=Paper= {{Paper |id=Vol-1197/paper14 |storemode=property |title=Automatic Defect Recognition in Corrosion Logging Using Magnetic Imaging Defectoscopy Data |pdfUrl=https://ceur-ws.org/Vol-1197/paper14.pdf |volume=Vol-1197 |dblpUrl=https://dblp.org/rec/conf/aist/GaibadullinaZB14 }} ==Automatic Defect Recognition in Corrosion Logging Using Magnetic Imaging Defectoscopy Data== https://ceur-ws.org/Vol-1197/paper14.pdf

Automatic Defect Recognition in Corrosion Logging
Using Magnetic Imaging Defectoscopy Data

Rita Gaibadullina, Bulat Zagidullin, Vladimir Bochkarev

Kazan Federal University, Kazan, Russia
{rita.gaibadullina,bulatza}@gmail.com, vbochkarev@mail.ru

Abstract. The Magnetic Imaging Defectoscopy is designed for detection of cor-
rosion zones in oil wells. Location of corrosion zones is a time-consuming pro-
cess, during which some defects can be missed. Therefore this process shall be
automated. This document describes an algorithm of automatic defect recogni-
tion based on maximum likelihood criterion and the use of wavelet threshold
processing for noise reduction and pre-conditioning of experimental data.

Keywords: Magnetic Imaging Defectoscopy (MID), wavelet filtering, maxi-
mum likelihood criterion.

1 Introduction

The Magnetic Imaging Defectoscopy can be used to identify defects, corrosion inter-
vals in oil wells. The tool generates an electromagnetic pulse and receives time-
related response of tubing and casing walls. The attenuation rate of the response de-
pends on the electromagnetic characteristics of the tube material and its thickness.
Metal loss due to corrosion causes a faster decay than non-corroded metal.
The Magnetic Imaging Defectoscope (MID) contains of two sensors: the short and
the long sensors. The short sensor is 120 mm in length designed to sense the tubing.
The sensor generates a short pulse (50 ms) of low amplitude and magnetises basically
the first barrier only, and then receives the response of 0.1 ms to 75 ms. Each decay of
the short sensor consists of 42 points. The long sensor is 320 mm long; it generates
pulses of greater amplitude and duration (250 ms) and takes the total response from
both the first and second barriers (tubing and casing) within 275 ms. Each decay of
the long sensor contains 51 points. Thus, the experimental data are presented by the
42 logs for the short sensor and 51 logs for the long sensor (Fig. 1).
Each log of the long or short sensors can be divided into the trend and drift
components. The trend means a log component slowly varying with depth (can be
found, for example, using a median filter). The drift means a component rapidly
varying with depth, which shows deviation of real log from the trend [1].
𝐴(𝑡)−𝐴𝑡𝑟𝑒𝑛𝑑 (𝑡)
𝐴𝑑𝑟𝑖𝑓𝑡 = (1)
𝑆𝑇𝐷

86
2000

1500
short sensor
response, mV

long sensor
1000

500

0
1 10 100
time, ms

Fig. 1. Responses of the short and long MID sensors.
A DRIFT panel shall be built for visualisation of a drift components normalised to the
standard deviation (STD). Generally, it is a three-dimensional graph, where the verti-
cal axis shows the depths, the horizontal axis shows the decay time and the colour
determines the signal amplitude (see Fig. 2). Gain in signal at depths of X835 ft and
X820 ft corresponds to the tubing and casing collars, respectively. Gain in signal at
depths of X855 ft also corresponds to the casing collars. Reduction in signal at the
depth of X837 ft displays casing corrosion, which can be detected through the tubing.
Nowadays, corrosion zones are detected during well log analysis, i.e. their location
is arbitrary. Moreover, the analysis of 6,000 – 9,000 ft wells consumes plenty of time,
during which the defects can be missed. Therefore solution to this problem is automa-
tion of corrosion interval detection.

2 Automatic Recognition of Corrosion Intervals

2.1 Wavelet Filtering of DRIFT Data
Data are pre-filtered to remove the noise components, which could affect the
performance of the recognition algorithm.
A two-dimensional wavelet decomposition is applied to DRIFT data. This wavelet
decomposition is designed for processing of two-dimensional pictures with
commensurable number of points in X- and Y-directions. In our case, the number of
counts in the vertical axis (i.e. well depth) has an order of thousands that ten and
hundred times greater than the number of values of horizontal axis (totally 42 and 51
time-related counts), therefore the two-dimensional wavelet decomposition is used
first and then the one-dimensional wavelet decomposition in X-direction. The
threshold value is calculated by two methods: Donoho and Birgé-Massart strategies
[2, 3].

87
3 Algorithm of automatic corrosion recognition

An algorithm of automatic defect recognition includes two main steps:
1. Construction of binary maps according to DRIFT panels;
2. Making a decision on significant deviation on binary maps;
1. Construction of binary maps using data from DRIFT panels. Statistical DRIFT data
𝜉𝑡,𝑑 , are converted into binary maps by some THR threshold. It is necessary to take
into account that increase of the signal corresponds to the presence of collar:
1 if 𝜉𝑡,𝑑 > 𝑇𝐻𝑅
𝜂𝑡,𝑑 = { , (2)
0 if 𝜉𝑡,𝑑 ≤ 𝑇𝐻𝑅

and decrease of the signal, on the contrary, corresponds to the presence of corrosion:
1 if 𝜉𝑡,𝑑 < −𝑇𝐻𝑅
𝜂𝑡,𝑑 = { , (3)
0 if 𝜉𝑡,𝑑 ≥ −𝑇𝐻𝑅

2. The automatic defect recognition process is based on the decision theory. There are
two hypotheses: Н0 – the defect is absent and Н1 – the defect is present.
𝑃(𝑥 = 1 / 𝐻0 ) = 𝛼, 𝑃(𝑥 = 0 / 𝐻0 ) = 1 − 𝛼, (4)

𝑃(𝑥 = 1 / 𝐻1 ) = 1 − 𝛽, 𝑃(𝑥 = 0 / 𝐻1 ) = 𝛽, (5)

where α - error of first kind, β - error of second kind.
Each hypothesis has its likelihood function. A value of the likelihood logarithm is
calculated for each hypothesis at each depth point.
The defect is absent:

𝑙(0) = ∑𝑡∈𝐼 𝜂𝑡,𝑑 ln α + ∑𝑡∈𝐼(1 − 𝜂𝑡,𝑑 ) ln(1 − α) (6)

The defect is present:

𝑙(1) = {∑𝑡∈𝐼 𝜂𝑡,𝑑 } ln(1 − β) + {∑𝑡∈𝐼(1 − 𝜂𝑡,𝑑 )} ln β. (7)

Then the two-decision statistical hypothesis is verified by the maximum likelihood
method:
𝑙(1) − 𝑙(0) ≷ 𝐻1
𝐻0С, (8)

where 𝐻1
𝐻0С = 0 - the decision threshold by the maximum likelihood criterion [4].
Figure 2 illustrates an example of corrosion in the casing, as the defect appears on the
LONG DRIFT panel. The algorithm correctly identified the presence of corrosion and
referred it to the corrosion of the first barrier, which is the tubing.

88
Fig. 2. Corrosion in the casing. Left to right: The DEPTH panel, TUBING CORROSION
shows corrosion in the tubing, SHORT DRIFT is the drift panel for the short sensor, WELL
SKETCH depicts well completion, LONG DRIFT is the drift panel for the long sensor, and
CASING CORROSION shows corrosion in the casing.

In order to verify the algorithm, data from 11 wells were processed. Corrosions
found during the well log analysis were compared with those processed by the algo-
rithm. The following results were obtained: the automatic defect recognition algo-
rithm accurately separates the defects of the 1st and 2nd barriers. When configuring
the algorithm to search for small intervals of corrosion (metal loss less than 10%), lots
of false defects are indicated, which complicates data processing. With such configu-
ration, the algorithm detects 89% of the first barrier corrosions and 93% of the second
barrier corrosions. Defects with metal loss less than 10% are not dangerous, unlike
major defects with metal loss greater than 10%. The algorithm is designed to find
major defects. When setting the appropriate algorithm parameters, all defects, includ-
ing a small number of false defects, are detected. Thus, the automatic defect recogni-
tion allows quick identification of probable corrosion zones, on which the well log
analyst should focus. This, in its turn, increases the speed and quality of data interpre-
tation.

References
1. Arbuzov, A.A., Bochkarev, V.V., Bragin, A.M., Maslennikova, Yu.S., Zagidullin,
B.A., Achkeev, A.A., Kirillov, R.S.: SPE 162054 Memory Magnetic Imaging De-
fectoscopy (2012).
2. Mallat, S.: A wavelet tour of signal processing. PP. 479-481. Mir Press, Moscow
(2005).
3. Birgé, L., P. Massart. From model selection to adaptive estimation, in D. Pol-
lard(ed), Festschrift for L. Le Cam, pp. 55–88, Springer (1997).
4. Lehmann, E.L.: Testing Statistical Hypotheses. P. 24. Nauka Press, Moscow
(1979).

89
Автоматическое обнаружение дефектов и коррозии
нефтяных скважин по данным магнитно-
импульсного дефектоскопа

Рита Гайбадуллина, Булат Загидуллин, Владимир Бочкарев

Казанский (Приволжский) Федеральный Университет
{rita.gaibadullina,bulatza}@gmail.com, vbochkarev@mail.ru

Аннотация. Магнитно-импульсная дефектоскопия предназначена для
выявления различных дефектов и интервалов коррозии. Высокочувстви-
тельные датчики, представляющие собой приёмно-возбуждающие катуш-
ки, позволяют анализировать отклик от окружающей среды в широком
диапазоне времен. В настоящее время определение зон коррозии осу-
ществляется интерпретатором, т.е. носит субъективный характер. Более
того, анализ 2.5–3 км скважины (около 300 трубок НКТ и колонны) — это
трудоёмкий процесс, в ходе которого часть дефектов может быть пропу-
щена. Для исключения такого рода ошибок необходимо автоматизировать
процесс поиска интервалов коррозии. В работе предложен алгоритм авто-
матического распознавания дефектов, позволяющий разделить типичные
и нетипичные отклики. Так же рассмотрено применение пороговой
вейвлет-обработки для подавления шумов и предварительной подготовки
экспериментальных данных к дальнейшей обработке.

Ключевые слова: магнитно – импульсная дефектоскопия, вейвлет-
фильтрация, критерий максимального правдоподобия.