Mimicking the biological neural system using electronic logic circuits

Transcription

1 Mimicking the biological neural system using electronic logic circuits G.R.Kirikera a, V. Shinde a, I. Kang a, M.J.Schulz *a, V. Shanov a, S. Datta a, D. Hurd a, Bo Westheider a, M. Sundaresan b, A. Ghoshal c a Smart Structures Bio-Nano Lab, Dept. Mechanical Engrg., Univ. Cincinnati, Cincinnati, OH b Department of Mechanical Engineering, North Carolina A&T State University, NC c Senior Scientist, United Technologies Research Center, CT 618 ABSTRACT Detecting and locating cracks in structural components and joints that have high feature densities is a challenging problem in the field of Structural Health Monitoring. There have been advances in piezoelectric sensors, actuators, wave propagation, MEMS, and optical fiber sensors. However, few sensor-signal processing techniques have been applied to the monitoring of joints and complex structural geometries. This is in part because maintaining and analyzing a large amount of data obtained from a large number of sensors that may be needed to monitor joints for cracks is difficult. Reliable low cost assessment of the health of structures is crucial to maintain operational availability and productivity, reduce maintenance cost, and prevent catastrophic failure of large structures such as wind turbines, aircraft, and civil infrastructure. Recently, there have also been advances in development of simple passive techniques for health monitoring including a technique based on mimicking the biological neural system using electronic logic circuits. This technique aids in reducing the required number of data acquisition channels by a factor of ten or more and is able to predict the location of a crack within a rectangular grid or within an arbitrarily arranged network of continuous sensors or neurons. The current paper shows results obtained by implementing this method on an aluminum plate and joint. The plates were tested using simulated acoustic emissions and also loading via an MTS machine. The testing indicates that the neural system can monitor complex joints and detect acoustic emissions due to propagating cracks. High sensitivity of the neural system is needed, and further sensor development and testing on different types of joints is required. Also indicated is that sensor geometry, sensor location, signal filtering, and logic parameters of the neural system will be specific to the particular type of joint (material, thickness, geometry) being monitored. Also, a novel piezoresistive carbon nanotube nerve crack sensor is presented that can become a neuron and respond to local crack growth. Keywords: Artificial Neural System, Structural Health Monitoring, Riveted Structures, Passive, Acoustic Emissions. 1. INTRODUCTION Structural Health Monitoring (SHM) of large structures requires the use of many sensors and actuators when using active SHM systems. Use of a large number of sensors to monitor critical parts of large structures such as wind turbines and aircraft requires substantial resources for data collection and storage, and subsequent signal processing. There have been advances in developing rugged, sensitive piezoelectric, MEMS, and optical fiber sensors [1-2]. These sensors can be embedded in/on the structure, but the problem of data acquisition becomes critical. One of the methods proposed to reduce the problem of data acquisition and processing of this information is the Artificial Neural System (ANS). The ANS is a method to connect n number of sensor nodes in series to form neurons, and the neurons in * Mark.J.Schulz@uc.edu. Figure 1. Analogous aircraft and Biological Neural Systems.

2 turn interact and pass signals akin to how the biological neural system functions [3-5]. The ANS is a passive health monitoring system that can reduce the required number of data acquisition channels by a factor of ten or more. Figure 1 shows the analogy of an aircraft with the ANS. The ANS consists of electronic logic circuits that have n number of channels as input from sensors, and only two channels of output. One of the output channels contains the crack location information and the other channel contains time domain information from the crack signals [6]. This proposed method is simple and could be part of an Integrated Vehicle Health Monitoring (IVHM) system, where the requirement of reducing the number of data acquisition channels is of critical importance. The ANS predicts the location of cracks within a grid of sensors. Previously reported testing of ANS monitoring of structures considered flat composite structures and simulated Acoustic Emissions (AE) using a pencil lead break, and some fatigue testing has also been done. This testing has shown that composite structures produce large acoustic emissions when the fibers break and when the layers delaminate, and the ANS was very successful in monitoring small test structures including beams and panels up to four feet square in size. The current paper considers predicting the location of acoustic emissions in complicated structures having rivets or varying geometrical profiles, and this aspect of SHM has been less investigated. For this paper, simple aluminum plates and also a riveted joint have been manufactured and tested under fatigue loading (bending and axial) using a MTS machine. The aluminum plates were connected to a simple two channel ANS analog processor box for predicting the location of cracks. Also, a new piezoresistive carbon nanotube nerve crack sensor is proposed for SHM. 2. THE ARTIFICIAL NEURAL SYSTEM The ANS has the capability of highly distributed sensing and parallel signal processing for in-situ real time SHM. The ANS can monitor complicated parts of a structure by listening for Acoustic Emissions (AE) or waves and any energy release due to damage in an aircraft fuselage section or wind turbine structure, including at joints. Shown in Figure 2 is an illustration of a fuselage of an aircraft that could be monitored using the ANS. The aircraft fuselage is split into sections to monitor for AE. In this concept, each stiffener will be monitored by a separate neuron. A neuron is a combination of long sensors connected in series. The n number of neurons will be fed into the ANS analog processor box. The ANS analog processor will have n channels as input and only two channels as output. One of the channels will predict the location of the acoustic emission, the other channel will capture the collective time domain signals caused by the straining of the sensor due to the AE. This time domain signal indicates the frequency content of the AE. The outputs from the ANS analog processor are fed into an onboard computer for data acquisition. Note that this system has several advantages because most of the signal processing is done in an analog fashion. This reduces the cost of maintaining and obtaining a large number of digital data acquisition channels, ultimately reducing the cost of maintenance. For laboratory testing, a small section of the unit cell depicted in Figure 2 has been replicated by constructing three aluminum plates joined with rivets as shown in Figure 2. Testing of this section is discussed next. Figure 2. Application of the ANS; Fuselage section of an aircraft, test section of a joint.

3 3. PROPAGATION OF ACOUSTIC EMISSIONS THROUGH A JOINT In complex built-up structures, each section is made of components joined using rivets or fasteners. To study how Acoustic Emissions (AE) propagate in a joint, three 775- T6 aluminum plates were joined by rivets. Figure 3 shows the geometry of the joint with rivets and sensors/actuators attached to the structure. Two of the base plates have dimensions of 6.2 X 6.2 X.12. The plate used to join the bottom two plates has dimensions 6.2 X 3 X.12. The plate was joined using 12 rivets. A piezoelectric (PZT) single sheet was purchased from Piezo Systems Inc. to form the individual sensor elements. The material was PSI-5A4E wih dimensions of 2.85 x 2.85 x.15. This sheet was cut manually using a sharp razor blade to.6 X.35 X.15 dimensions. Five monolithic PZT s were made. These PZT s were bonded onto the aluminum plate using superglue. Kapton was used as an insulating layer between the PZT and the aluminum plate. 3.1 Capacitance Effect of the Sensor A PZT acts as a capacitor that has charge attraction/repulsion due to strain. When using a PZT as a sensor or actuator on an aluminum structure, it is possible that the structure can serve as the ground or return circuit for the PZT by bonding the PZT directly to the electrically conductive structure. However, this approach will have current flowing in the structure to the PZT electrodes, and the PZT can generate 2 volts or more due to an impact to the structure or when driven to generate waves for active damage interrogation. If the ground circuit became Figure 3. Geometry of the riveted aluminum joint. degraded or interrupted due to corrosion or loosening, the sensor signal may become degraded and there could be arcing in the connection that in some situations might be dangerous. Also, if the aluminum were to become live due to a lightning strike or through faults in the vehicle electrical system, the senor system would possibly be damaged. Therefore, here the PZT is encapsulated in a kapton film that has a large electrical resistance that will prevent any current flow through the film. A thin kapton film is used to minimize shear lag that will reduce the strain in the PZT. It was found that the layer consisting of the kapton sheet and the glue between the aluminum plate and PZT acted as a capacitor and can cause small currents to flow in the structure. The impedance of this capacitor was calculated by using a simple impedance measurement circuit. Figure 4 shows the impedance measurement circuit that was used to calculate the capacitance of the kapton film and the bond layer. In the circuit the top electrode of the PZT sensor was not connected. This is because our primary interest is in measuring the capacitance value and impedance only of the bond layer. The thickness of the superglue and kapton sheet has been exaggerated for clarity in Figure 4. The resistors R1 and Ro refer to a known resistance of 1KΩ used for testing, and an oscilloscope resistance of 1MΩ, respectively. Figure 4 shows the variation of impedance of the bond layer as a function of frequency. The capacitance of the PZT sensor was measured using a capacitance meter and was found to be roughly 8 µf. The average capacitance of the bond layer was found to be in the order of.3 nf. The infinite impedance at low frequencies shown in Figure 4 is due to the high resistance of the kapton layer. An interesting result of studying the capacitive effect of multiple PZT sensors that are bonded to the aluminum and individually electrically insulated and connected to separate channels of an oscilloscope is that if the aluminum plate is not grounded, the electrical signal from one PZT due to strain is received at all the others immediately and this can obscure the later signal due to the acoustic waves traveling to the other sensors. The plate structure and the kapton film and the bottom electrode of the PZT form a capacitor and may cause very small currents to flow in the ground circuit. In practice, the operation of the neural system generates sensor voltages at high frequency less than one volt, and the power is very low. The capacitive coupling to the structure is only at high frequency and produces a very small current.

4 4.5 x Impedance (ohms) Frequency (Hz) x 1 4 Figure 4. Measurement of Impedance of the kapton and superglue layer; Impedance measurement circuit, Variation of Impedance with frequency 3.2 Pulse excitation The riveted aluminum plate was tested for propagation of AE with the aid of pulse excitation. This would tend to produce flexural waves in the plate. Figure 5 shows the excitation voltage used to excite the actuators embedded on the aluminum plate. The excitation voltage applied from the signal generator to the amplifier was a rectangular pulse. Because of the limited bandwidth of the amplifier, the signal shape has been distorted. Figures 5 and 6 show the actuation and response signals obtained from the PZTs attached to the structure. In Figures 5 and 6 in the legend, A refers to the PZT that is being used as an actuator. Similarly R refers to the response of the corresponding sensors. The geometrical location of PZTs on the riveted plate is shown on the top left of each figure. It can be seen from Figure 5 that the amplitude of the response of the sensors decreases as the waves propagate through the thick section of the lap joint. The amplitude of strain in the thick section is much less than in the thin section. This is expected because of the displacement being inversely proportional to the bending inertia. Figure 6 shows the response of sensors 2 and 4 for an excitation produced on the structure using PZT 3. It can be seen that most of the high frequency waves get damped out or are reflected due to travel from thicker section to thinner section. Figure 6 shows the response of sensors 2 and 3 for an excitation produced using PZT 5. It could be inferred from all the three graphs that most of the high frequency waves get reduced in amplitude when traveling from one section of the plate to the other section. 1 5 Voltage (Volts) Time (Sec) x 1-4 Figure 5. Excitation to the PZT; Pulse Excitation, Response of sensors 2, 3 and 4 for a pulse excitation to PZT 1.

5 Voltage (Volts) A3-R2 A3-R Time (Sec) x 1-4 Figure 6. Actuation through the joint; Response of sensors 2 and 4 for an excitation with PZT 3; Response of sensors 2, 3 and 4 for a pulse excitation from PZT FATIGUE TESTING OF AN ALUMINUM PLATE The study of propagation of AE through an aluminum plate was done using three experimental setups. The first consisted of loading the riveted aluminum plate joint on the MTS machine by applying a bending load. AE was provided by manually breaking a pencil lead on the surface and the location was determined using the ANS analog processor. To further investigate that the ANS analog processor can detect the original AE produced by the development of cracks, a second experimental setup was designed by providing a notch on an aluminum plate and loading it (bending load) in the MTS machine to produce a crack. The third experiment did not use the ANS and was to study AE in aluminum plates using conventional AE sensors and applying a tension load using a larger MTS machine. 4.1 Fatigue Testing of a Riveted Aluminum Joint The riveted aluminum plate previously tested for pulse excitation was further tested on an MTS machine. Figure 7 shows the setup of the riveted aluminum plate mounted on a MTS machine and fatigue loaded. An input of 3 Hz was provided to the aluminum plate joint. Sensors 2 and 4 (shown in Figure 3) were connected to the ANS analog processor box to verify if the ANS box does indeed predict the location of the acoustic emission. Data acquisition system Riveted Aluminum plate Four point bending fixture ANS box Figure 7. Setup of MTS machine to test the riveted aluminum plate.

6 With the plate being fatigue loaded, an acoustic emission was provided manually using a pencil lead break near sensor 4. The ANS electronic box had sensors 2 and 4 as inputs. Sensor 2 was pre-assigned a DC voltage of 3 volts and sensor 4 was pre-assigned with a DC value of 5 volts. The reference voltage of the ANS analog processor box was set at 3 volts. With the plate being fatigue loaded and an acoustic emission manually provided near sensor 4, the response of the ANS analog processor is shown in Figures 8 and 9. Figure 8 shows the initial amplitude of 5 volts indicating that the acoustic emission was located close to sensor 4. The firing of 3 volts shown in Figure 8 indicates that sensor 2 has also detected the AE. Figure 8 shows the time domain waveform of the simulated AE indicating the high frequency content of the waveform. Similarly Figure 9 shows the response of the ANS for an excitation located near sensor 2. The initial amplitude of 3 volts in Figure 9 indicates that the AE was located near sensor 2. Figure 9 indicates the time domain information of the straining of both the sensors. The initial peak in time domain waveforms indicates the frequency content of the AE. Note that even though the sensor is being excited by the excitation frequency of the MTS machine, the ANS analog processor automatically filters the low frequency data and outputs only the high frequency acoustic emission, and no fretting or noise interference of the MTS occurred. The joint is being used for sensor design studies and later joint sections will be tested to fatigue failure with the ANS attached. Figure 8. Response of ANS for an AE lead break located near sensor 4, Crack Information, and Time waveform of AE. Figure 9. Response of ANS for an AE lead break located near sensor 2, a) Crack Information b) Time waveform of AE. This testing indicates that the neuron of the ANS should be placed in the section of the structure where the AE occurs and that it is difficult to detect the AE wave after is travels in a joint from a section of one thickness to a section of another thickness. This is a

7 preliminary conclusion as testing in which the joint is fatigued and testing using long neuron sensors is needed. 4.2 Monitoring AE on an Aluminum Plate using the ANS The previous section showed that ANS analog processor can predict the location of simulated AE with only one channel of data acquisition. This architecture will be extended to have many simultaneous neurons. The ANS analog processor was next tested in a simulated real time environment. Figure 1 shows the geometry of a 775-T651 aluminum plate with a notch provided at the center span and side of the plate. The yield strength for this material as specified by the manufacturer is 73ksi. A thin aluminum plate was subjected to four point bending in an MTS testing machine. P a and P b are the loads applied from the MTS machine while R a and R b are the reactions from the supports. Two sensors were bonded to the aluminum plate using superglue. The narrow aluminum plate had 12 X 2.48 X.19 dimensions. Figure 1. Geometry of a narrow aluminum plate with supports. Figure 11 shows the crack that was generated due to the cyclic loading of the aluminum plate. In Figure 11 the starter notch is shown. Due to a high stress concentration near the notch, the crack starts progressing in the plate releasing strain energy due to bending of the plate. The plate in Figure 11 is straight and exhibited final brittle fracture without gross yielding. Notch Generated crack Figure 11. Crack generated due to loading the aluminum plate on the MTS machine (the plate is straight and finally fractured without bending). In a similar set-up to the riveted aluminum joint test, sensor 1 was provided with a constant DC voltage of 3 volts and Sensor 2 was provided with a constant DC voltage of 5 volts. A reference voltage of.75 volts was set for each sensor. Whenever each sensor response crosses.75 volts the ANS responds by firing the DC voltage corresponding to those sensors. Figure 12 shows the response of the ANS analog processor. Figure 12 shows that the crack is located close to sensor 1. The repeated firing seen in Figure 12 is due to the reflection of waves from the boundaries. It could

8 be noticed that the firing voltage of sensor 1 is not constant as predicted. This is because of the low bandwidth of the components and the low response time of the components used in the ANS analog processor box, which was designed for monitoring a composite material. The high frequency does not give enough time for the components in the ANS analog processor to fire the required voltage of 3 volts. Figure 12 shows the response of ANS analog processor box indicating the high frequency AE due to crack propagation in the structure. It was initially surprising to see that there was no response of Sensor 2 through the ANS analog processor. It was later found that sensor 2 was placed beside the loading point of the MTS shown as Pb in Figure 1, and most of the high frequency AE is damped when passing through the load point. For a few cases of AE due to crack growth that are not shown, there was a response from sensor 2 also. During these cases the amplitude of AE was large indicating a very large release of energy. In the beginning the AE levels were small and infrequent. As the crack grew the frequency of the bursts and the amplitude increased and the AE signal near failure was clipped during data acquisition. 1 8 Plot of Crack Information from ANS box Plot of Frequency Information from ANS box Voltage (Volts) 6 4 Voltage (Volts) Time (Sec) x 1-4 Time (Sec) x 1-4 Figure 12. Response of ANS analog processor; Crack Location Information, and Time domain waveform. This test showed that the ANS can detect AE if the sensor is close to the damage. Also, two different width sensors were tested. The wider sensor gave a larger voltage output but cut-off some of the high frequency signal. The narrower sensor captured the higher frequency signal, but the amplitude was lower partly due to the smaller area of the sensor. This sensor experiment indicates that a long narrow sensor is needed. The design proposed in [7] may be suitable. 4.3 Monitoring AE signals using Ultrasonic Transducers Fatigue testing was done on aluminum coupons using commercial barrel-type AE sensors. This testing was done in the tension mode. Acoustic emission cracks from a 12 long, 1 wide, and.125 thick 775-T6 and 224-T3 aluminum coupons were monitored using.25 diameter, 5 MHz damped ultrasonic transducers. The specimen geometry and the sensor locations are shown in Figure 13. The 5 MHz damped ultrasonic sensors were chosen for their wide band characteristics and their ability to reproduce acoustic emission waveforms with higher fidelity than the typical 15 KHz AE sensors. Figure 13. Specimen geometry and sensor locations. The acoustic emission signals were amplified by 34 db and recorded using a digitizing oscilloscope. A large number

9 of acoustic emission signals were collected during these experiments. AE waveforms from 775-T6-aluminum alloy coupon are shown in Figure 14. Frequency components in excess of 1 MHz are seen in the output of these sensors. A large number of such AE sensors could be connected to the ANS analog processor to predict the location of crack growth. The ANS can be configured using any sensor type. Figure 14. AE waveforms; Response of AE Sensor 1, Response of AE sensor Carbon Nanotube Nerve Crack Sensor A simple neuron can be formed using Single Wall Carbon Nanotubes (SWCNT). This is a piezoresistive sensor (electrical resistance changes with strain) consisting of SWCNT in a binder that is electrically insulated and bonded to the surface of the structure. The electrical resistance of the sensor changes due to strain. A crack propagating nearby may increase the strain of the sensor and if the crack reaches the sensor it will reduce the area or completely break the sensor. Thus, this is a very simple indicator of damage. A nanotube fiber sensor can detect sub-micron scale cracks due to the possible nanoscale size of the sensor. A nanotube mat sensor is shown in Figure 15. The structure shown was loaded on an MTS machine at an excitation sinusoidal frequency of 3 Hz (bending load) to generate and propagate cracks. As the structure bends due to load there is a change in resistance of the nanotube sensor. This change in resistance is converted to voltage using a wheatstone bridge, and the response is as shown in Figure 15. Once the crack penetrates the sensor the sensor snaps and the resistance goes to infinity. This would produce a DC voltage through the wheatstone bridge circuit. A fine mesh of nerves or a fiber mat formed using carbon nanotubes could be connected to the ANS analog processor to predict the location of cracks. This is a simple new approach for SHM..1.5 Voltage (Volts) Time (Sec) Figure 15. Carbon Nanotube Nerve Crack Sensor; Sensor is attached to the structure at failure, Response of the sensor due to bending of the aluminum beam before failure for an excitation frequency of 3 Hz.

10 5. CONCLUSIONS The laboratory testing has shown that ANS analog processor can detect and locate the acoustic emissions generated on a small aluminum panel using PZT sensor nodes. Extending the architecture, the ANS has the capability to reduce the required number of data acquisition channels by a factor of 1 or more for SHM applications. Using only one channel of data output, the location of an acoustic emission within a grid of sensors can be predicted. A second channel of data acquisition is used to measure the high frequency acoustic signal and this data is used to characterize the crack growth. The ANS could become part of an Integrated Health Monitoring System for metallic/composite structures including wind turbines and aircraft. The ANS is a massively parallel and highly distributed signal processing system that has capabilities similar to the human neural system. The future work includes replacing the components of the ANS analog processor with a higher bandwidth and much faster components for monitoring metallic structures. Also increasing the number of channels of input is being done. A higher sensitivity and higher frequency range are needed for crack detection in aluminum structures and joints, as compared to detecting acoustic emissions in composite materials. The new piezoresistive carbon nanotube never sensor is also being incorporated into the ANS. 6. ACKNOWLEDGEMENTS This material is based upon work supported in part by the National Renewable Energy Laboratory under subcontract number XCX Much of the help in making the prototype of ANS analog processor was provided by engineers from Texas Instruments. This support is gratefully acknowledged. 7. REFERENCES 1. M.M.Derriso, P.Faas, J.Calcaterra, J.H.Barnes, W..Sotomayer, Structural Health Monitoring Applications for Current and Future Aerospace Vehicles, 3 rd International workshop on Structural Health Monitoring, Structural Health Monitoring, The Demands and Challenges, Edited by Prof. Fu-Kuo Chang, CRC press, Pgs Prof. Fu-Kuo Chang, A summary report of the 2 nd workshop on Structural Health Monitoring, 3 rd International workshop on Structural Health Monitoring, The Demands and Challenges, Edited by Prof. Fu-Kuo Chang, CRC press. 3. W.N. Martin, A.Ghoshal, M.J. Sundaresan, M.J. Schulz, Structural Health Monitoring using an Artificial Neural System, book chapter in Recent Research Developments in Sound and Vibration, Transworld Research Network, Sundaresan, M.J., Schulz, M.J., Ghoshal, A., Pratap, P., "A Neural System for Structural Health Monitoring," SPIE 8 th International Symposium on Smart Materials and Structures, March 4-8, Patent: "Sensor Array System," (6,399,939 B1), June, 22, M. J. Sundaresan, A. Ghoshal and M. J. Schulz. 6. Invention Disclosure, Univ. of Cincinnati, , An Artificial Neural System, G.R.Kirikera, S.Datta, B. Westheider, M.J. Schulz, M.J. Sundaresan, W.N. Martin, Jr., A.Ghoshal, UC Invention Disclosure, Univ. of Cincinnati, , An Active Fiber Continuous Sensor, S. Datta, J. Hause, G. Kirikeria, D. Hurd, M.J. Schulz, M. Sundaresan, UC