See discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/328306729
High-speed data acquisition system for continuous acoustic emission
monitoring and real-time signal processing using FPGA-based platform within
a SHM framework
Conference Paper · October 2018
CITATIONS
READS
2
887
4 authors, including:
Sebastian Felix Wirtz
Adauto P. A. Cunha
University of Duisburg-Essen
Liebherr-Aerospace
16 PUBLICATIONS 84 CITATIONS
4 PUBLICATIONS 35 CITATIONS
SEE PROFILE
Dirk Söffker
University of Duisburg-Essen
536 PUBLICATIONS 3,335 CITATIONS
SEE PROFILE
All content following this page was uploaded by Dirk Söffker on 11 November 2018.
The user has requested enhancement of the downloaded file.
SEE PROFILE
More info about this article: http://www.ndt.net/?id=23250
9th European Workshop on Structural Health Monitoring
July 10-13, 2018, Manchester, United Kingdom
High-speed data acquisition system for continuous
acoustic emission monitoring and real-time signal processing
using FPGA-based platform within a SHM framework
Sebastian F Wirtz, Adauto PA Cunha, Nejra Beganovic and Dirk Söffker
Chair of Dynamics and Control, University of Duisburg Essen, Germany,
sebastian.wirtz@uni-due.de
Abstract
Acoustic Emission (AE) testing has emerged from the field of non-destructive testing as
a promising approach which is suitable for in-situ detection of damages in materials and
structures. Compared to competing approaches, advantages of the AE technique are the
suitability to detect incipient damages and that it can be applied passively. To record
AE, ultrasound stress waves resulting from the rapid release of elastic energy are detected by high fidelity transducers. The acquired waveforms, which can be related to distinct source mechanisms by suitable signal processing techniques, are characterized by
low amplitude, high frequency, and short duration. Due to the high frequency range
between 10 kHz and 1 MHz, fast data acquisition hardware is a key element to record
AE. Especially if detailed analysis of the AE is intended, sample rates of at least several
MHz are required. For the purpose of Structural Health Monitoring, continuous acquisition and processing of the full waveform data is desirable. This is still challenging due
to limitations related to disk I/O and processing speed. Therefore, most AE systems
provide only intermittent waveform acquisition over short periods of time i.e. using a
fixed threshold to trigger high-speed data acquisition and similar restrictions. In this
paper, the prototype of a new, integrated acquisition system is used. Instead of the typical setup based on a desktop computer and peripheral devices to enable I/O capabilities,
a system on chip is used as hardware platform. It comprises a dual-core ARM architecture running a Linux operating system and Field Programmable Gate Array (FPGA)
fabric, thus allowing software and hardware programmability. Using a suitable analogto-digital converter, data can be acquired continuously at a sample rate of 5 MHz. The
data are processed in real-time using FPGA-based implementation of Discrete Wavelet
Transform (DWT). Results of the DWT are stored immediately along with the original
measurement signal. The device can be accessed remotely via Ethernet to control the
acquisition process or to access the measurement data. Furthermore, the small form factor and low power consumption make this device ideally suited for field deployment.
The effectiveness of the measurement system is demonstrated at the example of detecting different damage mechanisms in composite material during indentation testing.
1. Introduction
The purpose of Structural Health Monitoring (SHM) is to provide statements regarding
the current state of components or systems (damages or undamaged) online or on demand to ensure safe and efficient operation of equipment. Due to increasing complexity
in many industrial applications, data-driven approaches are highly attractive. Here, sta-
Creative Commons CC-BY-NC licence https://creativecommons.org/licenses/by-nc/4.0/
tistical methods are used to conclude from sensor data. The process of implementing
SHM can be structured according to the Statistical Pattern Recognition (SPR) paradigm
(1). However, SPR approaches rely heavily on damage sensitive features, which have to
be extracted from the measurements using suitable signal processing techniques. However, online feature extraction becomes challenging if the data is generated at high
bandwidth. For instance, using Acoustic Emission (AE) for continuous monitoring,
large amounts of data are acquired in a short amount of time due to high sample rates.
Therefore, in this paper a novel measurement system is described, which is introduced
in (2). The design is inspired by the following principle ideas: i) leveraging data parallelism using FPGA-based hardware implementation to accelerate signal processing
tasks such as filtering and feature extraction and ii) placing computing power, storage,
and network capabilities close to data sources can lead to reduced latency for real-time
analytics (3). Key features are low cost, small form factor, and low power consumption
which makes this device ideally suited for field deployment and embedded implementation. Furthermore, several widely used I/O interfaces including USB and Ethernet are
available. An FPGA-based implementation of the Discrete Wavelet Transform (DWT)
which is suitable for real-time applications is used and the DWT coefficients are stored
immediately. The signal processing capabilities of the device are demonstrated at the
example of indentation testing of composite material.
The remainder of this paper is structured as follows. In section 2 an overview of the
related work is presented. In section 3, the developed measurement system is described
in detail. Experimental results at the example of indentation testing of composite material are presented in section 4. Finally, summary and conclusions are given in section 5.
2. Related work
In this section, a short overview of related work is given. This includes a summary of
previous works related to AE-based damage characterization in composite materials as
well as different hardware platforms which can be used for real-time signal processing.
Composite materials are composed of multiple constituents to achieve improved mechanical properties. Due to many advantages such as light weight, high strength, and
flexibility in design, composites are increasingly used for structural components e.g. in
aerospace and renewable energy applications. However, their use for critical loadbearing structures is still limited due to complex micro-mechanical damage mechanisms
which are currently not well understood. Recently, AE has been widely used to characterize and therefore to distinguish different damage mechanisms. The damage-related
ultrasound stress waves can be recorded passively by surface mounted transducers. Detailed investigation of the corresponding waveforms is conducted to draw conclusions
regarding the underlying source mechanisms. In an early study, Johnson and Gudmundsen (4) investigated AE signals in time domain to identify features that can be used to
correlate characteristic waveforms with different damage mechanisms including matrix
crack, delamination, and fiber breakage. However, the authors conclude that a parametric approach is not sufficient to classify different damage mechanisms reliably. Ni and
Iwamoto (5) used single fiber composite specimens in their study regarding the effect of
2
sensor distance on amplitude attenuation and frequency dependency of the AE transients. According to the experimental results, strong attenuation of the AE amplitude is
observed in time domain whereas peak frequencies remain unchanged at different sensor distances. Thus, it is pointed out that time-frequency domain analysis is suitable to
classify different damage mechanisms and to determine the time of occurrence (5).
Recently, time-frequency domain analysis has been extensively used to characterize
different damage mechanisms in composite materials based on AE measurements.
Gutkin et al. (6) used time-domain parameters to cluster AE waveforms. Subsequent
analysis of the peak frequencies indicates good agreement with the obtained clustering
result. Similarly, Kalogiannakis et al. (7) used DWT to establish a correlation between
AE waveforms and different damage mechanisms during tensile testing and wear testing
under frictional load. It is pointed out that similar peak frequencies are observed in both
cases. Furthermore, Marec et al. (8) used wavelet transform to cluster AE events observed during creep tensile tests. According to the results it is concluded that timefrequency domain descriptors provide improved separation of clusters. Moreover, Qi et.
al. (9) decomposed AE signals related to the material failure during static tensile loading
into multiple frequency bands using DWT. Here, analysis of different energy-based
descriptors indicates three potential damage modes. Similarly, Yousefi et al. (10) make
use of DWT to obtain energy of the AE signal in different frequency bands. Additionally, a cluster analysis is achieved. However, according to the results not all of the considered damage modes can be distinguished clearly using this approach (10). Recently,
Baccar and Söffker (11) identified different frequency ranges observed during static
indentation tests using continuous wavelet transform coefficients and established a correlation with distinct damage mechanisms.
For the purpose of SHM, continuous acquisition and processing of the full waveform
data is desirable. This is still a challenging task due to limitations related to disk I/O and
processing speed. Commercially available systems which allow continuous sampling
over long durations are usually based on PCI boards. However, besides high cost of
these systems, bulky equipment might be prohibitive for field deployment. Shateri et al.
(12) used a low-cost micro controller platform together with commercially available
sensors to record AE waveforms. Here, intermittent data acquisition is triggered by a
predefined threshold. Long term tests show that AE hits are detected reliably. However,
only a limited sampling rate of 667 kHz is achieved (13). For high-performance embedded implementations, hardware-software coprocessing architectures, i.e. an architecture
where the CPU is complemented with hardware accelerators for specific tasks, are frequently proposed. Compared to fully hardware-based implementations, these architectures have the advantage of faster development of initial prototypes, increased flexibility
for later changes, and ease of integration with peripheral components (13). For instance,
Cheng et al. (13) proposed a hardware-software coprocessing architecture for automatic
speech recognition. The system is implemented on FPGA fabric using soft processor
and hardware accelerators. Also, Virupakshappa and Oruklu (14) developed a hardware
architecture which is suitable for the embedded implementation of fault detection based
on ultrasonic A-scans. In contrast to Cheng et al. (13) a SOC-based platform is chosen
in (14), which comprises a CPU and FPGA fabric.
3
3. Materials and methods
In the sequel, the measurement system is described including hardware components and
the FPGA implementation of DWT for real-time use. The content presented here is
based on the work published in (2).
As a computational platform the ZedBoard (xc7z020clg484-1) used. This is an evaluation and development kit for Xilinx Zynq-7000 SOC which provides periphery for interfacing with additional hardware and storage including USB, Ethernet, and a SD card
slot. The dimensions of the board layout are 160 mm x 160 mm. The maximum power
consumption is 60 W. The SOC comprises two subsystems namely Processing System
(PS) with ARM Cortex-A9 dual-core processor and Programmable Logic (PL) fabric
running at clocks of 666 MHz and 100 MHz, respectively. Thus, this device allows efficient implementation of signal processing algorithms by leveraging both advantages of
FPGAs for fast signal processing and flexibility of software programmable devices to
implement higher level sequence control and communication interfaces. For data acquisition, Analog Devices AD7961 is used enabling AD conversion at a sampling rate of
5 MHz with a resolution of 16 bit. The chip is mounted on an evaluation daughterboard
and is connected to the device with the FPGA Mezzanine Card (FMC) connector.
The overall system is illustrated in Figure 1. The PL is used to implement the Discrete
Wavelet Transform (DWT) for real-time signal processing. The PS runs a Linux operating system, which is used to implement general functionality of the device. This includes loading drivers and enabling Ethernet at boot time, configuration of the Register
Bank, control of the data acquisition (start/stop), and storage. The measurement data are
stored either on the SD Card or in external memory, i.e. USB drive in binary format. For
communication between PS and PL, Advanced Extensible Interface (AXI) interconnect
is used. Furthermore, a Direct Memory Access (DMA) channel is established for data
transfer from PL to PS. High Performance Port (HPF) is used to achieve low latency
and an interrupt line is used initialized from the DMA to the PS. The data is stored in a
FIFO buffer. The implementation of the software and hardware logic is realized using
Xilinx Vivado and Software Development Kit (SDK) 2016.4. The Software is written in
C language. The configuration of the PL is defined using VHDL and Verilog.
The wavelet transform is a method for decomposition of non-stationary signals into
joint time-frequency domain using an orthogonal basis function referred to as wavelet.
In case of the DWT, decomposition of a signal can be achieved by using multirate filter
banks which are constructed from Finite Impulse Response (FIR) filters. Due to decimation of the input signal by passing through each filter bank, the DWT provides sparse
representation of the input signal. Nevertheless, DWT coefficients can yield a perfect
reconstruction of the original input signal (15). Furthermore, time complexity of the
algorithm is O(n). Therefore, DWT is well suited for denoising, data compression, and
feature extraction in real-time applications.
4
Storage
UART
Periphery
I2C
Interface
Ethernet
USB
DDR Controller
Linux OS
Processing system (PS)
AXI
DMA
Register bank
DWT
FIFO
ADC Driver
Programable Logic (PL)
FMC connector
Zynq-7000
AD7961FMCZ
AD7961
SMA connector
© SRS, 2018
Figure 1. Illustration of the measurement system architecture.
Different architectures for hardware implementation of the DWT algorithm are proposed, including pyramid and polyphase architectures (15). Regarding sample-wise calculation of DWT coefficients it has to be noted that to ensure suitable reconstruction of
the original signal equalization of delays along all filter paths is required (16). In the
sequel, polyphase realization using Quadrature Mirror Filter (QMF) pair as described in
detail by Cunha et al. (17) is adopted. The main advantage of this approach compared to
the classic implementation is the reduction of hardware resources required for synthesis
of the algorithm (i.e. adders, multipliers, and number of clock cycles) by a factor of two.
Multilevel DWT can be realized by cascading multiple QMF filter banks. In this case,
approximate coefficients are used as the input to the subsequent filter bank. In Figure 2
the implementation using cascading QMF filter banks is illustrated.
x[n]
d0
d1
QMF
d2
QMF
a2
QMF
a6
d7
QMF
a7
Figure 2. Schematic diagram of 8-level DWT implementation.
5
4. Experimental results
In this section, experimental results obtained from different experiments are presented.
Initial measurements are conducted during pencil lead break tests. Finally, results from
indentation testing of composites are considered.
The equipment used to perform AE measurements is illustrated in Figure 3. To record
AE waveforms a disc-shaped piezoelectric element (0.55 mm in thickness, Ø10 mm in
diameter, 3.6 MHz resonant frequency) is stiffly bonded to the surface of composite
material by means of cyanoacrylic glue. Furthermore, a preamplifier is used for signal
conditioning prior to digitalization. The raw waveforms are acquired by the ADC with a
sample rate of 5 MHz and 16 bit resolution. Additionally, DWT representation is calculated by the FPGA. Finally, both the raw waveform data and the corresponding DWT
representation are stored. For further analysis and visualization, the data can be accessed
via Ethernet. It is worth to mention that both the waveform data as well as the DWT
coefficients are acquired in real time.
Test specimen
and AE sensor
ZedBoard with
ADC daughterboard
Laptop
Figure 3. Illustration of AE measurement equipment.
4.1 Preliminary measurements
For initial tests of the measurement system, pencil lead break tests (PLB) were undertaken. Here, an AE source is simulated by breaking a pencil lead at the surface of a test
specimen (Hsu-Nielsen source). As a result, a strong broad band acoustic signal is generated. This test is routinely used to check functionality of AE equipment. The measurement results are presented in Figure 4 showing the raw measurement data and the
DWT coefficients. The waveform is decomposed into the DWT coefficients d0 – a7
using 8 cascaded filter banks. Due to subsequent downsampling within each filter bank
the bandwidth is reduced by a factor of 2. Thus highest frequencies can be observed in
level d0 whereas lowest frequencies are located in level a7.
4.2 Indentation testing of composite material
As previously reported by Baccar and Söffker, different micro-mechanical fracture
mechanisms are expected to occur during indentation testing of composites. These include delamination, matrix crack, debonding, and fiber breakage, which can be distinguished based on peak frequencies of the AE signal (11). The mechanical test rig which
is used for the experiments is shown in Figure 5.
6
Waveform
Amplitude
d0
10 -3
1
1
0
0
-1
-1
2000
4000
6000
8000
1000
10000
d1
10 -4
1
2000
3000
4000
5000
d2
10 -3
5
0
0
-5
500
1000
1500
2000
2500
-1
200
400
600
d3
10 -3
800
1000
1200
200
250
300
d4
0.01
2
0
0
-2
-0.01
100
200
300
400
500
600
50
100
150
d6
d5
0.02
0.01
0
0
-0.01
-0.02
50
100
150
20
40
60
80
30
40
a7
d7
0.05
0.05
0
0
-0.05
-0.05
10
20
Sample
30
40
10
20
Sample
Figure 4. Decomposition of PLB into 8 DWT levels.
Indentor
Clamping
system
CFRP
plate
Figure 5. Mechanical setup for indentation testing (11).
7
During indentation testing plate specimens of the dimension 425 mm x 425 mm x
1.8 mm are fixed using a clamping mechanism. Laminated cross ply material with woven fabric on the outer layers is used. During the test, the load is applied manually in
transverse direction until fracture of the specimen. A sharp, cone-shaped indentation
tool is used, which penetrates the surface of the material.
In Figure 6, the raw measurement data and cumulative energy at different levels of the
DWT representation are shown. Here, several AE bursts are visible, which are related to
the fracture of the material. In general the energy release rate (i.e. steps in cumulative
energy) correlates well with the AE bursts observed in time domain. However, different
distribution of energy among the frequency bands is evident. Furthermore, the bursts are
unevenly distributed in time due to the stochastic nature of AE. Nevertheless, groups of
successive bursts can be identified, which suggests a causal connection of their sequence.
Normalized amplitude
1
0.5
0
-0.5
-1
Scaled cumulative energy
1.2
1
0.8
0.6
d2
d3
d4
d5
d6
d7
0.4
0.2
0
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.1
0.11
Time [s]
Figure 6. Raw data and cumulative sum of energy in different frequency bands.
For instance after duration of 0.05 s, a group of bursts which follow each other closely
in time is visible. Considering the cumulative energy of each level, a rise of the red
curve (level d3) is detected during the first bursts, which indicates high peak frequencies. In contrast, the last burst in the group is related to a significant rise of the green
curve (level d6) and thus characterized by lower peak frequencies. Examples of the related signals are provided in the sequel. In Figure 7 (a), an example of high frequency
AE burst is presented. Here, energy is primarily distributed in the scales d3 and d2
which corresponds to the frequency bands 150 kHz – 312 kHz and 312 kHz – 625 kHz,
respectively. According to the literature, AE waveforms with peak frequencies above
350 kHz can be attributed to fiber breakage (11). Similarly, the low frequency burst is
shown in Figure 7 (b). Here, most of the energy is located in the level d6 with the respective frequency band of 19 kHz – 39 kHz. AE events with peak frequencies below
120 kHz are related to delamination of composites (11). Thus it may be concluded that
after fiber breakage subsequent delamination occurs due to stress redistribution.
8
Normalized amplitude
Normalized amplitude
0
-1
d0
d1
d2
d3
d4
d5
d6
d7
Level
Level
1
20
30
40
Time [ms]
50
1
0
-1
d0
d1
d2
d3
d4
d5
d6
d7
20
30
40
50
60
Time [ms]
(a)
(b)
Figure 7. Time and time-frequency domain representation of AE bursts.
5. Summary and conclusion
In SHM, data-driven approaches are frequently used to provide statements regarding the
current state of components or systems. It is well known that feature extraction is a crucial step in the development of a reliable classifier. However, if measurement data are
generated at high bandwidth, online feature extraction is challenging. In this context,
FPGAs are well suited to accelerate signal processing. In this paper a novel measurement system is used, which is based on a low-cost hardware platform. Due to a small
form factor and low power consumption it is well suited for embedded implementation
of AE monitoring. Furthermore, a FPGA-based implementation of DWT with 8 levels is
used to calculate DWT coefficients in real time. The DWT coefficients can be stored
along with the measurement data immediately. Waveforms are acquired with 16 bit resolution at a sample rate of 5 MHz. In this paper, experimental results from PLB test and
indentation testing are presented. In accordance with (11), different types of AE waveforms can be distinguished according to their frequency content. Furthermore, based on
the cumulative sum of energy in different frequency bands conclusions regarding the
particular evolution of damage mechanisms can be drawn. In future work, a classification scheme can be implemented (in software and/or programmable logic) to detect and
distinguish different damage mechanisms automatically.
References
1.
2.
3.
CR Farrar and K Worden, “An introduction to structural health monitoring”, Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences 365(1851), pp 303–315, 2007.
SF Wirtz, APA Cunha, M Labusch, G Marzun, S Barcikowski, D Söffker, “Development of low-cost FPGA-based measurement system for online processing of
acoustic emission data applied to control of pulsed laser ablation in liquids”, Sensors, 2018, submitted
G Georgakoudis, C Gillan, A Hassan, UI Minhas, I Spence, G Tzenakis, C Pattison, “NanoStreams: Codesigned microservers for edge analytics in real time”, 16th
International Conference on Embedded Computer Systems: Architectures, Modeling and Simulation (SAMOS), pp 180–187, 2016.
9
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
M Johnson and P Gudmundson, “Broad-band transient recording and characterization of acoustic emission events in composite laminates”, Composites Science and
Technology 60(15), pp 2803–2818, 2000.
QQ Ni and M Iwamoto, “Wavelet transform of acoustic emission signals in failure
of model composites”, Engineering Fracture Mechanics 69(6), pp 717–728, 2002.
R Gutkin, CJ Green, S Vangrattanachai, ST Pinho, P Robinson, and PT Curtis,
“On acoustic emission for failure investigation in CFRP: Pattern recognition and
peak frequency analyses”, Mechanical Systems and Signal Processing 25(4), pp
1393–1407, 2011.
G Kalogiannakis, J Quintelier, P De Baets, J Degrieck, and D Van Hemelrijck,
“Identification of wear mechanisms of glass/polyester composites by means of
acoustic emission”, Wear 264(3–4), pp 235–244, 2008.
A Marec, JH Thomas, R El Guerjouma, “Damage characterization of polymerbased composite materials: Multivariable analysis and wavelet transform for clustering acoustic emission data”, Mechanical Systems and Signal Processing 22(6),
pp 1441–1464, 2008.
G Qi, A Barhorst, J Hashemi, and G Kamala, “Discrete Wavelet Decomposition of
Acoustic Emission Signals from Carbon-Fiber Reinforced Composites”, Composite Science and Technology 57(4), pp 389–403, 1997.
J Yousefi, M Ahmadi, MN Shahri, AR Oskouei, and FJ Moghadas, “Damage Categorization of Glass/Epoxy Composite Material Under Mode II Delamination Using Acoustic Emission Data: A Clustering Approach to Elucidate Wavelet Transformation Analysis”, Arabian Journal for Science and Engineering 39(2), pp
1325–1335, 2014.
D Baccar and D Söffker, “Identification and classification of failure modes in laminated composites by using a multivariate statistical analysis of wavelet coefficients”, Mechanical Systems and Signal Processing 96, pp 77–87, 2017.
M Shateri, D Thomson, M Ghaib, and D Svecova, “System for Logging Acoustic
Emission Signals from FRP Reinforcing Rods that is Suitable for Field and Long
Term Laboratory Testing”, In: Chang, F.K.; Kopsaftopoulos (Ed.): Structural
Health Monitoring 2017, USA, Stanford, September 12-14, 2017, pp 1220-1227.
O Cheng, W Abdulla, and Z Salcic, “Hardware-software codesign of automatic
speech recognition system for embedded real-time applications”, IEEE Transactions on Industrial Electronics 58(3), pp 850–859, 2011.
K Virupakshappa, and E Oruklu, “A hardware/software co-design architecture for
ultrasonic flaw detection with Hidden Markov Model and Wavelet Transform”,
IEEE International Ultrasonics Symposium (IUS), pp 1–4, 2017.
M Bahoura and H Ezzaidi, “FPGA-implementation of discrete wavelet transform
with application to signal denoising”, Circuits, Systems, and Signal Processing
31(3), pp 987–1015, 2012.
M Bahoura, H Ezzaidi, “Real-time implementation of discrete wavelet transform
on FPGA”, IEEE 10th International Conference on Signal Processing Proceedings,
pp 191–194, 2010.
APA Cunha, SF Wirtz, D Söffker, N Beganovic, “Implementation of frequencybased classification of damages in composites using real-time FPGA-based hardware framework”, ASME 2017 International Design Engineering Technical Conferences & Computers and Information in Engineering Conference, Cleveland,
Ohio, USA, August 6-9, 2017, pp 1–8.
10
View publication stats