Go To On-Line Conferencing Forum

Interpretation of Electrochemical Voltage Noise in Corrosion of Mild and Stainless Steels

R. L. Motard, B. Joseph and X. D. Dai
Department of Chemical Engineering
Washington University
St. Louis, Missouri
email: motard@wuche.wustl.edu

Abstract

The results suggest that the wavelet transform could be effectively used to construct novel sensors to monitor and identify corrosion in fluid handling systems.

1. The corrosion monitoring problem

Corrosion is defined as the destruction or deterioration of a material because of reaction with its environment. Uniform attack is normally characterized by a chemical or electrochemical reaction that proceeds uniformly over the entire exposed surface or over a large area. The metal becomes thinner and eventually fails. This form of corrosion, however, is not of too great concern from the technical standpoint, because the life of equipment can be accurately estimated on the basis of comparatively simple tests.

Other forms of corrosion are considerably more difficult to predict. They are localized. The attack is limited to specific areas or parts of a structure. As a result, they tend to cause unexpected or premature failures of plants, machines, or tools. Crevice corrosion and pitting corrosion are commonly occurring examples of localized corrosion. The focus of this study is the monitoring of the presence and intensity of these two kinds of localized corrosion.

Intensive localized corrosion frequently occurs within crevices and other shielded areas on metal surfaces exposed to corrosives. This type of attack is usually associated with small volumes of stagnant solution caused by holes, gasket surfaces, lap joints, surface deposits, and crevices under bolt and rivet heads. The basic mechanism of crevice corrosion suggests that crevice attack proceeds steadily once started (Fontana, 1986).

Pitting results in holes in the metal. These holes may be small or large in diameter, but in most cases they are relatively small. Pitting is one of the most destructive and insidious forms of corrosion. It causes equipment to fail because of perforation with only a small percent weight loss of the entire structure. It is often difficult to detect pits because of their small size and because the pits are often covered with corrosion products. Pitting is particularly vicious because it is a localized and intense form of corrosion, and failures often occur with extreme suddenness. It is believed that during the initiation or early growth stages of a pit, conditions are rather unstable. Many new pits become inactive after a few minutes. It appears that pitting, though quite similar to crevice corrosion, deserves special consideration since it is a self-initiating form of crevice corrosion, which does not require a crevice. It is also known that all systems that show pitting attack are particularly susceptible to crevice corrosion, but the reverse is not always true: many systems that show crevice attack do not suffer pitting on freely exposed surfaces.

2. Previous work in corrosion monitoring

Corrosion of metals is an electrochemical phenomenon. Therefore, measurements of electrochemical parameters associated with corrosion processes may be used to estimate the corrosion attack. Methods that have been used in the past and are known and well documented include the linear polarization resistance method. Using the method, a DC signal is applied to a corroding cell consisting of two or three electrodes and the resulting DC polarization is monitored. Provided the applied current is small so that the potential shift is small, the response is linear in most cases and the measured resistance (called the polarization resistance) may be related inversely to the rate of uniform corrosion attack. Other techniques include the application of electrochemical impedance in which a sine wave current is applied, and concurrently the sine wave potential resulting from the applied current is monitored. These electrochemical techniques can only provide information on uniform corrosion conditions.

Various attempts have been made to monitor the localized corrosion. The use of electrochemical noise analysis (in the absence of an applied signal) has shown the most promising results (Silverman, 1995). Essentially, electrochemical noise has been referred to as spontaneous current and potential fluctuations between two or three, similar or identical metallic specimens (electrodes). It is thought that spontaneous electrochemical noise arising between identical electrodes can be a rich source of information concerning the processes simultaneously occurring at corroding interfaces (Roberge, 1993). Therefore, the study of spontaneous current or potential fluctuations for the characterization of corrosion processes has received considerable attention in recent years (Eden and Rothwell, 1992; Bertocci, 1989). The combinations of deterministic and stochastic events which make up the signal contents often complicate the extraction of useful information from the sensor.

An early study by Iverson (1968) sparked an increased interest in measurement of electrochemical noise and its relationship to localized corrosion (Roberge, et al. 1989). Simoes and Ferreira (1987) presented their qualitative observations in the study of crevice corrosion on stainless steel. The results showed no consistent correlations between noise transients and physical phenomena. Legat and Zevnik (1993) also reported their observations on the qualitative base, for the study of mild and stainless steel in various water solutions. Searson and Dawson (1988) proposed that the standard deviation of the potential noise fluctuations can be related to the corrosion rate and the noise spectra can give information on the nature of attack and differentiate between uniform and localized corrosion. However, Mansfeld and Xiao (1993) stated that Searson and Dawson's conclusion is not always correct. They found that the potential noise data did not agree with the known corrosion behaviors of the system by using time domain and frequency domain analysis methods. Roberge (1993) in his recent paper also expressed a similar concern that the claimed characteristic features of the noise power spectral density plot remains a problem, since much useful information is lost when real time data is converted into the frequency domain.

Although these spontaneous fluctuations between electrodes have been observed by several workers, consistent correlations between the noisy signal features and an actual process of localized corrosion have not been obtained. Some attempts made previously to study electrochemical noise have been concentrated in the time domain or frequency domain with the hope of identifying characteristic signatures of localized modes of attack. The time domain methods are mainly based on statistical analysis, such as the calculations of the mean, standard deviation and variance, etc. While the frequency domain methods focus on the observation of the Fourier transform results, such as the power spectral density (PSD) curves. It has been recognized that the presence of nonstationary phenomena can greatly complicate the final analysis currently employed. The confusion among the reported results may be due to the complexity of the corrosion process as well as the limitations of the analysis methods employed.

3.1 Experimental setup

The test data were obtained from an industrial laboratory. In the experiments, two identical coupons of metal were submerged in a given solution, and the electrochemical potential was measured by digital voltmeter (HP 3497 data logger connected to HP 3917 microcomputer, with 10-6 - 10-5 volt sensitivity) between the samples. No outside source DC or AC potentials are applied to the test electrodes in the system.

3.2 Test conditions

Various environmental conditions have been tested, including different metals, different solutions with various concentrations, different temperatures and pH values, etc. The test metals include stainless steel 304 and 316, mild steel and platinum, but mainly 304SS. The test solutions cover FeCl3, Fe3SO4 and NaCl, Fe2((SO)4)3 and NaCl, FeCl3 with added H2SO4, NaCl with added H2SO4, NaNO2, NaNO2 with added NaCl, Phosphate buffer, and two sets of plant condensate and process stream. The test temperature ranges from 35oC to 95oC, mainly 60oC. Measurements were hourly with 1024 sampled points (1 sample/sec, 1Hz) taken 18 times per day during the entire test period. The experiments may last up to six days. Table 1 summarizes the test data with experimental conditions.

3.3 Physical observation results

A visual inspection of the electrodes revealed two types of localized corrosion: crevice corrosion and pitting corrosion, as well as a baseline test that has very little or no corrosion attack. Microoptical observations revealed crevices as a localized area of attack on the top of the electrode or under an artificial crevice that was formed under the Telflon sleeve, and pitting as small pinpoints of penetration on the free metal surface. Some qualitative indices of severity of corrosion were recorded from the optical observations and cross-checked by two groups.

In the following section, we will show some typical results obtained by using traditional methods in time domain and frequency domain, and the proposed phase plane analysis method.

4. Noise analysis methods

Essentially, there are two major tasks for localized corrosion process monitoring:

(1) Identifying the type of localized corrosion, and

(2) Estimating the corrosion rate.

These are the fundamental conditions for further designing corrosion control and prevention methods effectively.

4.1 Statistical analysis in the time domain

Statistical analysis provides a simple and rapid method for the interpretation of the electrochemical noise data. For a test data set x(n), n=1, ..., N, we can easily calculate the mean Mx, the standard deviation Sx and the variance Vx.

Hladky (1986) claimed in his patent for corrosion monitoring that corrosion rate r is proportional to the standard deviation Sx of the potential noise data, with

(1)

where C is a constant.

Searson and Dawson (1988) also reported that the standard deviation of the potential fluctuations between the two electrodes corresponds to corrosion rate of the test sample in the following form:

(mpy, meter per year)(2)

Here, the correlation constant C = 10-5.

However, it is observed that the correlation data (corrosion rate r verse time t) has some degree of scatter. This scatter problem is not welcome in the analysis and cannot be satisfactorily explained based on the observed physical phenomenon. As a result, the corrosion rate estimation can only provide qualitative information about the process. It is difficult to obtain accurate quantitative estimation results. When the rate estimation data scatter severely, qualitative interpretation is also difficult.

Example 1: Consider the following test data, using time domain statistical analysis method.

(a) Test run B1: Phosphate buffer solution at 60 oC, with 304 Stainless steel electrodes. It is a baseline test, which should not show any corrosion activities in the test samples.

(b) Test run C1: 0.001M FeCl3 solution at 60 o C, with 304 Stainless steel electrodes.

(c) Test run C2: 0.01M FeCl3 solution at 60 o C, with 304 Stainless steel electrodes.

The same test material, 304SS, is used in the experiments, and the test temperature is uniform. By varying the concentration of the test solutions (corrosion intensity), the different degrees of corrosion attack are expected and can be compared. The test setup is the same as that described in the previous section. Among them, test run B1 is subjected to no corrosion attack, and test run C2 has a more severe corrosive environment than test run C1. Laboratory microoptical observation of the test coupons verified the above expectation.

The original sample data of B1, C1 and C2 in the time domain at time t = 25 hours are shown in Figures 1 - 3. The data show various trends in addition to dynamic "noise". Later, the dc trend of the data is removed to emphasize the dynamic portion of the signals. To measure the corrosion intensity (the corrosion rate) various statistical measures are calculated.

Figure 5 is the time dependence of the standard deviations S_B1, S_C1 and S_C2 The standard deviation plot shows the expected corrosion intensity sequence in general, i.e. the intensity of C2 is greater than C1, and the intensity of C1 is greater than that of B1. However, we can also see a lot of scatter in the plots. This scatter effect may confuse the qualitative judgement of the degree of corrosion intensity and affect the results of the quantitative estimation of corrosion rate using equation (1).

4.2 Frequency domain analysis methods

Another method of analyzing the electrochemical noise data is to transform the data into the frequency domain. The main tool here is the power spectral density (PSD) plot obtained either by nonparametric methods using the Fast Fourier Transform (FFT) or by model-based parametric methods.

Differentiating the types of corrosion is based on the observation of the shape of PSD curve, such as the roll-off point (frequency) in the curve and the slope of the curve. Hladky (1986) has patented his discovery that pitting corrosion has a typical roll-off rate in the power spectral density plot and crevice corrosion also has a characteristic feature in the PSD plot. He observed that different test materials, for instance Copper, aluminium and mild steel, have different frequency domain feature values. Searson and Dawson (1988) also indicated in their paper that the form of the noise spectra can give information on the nature of attack and differentiate between uniform and localized corrosion. Their study is also based on the roll-off rate of the PSD curve. However, the following example shows that these frequency features are not always reliable and can be affected by several factors, such as spectral leakage effects, the smoothing window of the transformation, and the length of the data set, etc.

Example 2: Using power spectral density curves obtained by nonparametric and model-based parametric methods, the same data sets as used in Example 1 are employed to which is added

(d) Test run P1: 0.001M FeCl3 solution at 60 o C, with 304 Stainless steel electrodes.

This group of data cover different types of corrosion attack, indicated by the final physical observation. Test run B1 is a baseline test, run C1 and C2 have crevice attack with different degree of corrosion intensity, and test run P1 has pitting corrosion attack on the test sample.

Power spectral density plots of the data obtained by nonparametric methods exhibit fluctuations known as spectral leakage effects. This effect is due to the finite length of the original data record. These undesirable effects are best alleviated by the use of windows that do not contain abrupt discontinuities in their time-domain characteristics. Different types of smoothing windows have been proposed. Figure 7 is the power spectral density plot for the data, using a Hanning window length of 64. We can observe that the curves become fairly smooth. However, we cannot find the characteristic features claimed for either crevice corrosion data sets C1 and C2, or pitting corrosion data set P1. The roll-off rate of these PSD curves fail to distinguish between the two types of corrosion.

A second approach is to obtain the PSD plot by the parametric method. The nonparametric power spectrum estimation methods used previously are relatively simple, well understood, and easy to compute via the FFT algorithm. However, these methods require the availability of long data records in order to yield the necessary frequency resolution that is required in many applications. Furthermore, these methods suffer from spectral leakage effects due to the windowing that is inherent in finite-length data records. The spectral leakage often masks weak signals that are present in data.

In the model-based approach, the spectrum estimation procedure consists of two steps. Given the data sequence x(n), n=1, ..., N, we estimate the parameters {ak} and/or {bk} of the model. Then from these estimates, we compute the power spectral density plot. The most popular model in use is autoregressive (AR) model. There are several approaches that can be used to estimated AR model parameters. Among them, the Burg power spectrum estimation procedure, often called the maximum entropy method (MEM), is one of the well known procedures in analyzing electrochemical noise data in the corrosion field. One of the most important factors that affects the PSD estimation is the selection of AR model order. As a general rule, if we select a model with too low an order, we obtain a highly smoothed spectrum. The model order will affect the shape of the PSD curve; an undesirable feature of the analysis. Figure 8 presents a typical result with a model order of 3. We find that the features of these PSD curves change with the model order used in the parameter estimation. Thus, the signal features of the PSD curves depend on the analysis procedure and are not reliable.

4.3 Phase plane analysis approach

As we have shown in the previous sections, both the time domain statistical analysis and frequency domain spectrum analysis can give some characteristic information of signal features. Neither the time domain nor frequency domain features are clear enough to represent the signatures of the phenomena under study. This confirms the complexity of the signal features of electrochemical noise. Therefore, we turn our attention to the joint time-frequency representation of the signal features, which combines the signal features in both the time domain and frequency domain. We expect to obtain a more comprehensive description of signal features, to characterize the "fingerprint" of the localized corrosion processes.

4.3.1 Visual Analysis

After selecting a proper wavelet packets transform basis and other parameters, we can perform the transformation and obtain a transform coefficient sequence. Then, this transform coefficient sequence can be represented and shown on the phase plane. Visual analysis can then be used to identify the time-frequency structures as well as intensity of the signal features on the phase plane. The theoretical and mathematical background for these computations are too lengthy to develop here.

Example 3: Consider the same group of corrosion test data sets B1, C1, C2 and P1 as in Example 2, by using the visual tool of the phase plane analysis.

Figures 9 - 12 present the signal patterns for the test data sets B1, C1, C2 and P1 respectively. In Figure 9, the baseline test data shows almost no activity on the phase plane, except for the very low frequency part of the signal pattern. In Figures 10 and 11, the signal patterns cover almost the entire phase plane. If we look at the signal patterns in the latter closely, we can find that they share the similar time-frequency structure, but have different signal intensities. The darker the patch shown on the phase plane, the stronger the signal intensity. The intensity of the signal pattern in 11 for run C2 is stronger than that in 10 for run C1. Figure 12, on the other hand, shows a totally different signal pattern with different time-frequency structures. Clear peaks can be identified on the phase plane in this case.

From these Figures, we can see clearly that,

• a difference in the time-frequency structure of the signal patterns is obvious for the different type of localized corrosions. The time-frequency structures of the signal patterns for B1, C1 and C2, P1 are quite different and very easy to separate from each other. Further studies have shown that they actually represent the signal features, the "fingerprint", for corrosion baseline, crevice corrosion and pitting corrosion respectively.
• the difference in intensity of the signal patterns with similar time-frequency structure can also be identified. The identified intensity difference for the signal pattern C1 and C2 matches the known corrosion test environment and the visual observations of the corroded test samples.
• the signal features in the two dimensional time-frequency domain are indeed more obvious than those either in the time domain or in the frequency domain alone for the same group of observed electrochemical noise data in the corrosion tests.

4.3.2 Feature Vectors

The visualization on the phase plane clearly indicates the patterns in the dynamic signals. Beyond that, the signal pattern can be quantified more precisely and extracted to form feature vectors for machine analysis. Using the signal feature extraction procedure described in Dai (1996), we can obtain a feature vector {V(k), k = 1, 2, ..., m.} from the phase plane representation of the transform results. This feature vector essentially summarizes the distribution of the energy in the patches on the phase plane for one set of test data. In other words, each element of the vector is the number of patches exhibiting a particular intensity for the phenomenon being observed through electrochemical activity.

Example 4: Consider the same group of test data in Example 3, B1, C1, C2 and P1, and their extracted feature vectors at one sampling period.

The length of the feature vector is chosen to be m = 14, which is sufficiently long to contain enough detailed feature information. Figures 13 - 16 are the extracted feature vectors for the test data B1, C1, C2 and P1 respectively. Figure 13 shows the feature vector with very low counts of the intensity levels, which is expected for the corrosion baseline test B1. The second and third figures, 14 and 15 indicate similar feature vector structures namely, the very high counts in the lower part of feature vector. The intensity counts in 15 for C2 are higher than those in 14 for C1. This matches the previous visual observations. In Figure 16, the feature vector for P1 shows a quite different structure namely, significant intensity counts for the higher portions of the feature vector.

Extending the feature vector collection to a full day of data (18 sample periods), Figure 17 to 20 are the feature maps for the corresponding four runs, B1, C1, C2 and P1. These figures confirm the characteristics of the feature vectors we observed previously at one sample period. Also, this presentation actually shows the trend of signal features varying in time. Figure 17 is the same baseline run B1 which shows almost no corrosion activities, 18 and 19 are the runs C1and C2 with crevice corrosion, and 20 for run P1 has the pitting corrosion signature. Also, we can easily observe the intensity differences for crevice corrosion with different conditions from the data in Figure 18 and 19, where C1 represents the run for 0.001M FeCl3 solution at 60 C, with 304 stainless steel electrodes and C2 for the test run of 0.01M FeCl3 solution at the same temperature and with the same stainless steel electrodes. The corrosive environment for C2 is obviously stronger than that for C1.

The overall impression gained from the feature maps for these experiments is that a class of corrosion, once initiated, continues unabated for as long as the environmental conditions persist.

4.3.3 Intensity measure

In order to correlate the signal intensity on the phase plane with the corrosion intensity (corrosion rate estimation), we define the signal intensity index I as follows.

Definition 1: The intensity index I is introduced to measure the relative signal intensity degree within each class (type) of signal patterns, and is defined in terms of the feature vector, the extracted form of the signal pattern:

(3)

where m is the length of the feature vector V(k).

Consequently, we can refine the estimation of corrosion rate r as,

(4)

where C is a constant.

Example 5: Intensity index for the data B1, C1, C2.

Figure 21 presents the sizable differences among the severe, minor and corrosion baseline for the data C2, C1 and B1 in this example, by using the corrosion intensity index I. Comparing this Figure with the Figures (1, 2 and 3) in Example 1, we can find that the intensity index defined in the time-frequency domain is a much better indicator for estimating corrosion intensity. The new index can clearly differentiate the corrosion intensity that agrees with the known corrosion test environment, and also shows much less scatter than time- or frequency-domain methods in isolation.

5. Sensitivity and robustness

Figure 22 presents another interesting case, which is a good example of the sensitivity and reliability of the intensity measure. In each of these test runs (R664 and R666), the solution in the first day was a 200ppm NaNO2 buffer solution which does not promote any form of corrosion. On the second day, the corrosive solution, 3wt% NaCl solution was added. Accordingly, the intensity index for both trials stayed at a very low level, equivalent to the no-corrosion case like a baseline trial during the entire first day and the intensity picked up exactly at the start point of the second day when the corrosive solution was added. Also, it was noticed that these two trials were conducted under the same controllable test conditions to verify the repeatability of the experiment. The two intensity index curves show in very good agreement, which indicates the reliability of this analysis method.

6. Detailed Classification

More than 1500 feature vectors were extracted for the total set of 23 trials. Generally, classification of corrosion type was straightforward and unequivocal for severe crevicing and for the baseline cases. When both crevicing and pitting are occurring it is more difficult to separate between minor crevice and pitting corrosion occurring in the same test. We found that one could improve classsification by generating up to five feature vectors per sample window rather than one; a sample window being 1024 voltage samples at one per second. The feature extraction algorithm was applied to frequency subregions on the phase plane to yield the additional vectors.

A summary of the final classification for all test runs are shown in Tables 2 - 4. The subscripts a through f indicate the date of the measurements since some experiments were conducted for up to six days. Group I (Table 2) and II (Table 3) results involving only severe crevicing or baseline corrosion could be classified with a single feature vector extracted per sample window. Group III (Table 4) experiments with mixed pitting and crevicing activity required subregion classification.

7. Conclusions

Using a group of typical observed electrochemical noise data with known corrosion test environment and final microoptical observations, we have compared the currently available time domain statistical analysis methods, frequency domain methods, and the phase plane analysis methods: the visual phase plane analysis approach, and the extracted feature vector analysis approach. The obtained results support the following observations:

• The earlier analysis methods in both the time domain and frequency domain can distinguish some signal features for the corrosion phenomena under study. However, the results are not definitive enough and are often obscured by analysis procedure parameters.
• The phase plane analysis approach based on Wavelet Transforms offers a clearer differentiation of the class of localized corrosion, the corrosion baseline, crevice corrosion, and pitting corrosion and their intensities. Both the visual signal pattern and the extracted feature vector give the same correct answers. The extracted feature vector reveals more precise quantitative information than the visual tool.
• As the proposed method matures, corrosion experiments may in fact require only a very few hours in order to identify even low level corrosive activity. The signal patterns and intensities established early in the process persist throughout the experiment.
• The phase plane analysis approach can be an effective tool to catch a process "fingerprint" in the time-frequency domain for any interesting process phenomenon, which are difficult to identify otherwise.

Bertocci, U. "Electrochemical noise analysis and its applications to corrosion." Corrosion 89, Paper No. 24, (1989).

Dai, X. D., "Wavelet Applications in Process Sensor Data Analysis", D. Sc. Dissertation, Department of Chemical Engineering, Washington University, St. Louis (1996).

Eden, D. A. and A. N. Rothwell "Electrochemical noise data: analysis, interpretation and presentation." Corrosion 92, Paper No. 292, (1992).

Fontana, M. G. Corrosion engineering, 3rd edition, McGraw-Hill Book Company, 1986.

Hladky, K. Corrosion monitoring, US Patent 4575678, (1986).

Iverson, W. P. J. electrochem. Soc. 115 (1968) : 617.

Legat, A. and C. Zevnik "The electrochemical noise of mild and stainless steel in various water solutions." Corrosion Science, 35, 5-8 (1993) : 1661-1666.

Mansfeld, F. and H. Xiao "Electrochemical noise analysis of iron exposed to NaCl solutions of different corrosivity." Journal of Electrochemical Society, 140, 8 (1993) : 2205-2209.

Roberge, P. R. "Analysis of spontaneous electrochemical noise for corrosion studies." Journal of applied electrochemistry. 23 (1993) : 1223-1231.

Roberge, P. R., R. Beaudoin and V. S. Sastri "Electrochemical noise measurements for field applications." Corrosion Science, 29, 10 (1989) : 1231-1233.

Searson, P. C. and J. L. Dawson "Analysis of electrochemical noise generated by corroding electrodes under open-circuit conditions." Journal of Electrochemical Society, 38, 2 (1988) : 1908-1914.

Silverman, D., Monsanto Company, Personal communication, 1995.

Simoes, A. M. P. and M. G. S. Ferreira "Crevice corrosion studies on stainless steel using electrochemical noise measurements." Br. Corros. J., 22, 1 (1987) : 21-25.