- Research Article
- Open Access
A Machine Learning Approach for Locating Acoustic Emission
© N. F. Ince et al. 2010
- Received: 18 January 2010
- Accepted: 20 October 2010
- Published: 31 October 2010
This paper reports on the feasibility of locating microcracks using multiple-sensor measurements of the acoustic emissions (AEs) generated by crack inception and propagation. Microcrack localization has obvious application in non-destructive structural health monitoring. Experimental data was obtained by inducing the cracks in rock specimens during a surface instability test, which simulates failure near a free surface such as a tunnel wall. Results are presented on the pair-wise event correlation of the AE waveforms, and these characteristics are used for hierarchical clustering of AEs. By averaging the AE events within each cluster, "super" AEs with higher signal to noise ratio (SNR) are obtained and used in the second step of the analysis for calculating the time of arrival information for localization. Several feature extraction methods, including wavelet packets, autoregressive (AR) parameters, and discrete Fourier transform coefficients, were employed and compared to identify crucial patterns related to P-waves in time and frequency domains. By using the extracted features, an SVM classifier fused with probabilistic output is used to recognize the P-wave arrivals in the presence of noise. Results show that the approach has the capability of identifying the location of AE in noisy environments.
- Acoustic Emission
- Support Vector Machine Classifier
- Acoustic Emission Signal
- Wavelet Packet
- Acoustic Emission Event
Rapidly changing environmental conditions and harsh mechanical loading are sources of damage to structures. Resulting damage can be examined based on local identification such as the presence of small cracks (microcracks) in a component or global identification such as changes in natural frequency of the structure. Continuous health monitoring process may involve both global and local identification. Generally, local damage, such as cracks in critical components, is inspected visually. This type of inspection is slow and prone to human error. Therefore, automated, fast, and accurate techniques are needed to detect the onset of local damage in critical components to prevent failure.
In this scheme, nondestructive testing and monitoring should be employed so that the damage can be inferred through analysis of the signals obtained from inspection. Acoustic emission (AE) events can serve as a source of information for locating the damage, particularly as caused by the initiation and propagation of microcracks [1–3]. The spatial distribution of AE locations can provide clues about the position and extent of the damage . In practice, the location of AE is estimated from the primary wave (P-wave), the first part of the signal to arrive at the sensor (see Figure 2(c)). However, the use of AE waveforms is often obscured by noise and spurious events, which may cause misinterpretation of the data. Even in controlled laboratory settings, it is difficult to account for all the sources of noise. Therefore, an AE system that automatically "learns" crucial patterns from the total AE data, as well as particular P-wave arrivals, may provide clues for distinguishing between real events and extraneous signals, thus improving the spatial accuracy of AE locations and reduce false alarms. Accurate detection of these events with appropriate signal processing and machine learning techniques may open new possibilities for monitoring the health of critical components; this offers the possibility for raising alarms in an automated manner if the degradation of structural integrity is severe.
The remainder of the paper is organized as follows. In the next section, the experiments and the AE data sets recorded from two specimens during controlled failure tests are described. Next, the signal preprocessing techniques used for enhancing the measured AE signals in the presence of noise and data acquisition imperfections are presented. This is followed by a description of a novel hierarchical clustering technique to group the AE events. The feature extraction and machine learning techniques for detecting P-waves are described in Section 4. Finally, the experimental results on the spatial distributions of AE events are provided and compared to the actual fracture locations.
AE events were recorded during a surface instability test that is used to examine failure near a free surface such as a tunnel wall. A photo representing the experimental setup is given in Figure 2. A prismatic rock specimen, wedged between two rigid vertical side walls and a rigid vertical rear wall, is subjected to axial load applied in the -direction through displacing rigid platens. The specimen is supported in the -direction such that compressive stress is generated passively. The rear wall in -direction ensures that the lateral deformation and failure (cracks) were promoted to take place on the front, exposed face of the specimen.
Four acoustic emission (AE) sensors were attached to the exposed face using cyanoacrylate glue, and their positions were measured. Four other AE sensors were fastened to the side walls of the apparatus. The AE data were collected with high-speed, CAMAC-based data acquisition equipment, consisting of four two-channel modular transient recorders (LeCroy model 6840) with 8-bit analog to digital converter (ADC) resolution and a sampling rate of 20 MHz. The data acquisition system was interfaced with eight piezoelectric transducers (Physical Acoustics model S9225), and eight preamplifiers with bandpass filters from 0.1 to 1.2 MHz and 40 dB gain were used for conditioning the raw AE signals. The frequency response of these transducers ranged from 0.1 to 1 MHz, with a diameter of approximately 3 mm. All channels were triggered when the signal amplitude exceeded a certain threshold on the first sensor. This sensor is referred to as the "anchor" sensor. AE data were acquired in a more or less continuous fashion until 128 Kbytes of a digitizer memory were filled; then the AE data were transferred to the host computer, with approximately four seconds of downtime. The entire waveforms were stored automatically and sequentially with a time stamp. This experiment was repeated twice using two very similar rock specimens with dimensions of 62 mm mm mm labeled as SR1 and SR2. A sample AE signal recorded with the system is presented in Figure 2(c). In total, 2176 and 1536 AE events were recorded in the experiments SR1 and SR2, respectively. This number includes both real AE and spurious (noise) events.
AE events related to a particular cluster with four members are shown in Figure 5. This step was followed by computing the averages of each cluster to obtain "super" AE signals. In this scheme, averaging is expected to reduce the uncorrelated noise in comparison with the repetitive AE signal component across the records of a given cluster, resulting in an amplitude SNR increase of at best , where is the number of events in a cluster. A similar approach has been utilized for processing gene expression profiles in ; it has been shown that averaged gene expression data within clusters have more predictive power than those from individual gene expressions. Thus, by increasing the SNR of the waveforms, AE locations will be more accurate.
In order to improve the amplitude SNR by a factor of two or more, clusters with at least four members were used in estimating the location of AE. Those clusters with large numbers of members increase the reliability of the location estimation step. We emphasize that the key assumption here, and one that has been observed experimentally, is the very low likelihood that, in practice, noise will also be highly correlated across multiple measurement records. Hence, it is expected that highly correlated signals (events) can only originate from a source such as microcracks.
4.1. Discrete Fourier Transform-Based Features
Based on the above observations on the frequency characteristics of P-waves and noise and within the spirit of , so-called Mel Scale, subband energy features were extracted from the spectrum of each time window using a fast Fourier transform. A Blackman-Tukey window was used during the estimation of spectra of segments. In total, five subbands were extracted. The widths of the subbands were not uniform and had a dyadic structure. The lowest two bands had the same bandwidth, and the following subbands were twice as wide as the preceding subbands. This setup focused more on the lower frequency bands since the energy of the signal was concentrated in this range. By concatenating the Mel Scale subband features from all three windows, a 15-dimensional feature vector was constructed. Generally, the noise (pseudo-P-waves) had jagged spectra. In contrast, the spectra of the P-waves were smooth. The variance of the derivative of the spectrum of each time window was also computed as another feature to capture this difference.
4.2. Discriminatory Wavelet Packet Analysis-Based Features
4.3. AR Model-Based Features
For each time point, computing the features described could be a demanding process. To reduce the number of candidate time points that need to be tested for P-wave arrival, first the signal was normalized, and then the envelope of the signal was computed with the Hilbert transform. When the envelope of the signal exceeded a predefined threshold, and then that time point was tested for P-wave arrival, it was found that a threshold value of 0.5 was good enough to determine most of the P-waves. The feature vectors for each method presented above were individually fed into a linear support vector machine classifier for the final decision . The main motivation for using an SVM classifier is based on its robustness against outliers and its generalization capacity in higher dimensions, which is the result of its large margin. Furthermore, the output of the SVM classifier was postprocessed by a sigmoid function to map the SVM output into probabilities. This was accomplished by minimizing the cross-entropy error function as suggested in . By using this procedure, we were able to assign posterior probabilities to SVM output which is later used as a confidence level to detect P-wave arrival. The SVM classifier was trained by selecting around 20 multichannel "super" AE events from each data set. Since each event includes AE data from 8 channels, this resulted in 160 P-waves to be tested in each dataset. This number included those clusters with low number of members. However, due to poor SNR, we were unable to visually identify the location of all P-waves in these data sets. Consequently, we selected those events which have a visible P-wave. The training feature vectors for P-waves and noise sets were constructed from this subset by manually marking the P-wave arrivals and noise events that exceeded the predefined threshold in each channel. The numbers of visually identified P-waves were 100 and 78 in datasets SR1 and SR2, respectively. The numbers of noise events were 155 and 162 for SR1 and SR2, respectively. The SVM classifier was trained on the features using the data set of one of the experiments and applied it on the other dataset. In this way, it was guaranteed that no test samples were used in training the classifier. In addition, using such a training strategy, it was investigated whether both data sets share similar patterns. The success of such a strategy can also validate the generalization capability of the classification system constructed.
We also examined the performance of a combination of feature sets. Interestingly, the features computed with WP method did not provide any better discrimination performance when they are combined with other features. For dataset SR1, the best performance was obtained with those features computed with WP method only. We note that the best separation performance was obtained with the combination of Mel Scale, AR model error, and spectrum variance features on the dataset SR2 (AUC = 0.483). Based on these observations, we trained the SVM classifiers either with only WP features or with the combination of Mel Scale, AR model error, and spectrum variance features. These classifiers were applied on the test samples we describe below.
In this study, it is desirable to have a system with low false positive rates since there exist several peaks in the baseline preceding the P-waves that can be potentially recognized as a P-wave. For this particular purpose, we used the probability output of the SVM classifier. We only accepted those points as P-Wave arrivals when the posterior probability exceeds a threshold of 0.9. The threshold can also be moved to more stringent levels. However, this may result in the classifier missing the P-waves which will yield low TP rates. One can also select that time as P-wave arrival point, where the posterior probability of the SVM classifier is maximum on the whole AE signal. However, this caused the system to miss the P-waves and identify those regions in the post-P-wave as they share similar characteristics. Therefore, we selected the first point as P-wave when the posterior probability exceeded the 0.9 threshold.
The locations of all detected clusters in SR1 spread over the specimen with a tendency towards the free surface (Figure 12(a)). This is an expected factor since those clusters with low number of members have lower SNR. It is also possible to capture noise by chance with a low number of members. In order to get around this problem, one can construct another decision system in order to discriminate between AE and noise. Observations indicate that keeping those clusters with large number of members automatically eliminates those recordings with noise or random nature. One can also increase the correlation threshold for identifying the clusters. However, there is a chance that a high correlation threshold may erase all possible clusters in the data, where the SNR is low. On the other hand, keeping it very low relaxes the constraints, where the chance of obtaining clusters with noise members is increased. The threshold can be adjusted depending on the quality of the available data.
Once the signal exceeds the threshold, it has to surpass the threshold at least 3 times in the subsequent 40-sample (40 × 50 ns = 2 μ s) window.
After 120 samples (i.e., 6 μ s) from the picked time mark, the signal has to exceed threshold at least once.
Novel approaches based on hierarchical clustering and support vector machines (SVM) are introduced for clustering AE signals and detecting P-waves for microcrack location in the presence of noise. Prior to feature extraction and classification process, spikes from the AE data are removed by employing a median filter. Clusters of AE events are identified by inspecting their pairwise correlation. After identifying clusters, an averaging step was implemented to obtain "super" AE with improved SNR. Characteristic features were extracted from the data in time and frequency domains to identify P-waves for time of arrival (TOA). SVM classifiers with probabilistic outputs were trained with these features to recognize P-waves for TOA determination. The location of each AE cluster was estimated accordingly.
The proposed machine learning technique with clustering analysis and SVM showed that the estimated clusters can successfully indicate the location of failure observed in surface instability tests, in which the cracks were promoted to occur close to the front free surface of the specimen. This approach, compared to the classic AE algorithm that gave a very disperse pattern and was not indicative of the region of failure, also presents the capability of filtering noisy signals and enhance the SNR to obtain more reliable AE cluster locations. The preliminary results show that the method has the potential to be a component of a structural health monitoring system.
Partial support was provided by the National Science Foundation, Grant no.CMMI-0825454. The authors express their appreciation for the constructive comments provided by the referees, which served to considerably improve the paper.
- Grosse C, Glaser SD, Krüger M: Wireless acoustic emission sensor networks for structural health monitoring in civil engineering. Proceedings of the European Conference on Non-Destructive Testing (ECNDT '06), 2006, Berlin, Germany 1-8.Google Scholar
- Golaski L, Gebski P, Ono K: Diagnostics of reinforced concrete bridges by acoustic emission. Journal of Acoustic Emission 2002, 20: 83-98.Google Scholar
- Emamian V, Kaveh M, Tewfik AH, Shi Z, Jacobs LJ, Jarzynski J: Robust clustering of acoustic emission signals using neural networks and signal subspace projections. EURASIPJournal on Applied Signal Processing 2003, 2003(3):276-286. 10.1155/S1110865703210027View ArticleGoogle Scholar
- Gong Z, Nyborg EO, Oommen G: Acoustic emission monitoring of steel railroad bridges. Materials Evaluation 1992, 50(7):883-887.Google Scholar
- Grosse CU, Finck F, Kurz JH, Reinhardt HW: Improvements of AE technique using wavelet algorithms, coherence functions and automatic data analysis. Journal of Construction and Building Materials 2004, 18(3):203-213. 10.1016/j.conbuildmat.2003.10.010View ArticleGoogle Scholar
- Hastie T, Tibshirani R, Friedman J: The Elements of Statistical Learning. Springer, New York, NY, USA; 2001.View ArticleMATHGoogle Scholar
- Iverson N, Kao C-S, Labuz JF: Clustering analysis of AE in rock. Journal of Acoustic Emission 2007, 25: 364-372.Google Scholar
- Theodoridis S, Koutroumbas K: Pattern Recognition Second Edition. Academic Press, New York, NY, USA; 2003.MATHGoogle Scholar
- Park MY, Hastie T, Tibshirani R: Averaged gene expressions for regression. Biostatistics 2007, 8(2):212-227.View ArticleMATHGoogle Scholar
- Cetin AE, Pearson TC, Tewfik AH: Classification of closed and open shell pistachio nuts using principal component analysis of impact acoustics. Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP '04), May 2004 677-680.Google Scholar
- Saito N: Local feature extraction and its applications using a library of bases, Ph.D. thesis. Department of Mathematics, Yale University, New Haven, Connm USA; December 1994.Google Scholar
- Ince NF, Goksu F, Tewfik AH, Onaran I, Cetin AE, Pearson T: Discrimination between closed and open-shell (Turkish) pistachio nuts using undecimated wavelet packet transform. Biological Engineering Journal, American Society of Agricultural and Biological Engineers (ASABE) 2008, 1(2):159-172.Google Scholar
- Burnham KP, Anderson DR: Model Selection and Multimodel Inference: A Practical Information-Theoretic Approach. 2nd edition. Springer, New York, NY, USA; 2002.MATHGoogle Scholar
- Platt JC: Probabilistic outputs for support vector machines and comparisons to regularized likelihood. In Advances in Large Margin Classifiers. Edited by: Smola A, Bartlett P, Schölkopf B, Schuurmans D. MIT Press, Cambridge, Mass, USA; 1999.Google Scholar
- Kruz JH, Koppel S, Linzer L, Schechinger B, Grosse CU: Source localization. In Acoustic Emission Testing: Basics for Research-Applications in Civil Engineering. Edited by: Grosse CU, Ohtsu M. Springer, Berlin, Germany; 2008.Google Scholar
- Shah KR, Labuz JF: Damage mechanisms in stressed rock from acoustic emission. Journal of Geophysical Research 1995, 100(8):15527-15539. 10.1029/95JB01236View ArticleGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.