Vessel Recognition in Induction Heating Appliances - A Deep-Learning Approach

Recei ed Decembe 8, 2020, accep ed Janua y 10, 2021, da e o publica ion Janua y 19, 2021, da e o cu en e sion Janua y 28, 2021. Digi al Objec Iden i ie 10.1109/ACCESS.2021.3052864 Vessel Recogni ion in Induc ion Hea ing Appliances—A Deep-Lea ning App oach JORGE VILLA 1, (Membe , IEEE), DENIS NAVARRO 1, ALBERTO DOMINGUEZ 2, JOSE I. ARTIGAS 1, AND LUIS A. BARRAGAN 1 1Depa men o Elec onic Enginee ing and Communica ions, I3A, Uni e si y o Za agoza, 50018 Za agoza, Spain 2Bosch and Siemens Home Appliances G oup, Depa men o Induc ion Technology o BSH Elec odomes icos, 50016 Za agoza, Spain Co esponding au ho : Jo ge Villa (j illal@uniza .es) This wo k was suppo ed in pa by he Spanish MICINN unde P ojec PID2019-103939RB-I00, in pa by he Spanish MICINN and AEI unde P ojec RTC-2017-5965-6 and G an PTQ-17-09045, in pa by he EU h ough FEDER P og am, in pa by he DGA-FSE, and in pa by he BSH Home Appliances G oup. ABSTRACT The selec ion o a essel by an induc ion-hob use has a signi ican impac on he pe o mance o he appliance. Due o he induc ion hea ing physical phenomena, he e exis many ac o s ha modi y he equi alen impedance o induc ion hobs and, consequen ly, he ope a ional condi ions o he in e e . In pa icula , he ype o essel, which is a sole decision o he use , s ongly a ec s hese pa ame e s. Besides, he e omagne ic p ope ies o he di e en ma e ials he essels a e made wi h, a y di e en ly wi h he exci a ion le el, and gi en ha mos o he domes ic induc ion hobs a e based on an ac-bus ol age a angemen , he exci a ion le el con inuously a ies. The algo i hm p oposed in his wo k akes ad an age o his ac o iden i y he equi alen impedance o he load and ecognize he po . This is accomplished h ough a phase-sensi i e de ec o ha was al eady p oposed in he li e a u e and he applica ion o deep lea ning. Di e en con olu ional neu al ne wo ks a e es ed on an augmen ed expe imen al-based da ase and he p oposed algo i hm is implemen ed in an expe imen al p o o ype wi h a sys em-on-chip. The p oposed implemen a ion is p esen ed as an e ec i e and accu a e me hod o cha ac e ize and disc imina e be ween di e en po s ha could enable u he unc ionali ies in new gene a ions o induc ion hobs. INDEX TERMS Con olu ional neu al ne wo k, home appliances, induc ion hea ing, neu al ne wo k applica ions, sys em-on-chip (SoC). I. INTRODUCTION Induc ion hea ing is a con ac -less hea ing me hod ha has been widely used in many applica ions [1]–[3]. The de el- opmen o semiconduc o de ices o powe elec onics and he con ac -less na u e o his hea ing echnology makes i p e e ed o e o he hea ing me hods due o i s highe e i- ciency. In he domes ic induc ion hea ing (DIH) case, i s main ad an age compa ed o he esis i e cook ops is ha while in he las ones he ho es componen is he esis o , in he o me , he ho es elemen is he bo om pa o he po . Logically, his o e s highe e iciencies while a colde cooking su ace leads o a sa e , easie - o-clean and mo e du able home appliance [4]. Howe e , gi en he cha ac e is ics o he equi alen load, he con ol o an induc ion hob is mo e complica ed han he The associa e edi o coo dina ing he e iew o his manusc ip and app o ing i o publica ion was Fanbiao Li . one o a esis i e cook op. The al e na ing magne ic ield ha hea s he bo om o he po due o induced cu en losses and magne ic hys e esis [5], [6] is usually gene a ed wi h a hal - b idge se ies esonan in e e . The equi alen impedance o he load is modeled as a esis ance, R, connec ed in se ies wi h an induc ance, L, [7], and he esonan beha io is achie ed by adding a esonan capaci o , C , in se ies wi h he load. To eed he in e e , he g id ol age, g id, is ull-wa e ec i ied and il e ed wi h a bus capaci o , leading o a high- ippled ol age called bus ol age, B, whose undamen al equency is wice he one o g id. Gi en he low-cos con ex o DIH, he swi ching equency, sw, which spans om 30 o 75 kHz, is usually gene a ed h ough insula ed ga e bipola ansis o s (IGBTs). The lowe limi o sw is imposed by he maximum equency o human hea ing, which is oughly es ablished a 20 kHz, while he uppe limi o he swi ching equency is se o educe he swi ching losses. Addi ionally, in he inpu o he home appliance a il e is placed o imp o e VOLUME 9, 2021 This wo k is licensed unde a C ea i e Commons A ibu ion 4.0 License. Fo mo e in o ma ion, see h ps://c ea i ecommons.o g/licenses/by/4.0/ 16053 J. Villa e al.: Vessel Recogni ion in Induc ion Hea ing Appliances—A Deep-Lea ning App oach he elec omagne ic compa ibili y (EMC) wi h he g id. This il e is mainly designed o comply wi h he s anda ds ha limi he emissions o adio- equency dis u bances in he equency ange om 9 kHz o 30 MHz. One o he d awbacks bu also challenges o DIH is ha he equi alen load a ies wi h many pa ame e s such as he swi ching equency, he exci a ion le el, he cha ac e is ics o he po (ma e ial, size, empe a u e), he misalignmen and dis ance be ween he induc o and he po , e c. Besides, many o hese pa ame e s depend exclusi ely on he use . Such a ia ion o he load o ces he con ol sys em o change and moni o he ope a ional condi ions o he in e e o gua an ee he maximum pe o mance and a sa e ope a ion o he elec onics. Fo his eason, much e o has been ocused on iden i ying he equi alen load and so is e lec ed in he li e a u e [8]–[15]. While hese me hods y o cap u e he equi alen impedance o he load, hey a e no able o classi y di e en po s and disc imina e be ween hem, i.e., i a use has en di e en po s, o be able o ecognize which one o hose en po s is used a a ime. Al hough some yea s ago his could be seen as an idealis ic unc ionali y, wi h he inc easing compu a ional capaci y o new sys em-on-chips (SoCs) and p ocesso s and he de elopmen o he in e ne - o - hings (many induc ion home appliances a e connec ed o he in e ne nowadays) hese unc ions a e ge ing close and close o a eal implemen a ion. The ecogni ion o he po could imp o e he use expe i- ence by enabling u he unc ionali ies such as he de ec ion o commonly used po s, a be e de ec ion o he sui abili y o a po o induc ion hea ing, he wa ning o a misaligned po , he con ol o he use and ageing o a speci ic po , e c. Mo e- o e , mo e accu a e he mal models could be achie ed [16] which would ul ima ely lead o a be e es ima ion o he empe a u e and he cooking p ocess. Besides, he appli- ance could use in o ma ion om p e ious cookings such as an app oxima ion o he powe -swi ching equency ans e unc ion, which po s a e used o which ypes o cooking p o- cesses and o how long, e c. This in o ma ion could simpli y he algo i hms ha dis ibu e he powe be ween di e en induc o s, lead o as e esponses o a change in he powe a ge and enable ad anced ea u es in new gene a ions o home appliances, e.g.: imagine placing a co ee po on he cooking su ace and he appliance au oma ically ecognizing ha his is you co ee po and ha you a e going o make a co ee, jus because you ha e done so in p e ious days. Howe e , o ob ain his in o ma ion, an accu a e iden i i- ca ion ha could se e as a inge p in o each po would be ce ainly help ul. In [17], he equi alen impedance o he load is analyzed om he pe spec i e o he elec omag- ne ic p ope ies o he ma e ial o he po . Apa om o he pa ame e s, he equi alen impedance depends on he elec i- cal conduc i i y and magne ic pe meabili y o he ma e ial. Besides, he magne ic pe meabili y a ies wi h he exci a ion le el and, o una ely, di e en ma e ials espond di e en ly o ha a ia ion o he exci a ion le el. Mo eo e , gi en he high ipple o he bus ol age, he exci a ion le el applied o he po a ies con inuously e en i a cons an swi ching equency is applied. The me hods p oposed in [8]–[14] whe he calcula e a single and a e aged alue o he equi alen impedance pe bus cycle (o o many swi ching cycles) o base hei wo k on a dc-bus ol age a angemen , meaning ha hey can only ex ac in o ma ion abou he a e age induc i eness o esis- i eness o he loads. None heless, no unde lying in o ma ion abou he ma e ials, ex ac ed om he a ia ion o hei elec omagne ic p ope ies due o he a ia ion o he exci- a ion le el, can be ob ained. On he con a y, in [15] his in o ma ion is p o ided hanks o an algo i hm ha acks he a ia ion o he equi alen impedance wi h he exci a ion le el along he en i e bus pe iod. Howe e , al hough he impedance di e ence be ween po s ha a e made o di e en ma e ials is shown, no au oma ic ecogni ion o po s is p oposed. On he o he hand, deep lea ning has also been used o he iden i ica ion o ce ain pa ame e s in induc ion hobs. Fo ins ance, in [18] a con olu ional neu al ne wo k based on he a e age powe , ms cu en and quali y ac o o he load is p oposed o es ima e he o e lap a ea be ween he po and he induc o s (a lexible induc ion hob is used, whe e many small induc o s lead o a lexible cooking a ea). In [19] he au ho s calcula e he powe ac o and he absolu e alue o he impedance, |Z|, o he i s ou ha monics o he swi ching equency and apply a neu al ne wo k o es ima e he size o he essel placed abo e he induc o . In his case, he algo i hm is implemen ed in a simple induc ion hob wi h a single induc o and a single-swi ch in e e . Likewise, in [20] a me hod o classi y di e en ypes o po s (o ma e ials) is p oposed. They use he same in o ma- ion as in [19], bu hey make su e he impedance |Z|and he powe ac o a e calcula ed a a bus ol age o 100 V. Then, di e en machine lea ning algo i hms a e es ed and p oposed. Howe e , al hough he au ho s a e awa e o he a ia ion o he equi alen impedance wi h he exci a ion le el ( his is why hey make su e he iden i ica ion is applied a a cons an bus ol age), hey do no ake ad an age o his phenomenon o ex ac mo e in o ma ion abou he ma e ials. In his pape , a me hod o iden i y he po ha is placed abo e he induc o is p oposed. The induc o is used as a sma senso so ha , compa ed o a adi ional induc ion hob, no addi ional ha dwa e is equi ed. Fu he mo e, no limi a- ions nei he on he swi ching equency (as long as i is con- s an du ing he en i e bus pe iod) no on he bus ol age a e imposed. The iden i ica ion o he equi alen impedance is ob ained wi h he algo i hm sugges ed in [15], and he ecog- ni ion o he essel is achie ed by means o a con olu ional neu al ne wo k (CNN) whose only inpu s a e wo ec o s which con ain he alues o Rand Ldu ing a bus pe iod. The neu al ne wo k is ained and alida ed o line, and inally implemen ed in o a p o o ype wi h sys em-on-chip (SoC). This pape is o ganized as ollows. In Sec ion II he me hod o iden i y he equi alen impedance as well as he da a acqui- si ion p ocess a e b ie ly explained. The p oposed neu al ne wo k, he da a augmen a ion and aining p ocesses and 16054 VOLUME 9, 2021 J. Villa e al.: Vessel Recogni ion in Induc ion Hea ing Appliances—A Deep-Lea ning App oach FIGURE 1. Simpli ied schema ic o he induc ion hob and i s main wa e o ms. g id is ec i ied and he small bus capaci o leads o a e y high ippled bus ol age which di ec ly eeds he in e e . Swi ches QHand QLgene a e he al e na ing cu en and snubbe capaci o s a e included o educe he swi ching losses. he main esul s a e desc ibed in Sec ion III. Then, he imple- men a ion o he whole sys em in a p o o ype is p esen ed in Sec ion IV. Finally, Sec ion Vconcludes his a icle. II. IDENTIFICATION METHOD AND EXPERIMENTAL DATA ACQUISITION In Fig. 1a simpli ied schema ic o he hal -b idge se ies esonan in e e is shown. The iden i ica ion o he equi - alen impedance o he load is based on he me hod p oposed in [15] because i is he one (among [8]–[15]) ha ex ac s mo e in o ma ion abou he load: i p o ides he alues o he R−Lequi alen impedance and i s a ia ion wi h he exci- a ion le el, wha in insically shows how he e omagne ic p ope ies o he ma e ials o he po s a y. To accu a ely ob ain he equi alen impedance wi h his me hod, i is necessa y o measu e he ol age d op o he load, L, and he cu en ha lows h ough he induc o , iL. The ol age Lis indi ec ly calcula ed om he sub ac ion o he ou pu ol age o he in e e , o, and he esonan capac- i o ol age, c, ( L= o− c) and iLis usually measu ed wi h a cu en ans o me . The iden i ica ion me hod is based on a phase-sensi i e de ec o (PSD), which gene a es he sinusoidal signals ha a e synch onized wi h he in e e sig- nals. A e some ope a ions and low-pass il e ing, he cosine and sine i s -ha monic componen s o Land iL, which a e deno ed as VL,c,VL,s,IL,c, and IL,s(subindex ‘‘1h’’ e e ing o i s ha monic is d opped ou o no a ion simplici y), a e ob ained and, wi h hese alues, and he angula swi ching equency, ωsw, he i s -ha monic equi alen impedance is ob ained as: R≈R1h=VL,cIL,c+VL,sIL,s I2 L,c+I2 L,s L≈L1h=VL,cIL,s−VL,sIL,c ωsw(I2 L,c+I2 L,s),(1) whe e i is assumed ha he phaso s o he load ol age and he cu en a e gi en by VL,1h=VL,c−jVL,sand IL,1h= IL,c−jIL,s, espec i ely. Fo mo e speci ic in o ma ion abou his iden i ica ion me hod, one may e e o [15]. Fo an app op ia e s udy o he p esen ed p oblem, a la ge expe imen al da ase ha includes a wide a ie y o po s is equi ed and, o eco d i , an expe imen al p o o ype wi h an au oma ic da a acquisi ion unc ionali y would be manda- o y. Fo his eason, he da a acquisi ion in he expe imen al p o o ype was ully au oma ed. To ob ain he da ase he esea che only has o place a po illed wi h wa e on he induc o and launch he au oma ic p ocess ha is explained in Algo i hm 1, which mus be epea ed as many imes as he numbe o po s o be iden i ied. This was accomplished hanks o a Di ec Memo y Access (DMA) con olle , which, o e e y swi ching equency, ans e s he alues o VL,c, VL,s,IL,c,IL,sand ωsw du ing a bus pe iod o he Double Da a Ra e Synch onous Dynamic Random-Access Memo y (DDR-SDRAM) a ailable in he p o o ype (mo e in o ma ion will be p o ided in Sec ion IV). This ansac ion is synch o- nized wi h he ze o-c ossing o he g id ol age and, o a g id pe iod o 50 Hz, a o al o 868 alues (pe swi ching e- quency) o he men ioned a iables a e sa ed. Addi ionally, hanks o his implemen a ion i would be possible o c ea e a eposi o y o s o age a ack eco d o old and new essels ha a e launched on o he ma ke . Then, he impedance alues a e calcula ed o line wi h (1), ob aining 868 alues o Rand L o e e y swi ching e- quency. Since o he cu en implemen a ion o Algo i hm 1, 1 sw =500 Hz, sw,max =75 kHz and he maximum powe o mos o he po s is eached a sw,@maxPowe below 40 kHz, a leas 70 cu es o R−La e eco ded; ha is mo e han 60 housand alues o Rand L o e e y po . The whole p ocess shown in Algo i hm 1 akes less han wo seconds VOLUME 9, 2021 16055 J. Villa e al.: Vessel Recogni ion in Induc ion Hea ing Appliances—A Deep-Lea ning App oach Algo i hm 1 Au oma ic Da a Acquisi ion sw ← sw,@maxPowe while sw ≤ sw,max do Launch DMA ansac ion (sa e ωsw and ec o s VL,c, VL,s,IL,cand IL,s) wai un il DMA ansac ion is comple ed sw ← sw +1 sw end while Send da a o lap op FIGURE 2. Resul o he au oma ic iden i ica ion algo i hm o an enameled po . E e y cu e co esponds o he R−L alues du ing a bus pe iod a a speci ic swi ching equency. The c es o he bus pe iod is a 5 ms. (a) Resis ance. (b) Induc ance. pe po . An example o he da a ob ained o one po is shown in Fig. 2. III. THE NEURAL NETWORK Deep lea ning is a widesp ead echnique ha has been inc easingly used du ing he las yea s. This is mo i a ed by i s abili y o sol e e y complica ed and non-linea p ob- lems by lea ning complex pa e ns om la ge da ase s [21], [22]. Mo eo e , unlike o he adi ional me hods, i does no equi e nei he he use o ex emely complex ma hema ical exp essions no a deep expe ise in he physics behind he p oblem o be sol ed. The e exis many ypes o neu al ne wo ks; one o hem, known as ully connec ed neu al ne wo k, connec each neu on in a laye o all he neu ons in he nex laye , wha exponen ially inc eases i s size and compu a ional complex- i y wi h he ne wo k dep h. On he o he hand, con olu- ional neu al ne wo ks ake ad an age o hie a chical pa e ns in da a o combine ha in o ma ion in o smalle and sim- ple pa e ns ha a e successi ely educed in he ollowing laye s. This educes he compu a ional complexi y and he numbe o coe icien s o be lea ned by he ne wo k, which a e wo c ucial aspec s o a easible eal- ime implemen- a ion in an embedded sys em (limi ed memo y and limi ed compu a ional capaci y). Fu he mo e, con olu ional neu al ne wo ks ha e been success ully applied o a wide a ie y o applica ions such as image ecogni ion o indus y [23] and medicine [24], speech ecogni ion [25], aul diagnosis in mechanical componen s [26], e c. A CNN is usually made o con olu ional laye s and pool- ing laye s. In he con olu ional laye , he inpu , in ou case he R−L ec o s, is con ol ed wi h a ke nel o il e which ex ac s he ea u es. Then, an ac i a ion unc ion is applied o ob ain nonlinea ans o ma ions. The nex ype o laye s, he pooling laye s, educes he dimension o he da a by ex ac ing he dominan ea u es. This p ocess is applied as many imes as he numbe o le els (dep h) he ne wo k has. Finally, a ully connec ed laye is applied o he esul o he las laye , which compu es he class sco es and ou pu s he inal esul . Apa om he low compu a ional complexi y o CNNs, he e is ano he eason o choose his ype o ne wo ks o his applica ion: i akes in o accoun he local connec i i y o he da a hanks o he connec ion be ween neu ons (a neu on in one laye is connec ed o some egions o he p e ious laye h ough he ke nel o he il e bank). This local connec i i y and neighbo ing ela ionship is also one o he easons why his ype o neu al ne wo ks pe o ms e y well in compu e ision applica ions [23], [24]. In his wo k, he ac i a ion unc ion which p o ides he nonlinea ans o ma ion is he ec i ied linea uni (ReLU) and i is de ined as: ReLU : (x)=max{0,x},(2) whe e xis he ou pu om he p e ious con olu ional laye . A. DATA AUGMENTATION AND TRAINING The expe imen al da a was ob ained o 19 di e en po s (di e en diame e s, ma e ials, manu ac u e s and b ands). These po s whe e placed on he induc o while ying o align he induc o and he po as good as possible and making su e he wa e was boiling when he au oma ic da a acquisi ion was launched (jus o s anda dizing he acquisi ion o da a). Besides, o 4 o hese 19 po s addi ional expe imen al da a was cap u ed a di e en empe a u es ( om ambien em- pe a u e o 100 ◦C) and a di e en posi ions (po aligned 16056 VOLUME 9, 2021 J. Villa e al.: Vessel Recogni ion in Induc ion Hea ing Appliances—A Deep-Lea ning App oach wi h he induc o and wi h a adial ho izon al displacemen induc o -po o 4 and 8 cen ime e s). The pu pose o his se o measu emen s is o es i he con olu ional neu al ne wo k is able, apa om dis- c imina ing be ween di e en po s, o de ec i a po is no p ope ly aligned wi h he induc o (no ma e he empe a- u e) o , wha is equi alen om he poin o iew o he load, i a oo small po is used. This is an impo an issue in induc ion hobs because when a po is no well aligned wi h he induc o he esis ance dec eases and, o supply he same powe , he cu en iLwould each such le els ha a e no wi hs ood by he powe elec onic de ices. I is wo h men ioning ha a adi ional esis i e cook op p o ides a cons an powe and i he po is smalle han he esis ance, he hea supplied o he po is au oma ically educed because o he smalle a ea o in e ac ion. Howe e , induc ion hobs need a mo e sophis ica ed con ol ha is able o educe he powe when hese smalle po s a e de ec ed. O he wise, he use would pe cei e ha smalle (o misaligned) po s a e o e hea ed. To sum up, aking in o accoun he 19 di e en po s and he addi ional measu emen s o 4 o hem, ha gi es us a o al o 27 di e en ags o iden i ie s (because o he mis- alignmen s) o be ecognized by he neu al ne wo k. A o al o 300 sweeps we e pe o med, and conside ing ha in e e y sweep app oxima ely 70 swi ching equencies a e analyzed, ha gi es us a o al o 20790 cu es o R−L o he whole se o po s. Since du ing he aining p ocess o he neu al ne wo k, he coe icien s o he ke nel ha e o be adjus ed, la ge da ase a e equi ed, so ha he p obabili y o o e - i ing is educed. The need o a la ge and di e se da ase is in ac one o he challenges o he success o neu al ne - wo ks. Al hough 20790 cu es migh be seen as a decen ly la ge da ase , da a augmen a ion is ca ied ou in his wo k. This is a commonly used echnique, which has al eady been used in o he applica ions [27], [28], when he amoun o expe imen al da a is no conside ed la ge o di e se enough. Fu he mo e, in he eal applica ion small de ia ions and noise could appea in he iden i ica ion, and his is some hing ha should be conside ed in he aining s age o he ne wo k. Thus, e e y R−L ec o is mul iplied by a andom ac o in he ange [0.9, 1.1], hen, a andom o se in he ange [0.9, 1.1] o he a e age alue o he ec o is added, and a andom noise o ±1% is added o e e y sample in he ec o . Fu he mo e, conside ing ha he ze o-c ossing synch oniza- ion o g id in he eal hob migh no be ideal, a ci cula shi o a andom numbe o samples in he ange [-10, 10] is applied o he ec o s. An example o he esul o his p ocess is shown in Fig 3. This p ocess is applied many imes o he da a sa ed o e e y po un il he e o a e ob ained o he aining da ase equals he one o he es ing da ase . Mo eo e , he da a augmen a ion algo i hm makes su e ha he same numbe o cu es a e gene a ed o e e y po , e en i in he expe imen al da a acquisi ion his condi ion did no FIGURE 3. Example o da a augmen a ion p ocess (only esis ance is shown). hold. This assu es ha he ecogni ion o all he po s will be simila ly weigh ed du ing he op imiza ion o he aining p ocess. The inal da ase is di ided in o wo equally sized se s o 653400 cu es. B. THE PROPOSED NEURAL NETWORK Be o e implemen ing he CNN in o he p o o ype, se e al ne wo ks we e ained and es ed o line. To do so, an exhaus- i e s udy changing he numbe o con olu ional laye s (#C), he s ide, he ke nel (K) size, he numbe o ke nels (#K) and he size o he pooling laye s was pe o med. Gi en ha he inal goal is he implemen a ion in an embedded de ice, he numbe o coe icien s was limi ed o a maximum alue o app oxima ely 20k. The numbe o il e s, n il,1, and ke nel size, ksize,1, o he i s con olu ional laye we e ob ained as: n il,1=2#K ksize,1=size(K) (3) Likewise, o he second con olu ional laye , n il,2and ksize,2we e calcula ed as: n il,2=2#K2ksize,2= bsize(K)/S idec,(4) Con olu ional neu al ne wo ks wi h one and wo con o- lu ional laye s we e conside ed because he es accu acy ob ained wi h wo laye s was al eady g ea . The aining and es ing we e pe o med wi h Tenso Flow and Ma lab using an NVIDIA GTX1085Ti GPU. The esul s o he pa ame ic s udy a e shown in Table 1, whe e he numbe o coe i- cien s (#Coe .), he equi ed mul iply-accumula e ope a ions (#MAC) and he e o s a e gi en o each es ed ne wo k. No e ha in CNNs wi h a single con olu ional laye no pooling laye s a e used. Apa om CNNs, long sho - e m memo y (LSTM) and ga ed ecu en uni (GRU) ne wo ks we e also es ed, bu hei esul s a e no discussed in his a icle because hey unde pe o med mos o he analyzed CNNs. Finally, he CNN ha o e ed he lowes e o in he es (see Table 1) was selec ed as he bes candida e o he eal- ime implemen a ion. I is a CNN wi h wo con olu ional VOLUME 9, 2021 16057 J. Villa e al.: Vessel Recogni ion in Induc ion Hea ing Appliances—A Deep-Lea ning App oach FIGURE 4. Con usion ma ix o he selec ed wo-laye CNN (numbe 17 in Table 1). The i s le e in he ag is he iden i ie o he b and/ ype o he po (whe he i is enameled, sandwich, mul ilaye , e c.), he nex numbe iden i ies di e en po s o a gi en b and/ ype, he numbe be ween he hyphens is he diame e o he po in cen ime e s, and he numbe a e he ‘‘M’’ le e is he adial misalignmen wi h espec o he induc o in cen ime e s. TABLE 1. Pa ame ic s udy o CNNs. laye s whose de ailed s uc u e, he size o he da a be ween laye s and he numbe o pa ame e s o each ke nel a e sum- ma ized in Table 2. The con usion ma ix o he selec ed ne wo k is shown in Fig. 4. I is wo h men ioning ha o he o iginal da a TABLE 2. In e nal s uc u e o he selec ed wo-laye CNN. (wi hou da a augmen a ion) he con usion ma ix showed no e o s. In Fig. 4i can be obse ed ha he CNN is able o ecognize he po mos o he imes and i a ely mis akes po s o di e en b ands o ypes. Some mis akes a e obse ed when he CNN ies o de ec a misalignmen (see D3-22-M0 and D3-22-M4) o when wo di e en po s o he same b and/ ype wi h he same diame e a e aligned wi h he induc o (see E1-22-M0 and E3-22-M0). Mo eo e , i can be no iced ha o a gi en ype o po , le ’s say ype ‘‘E’’, i a big po is misaligned 4 cm (E3-22-M4), he CNN some imes mis akes i o an aligned bu smalle po (E2-17.5-M0). In ac , his was an expec ed beha io because om he poin o iew o he load, hose si ua ions migh be seen as equi alen . IV. IMPLEMENTATION The expe imen al p o o ype is made o some pa s o a comme cial induc ion hob and wo p in ed ci cui boa ds 16058 VOLUME 9, 2021 J. Villa e al.: Vessel Recogni ion in Induc ion Hea ing Appliances—A Deep-Lea ning App oach FIGURE 5. Expe imen al p o o ype. (PCBs) ha we e designed o esea ch pu poses. One o he PCBs moun s he powe elec onics ci cui y and he o he one, which can be connec ed o he o me , moun s he condi ioning and acquisi ion ci cui s. Besides, he second PCB is compa ible wi h he ca ie boa d TE0703 om T enz Elec onics, which is connec ed o he sys em-on-module (SoM) TE0720 (also om T enz Elec onics) which, in u n, includes he sys em-on-chip Zynq-7020. This SoC includes in a single chip he ield-p og ammable ga e a ay (FPGA) ab ic and a dual co e ARM Co ex-A9 p ocesso . Besides, he SoM also includes a 1 GB DDR-SDRAM. The acquisi ion o he elec ical signals o, c,iLand g id ( his las a iable is used o he synch oniza ion wi h he ze o-c ossing o he g id ol age) is accomplished h ough 12-bi LTC-2315-12 analog- o-digi al con e e s (ADCs) whose sampling equency is se a 2.78 Msps. These ADCs a e con olled h ough a se ial pe iphe al in e ace (SPI) p o ocol. The modula o ha con ols he IGBTs, he iden i ica ion o he load [15], he con ol o he ADCs and he con- olle o he DMA ansac ions a e implemen ed on he FPGA ab ic whose clock equency is se a 100 MHz. The signals VL,c,VL,s,IL,c,IL,sand ωsw a e sen o he DMA con olle h ough he Ad anced eX ensible In e ace 4 (AXI4)-S eam p o ocol. Finally, he DMA con olle , which is con igu ed in s eam o memo y-map mode, w i es he da a in he DDR-SDRAM whene e he ARM (co e 1) p ocesso eques s i . A he same ime he ARM (co e 0), which uns he ope - a ing sys em Pe alinux, is in cha ge o he communica ion wi h a g aphical use in e ace (GUI) p og ammed in Ma lab and unning in a lap op. This communica ion is es ablished o e E he ne h ough a TCP/IP p o ocol and was o g ea impo ance o he au oma ic da a acquisi ion explained in Sec ion II. The con olu ional neu al ne wo k is implemen ed in loa - ing poin (single p ecision) in he co e 1 o he ARM p o- cesso . To do so, an in e up is gene a ed wi h one o he iple ime coun e s (TTCs), whose in e up handle launches a DMA ansac ion which sa es he ec o s VL,c,VL,s,IL,c, IL,sand ωsw o a bus pe iod in he DDR-SDRAM. When he ansac ion inishes, an in e up is gene a ed, and a e com- pu ing (1) o each o he 868 samples, he CNN is execu ed. This p ocess is execu ed 10 imes pe second, lea ing ime o he ARM (co e 1) o un o he asks such as con ol o powe , moni o ing o a iables, e c. The clock equency o he ARM is 667 MHz. I is wo h men ioning ha in he inal applica ion, depending on he equi emen s o he induc ion hob and he ea u es o i s SoC, i migh no be necessa y o ecognize he po con inuously o a leas no a such a as a e. The en i e code, once p og ammed in C, equi es 130 kB o memo y including he mo e han 85 kB o coe icien s. The execu ion o he con olu ional neu al ne wo k oge he wi h he calcula ion o he R−L alues (1) akes abou 2.55 ms in he ARM. I is wo h no ing ha he CNN is execu ed only once e e y 100 ms. A simpli ied block diag am o he sys em is shown in Fig 6. A ideo is a ached wi h his wo k whe e he pe o mance o he eal- ime implemen a ion in he p o o ype is shown. In his ideo, di e en po s a di e en misalignmen s a e FIGURE 6. Simpli ied block diag am o he sys em. Subindex ‘‘dig’’ e e s o digi al and has only been added o di e en ia e be ween analog and digi al signals. VOLUME 9, 2021 16059 J. Villa e al.: Vessel Recogni ion in Induc ion Hea ing Appliances—A Deep-Lea ning App oach placed on he induc o while he SoC execu es he p oposed algo i hm and p in s he ag o he p edic ed po h ough a e minal. V. CONCLUSION Deep lea ning has been e ol ing du ing he las yea s due o he con inuously inc easing compu a ional capaci y o new echnologies and i s amazing abili y o lea n complex pa e ns om la ge da ase s. In his wo k, a con olu ional neu al ne wo k is applied o ecognize he po ha is hea ed in a domes ic induc ion hob. The inpu da a o he neu al ne wo k a e wo ec o s o 868 alues each which con ain he a ia ion o he equi a- len esis ance and induc ance o he load du ing a bus pe iod. To ain he neu al ne wo k, he equi alen impedance o se e al po s was iden i ied du ing a bus cycle o di e - en swi ching equencies, empe a u es and misalignmen s. Th ough da a augmen a ion, a la ge da ase was ob ained, wi h which he neu al ne wo k was ained. Se e al ypes o neu al ne wo ks we e es ed o line by a ying he numbe o laye s, he s ide, he il e size and he numbe o il e s o each laye . The esul s ob ained du ing he es ing p ocess show ha he p oposed CNN is able o success ully ecognize he po s wi h an a e age e o a e o 0.7819%. Besides, he expe imen al esul s show ha he p oposed solu ion is also able o es ima e i , and how much, a po is misaligned wi h he induc o e en when he equi alen impedance is iden i ied a di e en po empe a u es. Mo eo e , wi h he p oposed me hod, he induc o is used as a sma -senso , so ha no ex a-ha dwa e is equi ed, p e- se ing he low-cos na u e o DIH and aking ad an age o he inc easingly highe compu a ional complexi y a ailable in embedded sys ems. Finally, he p oposed CNN was implemen ed in an expe - imen al p o o ype o show i s eal- ime pe o mance and he easibili y o i s inclusion in a eal hob. This is he i s imple- men a ion o a po iden i ie based on con olu ional neu al ne wo ks and he in o ma ion o he a ia ion o he equi - alen impedance wi h he exci a ion le el, which will help b oadening he esea ch and de elopmen a eas in domes ic induc ion hea ing and will enable new unc ionali ies in u u e gene a ions o induc ion hobs. REFERENCES [1] N. A. Ahmed, ‘‘High- equency so -swi ching AC con e sion ci cui wi h dual-mode PWM/PDM con ol s a egy o high-powe IH applica ions,’’ IEEE T ans. Ind. Elec on., ol. 58, no. 4, pp. 1440–1448, Ap . 2011. [2] V. Es e e, J. Jo dan, E. Sanchis-Kilde s, E. J. Dede, E. Mase , J. B. Ejea, and A. Fe e es, ‘‘Enhanced pulse-densi y-modula ed powe con ol o high- equency induc ion hea ing in e e s,’’ IEEE T ans. Ind. Elec on., ol. 62, no. 11, pp. 6905–6914, No . 2015. [3] J. Ace o, J. Bu dío, L. Ba agán, D. Na a o, R. Alonso, J. Ga cía, F. Mon e de, P. He nández, and S. L. I. Ga de, ‘‘Domes ic induc ion appliances,’’ IEEE Ind. Appl. Mag., ol. 16, no. 2, pp. 39–47, Ma . 2010. [4] O. Lucia, J. Ace o, C. Ca e e o, and J. M. Bu dio, ‘‘Induc ion hea ing appliances: Towa d mo e lexible cooking su aces,’’ IEEE Ind. Elec on. Mag., ol. 7, no. 3, pp. 35–47, Sep. 2013. [5] J. Ace o, C. Ca e e o, R. Alonso, and J. M. Bu dio, ‘‘Quan i a i e e alu- a ion o induc ion e iciency in domes ic induc ion hea ing applica ions,’’ IEEE T ans. Magn., ol. 49, no. 4, pp. 1382–1389, Ap . 2013. [6] H. Ku ose, D. Miyagi, N. Takahashi, N. Uchida, and K. Kawanaka, ‘‘3-D eddy cu en analysis o induc ion hea ing appa a us conside ing hea emission, hea conduc ion, and empe a u e dependence o magne ic cha - ac e is ics,’’ IEEE T ans. Magn., ol. 45, no. 3, pp. 1847–1850, Ma . 2009. [7] J. Ace o, O. Lucia, I. Millan, L. A. Ba agan, J.-M. Bu dio, and R. Alonso, ‘‘Iden i ica ion o he ma e ial p ope ies used in domes ic induc ion hea - ing appliances o sys em-le el simula ion and design pu poses,’’ in P oc. 25 h Annu. IEEE Appl. Powe Elec on. Con . Expo. (APEC), Feb. 2010, pp. 439–443. [8] B. A. Nguyen, Q. D. Phan, D. M. Nguyen, K. L. Nguyen, O. Du ieu, and P. Maussion, ‘‘Pa ame e iden i ica ion me hod o a h ee-phase induc ion hea ing sys em,’’ IEEE T ans. Ind. Appl., ol. 51, no. 6, pp. 4853–4860, No . 2015. [9] A. Dominguez, A. O in, I. U iza, L. A. Ba agan, D. Na a o, and J. I. A igas, ‘‘Load iden i ica ion o domes ic induc ion hea ing based on pa icle swa m op imiza ion,’’ in P oc. IEEE 15 h Wo kshop Con ol Modeling Powe Elec on. (COMPEL), Jun. 2014, pp. 1–6. [10] O. Jimenez, L. A. Ba agan, I. U iza, O. Lucia, D. Na a o, and J. I. A igas, ‘‘FPGA-based eal- ime ha monic impedance measu emen o se ies esonan loads by using lock-in algo i hm,’’ in P oc. 37 h Annu. Con . IEEE Ind. Elec on. Soc. (IECON), No . 2011, pp. 2808–2813. [11] O. Jimenez, O. Lucia, I. U iza, L. A. Ba agan, and D. Na a o, ‘‘Analysis and implemen a ion o FPGA-based online pa ame ic iden i ica ion algo- i hms o esonan powe con e e s,’’ IEEE T ans. Ind. In o ma ., ol. 10, no. 2, pp. 1144–1153, May 2014. [12] V. T. Kilic, E. Unal, N. Yilmaz, and H. V. Demi , ‘‘All-su ace induc ion hea ing wi h high e iciency and space in a iance enabled by a aying squi cle coils in squa e la ice,’’ IEEE T ans. Consum. Elec on., ol. 64, no. 3, pp. 339–347, Aug. 2018. [13] D. Puyal, C. Be nal, J. M. Bu dio, J. Ace o, and I. Millan, ‘‘Ve sa ile high- equency in e e module o la ge-signal induc i e loads cha ac e iza ion up o 1.5 MHz and 7 kW,’’ IEEE T ans. Powe Elec on., ol. 23, no. 1, pp. 75–87, Jan. 2008. [14] H. Sa nago, O. Lucia, and J. M. Bu dio, ‘‘FPGA-based esonan load iden i ica ion echnique o lexible induc ion hea ing appliances,’’ IEEE T ans. Ind. Elec on., ol. 65, no. 12, pp. 9421–9428, Dec. 2018. [15] J. Villa, L. A. Ba agan, J. I. A igas, D. Na a o, A. Dominguez Vicen e, and T. Cabeza, ‘‘SoC-based in-cycle load iden i ica ion o induc ion hea - ing appliances,’’ IEEE T ans. Ind. Elec on., ea ly access, Jul. 10, 2020, doi: 10.1109/TIE.2020.3007083. [16] C. F anco, J. Ace o, R. Alonso, C. Sagues, and D. Paesa, ‘‘Induc i e senso o empe a u e measu emen in induc ion hea ing applica ions,’’ IEEE Senso s J., ol. 12, no. 5, pp. 996–1003, May 2012. [17] J. Se ano, J. Ace o, I. Lope, C. Ca e e o, J. M. Bu dio, and R. Alonso, ‘‘Modeling o domes ic induc ion hea ing sys ems wi h non-linea sa - u able loads,’’ in P oc. IEEE Appl. Powe Elec on. Con . Expo. (APEC), Ma . 2017, pp. 3127–3133. [18] O. Lucia, D. Na a o, P. Guillen, H. Sa nago, and S. Lucia, ‘‘Deep lea ning-based magne ic coupling de ec ion o ad anced induc ion hea - ing appliances,’’ IEEE Access, ol. 7, pp. 181668–181677, 2019. [19] A. Bono-Nuez, C. Be nal-Ruíz, B. Ma ín-del-B ío, F. J. Pé ez-Cebolla, and A. Ma ínez-I u be, ‘‘Recipien size es ima ion o induc ion hea ing home appliances based on a i icial neu al ne wo ks,’’ Neu al Compu . Appl., ol. 28, no. 11, pp. 3197–3207, No . 2017. [Online]. A ailable: h p://link.sp inge .com/10.1007/s00521-016-2227-6 [20] A. Bono-Nuez, B. Ma in-del-B io, C. Be nal-Ruiz, F. J. Pe ez-Cebolla, A. Ma inez-I u be, and I. Sanz-Go acha egui, ‘‘The induc o as a sma senso o ma e ial iden i ica ion in domes ic induc ion cooking,’’ IEEE Senso s J., ol. 18, no. 6, pp. 2462–2470, Ma . 2018. [21] Y. LeCun, Y. Bengio, and G. Hin on, ‘‘Deep lea ning,’’ Na u e, ol. 521, no. 7553, pp. 436–444, May 2015. [Online]. A ailable: h p://www.na u e.com/a icles/na u e14539 [22] R. C. Gonzalez, ‘‘Deep con olu ional neu al ne wo ks [Lec u e No es],’’ IEEE Signal P ocess. Mag., ol. 35, no. 6, pp. 79–87, No . 2018. [23] X. Tao, D. Zhang, Z. Wang, X. Liu, H. Zhang, and D. Xu, ‘‘De ec ion o powe line insula o de ec s using ae ial images analyzed wi h con olu- ional neu al ne wo ks,’’ IEEE T ans. Sys ., Man, Cybe n. Sys ., ol. 50, no. 4, pp. 1486–1498, Ap . 2020. [24] A. H. Abdi, C. Luong, T. Tsang, G. Allan, S. Nou anian, J. Jue, D. Hawley, S. Fleming, K. Gin, J. Swi , R. Rohling, and P. Abolmaesumi, ‘‘Au o- ma ic quali y assessmen o echoca diog ams using con olu ional neu al ne wo ks: Feasibili y on he apical ou -chambe iew,’’ IEEE T ans. Med. Imag., ol. 36, no. 6, pp. 1221–1230, Jun. 2017. 16060 VOLUME 9, 2021 J. Villa e al.: Vessel Recogni ion in Induc ion Hea ing Appliances—A Deep-Lea ning App oach [25] Y. Qian, M. Bi, T. Tan, and K. Yu, ‘‘Ve y deep con olu ional neu al ne wo ks o noise obus speech ecogni ion,’’ IEEE/ACM T ans. Audio, Speech, Lang. P ocess., ol. 24, no. 12, pp. 2263–2276, Dec. 2016. [26] G. Jiang, H. He, J. Yan, and P. Xie, ‘‘Mul iscale con olu ional neu al ne wo ks o aul diagnosis o wind u bine gea box,’’ IEEE T ans. Ind. Elec on., ol. 66, no. 4, pp. 3196–3207, Ap . 2019. [27] Z. Meng, X. Guo, Z. Pan, D. Sun, and S. Liu, ‘‘Da a segmen a ion and augmen a ion me hods based on aw da a using deep neu al ne wo ks app oach o o a ing machine y aul diagnosis,’’ IEEE Access, ol. 7, pp. 79510–79522, 2019. [28] J. Salamon and J. P. Bello, ‘‘Deep con olu ional neu al ne wo ks and da a augmen a ion o en i onmen al sound classi ica ion,’’ IEEE Signal P ocess. Le ., ol. 24, no. 3, pp. 279–283, Ma . 2017. JORGE VILLA (Membe , IEEE) ecei ed he M.Sc. deg ee in indus ial enginee ing om he Uni e si y o Za agoza, Za agoza, Spain, in 2016, whe e he is cu en ly pu suing he Ph.D. deg ee in elec onic enginee ing. His main esea ch in e es s include esonan con e e s and digi al con ol o induc ion hea ing applica ions. M . Villa is a membe o he A agon Ins i u e o Enginee ing Resea ch (I3A), G oup o Powe Elec onics and Mic oelec onics (GEPM). DENIS NAVARRO ecei ed he M.Sc. deg ee in mic oelec onics om he Uni e si y o Mon pel- lie , F ance, in 1987, and he Ph.D. deg ee om he Uni e si y o Za agoza, in 1992. Since Sep embe 1988, he has been wi h he Depa men o Elec onic Enginee ing and Com- munica ions, Uni e sidad de Za agoza, whe e he is cu en ly a P o esso . In 1993, he designed he i s SPARC mic op ocesso in Eu ope. His cu en esea ch in e es s include CAD o VLSI, low-powe ASIC design, and modula ion echniques o powe con e e s. He is in ol ed in he implemen a ion o new applica ions o in eg a ed ci cui s. D . Na a o is a membe o he A agon Ins i u e o Enginee ing Resea ch (I3A). ALBERTO DOMINGUEZ ecei ed he Ph.D. deg ee in elec onic enginee ing om he Uni e - si y o Za agoza, Spain, in 2017. Since 2017, he has been wo king wi h BSH Home Appliances in he de elopmen o new domes ic induc ion cook ops. His main esea ch in e es s include modeling, con ol, and op imiza- ion o cons ained sys ems, especially de o ed o esonan in e e s in domes ic induc ion hea ing. JOSE I. ARTIGAS ecei ed he M.Sc. and Ph.D. deg ees in elec ical enginee ing om he Uni e - si y o Za agoza, Za agoza, Spain, in 1989 and 1996, espec i ely. He has been wi h he Depa men o Elec onic Enginee ing and Communica ions, Uni e si y o Za agoza, whe e he is cu en ly a P o esso . He has been in ol ed in di e en esea ch and de elop- men p ojec s. His main esea ch in e es s include signal acquisi ion, digi al con ol, and modula ion s a egies applied o powe con e e s. D . A igas is a membe o he A agon Ins i u e o Enginee ing Resea ch (I3A), G oup o Powe Elec onics and Mic oelec onics (GEPM). LUIS A. BARRAGAN ecei ed he M.Sc. and Ph.D. deg ees in physics om he Uni e si y o Za agoza, Za agoza, Spain, in 1988 and 1993, espec i ely. He is cu en ly a P o esso wi h he Depa men o Elec onic Enginee ing and Communica ions, Uni e si y o Za agoza. He has been in ol ed in di e en esea ch and de elopmen p ojec s on induc ion-hea ing sys ems o home appliances. His esea ch in e es s include modeling and digi al con ol applied o domes ic induc ion hea ing. D . Ba agan is a membe o he A agon Ins i u e o Enginee ing Resea ch (I3A), G oup o Powe Elec onics and Mic oelec onics (GEPM). VOLUME 9, 2021 16061