Automated oestrus detection method for group housed sows using acceleration measurements

Transcription

1 Auomaed oesrus deecion mehod for group housed sows using acceleraion measuremens C. Cornou and T. Heiskanen Deparmen of Large Animal Sciences, Faculy of Life Sciences, Universiy of Copenhagen, Groennegaardsvej, DK-1870 Frederiksberg C. Copenhagen, Denmark Absrac A pilo experimen was carried ou o define a mehod for oesrus deecion for grouphoused sows based on sows' acceleraion measuremens. Acceleraion daa was measured for five sows, in hree dimensions, four imes per second during weny days. The oesrus deecion mehod is based on hourly averages of acceleraion measuremens, modeled using a univariae Dynamic Linear Model combined wih a Kalman Filer. Alarms are produced when he observaions deviae oo widely from he model's esimaed level. The same deviaion paerns are observed for wo sows ou of hree sows ha came ino oesrus. The daa collecion and he deecion mehod are discussed and perspecives for improving he deecion model while scaling up his pilo experimen are presened. Keywords: oesrus deecion, dynamic linear model, conrol chars, acceleraion, group housed sows. Inroducion Group housing of sows resuls in difficulies idenifying and accessing he individual sow. The developmen of auomaed mehods for oesrus deecion aims o improve he deecion of he onse of oesrus and o faciliae he "managemen by excepion", by focusing he farmer's aenion o he sows o be checked for oesrus. Mehods for auomaed oesrus deecion in group-housed sows are reviewed in Cornou (006). These mehods can be based on sow's aciviy measuremens, using acceleromeers, infrared sensors, sow's aciviy near he boar pen and body emperaure measuremens. For hese mehods, variable analysis is mosly analysis of variance and Generalized mixed models. Oher mehods are found wihin deecion models for oesrus and diseases implemened for dairy cale. Examples of advanced ime series models are Muli Process Kalman Filer (Thysen, 1993) and mulivariae models (de Mol e al., 1999). In Denmark, he projec Hoghrob 1 aims o develop sensor nework echnology and an applicaion allowing farmers o rack changes in aciviy of group housed sows prior o oesrus o mae hem a an opimal ime. This paper presens he resuls of he pilo experimen of he projec Hoghrob, aimed o assess wheher i is possible o deec he onse of oesrus using acceleraion daa from five group-housed sows. The deecion mehod presened in his paper includes a univariae Dynamic Linear Model combined wih a Kalman Filer and wo ypes of conrol char. 1

2 Maerials and mehods Daa was colleced in a producion herd over a four weeks period (February 1 s - March 1 s 005). Sows, Housing and Feeding Five sows paricipaed in he experimen. They were all in heir hird or fourh pariy, had no leg disorder and had reproducion cycles in prior pariy of days. A weaning (day 0) and during he following week, he sows were housed in individual pens, in a naurally venilaed and non heaed maing secion and fed ad libium. The sows were esed for oesrus (weaning oesrus) bu no inseminaed. The es aimed o assure ha he sows acually came ino hea and o predic when o begin deecion of he second oesrus (experimenal oesrus). A day 8 (February 5 h ), he five sows were moved o he gesaion uni and inroduced ino a group of 100 oher sows. The group had access o wo feeding saions and wo nipple drinkers. Oesrus Deecion The Back Pressure Tes (BPT) (Willemse and Boender, 1966) was used o idenify he ime of firs and las sanding response as a sign of hea a boh weaning oesrus (BPTw) and experimenal oesrus (BPTe). The BPTw was performed hree imes a day (a 07:00, 14:00 and 1:00) from day 3 a 07:00 unil 4 hours afer he sow ceased o show sanding response a BPT. The BPTe was performed hree imes a day (a 07:00, 14:00 and 1:00) from day 1 a 07:00 unil 4 hours afer he sow ceased o show sanding oesrus by BPT. The es was performed nex o a easer boar. This deecion was performed o idenify he exac onse and end of oesrus, and consequenly assess wheher he acceleraion daa showed specific changes in acceleraion paerns a he onse of oesrus as compared o he anoesrus period. Aciviy measuremens In he experimenal pen, sows' aciviy was measured using wo ypes of acceleromeers: an analog acceleromeer (ADXL30 from Analog Devices), measuring he acceleraion in dimensions and a digial acceleromeer (LIS3L0DS from STMicroelecronics) measuring he acceleraion in 3 dimensions. The acceleromeers were fied on a neck collar and fixed on each experimenal sow. Daa was ransferred o wo PCs via an exernal Blueooh dongle hanging down from he ceiling in he middle of he pen. Sows' aciviy was recorded four imes per second, 4 hours a day, from day 11 o day 30. The analog acceleromeers were no calibraed, herefore only he resuls of he digial acceleromeer, which gave reliable measuremens, are presened. Video recordings The sows were video recorded 4 hours a day using four cameras covering mos of he pen from day 10 o day 30 (four picures per sec). The video recordings primarily aimed o assure ha he neck collars - where he acceleromeers were fied - did no fall off during

3 he experimen. Sows were individually marked on he back in order o idenify hem on video and arificial ligh was on 4 hours a day. Daa Colleced A calibraion problem wih one digial acceleromeer resuled in non valid daa. Therefore only resuls colleced for four sows (sow 1, 3, 4 and 5) are presened. Oesrus deecion Onse of oesrus is defined as 3.5 o 5 hours before he firs deecion of he sanding reflex, i.e. half way before he previous es; he end of oesrus period is defined as 3.5 o 5 hours afer he las deecion of he sanding reflex. Sow 1, 3 and 5 came ino oesrus respecively March he 10 h, 16 h and 14 h, corresponding o he inerval weaning oesrus - experimenal oesrus of 15.4, 18.4 and 19.4 days. The sow 4 did no show oesrus wihin his period bu came ino oesrus a week afer he end of he experimen ime. Aciviy measuremens The iniial ime series of measuremens (in vols) provided daa for he hree axes (x, y and z), which were convered o acceleraion unis (g). Lasly, he lengh of he vecor acceleraion was calculaed as x y z acc = acc + acc + acc (1) where acc x, acc y and acc z are he acceleraion values for he axes x, y and z. The ime series of he vecor lengh of acceleraion (acc) were analyzed from he 01/03/05 a 00:00 o he 1/03/05 a 00:00. The percenage of daa available was 53.34%, 6.43%, 60.57% and 71.3% respecively for sow 1, 3, 4 and 5. Transmission failures were he main reason for missing values: when lying, he sows covered mos of he ime he acceleromeer, which ended o slide under he neck. Periods corresponding o loss or change of he neck collars were deleed from he ime series and reaed as missing values (30 hours in oal, included 5 hours for sow 1). For all sows, he values of he 3 axis (x, y and z) ranged from - o and he values of he lengh of he vecor acceleraion acc range from o 0 o 3.46; 1 g corresponds o no acceleraion. Deecion Model A univariae Dynamic Linear Model (DLM) (Wes and Harrison, 1997) is implemened on he hourly averages of acceleraion daa (Y ). The general form of he DLM allows each sow o have her own model, characerisic of ha individual sow. The model's forecas errors are analyzed using wo ypes of conrol char. The DLM and he conrol chars are implemened in he sofware R (R Developmen Core Team, 005).

4 Dynamic Linear Model The general DLM is represened as a se of wo equaions (Wes & Harrison, 1997). The observaion equaion () defines he sampling disribuion for he observaion Y condiional on a sae vecor θ. θ consiss of he parameer µ describing he model level a ime. The sysem equaion (3) defines he ime evoluion of he sae vecor θ. F is a design marix and G is he sysem marix (defined as ideniy marix). The error sequences υ and ω are assumed inernally and muually independen. Observaion equaion Y = F' θ +ν, ν ~ N(0, V ), () Sysem equaion θ = G θ + ω, ω ~ N(0, W ), (3) 1 The DLM combined wih a Kalman Filer (KF) (Kalman, 1960) esimaes he underlying mean level (F' θ ) of he acceleraion daa a ime. The updaing equaions of he KF are shown in Equaions (4-11). The heory of he KF has been previously described e.g. in Thysen (1993) or de Mol e al. (1999). Prior mean a = G m (4) Prior Variance -1 R = G C G' + W (5) -1 One sep forecas mean f = F' a (6) One sep forecas variance F' R F S-1/n Q = + (7) Filered mean m = a + A e (8) -1 Adapive coefficien marix A = R F Q (9) One sep forecas error e = Y - f (10) Filered variance C = S / S-1 (R - AQA' ) (11) Wih = n 1 n -1 + S S-1 e = S n Q The model is iniialized by means of Reference analysis. I uses he firs observaions of he ime series o esimae he poserior disribuions afer a firs period. The evoluion variance W, is esimaed using a discoun facor δ. For each sep, he sysem variance W is defined as a fixed proporion of he model variance C. Missing values are reaed by he model, such as he poserior disribuions are equal o he priors, so he forecas errors e equal zero. Finally, he model observaions (hourly averages in our case) are weighed by he number of observaions per hour so ha he uncerainy Q becomes very large when he number of observaions is small, and reciprocally. The observaional variance (V) depends of he accuracy of he acceleromeers; V is assumed consan over ime. In normal condiions, i.e. anoesrus, we assume ha he sow does no change is behaviour: he rue underlying level is also assumed consan over ime, by seing he sysem variance (W ) o zero; his specific DLM refers o a saic model, wih consan parameers defined by uni discoun facor.

5 Figure 1 (a) shows he hourly averages of he acceleraion daa (Y ) and he filered mean (m ) esimaed by he KF for sow 5, afer adapaion of he Kalman Filer (he Adapive coefficien converges around March 8 h ) acc, in g Hourly average of acceleraion (dashed) and underlying level (plain) cusum Cusum of sandard. errors cusum Vmask (A=1, dela=1.5, alpha=0.01, bea=0.01) cusum Tabular cusum (H=1, K=1.5) (d) Time, in Days (a) (b) (c) Figure 1. Deecion model for sow 5: filered mean (underlying level) and hourly observaions (a), cumulaive sum of he sandardized forecas errors (b), applicaion of V- mask (c) and abular cusum (d); period from March 8 h o 0 h. Analysis of he forecas errors In normal condiion, i.e. anoesrus, he observaions are a random walk around he mean level of he model. Wide deviaions beween he observaions and he underlying level esimaed by he model are used as signal for he onse of oesrus. The model deviaions are esimaed by mean of he one sep forecas errors e, sandardized wih respec o heir scale parameers, i.e. he one sep forecas variance Q. The model's deviaions are aferwards moniored by means of a V-mask and a Tabular cusum conrol char (Mongomery, 1997).

6 Figure 1 (b) shows a cumulaive sum (cusum) of he sandardized errors (u ) for sow 5. From March 14 h (onse of oesrus), he cusum of he sandardized errors ends o increase during hree days, which indicaes ha he observaions end o be consanly greaer han he one sep forecass (f ). Figure 1 (c) illusraes he applicaion of a V-mask (Madsen e al., 005; Mongomery, 1997) on he cusum of he sandardized errors. The onse of oesrus measured by manual oesrus deecion, our "golden sandard", is indicaed by a black dash line, he 14h of March. Using he parameers A=1, dela=1.5, alpha=0.01 and bea=0.01 (Mongomery, 1997) wo alarms are produced wo o hree days before oesrus, and he V-mask caches he onse of oesrus. An alarm is given less han 4 hours afer he onse of oesrus, and anoher on he 18 h. The cusum is rese o zero afer each alarm. Figure 1 (d) shows he model's deviaions moniored using a abular cusum of he sandardized errors u. Alarms given by he abular cusum are indicaed by eiher a grey verical line (negaive alarm) or a black verical line (posiive alarm). Using he parameers H=1 and K=1.5 (equivalen o he V-mask parameers), a posiive alarm given by he abular cusum caches he increase of aciviy a he onse of oesrus. Three false alarms are produced he 10 h (posiive), 11 h and 0 h (negaive). The plo of he cusum of he sandardized errors (b) shows ha hese alarms respecively cach an increase (10 h ) and hen, decrease of aciviy (11 h ). A paern characerisic of he sow aciviy around oesrus, i.e. a decrease of aciviy around wo days before oesrus, followed by an increase of aciviy a he onse of oesrus, is observed boh for sow 1 and 5. Accuracy of he deecion model The oucome of an oesrus deecion es (in hea or no in hea) can be defined in erms of sensiiviy, specificiy and Error Rae (Cornou, 006). Due o he limied number of sows available, he accuracy of he mehod is esimaed by dividing he ime series ino blocks: during he anoesrus period, he "anoesrus blocks" are of 4 hours, whereas he "oesrus block" is defined as ±4 hours he onse of oesrus. Each of he defined blocks is checked for alarms. In he "anoesrus blocks", an alarm indicaes a FP and no alarm a TN; in he "oesrus block", an alarm indicaes a TP and no alarm a FP. To limi he repeiion of alarms, if an alarm is produced in a block, he nex 6 hours are checked. If a new alarm is produced, he block is exended by 6 hours and includes a single alarm. Table 1 shows he resuls in erms of Specificiy and Error Rae, boh for he Tabular Cusum and he V-mask. Table 1. Resuls of he deecion model in erms of True Posiive (TP), False Posiive (FP), Specificiy (SPEC) and Error Rae (ERR) for he five experimenal sows. Tabular cusum V-mask Sow TP FP SPEC ERR TP FP SPEC ERR The bes resuls were in sow 5, boh using he Tabular Cusum and he V-mask. Neiher of he deecion mehods allows caching he onse of oesrus for sow 3. Three and four alarms

7 are produced for sow 4. The false alarms for sow 4 may be parly explained by he influence of he oher sows coming ino oesrus and being more acive, reinforced by he fac ha he sows were always esed wih he BPT a he same ime. Moreover, he low percenage of daa available may have hindered he aemp o see a specific paern a he onse of oesrus. From he 8 h o he 0 h of March, he percenage of raw daa available is 36.7%, 37.% 9.3% and 49.5% respecively for sow 1, 3, 4 and 5. Discussion The Kalman filer presens major ineress o model sow's aciviy daa. As de Mol e al. (1999) sugges for dairy cows going o pasure, he model slowly adaps so ha new environmen condiions should no resul in alarms. This slow adapaion o seasonal variaion may also be an advanage for group-housed sows, where he aciviy in summer is expeced o be lower in case of high emperaure. Anoher ineresing feaure of he Kalman Filer is he possibiliy of implemening inervenions; a emporary smaller discoun facor helps he model o adap quicker o he change ha may have been caused by exernal facors, and hen limi he number of false alarms due o hese exernal effecs, like for insance, he inroducion of new sows in a large group. The abular cusum presens an ineres in visualizing upwards or downwards drifs: for wo sows ou of hree, he onse of oesrus was characerized by a downwards drif wo days before oesrus, followed by an upwards drif a he onse of oesrus; his feaure can be advanageous for idenifying false oesrus alarms. False oesrus alarms should however no be enirely discarded. A decrease of aciviy may be due, for insance, o illness or lameness, and such alarms should be used by he farmer o idenify sows ha should be given specific aenion. Besides, a major drawback of ransmission failures, which resuled in a large amoun of missing values, may have been a lack of low values of acceleraion daa (when a sow was sleeping). We can suppose ha his hindered a reducion of he general aciviy level, since daa were ransmied mosly during he periods of aciviy. Improvemens in daa ransmission should be considered while scaling up he mehod in a furher experimen. Regarding as a furher developmen of he model, anoher ineresing feaure of he Kalman Filer is he possibiliy of combining oher variables in a mulivariae Kalman Filer. Poenial variables are he sow's visi a he feeding saion and he aciviy near he boar pen, measured for insance by a icke window (Bressers e al., 1991). Anoher alernaive for deecing a sudden shif is a Muli Process Kalman Filer, as developed in Thysen (1993), where i is assumed ha a ime series follows an unknown combinaion of specific saes (e.g. normal evoluion, ouliers and level shif). Such mehod could be applied o hese ime series in order o disinguish beween aciviy/no aciviy or beween differen ypes of aciviies, i.e. walking, sleeping ec. For his purpose, he video recordings should cover he enire pen (which was no he case in his experimen), in order o be able o associae an aciviy paern o a specific sow behaviour. Lasly, i is suggesed ha combining informaion abou he previous oesrus cases would improve he deecion mehod's accuracy. The number of false alarms, which is a major drawback while increasing he accuracy of auomaic oesrus deecion mehods (Cornou, 006) is expeced o be reduced if he likelihood of a sow coming ino oesrus a a cerain ime is very low.

8 Conclusions This pilo experimen helped in defining an oesrus deecion mehod for group-housed sows, based on a univariae DLM and wo ypes of conrol char. This mehod allowed deecing he onse of oesrus for wo ou of he hree sows ha came ino oesrus and should be applied in a furher experimen wih a greaer number of sows. Perspecives for developmen of his mehod include he implemenaion of a Muli Process Kalman Filer or a mulivariae approach. Acknowledgemens Thanks o he Deparmen of Compuer Science from Copenhagen Universiy (DIKU) for he developmen of he acceleromeers and heir help o collec he daa along he experimen. Funding was provided by he Danish Research Agency. References Bressers, H. P. M., J. H. A. Te-Brake, and J. P. T. M. Noordhuizen Oesrus deecion in group-housed sows by analysis of daa on visis o he boar. Applied Animal Behaviour Science 31: Cornou, C Auomaed oesrus deecion mehods in group housed sows: Review of he curren mehods and perspecives for developmen. Livesock Science 100:1-11. de Mol, R. M., A. Keen, G. H. Kroeze, and J. M. F. H. Achen Descripion of a deecion model for oesrus and diseases in dairy cale based on ime series analysis combined wih a Kalman filer. Compuers and Elecronics in Agriculure : Kalman, R A new approach o linear filering and predicion problems. Journal of Basic Engineering 8: Madsen, T. N., S. Andersen, and A. R. Krisensen Modelling he drinking paerns of young pigs using a sae space model. Compuers and Elecronics in Agriculure 48: Mongomery, D Inroducion o Saisical Qualiy Conrol. 3rd ed. Wiley J. and Sons, USA. R Developmen Core Team R: A language and environmen for saisical compuing. Thysen, I Monioring Bulk Tank Somaic Cell Couns by a Muli-Process Kalman Filer. Aca Veerinaria Scandinavica Secion A. Animal Science: Wes, M. and J. Harrison Bayesian Forecasing and Dynamic Models. nd ed. Springer, New York, USA. Willemse, A. H. and J. Boender A quaniaive and qualiaive analysis of oesrus in gils. Tijdschr. Diergeneesk. 91: