Experimental Study on Feature Selection Using Artificial AE Sources

Transcription

1 3th European Conference on Acoustic Emission Testing & 7th International Conference on Acoustic Emission University of Granada, September Experimental Study on Feature Selection Using Artificial AE Sources Gabriel Cséfalvay, Petr Sedlák Department of Physics, Faculty of Electrical Engineering and Communication, Brno University of Technology, Technická 1, 616 Brno, Czech Republic, Abstract Acoustic emissions in materials might be generated by several different physical processes. Distinction of these processes can be supported by analysis of the recorded signals. An AE signal itself is not directly related to its origin. Signals are affected by many factors such as the tested specimen, waveguide transfer function given by structure and homogeneity of material, transfer function of used sensors, etc. This work aims to identify signal features that are correlated with acoustic emission source type by analyzing artificial acoustic emission events. Signals are generated using a piezoelectric transducer and a signal generator that emits signals with various parameters. Keywords: artificial sources, source emulation, AE signal features, correlation 1 Introduction Acoustic Emission (AE) is a natural effect whereby elastic waves suddenly appears in a material under mechanical stress. The elastic waves arise from the energy released during displacement in the structure of the material. Acoustic emissions in materials might be generated by several different physical processes. Distinction of these processes can be supported by analysis of the recorded signals. An AE signal itself is not directly related to its origin. Signals are affected by many factors such as the tested specimen, waveguide transfer function given by structure and homogeneity of material, transfer function of used sensors, etc. It contains a large amount of information, thus, it is more difficult to extract the required information. Characteristic information, namely features, can be extracted from acquired signals and used to analyze the AE source [1, 2]. The features are related indirectly to physical parameters of the generating mechanical process, such as crack size or dislocation type. The objective of this paper is motivated by the fact that one of the general aims in acoustic emission is to find suitable features that would distinguish sources in the examined material or construction. The paper tries to discover the influence of signal source parameters on individual signal features. 2 Experiment setup In this experiment, the acoustic emission source was emulated by a piezoelectric transducer made of PZT-27 piezoceramics cast-in in a mortar block (Fig. 1). This transducer was excited by signals of several different shapes (Fig. 2) with variable amplitude and width. The excitation signals were generated by signal generator Agilent 3322A. Three wideband piezoelectric

2 Excitation parameters Signal generator Piezoelectric sensor PC LNA Mortar block Piezoelectric transducer A/D Signal acquisition Figure 1: Experiment setup sensors of type 3S Sedlak SHS-WB were used for sensing elastic waves. Their electric output signals were amplified by low noise amplifiers of type 3S Sedlak PA-31 and recorded by a digital acquisition card TiePie HandyScope HS4. The amplitude of the excitation pulse is intended to reflect the attenuation in various situations depending on the distance between the transducer and sensor, their mutual orientation, the amount of reflected waves from the specimen boundaries and the sensor position itself. The sensor position highly affects the transfer function from the source to the sensor and practically it s not predictable. Because of that mainly in small specimens, the amplitude is affected by many factors not related to the AE source, we consider it in this experiment irrelevant regarding the source type. The pulse width variation and its shape may represent different source types. The widths of the excitation pulse were 5, 1, 2 and 5µs and the amplitudes were 1, 2, 4 and 8V. The shortest rectangular pulse of 5µs has base harmonic of 2kHz. Its fifth harmonic is 1MHz, it is expected to generate signals up to this frequency. The ramp pulse will have slightly less harmonics and the sine pulse will have very low harmonics. The choice of excitation pulse shape is based on our observations of electromagnetic emission signals [3, 4, 5, 6] except of the first type (rectangular). It is intended to excite the system by a Dirac impulse in a real-world situation. The second type represents gradual crack propagation [5]. The third type substitutes sudden structure change followed by gradually slowing crack propagation. The last one is characterized as crack bouncing [5, 7]. Electromagnetic emission has a wide variety of shapes [5, 6, 7] but only these four basic excitation pulse shapes were selected for this experiment. Figure 2: Excitation pulse shapes: rectangular pulse, ramp up, ramp down, sinus The acquired events were analyzed and signal features were calculated. The following signal features were used for event analysis. From the continuous wavelet transformation (CWT) of the event, two features are calculated: the frequency of the highest peak and the frequency of

3 the first peak. The signal is divided into 6 frequency bands and, subsequently, the energies of the signal in these bands are calculated (P1 P6). The centrum of gravity of the Fouriertransformation is also a frequency-domain feature. The amplitude of the first peak and the third and fourth statistical moments are time-domain features that were also used. The remaining features are defined in EN Channel 1 Channel 1 Amplitude [V].1.1 Frequency [khz] Channel Channel 2 Amplitude [V].1.1 Frequency [khz] Channel Channel 3 Amplitude [V].1.1 Frequency [khz] Figure 3: Sample of acquired signals for each channel and their wavelet transformation The AE features shows interesting correlations to excitation signal parameters (amplitude, width and shape). AE features that have low correlation with amplitude and high correlation with width and shape might be denoted as the most suitable ones for acoustic emission source type identification. In addition, the relationship with the excitation pulse type has to be considered. Both pulse type and its width essentially affects the spectral composition of the AE signal. This effect can be used to emulate the differences between real AE source types, because different sources have different spectral composition of their emitted signals [8, 9, 1].

4 3 Evaluation results Events were generated for all of the combinations of the following parameters: 4 excitation pulse shapes, 4 pulse widths, 4 amplitudes. All of these combinations were used 4 times, resulting in 256 events. The events were sensed by 3 sensors, resulting in the total number of recorded signals of 768. Correlation with pulse width Correlation with amplitude ρ [ ] CWT max. peak f CWT 1st peak f P kHz P kHz P kHz P kHz P kHz P1 1 11kHz Centrum of gravity FFT max. f FFT max. A 4th moment 3rd moment Energy Avg. freq. Zero crossings Count to max. Count 1st amplitude Rise angle Rise time Duration Max. ampl ρ [ ] Figure 4: Correlation coefficients of each signal feature with the excitation pulse width (left) and amplitude (right) For all recorded signal, signal features were calculated. Each of these features were investigated for the relationship with the amplitude and width of the excitation pulse by calculating the correlation coefficient for each feature with amplitude and width [11, 12]. The result of this analysis is shown in fig. 4. Also, the coefficients were calculated for signals from each channel individually and each excitation type. The results for these selected signal sets were similar to each other. On the base of this result and the above considerations, it can be concluded that the most information of the source type is carried by the signal features that have correlation with the pulse width and have low correlation with its amplitude. These features are the rise time, 3rd statistical moment, frequency of the maximal peak in FFT (fast Fourier transformation), center of gravity of FFT, energies in the first 3 frequency bands and the two wavelet transformation based features: the frequency of the first and of the highest peak in CWT (continuous wavelet transformation). These features can be assumed to carry the most information on the source type. Acquired signals Parameter calculation Normalization Dimension reduction Analysis of results Figure 5: Analysis workflow

5 In order to verify this conclusion, dimensionality reduction was applied for the features of the recorded events [13]. The calculated features were normalized, and after that 2 dimensionality methods were applied: Principial Component Analysis (PCA) and Stochastic Neighbor Embedding with t-distribution (tsne) (fig. 5). These methods in an ideal case would group the data points based on their mutual similarity. In our case this similarity is the source type. This method was applied for the full feature set as well as only for the above selected features. The results are represented in figure PCA tsne Rect 5us Rect 1us Rect 2us Rect 5us Ramp up 5us Ramp up 1us Ramp up 2us Ramp up 5us Ramp down 5us Ramp down 1us Ramp down 2us Ramp down 5us Sin 5us Sin 1us Sin 2us Sin 5us (a) PCA tsne (b) Rect 5us Rect 1us Rect 2us Rect 5us Ramp up 5us Ramp up 1us Ramp up 2us Ramp up 5us Ramp down 5us Ramp down 1us Ramp down 2us Ramp down 5us Sin 5us Sin 1us Sin 2us Sin 5us Figure 6: The results of dimension reduction methods for (a) the full feature set and (b) selected features As we can see, in the first case (all features used) this separation did not happen very well. The result is probably distorted by the amplitude information, while in the second case, using only the selected features, the source types are separated well from each other. Since the spectral composition is affected by the excitation pulse type as well as by the pulse width, almost all of their combinations are recognized as different source types, as we can see in fig. 6(b). This second case, compared to the first, shows significantly better results that are yet suitable for further analysis (e.g. clustering). 4 Conclusions This work analyzed the suitability of common and some uncommon signal features concerning acoustic emission source type identification. Some signal features have high correlation with

6 the excitation signal shape but lower correlation with its amplitude. These features carry the most information regarding the acoustic emission source type. On the basis of this investigation we can state that rise time, 3rd statistical moment, frequency of the maximal peak in FFT, center of gravity of FFT, energies in the first 3 frequency bands and the two CWT based features are the most suitable features for AE source analysis, since those carry the most information about AE source type. This conclusion was verified by applying dimensionality reduction methods on the calculated features. It was shown that the named features alone provide better results than the full feature set. Acknowledgement This research has been supported by the Grant Agency of the Czech Republic within the framework of the project GAČR 12/9/H74 Diagnostics of material defects using the latest defectoscopic methods. References [1] AE Signal Features, [online] EducationResources/CommunityCollege/Other%2Methods/AE/ AE_Signal%2Features.htm. cit [2] B. Muravin, Acoustic Emission Method, [online] ae/muravin%2-%2acoustic%2emission%2method%2-%2short% 2presentation%2for%2students.ppt. cit [3] P. Koktavy, J. Pavelka and J. Sikula, Characterization of acoustic and electromagnetic emission sources, Measurement Science and Technology 15 (5), pp , 24. [4] P. Sedlak, J. Sikula, T. Lokajicek, Y. Mori, Acoustic and electromagnetic emission as a tool for crack localization, Measurement Science and Technology 19 (4), 4571, 28. [5] P. Koktavy, Experimental study of electromagnetic emission signals generated by crack generation in composite materials, Measurement Science and Technology 2 (1), 1574, 29. [6] P. Sedlak, M. Enoki, T. Ogasawara, J. Sikula, Electromagnetic and Acoustic Emission in PEEK/Carbon Nanotube Composites. In Proceedings of 29th European Conference on Acoustic Emission Testing 21. Vienna: NDT.net, 21 [7] V. Frid, A. Rabinovitch, D. Bahat, Fracture induced electromagnetic radiation, Journal of Physics D: Applied Physics 36, 162, 23. [8] J. Bohse, Acoustic emission characteristics of micro-failure processes in polymer blends and composites, Compos. Sci. Technol. 6, pp , 2 [9] S. Huguet, N. Godin, R. Gaertner, L. Salmon, D. Villard, Use of acoustic emission to identify damage modes in glass fibre reinforced polyester, Compos. Sci. Technol. 62, pp , 22

7 [1] X. Li, C. Ramirez, E. L. Hines, M. S. Leeson, P. Purnell, M. Pharaoh, Pattern recognition of fiber-reinforced plastic failure mechanism using computational intelligence techniques, IEEE World Congress on Computational Intelligence, Neural Networks, pp , 28 [11] G. Cséfalvay, Correlation Analysis of Artificial Acoustic Emission Signal Parameters and Features, Proceedings of the 18th Conference STUDENT EEICT 212 Volume 3. Brno: LITERA Brno, 212. s ISBN: [12] T. Trčka, P. Koktavý, Electromagnetic and acoustic emission signals continual measurement and real time processing, ElectroScope - cz, 21, no. 3, ISSN: [13] G. Cséfalvay, M. Klampár, Dimensionality Reduction Methods for Classification of Acoustic Emission Sources, Proceedings of the 17th conference STUDENT EEICT 211 Vol. 3, No. 1, Brno: NOVPRESS s.r.o., pp , 211, ISBN: