Computers and Electronics in Agriculture 81 (2012) 33–44 Contents lists available at SciVerse ScienceDirect Computers and Electronics in Agriculture journal homepage: www.elsevier.com/locate/compag Practical considerations for wireless sensor networks in cattle monitoring applications Kae Hsiang Kwong a, Tsung-Ta Wu a,⇑, Hock Guan Goh a, Konstantinos Sasloglou a, Bruce Stephen b, Ian Glover c, Chong Shen d, Wencai Du d, Craig Michie a, Ivan Andonovic a a Centre for Intelligent Dynamic Communications, Department of EEE, Royal College Building, 204 George Street, Glasgow, United Kingdom Institute for Energy and Environment, Department of EEE, Royal College Building, 204 George Street, Glasgow, United Kingdom Institute for Image and Signal Analysis, Department of EEE, Royal College Building, 204 George Street, Glasgow, United Kingdom d College of Information Science & Technology, No. 58, Renmin Avenue, Haikou, Hainan Province 570228, PR China b c a r t i c l e i n f o Article history: Received 20 August 2010 Received in revised form 16 October 2011 Accepted 22 October 2011 Keywords: Cattle monitoring Wireless sensor networks Data collection Delay tolerant networking Hardware and software design Real-time communication a b s t r a c t The paper presents an investigation into wireless sensor networks (WSNs) for cattle monitoring. The proposed solution fulfils the requirement for intensive condition monitoring of individual animals, aggregation and timely reporting of data to the farm manager. The core contribution of this study is a wireless communication solution designed for both loose house dairy cattle and free ranging beef cattle. The design target utilises inexpensive, low power consumption sensor nodes as the base elements of a data gathering and communication infrastructure. This platform facilitates real-time data download for loose housed dairy cattle and non real-time communication for free ranging beef cattle where the former is more challenging. In order to meet the target objectives, both the hardware and software are designed to adapt to the deployment challenges which include mobility, radio path interference, short transmission range of sensor nodes and limited resources in terms of energy and storage. These challenges have been analysed and addressed. Laboratory experiments and farm trials have been carried out to evaluate the performance of the platform communication protocol. The results of experiments demonstrate that the platform performs efficiently while conforming to the limitations associated with WSN implementations. Ó 2011 Elsevier B.V. All rights reserved. 1. Introduction The challenges faced by modern agriculture have never been greater. Increasing feedstock and labour prices, coupled with pressure from retailers to keep food prices low, are squeezing profit margins to a point where farmers cannot grow their business to a sustainable size. Recent high profile welfare threats such as the Bovine Spongiform Encephalopathy (BSE) and Food-and-Mouth Disease (FMD) outbreaks in the UK have further weakened the financial position of many in the industry. The application of novel technology based solutions such as Expert Systems and Machine Learning technologies, are showing real potential to deliver productivity benefits (Esslemont and Kossaibati, 2002; Firke et al., 2002). As the cost and availability of digital storage and communication decreases, the use of such technologies are becoming widespread. ⇑ Corresponding author. Tel.: +44 1415482082; fax: +44 1415524968. E-mail addresses: kwong@eee.strath.ac.uk (K.H. Kwong), twu@eee.strath.ac.uk, wgd8700@hotmail.com (T.-T. Wu), goh.guan@strath.ac.uk (H.G. Goh), ksasloglou@ eee.strath.ac.uk (K. Sasloglou), bruce.stephen@eee.strath.ac.uk (B. Stephen), ian. glover@eee.strath.ac.uk (I. Glover), cshen@hainu.edu.cn (C. Shen), wencai@hainu. edu.cn (W. Du), c.michie@eee.strath.ac.uk (C. Michie), i.andonovic@eee. strath.ac.uk (I. Andonovic). 0168-1699/$ - see front matter Ó 2011 Elsevier B.V. All rights reserved. doi:10.1016/j.compag.2011.10.013 It is now feasible to monitor and capture measurements of the condition of individual animals or processes in levels of detail that has hitherto been impossible. Mobile wireless sensors allow continuous, round the clock, welfare monitoring with greater robustness than would be practical using human observation alone. The application of Global Positioning System (GPS) technology is one such example; GPS enabled collars can capture the grazing habits of free ranging cattle (Hiroaki and Takaaki, 2007), allowing farmers to make better informed decisions on the efficient use of land. In order for sensory systems to be of value, data must be communicated back to the farmer. Although real-time communication presents several challenges – most notably battery conservation – many welfare threatening conditions require timely notification as a consequence of the rate at which deterioration can occur. Mayer et al., 2004, have demonstrated that real time data retrieval from animal mounted devices can be implemented via the Global System for Mobile (GSM) infrastructure. While interesting as a research tool, it is not a solution for practical applications. Apart from battery lifetime concerns, GSM enabled collars are prohibitively expensive when monitoring large numbers of animals i.e. the typical cost of a collar is approximately €1700 (Environmental Studies, 2009). 34 K.H. Kwong et al. / Computers and Electronics in Agriculture 81 (2012) 33–44 This paper proposes several solutions that implement real-time or opportunistic health monitoring using alternative, low cost and low power consumption wireless sensor nodes. In order to achieve this goal, an antenna diversity collar was designed to optimise the performance of radio coverage in typical farming environments. In addition, in contrast to the traditional store and forward mechanism, a routing protocol is presented to facilitate multi-hop connectivity. The protocol obviates the need to create and maintain an explicit routing path resulting in shorter packet delivery delay. To the best of our knowledge, no routing scheme is currently developed with the explicit view of supporting animal monitoring. An alternative approach utilises a portable collection device mounted on a vehicle or carried by a farmer to collect data when the data is needed e.g. during disease outbreaks. A router scheme has also been investigated that provides an alternate feasible solution to improving network performance. The remainder of the paper is organized as follows. Deployment challenges are outlined in Section 2. In Section 3, cattle mobility is described and related issues on wireless communication are discussed in Section 4. The influence of animal mobility is further analysed and presented in Section 5. Section 6 presents two opportunistic data collection schemes: the data collector and router schemes. Section 7 presents a real-time data download scheme. Section 8 presents conclusions. 2. Deployment challenges Wireless sensor nodes are characterised by limited resource in terms of energy, computational power, memory and low data transmission rate radio. One example, operating in the Industrial, Scientific and Medical (ISM) band is the MICAz node, powered by two alkaline AA batteries, comprising one 4 MHz processor with 512 kb of ROM and 4 kb of RAM (Crossbow Technology, 1995). These components are representative of the capability of low power wireless sensor devices. Given these limitations, the implementation of a cattle monitoring solution raises specific and severe challenges. 2.1. Mobility Mobility presents a major challenge for sensor nodes mounted on cattle since they are subject to frequent changes in location. The network topology and routing paths should therefore be dynamic, able to respond to frequent animal movement while optimising packet delivery. 2.2. Radio interference caused by the animals Dairy cattle are generally kept in herds. Since the body of the animal seriously attenuates radio signals, they can have a dramatic influence on the quality of radio propagation links (Gabriely et al., 1996; Dielectric Properties of Body Tissues: HTML clients, 2009). The hardware design should thus take into consideration this issue, addressed in Section 4.3. 2.3. Limitations in data storage The use of low cost sensor nodes by definition limits the overall capacity that the node has to store data. Thus, only a limited number of data packets can be buffered when a ‘store and forward’ approach is used to facilitate non real-time communication. To prevent excessive packet losses caused by buffer overflow, a robust protocol is required to efficiently forward the buffered data to the base station whenever communication is possible, discussed in Section 6. 3. Design consideration Domestic cattle breeds are essentially descended from prey species from which they inherit herding traits that benefits them from both a social and a general welfare perspective. Herds of animals do not uniformly distribute themselves and may not always move as a collective. Herding animals often have different social hierarchies within their number which influences the animals they are most likely to associate with and how they behave with their associates. As a consequence, the herd may break up into independent sub-herds each with their own dynamics. This raises an additional design issue for a WSN implementation: how rapidly are network topologies likely to change and what are the probabilities of nodes (animals) moving out of range of each other and of base stations? In order to anticipate the extent and rate of such changes, the behaviour of herds needs to be captured and modelled. In previous work, this has been attempted using collar mounted GPS transponders; for example Hiroaki and Takaaki, 2007 utilised GPS to record the behaviours of 14 free ranging Zebu cows in Western Niger. Samples of position were taken at 0.1 Hz rate so that displacement and rate of position change could be used to determine grazing coverage. The constraints within the farming environment in the UK are different and therefore a similar experiment was undertaken to determine whether or not meaningful data could be obtained that would facilitate the design of the monitoring platform. Two days of GPS data was recorded at 3 min intervals from a herd of 14 Limousin and Angus crosses equipped with collar mounted transponders, free-ranging on a farm in West Lothian, Scotland (Kwong et al., 2008). Days were selected at random from a set where the herd had good satellite coverage with no major interventions from farm staff. This data set was used to investigate two main issues that determine the viability of the WSN deployment: the range of distances between individual animals and a base station and the range of separations between animals. Rather than rely on a point estimate of the expected value of these quantities of interest, which would obscure skewed or multimodal data, the entire distribution was used. An accurate picture of a probability density function (pdf) can be obtained using a kernel density estimator such as a Parzen Window (Parzen, 1962). In kernel density estimates, the probability distribution of the distance of the herd from the base station (on 4th August) is shown in Fig. 1(a) and distance of the herd from the base station (on 8th August) is shown in Fig. 1(b). On the 4th August, the most likely distance from the base station is around 50 m with a second, smaller mode at around 90 m; on the 8th August, the herd is entirely concentrated around 80 m. With low power wireless sensors, the range from the base station can be an issue even in fields of the size of those on UK farms. Although animals have been found to stray to ranges of up to 300 m from a base station, the most likely distances lie between 50 and 90 m (Kwong et al., 2008). The density functions show the minimum inter-cow distance on the 4th August in Fig. 1(c) with the minimum inter-cow distance on the 8th August given in Fig. 1(d). In both cases the majority of observations are below 40 m with the maximally likely distance being at around 10 m. Therefore, WSNs must implement a multi-hop strategy for data collection. 4. Wireless communication issues In cattle monitoring systems, the animal is free to roam. Wireless technology is considered the only feasible method to establish and maintain communications between a base station and collars attached to cattle. Access to the majority of radio frequency bands 35 K.H. Kwong et al. / Computers and Electronics in Agriculture 81 (2012) 33–44 Fig. 1. (a) Probability distribution of the distance of herd from base station 4th August; (b) Probability distribution of the distance of herd from base station 8th August; (c) Minimum inter-cow distance 4th August and (d) Minimum inter-cow distance 8th August. is constrained by various standards and regulatory bodies e.g. the International Telecommunication Union (ITU-R) and/or OFCOM in the UK. Consequently, WSNs tend to use unlicensed bands and, in particular, the Industrial, Scientific and Medical (ISM) bands, originally reserved internationally for non-commercial use. The remainder of this section analyses the issues of radio signal penetration, base station antenna, collar antenna design, and the performance at ISM-bands for different data loads. Table 1 Tissue penetration depth (m) at major ISM band frequencies (Dielectric Properties of Body Tissues: HTML clients, 2009). Blood Bone cortical Fat Muscle Skin (dry) Skin (wet) 315 MHz 433 MHz 868 MHz 915 MHz 2.4 GHz 0.036515 0.23495 0.3225 0.55463 0.060904 0.063104 0.033685 0.20579 0.30421 0.051692 0.053641 0.056858 0.028143 0.13535 0.24793 0.042904 0.040842 0.044047 0.027682 0.12985 0.24234 0.0421 0.039951 0.043032 0.016407 0.046992 0.11956 0.022785 0.022956 0.022471 4.1. Signal penetration depth An estimate of signal penetration through an animal can be made using the electrical properties of body tissues (Gabriely et al., 1996; Dielectric Properties of Body Tissues: HTML clients, 2009). The properties of mammalian tissue are expected to be similar between species; therefore the above formulation represents a reasonable approximation to cattle. Table 1 summarises the penetration depth at major ISM frequency bands. Penetration at 2.4 GHz is less than 0.025 m in fleshy tissues (skins or muscle). Although the signal has a better penetration at 315 MHz, the width of a cow’s neck is approximately 0.25 m. It is unlikely, therefore, that using a collar mounted transmitter, that there will be sufficient penetration of radio signals through cattle to maintain reliable network connectivity. by frequency, transmitted power, antenna characteristics and the radio propagation channel. A single, line-of-sight, path between transmitter and receiver seldom exists in a real-world environment. Furthermore, in open environments, received signal strength may be very sensitive to the strength of the ground reflected propagation path. In the case of a strong ground reflection, receive and transmit, antenna heights (above ground level) have a large impact on received signal strength depending on whether interference between direct and reflected paths is constructive or destructive. A simple two-path model can be used to describe (and predict) this effect (Glover and Grant, 2003). Using this model the received signal power, Pr, at a large distance d is given by: 2 2 Pt Gt Gr ht hr 4.2. Base station antenna optimisation Pr ðdÞ ¼ Link quality between transmitter and receiver plays a major role in the performance of any radio network. The range is determined where Pt is transmitted power, Gt and Gr are transmit and receive antenna gains (as ratios, not in dB), ht and hr are the heights of 4 d L ð1Þ 36 K.H. Kwong et al. / Computers and Electronics in Agriculture 81 (2012) 33–44 transmit and receive antennas above ground, d is the distance between transmit and receive antennas and L accounts for any losses not represented by the two-path model. ‘Large’ distance refers to a receiver range from the transmitter greater than 4hthr/k, where k is wavelength. Pt, hr, Gt and Gr are determined by selection and configuration of the communications hardware whereas ht varies according to animal size. On a farm, the transmit antenna height is approximately 1.2 m (for an adult standing animal of average height). In principle, if the range is only slightly less than 4hthr/k the height for the base station (hr) antenna could be optimised (over the expected range of d providing this range is not too large) to ensure interference between direct and ground reflected signals is as close to constructive as possible. This optimum height is then given by: hr ¼ n kd 4ht ð2Þ where k is wavelength and n is an odd integer which would normally be chosen to be one. The antenna height requirements are shown in Fig. 2 as a function of position from the transmission point. Clearly, it is desirable to have a receiver antenna which is significantly greater than the transmit antenna height (1.2 m). If this is not the case, then despite the fact that the antenna will perform well in free space, the attenuation produced by the rest of the herd will significantly compromise system performance. To gain a clear view of the transmitting collar, the base station is best situated at a height of 2 m or more. Fig. 3 shows the received power as a function of a range of distances (40–50 m) at different base station height (0.1–10 m) for a pasture (er = 6 and r = 0.1). The difference between the optimum and minimum base station antenna height can reach 20 dB. This difference is higher for lower antenna heights. Fig. 4 shows the two dimensional projection of Fig. 3 to emphasize the dependence of received power at different base station heights, within the range of 40–50 m. When the base station antenna is at approximately 0.75 m, the received power is independent of the distance but this is not a practical solution since signals would be severely attenuated by cow body mass. A compromise condition is obtained when the base station antenna height is increased above 4 m since at this height, the received power becomes less sensitive to distance. 4.3. Collar antenna placement If a single antenna is attached to the collar, although an animal may be in-range of a base station, nevertheless the collar may not be able to relay information due to shadowing (by the animal itself) (Fig. 5(a)). Deploying two antennas to a collar provides limited spatial diversity. A diversity scheme was examined in which a pair of antennas (located at top left and top right of the collar as shown in Fig. 5(b)) are used to improve radio coverage. The locations on the collar represent a compromise. Locations on the side of the neck would minimise the effect of shadowing by the animal wearing the collar; however, it would also be susceptible to shadowing from other animals in the immediate vicinity. The locations chosen allow energy to propagate over the top of nearby animals. Fig. 5(c) shows the antenna location on the collar and Fig. 5(d) illustrates, schematically, the favoured signal propagation directions. The antennas used were ceramic patches (CABPB1240A) (TDK Corporation, 2009), small, low directivity devices providing a gain of about 2 dBi. The platform used to implement the communication link was based on MICAz; the RF switch (HMC197) (Hittite Microwave Corporation, 2009) selects one of the antennas under MICAz control. An experiment with diversity-equipped collars was carried out in an open environment. Fig. 6(a) and Fig. 6(b) show the Received Signal Strength Indication (RSSI) at the base station. Fig. 6(a) plots the reading of RSSI at the base station with the collar oriented such that one antenna points towards the base station whilst the other directly away (representing a cow standing side on). The signal from the antenna facing the base station (Antenna 1) is consistently higher by at least 10 dB. Fig. 6(c) shows the received packet rate performance with the collar in the same orientation. Again Antenna 1 outperform Antenna 2 by 10–20% consistently up to a range of more than 50 m. The experiment is repeated (Fig. 6(c and d)) for cases where each antenna radiates perpendicular to the direction of the base station, equivalent to the cow directly facing the base station. 4.4. Bandwidth, data load, and power consumption Battery life is a limited resource in any wireless platform, the target invariably being an operational life for as long as possible. For cattle monitoring, a collar is expected to operate up to at least Fig. 2. Optimum antenna height (n = 3, 5). K.H. Kwong et al. / Computers and Electronics in Agriculture 81 (2012) 33–44 37 Fig. 3. Received power along the distance with different base-station height. Fig. 4. Two dimensional projection of Fig. 3. five years without battery replacement. Commonly, battery conservation is achieved by reducing power consumption owing to radio transmission and implementing a low duty cycle (Ye and Heidenmann, 2004). This paper evaluated the impact of radio capacity on battery lifetime. The radio element is switched into low power mode as soon as radio transmission is complete, that time being dictated directly by radio channel capacity. Table 2 provides a battery lifetime comparison for a number of data download capacities. Nodes are assumed to transmit 100 kb, 1 Mb or 10 Mb of data to a base station on a daily basis. The radio is turned off as soon as data download is complete. 5. Analysis of animal mobility The animal monitoring system must be able to support animal mobility; wireless sensors are used to monitor the health condition 38 K.H. Kwong et al. / Computers and Electronics in Agriculture 81 (2012) 33–44 Fig. 5. (a) Radio shadow cast by animal; (b) enhanced collar design with two antennas; (c) antenna location on collar and (d) RF propagation from redesigned collar. Fig. 6. (a) Antenna 1 facing towards base station vs. Antenna 2 facing away from base station; (b) Antenna 1 facing to right perpendicular angle vs. Antenna 2 facing left perpendicular angle; (c) Antenna 1 facing towards base station vs. Antenna 2 facing away from base station and (d) Antenna 1 facing to right perpendicular angle vs. Antenna 2 facing left perpendicular angle. Table 2 Battery lifetime comparison between low and high radio capacities with different data loads (Analog, Embedded Processing, Semiconductor Company, Texas Instruments, 2009). Data load 100 kB Specification Radio chip Band Data rate (kbps) Current transmit (mA) Current stand by (mA) Radio cycle Current transmit (mA) Current stand by (mA) Total (mA) Battery capacity 1000 mAh 2000 mAh 5000 mAh CC1000 Sub 1 GHz 76.8 10.4 0.0002 Data load 1 MB CC1150 Sub 1 GHz 500 15.9 0.0002 CC2420 2.4 GHz 250 17.4 0.00002 4.514 0.720 5.234 1.060 0.720 1.780 2.320 0.072 2.392 6.369 12.738 31.844 18.727 37.453 93.634 13.935 27.871 69.677 CC1000 Sub 1 GHz 76.8 10.4 0.0002 Data load 10 MB CC1150 Sub 1 GHz 500 15.9 0.0002 CC2420 2.4 GHz 250 17.4 0.00002 45.139 0.719 45.858 10.600 0.720 11.320 0.727 1.454 3.634 2.945 5.889 14.723 of animals moving freely around open fields. Since the links between sensors and the data sink are sporadic due to animal movement, a farm trial was carried out on a farm in Lothian, Scotland to characterise animal mobility. The movements of 13 free ranging CC1000 Sub 1 GHz 76.8 10.4 0.0002 CC1150 Sub 1 GHz 500 15.9 0.0002 CC2420 2.4 GHz 250 17.4 0.00002 23.200 0.072 23.272 451.39 0.711 452.10 106.000 0.719 106.719 232.000 0.072 232.072 1.432 2.865 7.162 0.074 0.147 0.369 0.312 0.625 1.562 0.144 0.287 0.718 cows were recorded using GPS equipped collars over two days. The captured position data was then analysed from the perspective of network connectivity. The network connectivity in this context is defined as the ability a network to download data from individual K.H. Kwong et al. / Computers and Electronics in Agriculture 81 (2012) 33–44 animals in real-time, related to its performance in terms of packet download delay and buffer utilisation. The evaluation was based on the following two assumptions First, each cow was equipped with a node with a capability for data forwarding within a transmission range of around 30 m; second, a base station can collect data from each individual node within a transmission range of around 30 m. Fig. 7 shows the cows’ distributions. It is clear that herds do not distribute themselves uniformly and may not always move as a single collective. Instead, the herd may break up into independent sub-herds each with their own dynamics. Such behaviour leads the need for the network topology to change dynamically. Fig. 7 also shows that data transmission is established either when the cows move within radio range of a base station or is facilitated by multihop. To further quantify animal mobility, an analytical metric for network connectivity is proposed. Network connectivity, NCi is defined as the fraction of time during which the wireless node on cow i is able to communicate with a base station. Sources of interference such as noise and collisions are ignored. The connectivity over a typical day was simulated and 24 h GPS data from 13 cows was recorded. All cows were allowed to move unconstrained in a six hectare open field. The average network connectivity represents the real-time download capability of the network, obtained by: NC ¼ PT e PM i¼1 NC i t 1 ¼0 M ð3Þ where Te is the evaluation period and M is the number of mobile nodes. Fig. 8 shows the average network connectivity as a function of transmission coverage of a base station. The network approaches full connectivity at a range in excess of 100 m. Wireless sensor nodes are typically power constrained which translates into a practical coverage range of <30 m; the network connectivity is consequently in the order of 10% at that range. This low connectivity compromises network performance in terms of delay and buffer utilisation. 39 There are two routes to improving network connectivity. The first option is to increase the transmission power of the nodes; however, more energy resource is consumed on transmission and consequently, network lifetime will decrease significantly. Here an alternative strategy to enhancing radio converge is adopted through the deployment of additional access points or mobile collectors. The second option is multi-hop communication, only practical for large herds; in Section 7, a protocol that facilitates multi-hop routing is presented together with initial results obtained from experiments. 6. Data collection Although a number of optimisation approaches are proposed and proven to be effective in enhancing radio link quality, these solutions alone are not sufficient to mitigate the impact of mobility. The following Sections introduce two data transport schemes which facilitate data download within dynamic environments such as those encountered with animal movement supported by results from a farm trial for performance comparisons. 6.1. Opportunistic data collection A data collector, also known as a data mule, is a mobile/portable device that can be brought into the field for the soel purpose of download (Anastasi et al., 2007; Min et al., 2006; Shah et al., 2003). The scheme is useful when animals are beyond the transmission range of an access point. A data collector relays the information back to a data sink. The hardware structure of a data collector comprises processing, memory, radio transmission and power units. There is no power constrain on the collector and in order to cover a large area, a high gain antenna and large battery pack can be used. With these additional add-ons, the collector can continuously download the data from nearby nodes (Fig. 9). When the data on nodes cannot be forwarded to the base station, it will temporally be buffered into onboard flash memory Fig. 7. A snapshot of network connectivity during the farm trial. The figure shows the position of the GPS equipped herd in the field with the axes representing relative northings and eastings with respect to a base station, denoted by a black X. The labelled squares represent animal fixes at a given instant in time. Each fix is surrounded by a circle representing the wireless range which will indicate which animals are within communication range when these intersect with those of another animal. The heavy dashed line around certain animals represents a wireless connection via multi-hop back to the base station. 40 K.H. Kwong et al. / Computers and Electronics in Agriculture 81 (2012) 33–44 Fig. 8. Percentage of connected nodes from various ranges for two different types of connection. Fig. 9. Operation strategy for data collector scheme. in a FIFO (First In First Out) fashion. Periodically, the data will be rebroadcasted to search for the base station or relay nodes. However if the node is isolated for an extended period, the data will be dropped whenever the buffer becomes full. In a farm environment, a collector can be carried by a farmer or a trained dog. The communication protocol of the data collector can be divided into two phases: the pickup and dispatch phases. During pickup, the collector will handshake with nearby sensor nodes. Whenever the collector enters the transmission coverage of the node, it will download data from that node. If the packet is successfully received, it will be stored in the memory of the collector. An acknowledgement packet (ACK) will be sent back to the node. Once the node receives the ACK, it will remove the sent data from its memory. The dispatch process occurs between the collector and base station. During the period when the collector downloads data from nodes, it can broadcast its own buffered data periodically to a base station until its memory is empty. Otherwise, it defers data transmission for T sec and rebroadcasts the data. The detailed flow chart of collector scheme is provided in Fig. 10. 6.2. Router scheme Here several routers are deployed at fixed geographical locations to extend the transmission coverage of a base station, in so doing improving system performance. Since each router is situated on the ground, there are no constraints on the size of the hardware component. Therefore, a large battery pack can be deployed to K.H. Kwong et al. / Computers and Electronics in Agriculture 81 (2012) 33–44 extend its lifetime and a directional antenna to increase the transmission range of each router can be used (Fig. 11). The communication protocol of the router scheme also includes a pickup and a dispatch phase. During pickup, it operates exactly the same as with the collector scheme. When the fixed router captures data from a node, it will reply with the ACK packet to confirm that the packet has been successfully transmitted. Nodes will upload data until the buffer is empty. During dispatch, as a consequence of the pre-designed locations of routers, each will unicast data to either the relay router or base station. As soon as a router receives packets, it will forward them to the next hop and wait for an acknowledgement. Since the position of base station and routers are fixed and predefined, the radio link between them are robust. Therefore the need to defer T sec before data re-transmission is removed. The detailed flowchart for this scheme is provided in Fig. 12. 41 6.3. Experiment and results An 8-day period farm trial on a farm in Woodhouselee, Scotland was conducted to study the performance of the proposed schemes. Data was recorded using collars based on the the MICAz platform with GPS (Global Positioning System) attached on the neck of eight mature Limousin and Angus Crosses (Fig. 13). The location of each animal was recorded every 8 s. Fig. 14 shows typical location data over a period of 24 h. The field size was approximately 200 m  200 m with fencing dividing it into three parts. All three areas were accessible via gates. Three water troughs were placed within the field and the base station was deployed near one of them. The experiment investigated three different schemes. Initially the single base station scheme was utilised over a two days; data can only be uploaded when the collar falls within the range of Fig. 10. (a) Flowchart of pickup phase; (b) Flowchart of dispatch phase. Fig. 11. Operation strategy for router scheme. 42 K.H. Kwong et al. / Computers and Electronics in Agriculture 81 (2012) 33–44 Fig. 14. A 12 h trace of a cow and the layout of the farm. Fig. 12. Flow chart of dispatch phase for router scheme. Table 3 Buffer utilisation and average delay comparison between different communication schemes. Collar ID Pure base station scheme Data collector scheme Router scheme Buffer utilisation AVG Buffer delay (s) utilisation AVG delay (s) Buffer utilisation AVG delay (s) 709.334 262.733 198.611 276.748 357.314 360.511 463.457 851.583 5859.64 2233.48 1639.03 3201.1 2972.04 3024.16 4023.57 6946.94 594.266 108.49 141.865 209.79 301.836 315.605 606.487 615.355 4874.56 958.748 1192.24 1923.82 2529.5 2636.93 5290.03 5073.38 558.42 241.59 205.85 138.81 259.68 272.95 290.67 328.26 5137.37 2179.18 1714.93 1242.8 2128.62 2394.07 2586.88 2761.37 Average 435.036 3737.49 361.712 3059.9 287.03 2518.15 1 2 3 4 5 6 7 8 Fig. 13. Cattle wearing collars. the base station. Therefore, network connectivity is highly dependent on the movement of each individual cow. Subsequently the router approach was evaluated over a four days. Two routers were deployed in the middle of the pasture, separated at a distance of around 50 m. A 9 dBi directional antenna was installed to extend the transmission range. The routers collect data from nodes within their transmission range and relay them to the base station. The network performance of this scheme still depends on the animal distribution across the pasture, but the range for collecting data is extended. Finally, the data collector scheme was implemented over another two days. The collector was deployed four times every day (two times in the morning and two times in the afternoon). The route for the collector was planned and took approximately 10 min to access all remote cattle and download their data. Network connectivity is analysed in terms of buffer utilisation and average delay. The buffer utilisation is the number of packet stored in the buffer when no direct link exists between nodes and any access points. The value of average delay is the length of time from the packet generated at the node to the packet received at the base station. In Table 3, the values represent the average data downloaded over 24 h for each collar. Results show that the buffer utilisation in the single base station scheme is consistently at a high level of around 435. In contrast, the average value was re- duces to around 361 for the collector scheme and around 287 for the router scheme. The average delay for each packet follows the same trend, decreasing by 18.1% and 32.6% for the collector and router scheme, respectively. The network connectivity NCi metric introduced in Section 5 is defined as the fraction of time the node on a cow is able to communicate with a base station. In the field experiments, a timestamp is generated whenever the node falls within range of an access point. Table 4 shows the time period during which the cow was out of the range of base station over 24 h. The average time for the pure base station scheme was around 79,493 s, which means that for 22 out of 24 h, nodes had no connection to the base station. Thus the network connectivity (NCavg) is equal to 8% according. For the other two schemes, network connectivity is equal to 14% for the data collector and 38% for the router, respectively. The router scheme also outperforms the collector scheme in respect of buffer utilisation which decreases by 20.4% and the average delay which reduces to 17.7%. Note that the number of routers affects the performance of the network; as the number of scattered router increases, the radio coverage of a base station extends further. For the collector scheme, network connectivity is also influenced by how frequent the collector is deployed. A trade-off between system performance and the additional cost of installation and farm management exists. K.H. Kwong et al. / Computers and Electronics in Agriculture 81 (2012) 33–44 43 Table 4 Length of time when the cows beyond transmission range for different communication schemes. Collar ID Pure base station scheme Data collector scheme Router scheme 1 2 3 4 5 6 7 8 83,309 74,584 74,096 75,535 79,448 79,763 80,754 88,459 83,186 55,866 67,209 71,496 79,159 76,660 87,580 76,021 77,750 65,479 65,355 55,451 65,646 64,597 65,646 70,425 Average 79493.5 74647.125 66293.625 Fig. 16. Received packet rate. 7. Real-time collection In the section, an Implicit Routing Protocol (IRP) is defined and analysed for cattle monitoring systems. The proposed IRP operates according to the following two phases: the configuration and the data forwarding phase. During configuration, the base station periodically floods a TIER message throughout the network. This TIER message contains a base station’s ID field, and a hop count field, numbered one, two, three, . . . x, starting from the base station. On receipt of a TIER message, each node records the received hop count field and increases the hop count by one before forwarding the TIER message. The hop count field is used to track the number of hops the TIER message has traversed from the base station e.g. a node with a hop count field, h, represents the h-th tier away from the base station. This critical field is also referred as the TIER ID. As animals move, the base station is required to send TIER messages periodically at intervals of T sec to maintain the correct configuration. At the data forwarding phase, if the collar has data to download to the base station, it will form a packet comprising its current TIER ID and measurement data. This packet is then broadcast. Only receiving collars with lower TIER IDs are required to respond with an ACK packet. The collar that sends out the first ACK packet will notify the remaining collars to discard their ACK packets. After acknowledging the source collar, the collar will broadcast the packet. Receiving collars with equal or higher TIER IDs will discard the received data immediately. This forwarding rule repeats until the data arrives at the base station. Data moves one hop closer to the base station at each forwarding stage. Experiments were conducted to study the performance of IRP. The IRP was implemented on the MICAz node using TinyOS (TinyOS Community Forum, 2000); the test-bed is a three hop network with one source node, one base station and N pairs of intermediate relay nodes. Fig. 15 depicts the average packet delay and Fig. 16 the received packet rate of the test bed configuration N = 4, and in each tier there are four relay nodes. During each experiment, the source node generates 10,000 packets at an interval of 250 ms, each packet containing 85 bytes of payload. In order to simulate movement, an asynchronous random ‘‘on/off’’ mechanism was implemented. A sensor node in ‘‘off’’ mode represents cow movement out of communication range; and when a sensor node is switched to ‘‘on’’ mode, it represents cows entering the range. This ‘‘on/off’’ mechanism is characterised by an ‘‘off’’ probability Poff viz. probability that the sensor node’s stay out of range. Figs. 15 and 16 show that network performance is severely impacted with increasing Poff; the performance improves as the number of sensor nodes in each tier increases. In farm environments, the performance of IRP is impacted by animal movement, e.g. how frequent the cow changes its physical location and how many neighbouring cows remain close. Experimental results imply that the proposed IRP is suitable for cattle monitoring specifically for loose housing of dairy cattle. Since dairy cows spend about 40–50% of the day resting (Christian et al., 1997) they do not move frequently. Over winter, dairy cows are commonly reared in pens, ideal for the IRP routing scheme. This scheme is most effective for farms stocking a large number of cows. The density of animals governs the probability of establishing viable links between cows and in turn a robust path, allowing data from a cow to be successfully relayed to the base station in a hop-by-hop manner. From experimental results, the protocol has been shown to be self-forming and resilient to changes of topology. 8. Conclusions This paper has presented an animal monitoring platform using wireless sensor networks addressing the challenges in managing livestock. Challenges in adapting wireless sensor networks to cattle monitoring stem from animal movements compromising node download capability. An analysis of the distribution of a herd comprising 14 Limousin and Angus crosses in a working farm was recorded to evaluate the feasibility of maintaining a connection between individual animals and a base station. Using that knowledge of herd mobility as the foundation, three tailored networking schemes are proposed facilitating opportunistic or real-time data download. These schemes enable up-to-date individual animal status to be fed back to farmers improving animal welfare and operational efficiency via more informed decision-making. References Fig. 15. Average packet delay. Anastasi, G., Conti, M., Monaldi, E., Passarella, A., 2007. An adaptive data-transfer protocol for sensor networks with data mules. In: Proceedings of the IEEE Symposium on Computers and Communications (ISCC 2008), Marrakech, Morocco, July. 44 K.H. Kwong et al. / Computers and Electronics in Agriculture 81 (2012) 33–44 Christian, D., Veermac, I., Praks, J., Poikalainen, V., 1997. Behaviour of dairy cows in the waiting area of large uninsulated cowsheds. In: Proceedings of Ninth International Congress on Animal Hygiene, Helsinki, Finland. Crossbow Technology., 1995. Available at: <http://www.xbow.com> (accessed 10.12.09). Dielectric Properties of Body Tissues: Html Clients. Available at: <http:// niremf.ifac.cnr.it/tissprop/htmlclie/htmlclie.htm> (accessed 01.09.09). Environmental studies. Available at: <http://www.environmental-studies.de/> (accessed 01.09.09). Esslemont, R.J., Kossaibati, M., 2002. The costs of poor fertility and disease in UK dairy herds (Trends in DAISY herds over 10 seasons). DAISY Research Report No. 5. Intervet. Firke, R., Stamer, E., Junge, W., Krieter, J., 2002. Automation of oestrus detection in dairy cows: a review. Livestock Production Science 72, 219–232. Gabriely, S., Lau, R.W., Gabriel, C., 1996. The dielectric properties of biological tissues: III. Parametric models for the dielectric spectrum of tissues. Physics in Medicine and Biology 41, 2271–2293. Glover, I., Grant, P., 2003. Digital Communications, second ed. Prentice Hall. Hiroaki, M., Takaaki, N., 2007. Grazing behavior of cows measured by handheld GPS and bite counter collar: a case of Fazenda Baia Bonita in South Pantanal, Brazil. Japanese Journal of Human Geography 59 (1), 30–42. Hittite Microwave Corporation. Available at: <http://www.hittite.com/> (accessed 01.09.09). Kwong, K.H., Goh, H.G., Michie, C., Andonovic, I., Stephen, B., Mottram, T., Ross, D., 2008. Wireless sensor networks for beef and dairy herd management. The 2008 American Society of Agricultural and Biological Engineers Annual International Meeting, Providence, Rhode Island, United States. Mayer, K., Ellis, K., Taylor, K., 2004. Cattle health monitoring using wireless sensor networks. In: Proc. of the Second IASTED International Conference on Communication and Computer Networks, Cambridge, Massachusetts, USA. Min, C., Taekyoung, K., Yong, Y., Victor, C.M.L., 2006. Mobile agent based wireless sensor networks. Journal of Computers 1 (1), 14–21. Parzen, E., 1962. On estimation of a probability density function and mode. Annals of Mathematical Statistics 33 (3), 1065–1076. Shah, R.C., Roy, S., Jain, S., Brunette, W., 2003. Data MULEs: modeling a threetier architecture for sparse sensor networks. In: Proc. of the First IEEE International Workshop on Sensor Network Protocols and Applications, pp. 30–41. TDK Corporation. Available at: <http://www.tdk.com/> (accessed 01.09.09). TinyOS Community Forum, 2000. An open-source OS for the networked sensor regime. Available at: <http://www.tinyos.net/>. (accessed 01.09.09). Ye, W., Heidenmann, J., 2004. Medium access control with coordinated adaptive sleeping for wireless sensor networks. IEEE/ACM Transactions on Networking 12 (3), 493–506.