Backup Sensor Placement with Guaranteed Fault ...

2011 The Sixth Beijing-Hong Kong International Doctoral Forum (IDF'11)

Backup Sensor Placement with Guaranteed Fault Tolerance for Structural Health Monitoring Md Zakirul Alam Bhuiyan*+, Jiannong Cao+, and Guojun Wang*

*School of Information Science and Engineering, Central South University, Changsha, Hunan Province, China, 410083 +

Department of Computing, Hong Kong Polytechnic University, Kowloon, Hong Kong 410083

Abstract—Wireless Sensor Networks (WSNs) are being increasingly deployed for data-intensive structural health monitoring (SHM). The most critical problem for a structural health monitoring (SHM) system is sensor placement. In this work, we propose to deploy a set of backup sensors for SHM by finding locations around the possible network failure points, also known as critical points. The processes of finding critical points and locations for the backup sensors are performed through decentralized clusters. We found that the sensors in the clusters are weakly connected because civil engineering does not consider connectivity between sensors but ensures the quality of the sensor placement. The objective of our backup sensor placement (BSP) is to guarantee that the network will remain connected in the event of a single or multiple sensors failure during the SHM operation and thus prolong the network lifetime under the connectivity and data delivery constraints. We extensively evaluated our approach under a real system implementation and discussed its performance in the context of both WSN and SHM requirements. Experimental results indicated that the fault tolerance can effectively prolong the network lifetime.

I. I NTRODUCTION Wireless sensor networks (WSNs) have been widely used in many monitoring applications including battlefield surveillance, environmental monitoring and biological detection. Structural health monitoring (SHM) using WSNs have drawn a significant attention recently from both computer science and civil engineering researchers. Example includes Guangzhou New TV Tower (GNTVT) [1] and torre aquila [2]. From a civil engineering (CE) perspective for SHM, the quality of data about the health of a structure is affected by such optimal sensor placement; however, from computer science (CS) perspective for WSN, quality of sensor data is affected by sensor faults. The challenge with monitoring the health status of structures is that if a sensor fails at some moment at a location, data about physical changes (e.g., damage) from the location cannot be collected so that the changes cannot be understood accurately. It means that at some moment we are unable to know: is a bridge going to crash? It emphasizes that tackling sensor fault and failure at the optimal location should be an integral part of SHM applications, which is surprisingly ignored by CE. Probably one main reason is that CE still enjoys using wired sensors. However, this cannot be ignored by CS because wireless sensors are prone to failure. Thus, we consider a set of N primary sensors is already placed in a set of locations of a structure by using the most widely accepted sensor location optimization approach EFI(Effective independence) or EFI-DPR [1], [3] from CE.

EFI-based approach gives EFI values as location quality. Sensors are placed at locations with high EFI values. Our main concern is about the remaining set of locations after placement of the set of N sensors, which can also be effective for future use, e.g., a number of additional sensors that can be placed in a subset of the remaining/unused locations for consideration of sensor communication and failure of WSNs. The most stimulating aspect of presenting backup sensor placement (BSP) is the idea of placing backup sensors around the critical points, which is relatively simple yet reliable for identifying candidate locations. Critical points are the points where the sensors or communication link failures are highly possible, which are separable points, isolated points, and critical middle points. Therefore, in order to search highly possible critical points, we think of this searching in a decentralized manner and limit the searching to clusters.The searching algorithms identify all the critical points in the network and then place backup sensors around the critical points. After placement of all sensors, when the WSN runs for SHM, if a sensor fails, connectivity in the network becomes weak. The network connectivity should be improved in order to identify the structural health without interruption. The technique is to activate a placed backup sensor which is in sleeping state at the location of the failed sensor. An specialized hardware, for radio-triggered wakeup module, is used to wake the backup sensor. This results in a goal-directed approach for instant maintenance of the network. Our contribution in this paper is the following: 1) We are the first to design and implement an approach of placing a small set of backup sensors in unused locations of a structure by finding network failure points. 2) We guarantee that if a single or multiple sensors fail during SHM operation, our system has the ability to tolerate it and continue operation. 3) We evaluate our sensor placement approach extensively in in a real system implementation. We compare our placement approach with the state-of-the-art placement approaches. The rest paper is organized as follows: Section II describes our proposed approach. We perform a brief study of sensordata clustering algorithms in Section III. Proposed BSP algorithm is given in Section IV. Evaluation are described in Sections V. Finally, Section VI concludes the paper. II. P ROPOSED S ENSOR P LACEMENT FOR SHM An essential problem in SHM is the selection of sensor locations. There are many sensor placement techniques avail-

Rsr

BS Ci

Ci+1

Ci+2

Ci+3

Fig. 1. Finite element model (FEM) of a bridge structure, showing possible bridge elements and locations, and progressive clustering process

able in CS, such as effective independence (EFI) [1], [3]. EFI technique offers Fisher information matrix (FIM) determinant, which help to calculate EFI values. The EFI value is referred as placement quality indicator. The larger EFI value, the better the placement [1]. The EFI value is estimated by the mode shape. Mode shape is the vibration pattern at a specific frequency. A. Problem Formulation Consider a set P of N primary sensors, and a set B of R backup or redundant sensors to be placed for SHM. A structure F has a set of M feasible locations for sensor placement. N (< M ) primary sensors are placed by finding candidate locations out of M locations using EFI approach. The location of a base station (BS) is suitably located, may be far away from the sensing structure. For simplicity, we assume that a WSN with the placed N primary sensors P = {li = (xi , yi ) ∈ F, 1 ≤ i ≤ N } is weakly connected (i.e., k-connected, (k ≥ 1) and prone to failure, where sensor i is placed at li . Objective Function. The problem is to place a set B of R backup sensors into a network with N primary sensors by optimally finding locations out of (M − N ) remaining/unused locations such that: (i) the network is guaranteed to be fault tolerant to the failure of up to k − 1 sensors, and (ii) Netwrok lifetime (T ) is prolonged under constraints k-connectivity and data delivery. III. S ENSOR -DATA C LUSTERING Decentralized SHM system is developed basing upon the cluster based modal analysis and processing locally, i.e., the major decision can be made about the health of the structure at the cluster head. The present SHM system utilizes Imote2 hardware, TinyOS and damage detection algorithms [4]. We devise a location-based sensor-data clustering algorithm in a progressive manner on a bridge structure model to form a hierarchical network (see Fig. 1). When clustering process is done, the placement of backup sensor module starts. Before the network starts for operation, some parameters (e.g., sensor measurement directions according to the structure type (bridge, building or others) and data collection rate, damage-sensitivity indicator are forwarded to the cluster members (CMs) through the cluster head (CH). A CM transmits an intermediate result about mode shape. However, CMs also sends a final result if there is any damage around its location of the structure. Cluster as a Subgraph. The sensor fault tolerance has to be achieved by all independent CH and CM. We transform clusters into subgraphs, i.e., a cluster is considered as a subgraph G′ = (E ′ , V ′ ) of the network graph G = (V, E), where V is a set of placed sensors and E is the set of edges in

the network. We need to ensure that if one or more sensors fail, even k −1 sensors fail (or removal) in a subgraph, the network can still monitor the structure properly. However, each cluster in the network with the primary sensors weakly connected (k ≥ 1). This connectivity should be improved to monitor SHM for a prolonged period of time. IV. BSP: BACKUP S ENSOR P LACEMENT In this section, we give a BSP algorithm for SHM to ensure that the network is strongly connected with a adjustable transmission range and fault tolerant. The placement of backup sensors is performed through the subgraphs. BSP algorithm is relatively simple: finding locations to place the backup sensors and improving weakly k-connected to strongly k-connected subgraphs. First, BSP algorithm finds all the critical points (CPs for short) step by step and places backup sensors until finding all CPs or all backup sensors placed through another three algorithms, namely, BSP1, BSP2, and BSP3. All of these three algorithms call another algorithm Search-and-Place for finding locations by each CP. We will illustrate all the three algorithms in this paper but we will briefly describe them. In BSP algorithm, we have a set B of R backup sensors, k as connectivity requirement, all subgraphs of the network G as inputs. The output is the placement of R backup sensors and strongly k-connected network. BSP algorithm has three steps. In Step 1, BSP algorithm first calls the three algorithms, namely, BSP1, BSP2, BSP3, for finding critical points (CPs).When running any of the algorithms, if there is no CP in any G′ , the algorithm stops searching and goes to next algorithm. When the three algorithms are executed, Step 2 continues. We consider |B1 |, |B2 |, and |B3 | (|B1 |+|B2 |+|B3 | ≤ |B|) are the number of backup sensors to be placed during BSP1, BSP2, and BSP3 algorithms runtime, respectively. We consider Step 2 as an option and open. If there are still some backup sensors available to be placed, we can place them or save them. We assume that the network can be sufficiently connected after placing backup sensors. Step 2 checks all the primary sensor locations through each subgraph one after one and counts how many backup sensors at each location. If a location is with only one primary sensor (i.e., no backup sensor placed), we place a backup sensor at that location. In step 3, BSP algorithm maintain connectivity of the network, which starts with a G′ (the cluster is static in BSP) of network G. The purpose of this algorithm is to improve connectivity in order to achieve a fault tolerant netwotk. Firstly, the algorithm repeatedly adds connection in each subgraph in increasing order by communication range until G′ is strongly k-connected. Secondly, the algorithm repeatedly attempts to delete edges from G′ in decreasing order but putting the connection back if it is necessary for required kconnectivity. The resulting G′ is therefore a k-connected G′ of G. The value of k can be a fixed, such as k=3. For the case of k ≥ 2, the minimum weight k-connected subgraph is known to be NP-hard. Related k-connectivity theorems and proofs can be found a lot in literature [5].

radio-triggered wakeup & synchronization module The SHM Mote sensor board module Imote2

Fig. 2. Finding separable points as CPs. Half-black vertices are CPs. There is only one connection from v1 to v5 , v13 to v12 , and v9 is a connector of multiple neighbors.

input impedance Trigger signal

energy storage circuit receiving module output impedance

Fig. 5.

low-power amplifier Imote2 trans. module wireless transmitter

The SHM mote used in our experiment

D. Searching Locations and Placing Backup Sensors Fig. 3.

Finding critical middle points as CPs

A. BSP1: Finding Separable Points Assume that G′ be a weakly k-connected subgraph of G. One most likely asks a question: is G′ separable? That is, is there one or more separable points in G′ . A separable point or sensor of G′ is a CP which is only the connecting point of other several sensors and whose removal results in a disconnected G′ (see Fig. 2). We place a backup sensor at the separable points. B. BSP2: Finding Critical Middle Points A middle point between two sensors u and v of G′ is a CP, which is with the longest transmission distance and the link between u and v is vulnerable (see Fig. 3). In order to balance in transmission distance, we place a backup sensor in between u and v. The BSP2 finds such CPs in each G′ . The following two cases are considered in BSP2: (1) duv is largest. If v is a CH, data sent by sensor u to v may be lost or the receiver of v fails in the presence of environment interference. (2) There may have obstacles. C. BSP3: Finding Isolated Points When a sensor does not have a path or communication to another sensor in any G′ , or may receive broken messages, the sensor is an isolated sensor. We consider the sensor location as a CP. If there is only one potential sensor in a G′ , it would be automatically chosen as an isolated sensor and a CH. Our investigation on EFI-based sensor placement shows that a small part of sensors do not join in the cluster as shown in Fig. 4. Thus, we need to make good connection with the isolated sensors, we require placing backup sensors to support all isolated sensor locations. We assume that long distance single-hop transmission is not reliable.

Fig. 4.

Finding isolated points

In this subsection, we provide this algorithm called Searchand-Place, which is used by all the above three algorithms, namely, BSP1, BSP2, and BSP3 for searching the locations of CPs and placing backup sensors around the CPs. We have M − N feasible locations and a set B of R given backup sensors. The location of the primary sensors which can be the same or around the location coordinate of CP. The Searchand-Place checks whether there is available sensor location from M − N according to the EFI values. After placement of each backup sensor, the algorithm orders rest of the location of M −N . We need a suitable location around a CP. However, there may be several sensor locations around a CP. In that case, we prefer the location which is along with a CP. If there are several locations available along with a CP, then the location with the larger EFI value is a better choice. Then, the output can be locations near the CP. The number of locations found near the CP can be zero to many, which means that there is a chance of having more than one location or no available location. If a backup sensor is available but the location is not available near a CP, the algorithm places a backup sensor at the same sensor location which is a CP. V. S YSTEM I MPLEMENTATION The purpose of implementing this system is to validate the accuracy of our BSP approach including BSP algorithm with two objectives: 1) fault tolerance, and 2) prolongation of the network lifetime. In our system implementation, the wireless sensor nodes adopted are called SHM Mote as shown in Fig. 5. A multimetric and specialized SHM mote with onboard signal processing and local processing specifically planned for general SHM applications has been designed. Each SHM mote is mainly integrated by three main hardware components: a sensor board, an Imote2, and a radio-triggered wakeup together with synchronization module. The Imote2 [6] is with an external 32Mb non-volatile memory chip, an AM radio receiver for synchronized sensing, and a RF amplifier. The Imote2 is built around the low-power PXA271 XScale processor and 802.15.4-compliant radio hardware. A twelve-story shear frame structure, made from steel, is employed as shown in Figure 6(a). The lateral stiffness of each floor originates from the four vertical steel columns, 1.5 in. by 1.5 in. We instrumented the structure for the first time by

Fig. 7. Identified eight natural frequencies and network lifetime of different sensor amount (Failure free condition)

Fig. 6. Twelve-story test structure and the placement of 14 SHM motes are on the test structure

placing the wireless SHM motes according to the placement approach we described in this paper. We place the motes and form a WSN, verify our algorithms for several rounds in several days. There are at least 62 locations (=M ) in the structure. We have 14 SHM motes. We first place 10 SHM motes (=N ) according to the EFI described in Section 3. We then find 4 appropriate locations out of remaining locations (50 locations) for the 4 backup sensors (=R). To produce a sizable vibration response of the test structure, the only excitation source selected is an impact blow to the span from a modal hammer. We adjust the communication range by estimating height of the test structure and each floor. Under adjustable range, the k-connectivity between motes are maintained. We use a BS mote which is connected to a computer for the control purpose. The SHM motes run modified TinyOS and are configured to sample the accelerometers in a synchronized manner at a maximum frequency of 512Hz. Through the BS mote connected to the computer, two clusters are mainly constructed as shown in Fig. 6(c). Each SHM mote extracts identified the the frequency set [1.1223 Hz, 18.412 Hz] and sends back to the corresponding CH and the CH send back to the BS. A. Results Figure 7(a) illustrates the identified frequency sets in the two clusters, which indicates the output frequencies at different sensor locations (failure free condition). As expected, the network lifetime (as shown in Figure 7(b)) in BSP is much better than SPEM under different amount of sensors. Fig. 8(a) demonstrates the standard deviation in the first 6 iterations of the experiments. It can be seen that the standard deviation is abruptly ascent at 2nd and 4th iterations when we remove (i.e. switch off) the 3rd and 7th sensors, while remaining stable after each sensor removal (note that 0 (zero) is considered as a identified frequency at the location of a failed sensor). The results show that BSP is able to clearly overcome the situation of sensor removal. Furthermore, we observe from these experiments that the neighboring sensor data collection is slightly influenced by sensor removal. Fig. 8(b) clearly shows that the mode shape is affected by the sensor failure in SPEM, where there are no backup sensors available at the locations of the failed sensors. It indicates

Fig. 8.

Impact of the sensor removal on the network and SHM

that EFI-based approaches do not have the ability to tolerate a sensor failure. A closer look reveals that the identified mode shape is not completely distorted in SPEM, because neighboring sensors can still capture some amount of data from the location of a failed sensor. VI. C ONCLUSION A high quality WSN design can guarantee the structural health monitoring (SHM) accurately only, when sensors are placed at optimal locations by extensively satisfying SHM requirements and the sensor fault tolerance is considered. To the best of our knowledge, this was the first work of finding possible network failure points and placing backup sensors around the locations of the failure points. We implemented our approach on specialized SHM mote which includes the Intel Imote2 hardware platform and the TinyOS operating system. The evaluation results were presented which validated the effectiveness of our approach. ACKNOWLEDGMENT This work is supported by the National Natural Science Foundation of China under grant number 61073037 and the Ministry of Education Fund for Doctoral Disciplines in Higher Education under grant number 20110162110043. This work is also supported by Frace / HK Collaborative Research Grant F-HK 25/10T and HK PolyU Niche Area Fund 1-BB6C. R EFERENCES [1] B. Li, D. Wang, F. Wang, and Y. Q. Ni, High quality sensor placement for SHM systems: Refocusing on application demands. Proc. of. INFOCOM, 2010. [2] M. Ceriotti, L. Mottola, G. P. Picco, A. L. Murphy, S. Guna, M. Corra, M. Pozzi, D. Zonta, and P. Zanon, Monitoring heritage buildings with wireless sensor networks: The torre aquila deployment. Proc. of IPSN, 2009. [3] R. X. Gao, R. Yan, S. Sheng, and Z. Li, Sensor placement and signal processing for bearing condition monitoring. Springer Series in Advanced Manufacturing, 2006. [4] X. Liu, J. Cao, S. Lai, C. Yang, H. Wu, Y. Xu, ”Energy Efficient Clustering for WSN-based Structural Health Monitoring”, Proc. of INFOCOM, 2011 [5] L. J. Bredin, E. D. Demaine, M. T. Hajiaghayi, and D. RusDeploying, Sensor Networks With Guaranteed Fault Tolerance, IEEE/ACM Transaction aon Networking 18 (1): 216-230, 2010. [6] Crossbow Technology, Inc. Imote2 Hardware Reference Manual, 2007.