1. Structural Health Monitoring Sukun Kim, David Culler James Demmel, Gregory Fenves, Steve Glaser Thomas Oberheim, Shamim Pakzad UC Berkeley NEST Retreat – Jun 4, 2004

2. Structure Monitoring Data Acquisition Data Collection Processing & Feedback

3. Overview Low cost structure monitoring - Monitor structure, and analyze the health of structure based on sensed data at low cost Low cost structure monitoring - Monitor structure, and analyze the health of structure based on sensed data at low cost For Golden Gate Bridge, monitor vibration of bridge, and detect unusual behavior by wind, earthquake, or local damage For Golden Gate Bridge, monitor vibration of bridge, and detect unusual behavior by wind, earthquake, or local damage Extend reach of Wireless Sensor Network in a different direction – high fidelity sampling Extend reach of Wireless Sensor Network in a different direction – high fidelity sampling High accuracy, high frequency with low jitter, large amount of data High accuracy, high frequency with low jitter, large amount of data

4. Challenges Data Acquisition Data Acquisition Accelerometer Board Accelerometer Board High Frequency Sampling & Jitter High Frequency Sampling & Jitter Data Collection Data Collection Large-scale Reliable Data Transfer Large-scale Reliable Data Transfer Signal processing & System Identification Signal processing & System Identification

5. Accelerometer Board ADXL 202ESilicon Designs 1221L Range-2G ~ 2G-0.1G ~ 0.1G System noise floor200(μG/√Hz)30(μG/√Hz) Price$10$150 Both accelerometers for two axis Both accelerometers for two axis Thermometer Thermometer 16bit ADC 16bit ADC

7. HighFrequencySampling Jitter Histogram 0μs0μs10μs Sampling Other job Non-preemptible portion (atomic section) Preemptible task portion

8. Large-scale Reliable Transfer Explicit open handshake - Data description and size of cluster is sent as a transfer request Explicit open handshake - Data description and size of cluster is sent as a transfer request Data transfer is composed of multiple rounds. In each round, sender sends packets missing in the previous round Data transfer is composed of multiple rounds. In each round, sender sends packets missing in the previous round Tear-down is implicit Tear-down is implicit Sender Receiver Open Ack for Open Ack for Data Block 1 Data Block 2 Data Block 3 Data Block 4 DONE

9. Considering loss rate of 3%, actual relative throughput is 91%, which is higher than 85% of channel utilization ratio. This is because control packets do not follow 10 packets/s. Considering loss rate of 3%, actual relative throughput is 91%, which is higher than 85% of channel utilization ratio. This is because control packets do not follow 10 packets/s. Throttle for data packet is fixed at 10 pkt/s Throttle for data packet is fixed at 10 pkt/s Optimal case: window size is infinite Optimal case: window size is infinite For the case with window size 16, throughput is 88% of optimal case. For the case with window size 16, throughput is 88% of optimal case.

10. Status Measure acceleration from multiple boards synchronously Measure acceleration from multiple boards synchronously Sather tower Sather tower PowerBar building PowerBar building Data is available on the web Data is available on the web

11. Questions

12. Signal Processing and System Identification Signal Processing Signal Processing Analog low-pass filter with threshold frequency 25Hz is used Analog low-pass filter with threshold frequency 25Hz is used Averaging is used. If noise follows Gaussian distribution, by averaging N numbers, noise decreases by a factor of sqrt(N) Averaging is used. If noise follows Gaussian distribution, by averaging N numbers, noise decreases by a factor of sqrt(N) System Identification System Identification Identifying model of target system Identifying model of target system By matching input to system and output from system, construct a mathematical system model (Box-Jenkins multi-input multi-output model) By matching input to system and output from system, construct a mathematical system model (Box-Jenkins multi-input multi-output model)

13. Conclusion New challenges are analyzed which are brought by structure monitoring to wireless sensor network New challenges are analyzed which are brought by structure monitoring to wireless sensor network High accuracy accelerometer, high frequency sampling with low jitter, low-pass filter, averaging, large-scale reliable data collection High accuracy accelerometer, high frequency sampling with low jitter, low-pass filter, averaging, large-scale reliable data collection Temperature Gravity Variation Accelerometer variation Acoustic Noise nG μGμG mG G Challenges versus Accuracy Local Damage Detection Large Scale Earthquake Nuclear Test Detection Traffic Identification nG μGμG mG G Possible Applications versus Accuracy

14. Table of Contents Overview Overview Data Acquisition Data Acquisition Accelerometer Board Accelerometer Board High Frequency Sampling & Jitter High Frequency Sampling & Jitter Data Collection Data Collection Large-scale Reliable Data Transfer Large-scale Reliable Data Transfer Signal processing & System Identification Signal processing & System Identification Conclusion Conclusion Challenges & Future Work Challenges & Future Work

15. HighFrequencySampling Made by David Gay Made by David Gay Up to 6.67KHz with 4 bytes sample Up to 6.67KHz with 4 bytes sample MicroTimer – Supports one timer, micro second level granularity MicroTimer – Supports one timer, micro second level granularity BufferLog – Has two buffers. One is filled up by upper layer application while the other buffer is written to flash memory as a background task BufferLog – Has two buffers. One is filled up by upper layer application while the other buffer is written to flash memory as a background task

16. Jitter Test (1KHz, 5KHz, 6.67KHz) Peak to Peak is time to fill up buffer Peak to Peak is time to fill up buffer Spiky portion is time to write buffer to flash Spiky portion is time to write buffer to flash Can sample as long as the former is larger than the latter Can sample as long as the former is larger than the latter

17. Jitter Test Histogram (1KHz, 5KHz, 6.67KHz) Jitter is within 10µs Jitter is within 10µs Peak at 625ns – Wakeup time from sleep mode Peak at 625ns – Wakeup time from sleep mode

18. Jitter Analysis Sampling Other job Non-preemptible portion (atomic section) Preemptible task portion Jitter Sample CC+T(k1)C+T(k2)... F(k2) F(k3)

19. Table of Contents Overview Overview Data Acquisition Data Acquisition Accelerometer Board Accelerometer Board High Frequency Sampling & Jitter High Frequency Sampling & Jitter Data Collection Data Collection Large-scale Reliable Data Transfer Large-scale Reliable Data Transfer Signal processing & System Identification Signal processing & System Identification Conclusion Conclusion Challenges & Future Work Challenges & Future Work

20. Large-scale Reliable Data Transfer 4Byte of data and 4Byte of time stamp at 100Hz in 100 nodes, transfer 40pkt/s – Sample data for 5 minutes, and collect data for more than 5 hours!!! 4Byte of data and 4Byte of time stamp at 100Hz in 100 nodes, transfer 40pkt/s – Sample data for 5 minutes, and collect data for more than 5 hours!!! Efficient and reliable data transfer is crucial Efficient and reliable data transfer is crucial RAM to RAM one-hop transfer is implemented as a building block - LRX RAM to RAM one-hop transfer is implemented as a building block - LRX

21. LRX component (continued) Explicit open handshake - Data description and size of cluster is sent as a transfer request Explicit open handshake - Data description and size of cluster is sent as a transfer request Data transfer is composed of multiple rounds. In each round, sender sends packets missing in the previous round Data transfer is composed of multiple rounds. In each round, sender sends packets missing in the previous round Tear-down is implicit Tear-down is implicit Sender Receiver Open Ack for Open Ack for Data Block 1 Data Block 2 Data Block 3 Data Block 4 DONE

22. Considering loss rate of 3%, actual relative throughput is 91%, which is higher than 85% of channel utilization ratio. This is because control packets do not follow 10 packets/s. Considering loss rate of 3%, actual relative throughput is 91%, which is higher than 85% of channel utilization ratio. This is because control packets do not follow 10 packets/s. Throttle for data packet is fixed at 10 pkt/s Throttle for data packet is fixed at 10 pkt/s Optimal case: window size is infinite Optimal case: window size is infinite For the case with window size 16, throughput is 88% of optimal case. For the case with window size 16, throughput is 88% of optimal case.

23. As loss rate increases, retransmission increases, and throughput decreases As loss rate increases, retransmission increases, and throughput decreases

24. Channel Utilization TOS_MsgLRX (only data)LRX (Window Size 16) Total Data (bytes) 36 613 Meta Data (bytes) 710197 Real Data (bytes) 2926416 Channel Utilization (%) 78.3872.2267.86 Comparison to TOS_Msg (%) 10089.6684.24 LRX (data only) is the theoretical limit of LRX (when window size is infinite) LRX (data only) is the theoretical limit of LRX (when window size is infinite) Usage LRX lowers channel utilization by 15% Usage LRX lowers channel utilization by 15%

25. Table of Contents Overview Overview Data Acquisition Data Acquisition Accelerometer Board Accelerometer Board High Frequency Sampling & Jitter High Frequency Sampling & Jitter Data Collection Data Collection Large-scale Reliable Data Transfer Large-scale Reliable Data Transfer Signal processing & System Identification Signal processing & System Identification Conclusion Conclusion Challenges & Future Work Challenges & Future Work

26. Signal Processing As an analog signal processing low-pass filter is used, which filters high frequency noise As an analog signal processing low-pass filter is used, which filters high frequency noise For accelerometer board, low-pass filter with threshold frequency 25Hz is used. Then ADC should sample at frequency much higher than 50Hz by Nyquist theorem, and imperfect low- pass filter For accelerometer board, low-pass filter with threshold frequency 25Hz is used. Then ADC should sample at frequency much higher than 50Hz by Nyquist theorem, and imperfect low- pass filter As a digital signal processing, averaging is used. If noise follows Gaussian distribution, by averaging N numbers, noise decreases by a factor of sqrt(N) As a digital signal processing, averaging is used. If noise follows Gaussian distribution, by averaging N numbers, noise decreases by a factor of sqrt(N)

27. System Identification Identifying model of target system Identifying model of target system By matching input to system and output from system, we can construct a mathematical system model. By matching input to system and output from system, we can construct a mathematical system model. Usual process is (1) fitting a general Box-Jenkins multi-input multi-output model to sampled data. (2) And natural frequencies, damping ratios and mode shape are then estimated using the estimated Box-Jenkins model. Usual process is (1) fitting a general Box-Jenkins multi-input multi-output model to sampled data. (2) And natural frequencies, damping ratios and mode shape are then estimated using the estimated Box-Jenkins model. Most part of system identification is under development on civil engineering side. Most part of system identification is under development on civil engineering side.

28. Table of Contents Overview Overview Data Acquisition Data Acquisition Accelerometer Board Accelerometer Board High Frequency Sampling & Jitter High Frequency Sampling & Jitter Data Collection Data Collection Large-scale Reliable Data Transfer Large-scale Reliable Data Transfer Signal processing & System Identification Signal processing & System Identification Conclusion Conclusion Challenges & Future Work Challenges & Future Work

29. Conclusion New challenges are analyzed which are brought by structure monitoring to wireless sensor network New challenges are analyzed which are brought by structure monitoring to wireless sensor network High accuracy accelerometer, high frequency sampling with low jitter, low-pass filter, averaging, large-scale reliable data collection High accuracy accelerometer, high frequency sampling with low jitter, low-pass filter, averaging, large-scale reliable data collection Temperature Gravity Variation Accelerometer variation Acoustic Noise nG μGμG mG G Challenges versus Accuracy Local Damage Detection Large Scale Earthquake Nuclear Test Detection Traffic Identification nG μGμG mG G Possible Applications versus Accuracy

30. Table of Contents Overview Overview Data Acquisition Data Acquisition Accelerometer Board Accelerometer Board High Frequency Sampling & Jitter High Frequency Sampling & Jitter Data Collection Data Collection Large-scale Reliable Data Transfer Large-scale Reliable Data Transfer Signal processing & System Identification Signal processing & System Identification Conclusion Conclusion Challenges & Future Work Challenges & Future Work

31. Challenges & Future Work Calibrating acceleration value to temperature Calibrating acceleration value to temperature Time synchronization – RBS, TPSN Time synchronization – RBS, TPSN To maximize utility of channel, we need to monitor channel quality (loss rate), and throttle packet injection rate accordingly To maximize utility of channel, we need to monitor channel quality (loss rate), and throttle packet injection rate accordingly Using LRX as a building block, multi-hop data collection need be implemented Using LRX as a building block, multi-hop data collection need be implemented TASK TASK

32. Backup Slides

33. Cost Comparison Conventional piezoelectric accelerometer with PC system costs $40,000 Conventional piezoelectric accelerometer with PC system costs $40,000 Budget for structure monitoring budget is $1,000,000 level Budget for structure monitoring budget is $1,000,000 level Wireless sensor network with MEM accelerometer costs $500 Wireless sensor network with MEM accelerometer costs $500 Cheaper by a factor of 100 Cheaper by a factor of 100

34. Shaking Table Test Silicon Design 1221L is more quite, but less sensitive to dynamic movement Silicon Design 1221L is more quite, but less sensitive to dynamic movement

35. Noise Floor Test Blue – Seismic Vault Blue – Seismic Vault Red – McCone Hall Red – McCone Hall

36. Jitter Analysis (continued) Assume that the probability of timer event occurring at any point in atomic section i is same, then jitter will follow C+X(i). Assume that the probability of timer event occurring at any point in atomic section i is same, then jitter will follow C+X(i). Since jitter distribution of every atomic section begins from C, the frequency is highest near C and decreases as moving farther. And frequency drop at C+T(i) by F(i), since atomic section i will not have any distribution beyond C+T(i). Since jitter distribution of every atomic section begins from C, the frequency is highest near C and decreases as moving farther. And frequency drop at C+T(i) by F(i), since atomic section i will not have any distribution beyond C+T(i). Actually there is a peak at C, because when program is in preemptible section, it will immediately service timer event after context switch time C. Actually there is a peak at C, because when program is in preemptible section, it will immediately service timer event after context switch time C. Jitter Sample CC+T(k1)C+T(k2)... F(k2) F(k3) T(i): execution time of atomic section i T(i): execution time of atomic section i X(i): a random variable uniformly distributed in [0, T(i)] X(i): a random variable uniformly distributed in [0, T(i)] C: context switch time C: context switch time F(i): frequency of occurrence of atomic section i F(i): frequency of occurrence of atomic section i

37. Calculation of Transfer Timer Let us assume each node store 4Byte of data and 4Byte of time stamp at 100Hz. And assume there are 100 nodes, radio throughput is 1.2KB/s, and data is collected to one base station. If acceleration data worthy 5 minutes is collected, each node will transfer 240,000Bytes. 100 nodes will transfer 24,000,000Bytes. Since the end link to base station is a bottleneck, it will take more than 5 hours. We can see bandwidth is narrow compared to aggressive data sampling. Even if we alleviate this problem using multi-channel or multi- tier network, still we will be in short of bandwidth. Let us assume each node store 4Byte of data and 4Byte of time stamp at 100Hz. And assume there are 100 nodes, radio throughput is 1.2KB/s, and data is collected to one base station. If acceleration data worthy 5 minutes is collected, each node will transfer 240,000Bytes. 100 nodes will transfer 24,000,000Bytes. Since the end link to base station is a bottleneck, it will take more than 5 hours. We can see bandwidth is narrow compared to aggressive data sampling. Even if we alleviate this problem using multi-channel or multi- tier network, still we will be in short of bandwidth.

38. LRX component Transfers one data cluster, which is composed of several blocks. Transfers one data cluster, which is composed of several blocks. One block fits into one packet, so the number of blocks is equal to window size. One block fits into one packet, so the number of blocks is equal to window size. Each data cluster has a data description. After looking at data description, receiver may deny data (receiver already has that data, or that data is not useful anymore). Each data cluster has a data description. After looking at data description, receiver may deny data (receiver already has that data, or that data is not useful anymore).

39. Sender

40. Receiver

41. Sender Receiver Open Ack for Open Ack for Data Block 1 Data Block 2 Data Block 3 Data Block 4 DONE

42. Sender Receiver Open Data Block 1 Data Block 2 Open Ack for Open

43. Sender Receiver Open Data Block 1 Data Block 2 Open Ack for Open Ack for Open

44. Sender Receiver Ack for Data Block 1 Data Block 2 Data Block 3 Data Block 4 DONE Data Block 2 Ack for Data

45. Sender Receiver Ack for Data Block 1 Data Block 2 Data Block 3 Data Block 4 DONE Data Block 4 Ack for Data

46. Sender Receiver Data Block 1 Data Block 2 Data Block 3 Data Block 4 DONE Data Block 4 Ack for Data DONE

47. Sender Receiver Ack for Data Block 1 Data Block 2 Data Block 3 Data Block 4 Data Block 2 Ack for Data Block 3

48. Why Sender times out There are two reasons why only sender times out and stimulate receiver for Ack. The first reason is shown in Figure 16. If sender doesn’t time out, for a receiver to make sure Ack is delivered to sender, receiver should get acknowledgement from sender for Ack itself. This is not good. So it is clear that sender should timeout. Given that sender times out, timeout of receiver makes no difference except that channel is wasted by unnecessary Ack from receiver. So timeout in only sender side is desirable. As a second reason, if receiver times out, in case like Figure 18 (if first Data after Ack is lost), second Data always collide with resent Ack of receiver. This is not a good phenomenon. Therefore, after sending last packet in each round, if acknowledgement does not come, sender sends the last packet in that round again to stimulate acknowledgement. However, this does not mean receiver has no timeout. Receiver waits sufficient amount of time, and if nothing happens, it regards the situation as a failure. There are two reasons why only sender times out and stimulate receiver for Ack. The first reason is shown in Figure 16. If sender doesn’t time out, for a receiver to make sure Ack is delivered to sender, receiver should get acknowledgement from sender for Ack itself. This is not good. So it is clear that sender should timeout. Given that sender times out, timeout of receiver makes no difference except that channel is wasted by unnecessary Ack from receiver. So timeout in only sender side is desirable. As a second reason, if receiver times out, in case like Figure 18 (if first Data after Ack is lost), second Data always collide with resent Ack of receiver. This is not a good phenomenon. Therefore, after sending last packet in each round, if acknowledgement does not come, sender sends the last packet in that round again to stimulate acknowledgement. However, this does not mean receiver has no timeout. Receiver waits sufficient amount of time, and if nothing happens, it regards the situation as a failure.

49. Imperfect Low-pass Filter Frequency Amplitude Filtering threshold

50. Time Synchronization Temporal jitter is handled by high frequency sampling component. Spatial jitter should be solved by time synchronization. ITP [8] is a time synchronization protocol widely used in Internet. In wireless sensor network, there were several studies. In RBS [9], synchronization is done among receivers, eliminating sender’s jitter in media access. TPSN [10] put time stamp after obtaining channel. This gives even better synchronization accuracy than RBS (10μs compared to 20μs). Still there is a source of jitter at receiver side. As we saw in jitter for sampling, handling interrupt by radio can be delayed by atomic section of other activity. As suggested in [10], putting time stamp at MAC layer in receiver side will eliminate this jitter. Temporal jitter is handled by high frequency sampling component. Spatial jitter should be solved by time synchronization. ITP [8] is a time synchronization protocol widely used in Internet. In wireless sensor network, there were several studies. In RBS [9], synchronization is done among receivers, eliminating sender’s jitter in media access. TPSN [10] put time stamp after obtaining channel. This gives even better synchronization accuracy than RBS (10μs compared to 20μs). Still there is a source of jitter at receiver side. As we saw in jitter for sampling, handling interrupt by radio can be delayed by atomic section of other activity. As suggested in [10], putting time stamp at MAC layer in receiver side will eliminate this jitter.

51. Table of Contents Overview Overview Data Acquisition Data Acquisition Accelerometer Board Accelerometer Board High Frequency Sampling & Jitter High Frequency Sampling & Jitter Data Collection Data Collection Large-scale Reliable Data Transfer Large-scale Reliable Data Transfer Signal processing & System Identification Signal processing & System Identification Conclusion Conclusion Challenges & Future Work Challenges & Future Work

52. Acknowledgement This work is supported, in part, by the National Science Foundation under Grant No. EIA- 0122599. This work is supported, in part, by the National Science Foundation under Grant No. EIA- 0122599.