimization of the difference between the experimentally collected waveforms y(t) and their respective modeled Gaussian pulses was a non-linear least square scheme (in time domain):
Fig. 4. Typical results obtained from Response Surface Method. (a) 4th order polynomial, (b) 8th order polynomial, and (c) 10th order polynomial. The objective is reducing the chi-square value as much as possible.


min kyt x01 t x02 t k2

4. Model-based estimation (MBE) technique 4.1. Gaussian model for the reference pulse A linear ultrasonic systems output can be expressed as a convolution of the transfer function x(t) of the test piece (a discrete sum of sharp spikes), and the material and electronic systems impulse response h(t) [17]:

yt xt ht et

The term e(t) represents the random additive noise assumed to vanish by temporal averaging; it is not to be considered any further in this study. For the case of thickness measurement, the test piece transfer function x(t) would ideally consist of a sequence of spikes that are equally spaced in time. The total echo signal is distorted by the response of the correlated noise, xb(t), originating from grain boundaries, reections within the transducer and coupling layer, and electronics such that y(t) is now given by:

where x01 and x02 are the FW and BW echoes, and P is the number of waveforms in the selected time interval. The scheme uses the LevenbergMarquardt algorithm to provide a numerical solution to the desired non-linear minimization problem over a space of parameters of the function x01 x02 [19]. This is a complex task which, depending on the size of the data set, can take quite a long time to be processed. Moreover, if the input parameters to the algorithm are not carefully selected, the process can easily destabilize. This is however, not the case here, since the input parameters are extracted from the measured signal. Fig. 5a shows the Gaussian pulse tted to a FW echo and its corresponding frequency spectrum compared to that of the measured echo, Fig. 5b. In the time domain, the Gaussian pulse ts the measured echo well, with some deviation apparent at the trailing edge. The time-delay between the FW and BW pulses is found by differentiating the model curve in the neighborhood of each echo, and to nd the location of its maximum/minimum value. The time difference between the locations of the maximum/minimum is the time-delay between the two echoes. 4.2. Temperature compensation scheme Ultrasonic thickness measurements are affected by temperature in two ways: (1) change of specimen thickness due to thermal

yt xt xb t ht

F. Honarvar et al. / Ultrasonics 53 (2013) 12511258


Fig. 5. (a) A typical FW echo and its corresponding tted Gaussian pulse, and (b) frequency spectra of the two echoes.

expansion or contraction, and (2) change in ultrasonic wave velocity within the material. The thermal expansion of materials due to temperature variation is described approximately by a linear thermal expansion coefcient.

This is consistent with the earlier results reported by Mak and Gauthier [22]. In this paper, the parameters of Table 1 have been used for the steel piping. 5. Experiments 5.1. Test setup Fig. 6 shows the test setup which uses the Electro-Chemical Machining (ECM) principle to remove material from a delineated area on the inside surface of a section of a pipe. A pump produces a ow of electrolyte between a cathode and the corroding surface, with the rate of corrosion controlled by the voltage drop across the two surfaces. The metal ions removed from the pipe wall form a precipitate in the owing electrolyte, and are washed away such that they do not end up collecting on the cathode; they are ltered out downstream of the metal removal area. Faradays Law is assumed to govern the material removal rate, i.e. the rate of material removal is linearly proportional to the electrical current. The test setup was designed to produce a material removal rate of up to 3 lm per minute at maximum current. This implies a time acceleration factor of approximately 104 compared to the maximum erosion rate of about 150 lm/year found in the power generation industry.

d d 1 aT T 0

where d and d0 are the current and reference (at temperature T0) thicknesses, respectively, and a is the thermal expansion coefcient of the material. The longitudinal wave velocity can be written as a function of the Youngs modulus, E, and Poisson ratio, m, as follows [20]:


s E1 m q1 m1 2m


It can be assumed that E and q are temperature dependent but m is not [3,21]; this yields the following equation governing the dependence of cL on temperature:

s 1m 1 aT T 0 3 c L T E0 bT T 0 1 m1 2m q


where E0 is Youngs modulus at the reference temperature T0. The coefcient b which governs the temperature dependence of E is dened as:

E E0 T T0


Although Eq. (11) is non-linear in T, Rommetveit et al. [21] showed that for carbon steel, at temperatures below 100 C, the relationship between velocity and temperature is nearly linear.
Table 1 Parameter values for carbon steel piping used in calculations [21]. Variable E0 Value 210 GPa 7872 kg/m3 0.295 47.7 MPa/K 12.8 ppm/K 25 C

q m


Fig. 6. Accelerated corrosion test setup.


F. Honarvar et al. / Ultrasonics 53 (2013) 12511258

The material removal in the test setup was performed inside the designated ECM cell, which was enclosed by the pipe [23]. A plastic manifold was inserted inside the pipe which provided for the passage of the electrolyte against the pipe inside surface over a small well-dened area; the rest of the pipes inside surface was masked. A stainless steel cathode formed the opposite side of the ow passage. The cathode was machined to be parallel to the pipe inside surface. The cathodeanode gap at the start of the test was around 5.3 mm; the width of the ow channel was approximately 38 mm and its length was 50 mm. In order to produce uniform corrosion over the test area, the test setup was equipped with a pulsed current source. Using a short pulse of current of the order of a few microseconds for removing the material, the ow would have appeared stagnant from the viewpoint of the released ions and led to a reduced role by the electrolytes ow turbulence on the resulting surface texture. The power supply delivered current pulses of up to a peak of 500 A over a duration of 15 ls. The average current (i.e. electron charge rate) was maintained constant at selectable levels and time periods by varying the current pulse repetition rate between 0 and 120 Hz. The time-averaged current could be varied from 0 to 380 mA, and held constant by a feedback loop. A 4 MHz high temperature piezoelectric transducer having a diameter of 0.5 inches was used for the experiments. The transducer had an integrated 25 mm long titanium delay line which was mounted on the outside surface of the pipe using a soft silver foil shim as coupling agent. A clamping mechanism pressed the transducer rmly against the shim and pipe wall. The clamping force was large enough to ensure good coupling between the delay line and pipe. 5.2. Test procedure An experimental test started with a measurement of pipe wall thickness using a deep-throat micrometer. Thirty-two measurements were performed on random spots over an area of one square centimeter. The large number of measurements was important because of the very slight non-uniformity in material removal over the inspection area. The mean value of the thickness was xi = 11.266 mm with a 95% condence interval (CI) of 2 lm. Using manufacturers data, the systematic error for the deep throat micrometer was estimated as 2.30 lm at 95% CI. By combining this value with the random uncertainty obtained from measurements, the total uncertainty was found to be 2.9 lm for the initial measurement of pipe wall thickness [24]. A 20% solution (%Weight) of table salt in distilled water was used as the electrolyte. The system was warmed up to the steady-state temperature by running the pump with no material removal (no current). The pulsed electrical current was then turned on to allow the accelerated corrosion to commence. Waveform data were collected during the accelerated corrosion test at a rate of approximately one waveform every 30 s. Each recorded waveform was the result of averaging 3000 measured waveforms; this effectively eliminated the effects of random electrical noise. Depending on the desired corrosion rate, the average current level was varied in steps between 0 mA and 109 mA. The experiment reported here took more than six hours, during which the average current setting (thinning rate) was adjusted to nine different levels, as shown in Fig. 7. Since every 30 s of the accelerated experiment corresponds to one simulated day, the total number of simulated days for our experiment is 800. At the end of the experiment, the pipe was removed from the manifold and washed thoroughly. The micrometer was used again to measure the thickness at 32 random spots over the same one square centimeter area. The average thickness was xi = 11.202 mm with a random uncertainty of 2.9 lm at 95% CI. By combining this random uncertainty with the systematic uncertainty of the

Fig. 7. Settings of average current levels during the accelerated corrosion experiment.

micrometer, the total uncertainty was found to be 3.7 lm. Based on the initial and nal thickness measurements, there was an average uncertainty of 3.3 lm in the amount of wall material removed during the experiment. The average pipe wall thicknesses before and after the test were correlated with the time-integrated current over the entire experiment; this result was then applied to the currenttime prole of Fig. 7 to estimate the average measured thinning rate for each current level that was used in the test. The resulting 95% condence interval of the average measured thickness was used to calculate the 95% condence interval of the measured erosion rate at each current level, assuming that corrosion rate varies linearly with current. 5.3. Experimental Results and Analysis The experimental waveform data les were processed by both the cross-correlation and model-based estimation algorithms, including corrections for changes in pipe wall temperature. The signal already shown in Fig. 2 was a typical waveform measured in our experiments. The pipe wall temperature was measured by a calibrated thermocouple during the experiment with a resolution of 0.01 C. The range of temperature variations during the experiment was 2.12 C as shown in Fig. 8. For the MBE algorithm, the waveforms within each moving time period were used for estimating the Gaussian function parameters. The thinning rate was assumed to be constant within that time period. Fig. 9 shows the pipe wall thickness as a function of time

Fig. 8. Variation of pipe wall temperature during the experiment.

F. Honarvar et al. / Ultrasonics 53 (2013) 12511258


Fig. 9. Reduction of pipe wall thickness during the accelerated corrosion experiment obtained from cross-correlation technique.

obtained by the cross-correlation technique. The uncertainty of thickness measurement was estimated to be 9 lm with 95% condence. This value was measured when there was no corrosion in the pipe (zero thinning rate). It should be noted that the thickness data reported in Fig. 9 is compensated for temperature changes during the experiment. Temperature compensation is also applied to all thinning rate data that will be reported later in this paper. Using the cross-correlation technique, the thinning rates were calculated using moving time periods of 15 days and plotted in Fig. 10 (solid line). For data analysis, the pipe wall thickness was calculated based on the time delay measured between the front and backwall echoes and corrected for temperature variations. The thinning rates were then estimated from the slope of the thickness data analyzed over the desired time period (15 days in Fig. 10). Fig. 10 shows good correlation between thinning rate levels calculated from the ultrasonic data using the cross correlation algorithm, and the measured corrosion rate as inferred from the electric currenttime prole of the test system displayed in Fig. 7. Complementing Fig. 10, the thinning curve obtained from the MBE technique is also plotted in Fig. 10 with dashed lines. In Fig. 10, the consistency of rate measurement at removal rates as low as 10 lm/year is noteworthy. The corrosion rates displayed in Fig. 10 has a 95% condence interval of approximately 1.5 lm/ year. The thinning rates as measured by both techniques fall

slightly outside the micrometer 95% condence intervals (shown by dashed lines in Fig. 10 at higher removal rates. It is speculated that the removal rate may deviate somewhat from proportionality to the injected electrical charge due to changes in the electron valance of the removed ions at higher current levels and as the electrolyte ages during the experiment. This is considered to be a subject that needs further investigation. For the results displayed in Fig. 10, step changes in the true thinning rate appear as a time convolution of the actual thinning rate with a window of rate determination time period (15 days here). This means that in order to evaluate a step change in the true thinning rate, at least 15 days of measurements at the new thinning rate need to be made. As a result, the step changes in the thinning rate curve shown in Fig. 10 appear as ramp-up/ramp-downs smeared over a time period of about 15 days, i.e., the data are effectively processed through a lag lter with a time constant of the order of 15 days. The use of a smaller number of days for thinning rate estimation can reduce the time constant of this lag lter, but more uctuations will then be seen in the thinning rate curve. Fig. 11 shows the thinning rates estimated based on moving samples of 5 days of data collection by the cross-correlation and MBE techniques,

Fig. 10. Thinning rate estimated based on a moving period of 15 days of data collection using cross-correlation technique (solid line) and model-based estimation technique (dashed line). The dotted and dash-dotted lines show the 95% condence interval of micrometer measurements.

Fig. 11. Thinning rate estimated based on a moving period of 5 days of data collection (a) using cross-correlation technique, and (b) using model-based estimation. The dotted lines show the 95% condence interval of micrometer measurements.


F. Honarvar et al. / Ultrasonics 53 (2013) 12511258 [4] F.W. Mauldin Jr, F. Viola, W.F. Walker, Reduction of echo decorrelation via complex principal component ltering, Ultrasound Med. Biol. 35 (8) (2009) 13251343. [5] Corrpro Companies Inc., H., TX 77040, USA <>. [6] Corrocean S.r.l., R., Italy <>. [7] D. Morison, Sensor system monitors corrosion damage, Mater. Perform. (2006). [8] D.H. Lister, J. Slade, N. Arbeau, Accelerated corrosion of CANDU outlet feeders observations, possible mechanisms and potential remedies, in: Annual Conference of Canadian Nuclear Association, 1997. [9] H.S. Chung, A review of CANDU feeder wall thinning, Nucl. Eng. Technol. 42 (5) (2010) 568575. [10] D. Brook, R.J. Wynne, Signal Processing Principles and Applications, Edward Arnold, 1988. [11] S.J. Orfanidis, Optimum Signal Processing An Introduction, McMillan Publishing Co., 1988. [12] R. Zahiri-Azar, S.E. Salcudean, Time-delay estimation in ultrasound echo signals using individual sample tracking, IEEE Trans. Ultrason. Ferroelectr. Freq. Control 55 (12) (2008) 26402650. [13] C. Pantea et al., Digital ultrasonic pulse-echo overlap system and algorithm for unambiguous determination of pulse transit time, Rev. Sci. Instrum. 76 (11) (2005) 114902. [14] F. Honarvar et al., Estimation of erosion/corrosion rate in pipe walls by crosscorrelation technique, in: Review of Progress in Quantitative Nondestructive Evaluation, Burlington, Vermont, USA, 2011. [15] J.F. Kenney, E.S. Keeping, Mathematics of Statistics, Pt. 2, 2nd ed., Van Nostrand, Princeton, NJ, 1951. [16] J. Antony, Design of Experiments for Engineers and Scientists, ButterworthHeinemann, 2003. [17] F. Honarvar et al., Improving the time-resolution and signal-to-noise ratio of ultrasonic NDE signals, Ultrasonics 41 (9) (2004) 755763. [18] R. Demirli, J. Saniie, Model-based estimation of ultrasonic echoes part I: Analysis and algorithms, IEEE Trans. Ultrason. Ferroelectr. Freq. Control 48 (3) (2001) 787802. [19] D. Marquardt, An algorithm for least-squares estimation of nonlinear parameters, SIAM J. Appl. Math. 11 (1963) 431441. [20] D.E. Bray, R.K. Stanley, Nondestructive Evaluation: A Tool in Design, Manufacturing and Service, CRC Press, Inc., 1997. [21] T. Rommetveit, R. Johnsen, . Baltzersen, Using ultrasound measurement for real-time process control of pipelines and process equipment subjected to corrosion and/or erosion, in: Corrosion, New Orleans, LA, 2008. p. Paper 08285. [22] D.K. Mak, J. Gauthier, Ultrasonic measurement of longitudinal and shear velocities of materials at elevated-temperatures, Ultrasonics 31 (4) (1993) 245249. [23] A.N. Sinclair, V. Safavi, F. Honarvar, Ultrasonic measurement of erosion/ corrosion rates in industrial piping systems, in: D.O. Thompson, D.E. Chimenti (Eds.), Review of progress in quantitative nondestructive evaluation, AIP Conf. Proc., 2011, pp. 13331340. [24] S. Richard, R.S. Figliola, D.E. Beasley, Theory and Design for Mechanical Measurements, John Wiley & Sons, 2006.

respectively. These thinning rate curves appear to be noisier than those in Fig. 10. The uctuations of thinning rates at each step in Fig. 11 were found to be less than 4 lm/year with a condence of 95%. Since overlapping of the two curves, as done in Fig. 10, makes the comparison of the two curves difcult, the two curves are plotted separately in Fig. 11a and b. 6. Conclusion Two signal processing techniques were used for measuring the erosion/corrosion thinning rate of pipe walls. The data collected from an accelerated corrosion apparatus were analyzed by these techniques. Both techniques were able to estimate thinning rate changes as small as 10 lm/year within 15 days with a 95% condence interval of 1.5 lm/year. The corresponding 95% condence interval for data collection periods of 5 days was less than 4 lm/ year. The MBE technique provides results that are marginally more accurate than those of the cross-correlation algorithm. However, the MBE technique is more complex, less stable numerically and more expensive in terms of computation time. For industrial applications, therefore, the cross-correlation technique is the preferred choice. Field testing of these two signal processing techniques is currently underway. Acknowledgements The authors gratefully acknowledge nancial sponsorship of this project by the Advanced Measurement and Analysis Group (AMAG), and the Canadian Natural Sciences and Engineering Research Council (NSERC). References
[1] V.S. Agarwala, S. Ahmad, Corrosion Detection and Monitoring A Review in Corrosion, Orlando, FL, USA, 2000. [2] M.J. Cohn, J.W. Norton, Case studies of pulsed eddy current to measure wall loss in feedwater piping and heater shells, in: N.P. Odowd (Ed.), Proceedings of the Asme Pressure Vessels and Piping Conference 2008, vol. 6, Pt a and B, 2009, pp. 721731. [3] T. Rommetveit, T.F. Johansen, R. Johnsen, A combined approach for highresolution corrosion monitoring and temperature compensation using ultrasound, IEEE Trans. Instrum. Meas. 59 (11) (2010) 28432853.