SUMMARY

Acoustic Emission (AE) has been extensively applied to vessel testing during pressurization, and has proven to be a fast and economic method for structural integrity assessment. High-pressure hydrogen cylinders are commonly tested with AE in-service, pressurizing with product, to verify fitness for service and to detect developing flaws at early stages. Testing large numbers of such cylinders and the presence of extraneous noise due to filling/pressurization processes and other sources can make AE data analysis difficult and time consuming. This paper presents work carried out for the analysis of hydrogen cylinder AE data, utilizing Supervised Pattern Recognition (SPR) techniques from modern AE software, in order to discriminate developing crack data from other data (e.g. noise) in an effort to automate the analysis process for large scale testing and increase confidence of the results. This work has led to the successful creation of a classifier to automate the discrimination between AE data from developing cracks and other indications, as diagnosed by follow-up NDT for verification, leading to minimum data analysis time, increased confidence on final result through the minimization of operator dependency.

INTRODUCTION

Pressure vessels are installed in most process and storage sites and are used for a large variety of products. In most cases the nature of the contents render these vessels critical to operations and safety. Vessels of this type have been tested with Acoustic Emission for more than two decades [1] using industry developed procedures [2] and standard codes [3] with good results. The main advantage of AE tests is the speed of the test and the ability to test vessels in-service with minor disruption to operations (no need to remove the vessel from service, empty and clean for inspection). AE provides complete coverage of the vessel with one load test and is sensitive to active defects for the loads achieved.

Liquid hydrogen storage cylinders are common pressure vessels and are usually installed in clusters which may contain a large number of units. Due to their product and operating conditions these vessels need inspections for their structural integrity. Hydrostatic pressure tests and other various NDT methods are used but AE offers fast in-service tests that save time and money for the vessel operator/owner. The large number of such vessels has led to attempts to automate the analysis and recommendations of the AE test. This need has led to the work presented here which takes advantage of latest tools in AE data analysis to achieve an automatic sorting of vessels with possible active defects and vessels with no active defects. The present work outlines the development of such a methodology based on advanced AE data processing and Pattern Recognition analysis [4]-[7].

PRESSURE TESTS

The tests were performed on high-pressure liquid hydrogen cylinders with sizes varying from 2.9 to 3.6m length and 0.4 to 0.6m diameter. The cylinders were pressurized by the addition of liquid hydrogen at pressures slightly above their normal operating conditions (usually about 150-200 kg/cm²). During pressurization and load hold periods each cylinder was monitored by two acoustic emission sensors placed near the ends.

Liquid hydrogen pressurization produced some noise at the inlet end of the cylinders. Acoustic emission was recorded with real-time linear location and correlation graphs.

Fig 1: Cylinder schematic with typical sensor positions.

DATA

The data used for this work is from pressure tests on 4 different cylinders. Two of the cylinders gave AE, which was confirmed by follow-up NDT to be due to cracks in the cylinder. Another two cylinder data were used as typical examples of non-relevant emission (e.g. noise due to filling and other sources) and follow-up NDT did not indicate any defect in these cylinders.

During pressurization some noise existed near the pressurization end for all cylinders. Traditional analysis was used to process the data and decide on the existence of relevant indications. This analysis is usually based on signal feature correlation plots to locate and remove noise with a number of other considerations taken into account, such as sensor position, pressurization means statistical feature distribution etc. Linear location is used on the processed data to investigate concentrate sources and improve the confidence for the existence or not of defects. Table 1 gives a brief history of each cylinder test with AE indications and follow-up results.

Cylinder	Test Type	Indications	Follow-up Results
A	Pressurization by liquid hydrogen	Linear location indications at one end.	Confirmed existence of cracks in cylinder.
B	Pressurization by liquid hydrogen	Linear location indications at both ends.	Confirmed existence of cracks in cylinder.
C	Pressurization by liquid hydrogen	Minor linear indications at pressurization end. High data rate during pressurization. Inlet noise suspected.	No findings
D	Pressurization by liquid hydrogen	Insignificant linear indications. Low activity.	[No findings]
Table 1: Cylinder History and Test Results

Fig 2a: Linear location indications from cylinder A.

Fig 2b: Linear location indications from cylinder B.

CLASSIFIER TRAINING

Fig 3: Typical correlation and distribution plots of normalized counts and amplitude.Data are highly mixed (orange are crack signals, green are non crack data).

The traditional data analysis described briefly in the previous paragraph provided good results but the time needed becomes a problem when large volumes of such tests are performed, which is usually the case. For example, in a single plant installation more than one hundred of such cylinder may be tested.

Using the results and knowledge from follow-up actions an automated data discrimination technique was developed based on Supervised Pattern Recognition from commercially available software. Noesis v3.3, Pattern Recognition and Neural Networks [6] software was used for the analysis.

Data from cylinders B and D only, were used as the known data. Since, at least some, noise was present in each test, data from active defects could not be clearly distinguished from other, non-relevant, indications (Figure 3). More specifically, hit signature overlapping between crack and non-crack data is observed in typical counts-amplitude and duration-energy scatter plots (see Figure 3 left) and on respective amplitude and counts max-min distribution plots (see Figure 3 right).

To investigate how well separated data from active defects are from noise or non-relevant data, Principal Component Axes (PCA) were created based on Covariance analysis [5],[6]. The method creates a projection to a set of orthogonal axes using the eigenvectors of either the data’s Covariance (Variance-Covariance Analysis) or Correlation Matrix (Correlation Analysis). In the present case the Variance-Covariance analysis was used. The variance of the projected data on each new axis equals the corresponding eigenvalue. To determine how well the new axes will fit the data, a Degree-Of-Fit (DOF) can be used (see Figure 6).

Fig 6: Principal component analysis. Axis 4 was dropped as the Degree of Fit remained at 100%.

It was observed that data from cracks with data form noise or other non-relevant indication were now easier separable (Figure 4). A portion of the data remained mixed and this led to the creation of a new group of data (class) marked as "inconclusive" (Figure 5).

Fig 4: Principal Component Axes, Cavariance analysis. Data have better separation.

Fig 5: Overlapping data are grouped and mark as "inconclusive".

Fig 7: Neural Network topology. 4 input dimension, into 2x4D hidden layers and a 3 class output.

Having set-up the original three data classes (groups of similar signals) as "crack" (orange color in graphs), "non crack" (green color in graphs) and "inconclusive" (blue color in graphs) from known data, a Back Propagation Neural Net [6]-[7] was trained to classify data in these classes. Half of the original data (1014 hits) were used to train the classifier and half (1014 hits) were used to test the effectiveness of the training so as to get a, relatively, objective perspective on the accuracy and effectiveness of the training process. The training and testing set were composed with a random generator. The Back Propagation Net used 4 input dimensions, 2x4D hidden layers and a 3 dimensional output. Figure 7 displays the network architecture.

The error, when applied to the test data, was 1.18%. Closer inspection of the error revealed that there was no mixing of "crack" with "non-crack" data and minor mixing between "crack" and "inconclusive" and "non-crack" and "inconclusive" data. This increases the confidence of the result as the classes of interest are the "crack" and "non-crack" which have been clearly separated with no mixing.

This classification and data discrimination strategy resulted in the following criteria logic:

• Cylinders that produce data that contain even small percentages of "crack" data are suspect for active defects and will require closer investigation.
• Cylinders that contain no "crack" data are not suspect of active defects and, unless other special or unusual indications are present, no further action is required.

AUTOMATED DATA DISCRIMINATION

The trained classifier was applied to unknown data (cylinders A and C) and classified them in the above categories (classes). The criteria logic will be used for a final recommendation.

Application of the method to the data from cylinder A produced 2.0% "crack" and 17% "non-crack" data with the rest being "inconclusive". The result is consistent with the logic described above: A cylinder with confirmed cracks was automatically recognized by the classifier, despite the low crack data percentage.

Application of the method to the data from cylinder C produced 0% "crack" and 9.1% "non-crack" with the rest of the data being classified as "inconclusive". This cylinder can thus be said to contain no cracks as was confirmed by follow-up.

Table 2 summarizes results from all cylinder data files.

Class Name
Cylinder	Crack	Non-Crack	Inconclusive
A	2.0	17.0	81.0
B	9.6	0.1	90.4
C	0.0	9.1	90.9
D	0.0	16.9	83.1
Table 2: Percent (%) Data in each class for all cylinders.

Closer investigation of the results from the classifier showed that the data could be further analyzed with some interesting results. As can be seen in Figure 7 "crack" data appear on channel 8, which was the sensor at the far end of the pressurization inlet. Follow-up inspection showed that a crack was present in that area. In addition, all "non-crack" data, which are expected to contain noise, appeared on channel 7, which was at the pressurization end. This channel is susceptible to noise due to its proximity to the pressurization inlet that usually generates some noise. The "inconclusive" class of data may contain both "crack" and "non crack" data. Further investigation using linear location techniques showed that this was indeed the case for Cylinder A where some secondary hits from crack events were classified in the "inconclusive" class.

Fig 8: Automatic classification of cylinder A data. Duration-Energy.Automatic classification of cylinder A data. Duration-Energy.

Fig 9: Channel distribution of total AE energy(Cylinder A).

DISCUSSION

In this work the effectiveness of modern tools for AE data analysis has been demonstrated.

In many cases AE test data are difficult to analyze and interpret via traditional analysis techniques. The alternative analysis presented in this work can increase the confidence of the conventional analysis and leads to a criteria logic for an immediate test result.

Modern AE analysis software include improved features for traditional analysis (statistical, correlation etc) and provide new tools such as Pattern Recognition.

As demonstrated in this work, a small number of control tests with known results and follow-up verification, can provide the necessary information to develop a Supervised Pattern Recognition classifier that will automatically classify data from tests similar to the control ones. This process of automating the analysis and having direct results that may include significant information (e.g. determination of existence of cracks) is very effective as regards to time and confidence for large numbers of tests of similar equipment.

During the development of the methodology presented, other Pattern Recognition analysis schemes were tested with moderate results. These techniques were based on the original (recorded) AE signal features (amplitude, energy etc) and, as demonstrated earlier, could not provide data separation for the classifier to be effective. Thus, Principal Component Analysis was used. An investigation of the possibility to separate highly mixed data ("inconclusive" class) and classify them to the appropriate class (e.g. secondary hits from crack events) proved complicated and produced ambiguous results. There were some encouraging indications, though, that may lead to the detailed classification of all signals.

Careful control tests and proper classifier development can lead to a sound, case-specific AE analysis tool for hydrogen cylinders, provided that the database and training set will be enriched with additional cracked or otherwise damaged cylinder data, with follow-up NDT confirmation. Since the analysis of even the most well developed classifier will ultimately depend on the data acquired, a continuous scheme for corrective learning as a function of database size should be foreseen. At the present state of development the SPR analysis can help clarify results when data are contaminated by noise but it cannot give solutions where solutions do not exist.

REFERENCES

1. S. J. Vahaviolos, "Acoustic Emission: A new but sound NDE technique and not a panacea" in NDT, eds. D. Van Hemelrijck & A. Anastassopoulos, 1996, Balkema, pp. 101-115.
2. T. J. Fowler, J. A. Blessing, P. J. Conlisk and T. L. Swanson, "The MONPAC System", Journal of Acoustic Emission, Volume 8, Number 3, 1989.
3. ASME, Section V, Article 12, "Acoustic Emission Examination of Metallic Vessels During Pressure Testing", 1988.
4. A. A. Anastassopoulos, T. P. Philippidis, " Clustering Methodologies for the evaluation of AE from Composites ", J. of Acoustic Emission, Vol. 13, Noa 1/2, 1995, pp 11-21.
5. T. Philippidis, V. Nikolaidis, A. Anastassopoulos, "Damage Characterisation of C/C laminates using Neural Network Techniques on AE signals", NDT&E International, Vol. 31, No 5, Elsevier 1998, pp. 329-340.
6. Envirocoustics ABEE, "Noesis v3.3 Professional, Reference Manual", Rev.4, February 2001.
7. A. A. Anastassopoulos, V. N. Nikolaidis and T. P. Philippidis, "A Comparative study of Pattern Recognition Algorithms for Classification of Ultrasonic Signals", Neural Computing & Applications, 1999, Vol. 8, pp 53-61. (Initially Accepted April 1996 in J. of Systems Engineering).

© NDT.net - info@ndt.net

|Top|