CRRN don Gano Abubuwan da ba su dace ba a Lokaci da Wuri a Binciken Kullin Solder

1. Gabatarwa & Bayyani

Wannan takarda tana magance kalubalen ingancin inganci mai mahimmanci a cikin Fasahar Dutsen Saman (SMT) don kera Allon Kewayawa na Buga (PCB). Babban yanki (50-70%) na lahani na PCB ya samo asali ne a matakin buga kullin solder. Hanyoyin bincike na gargajiya, kamar Binciken Kullin Solder (SPI), sun dogara da ƙofofin ƙididdiga suna ɗaukar rarraba al'ada na yawan kullin solder. Wannan hanyar ta kasa lokacin da lahani na bugu suka karkata rarraba bayanai bisa tsari.

Marubutan sun ba da shawarar Cibiyar Sadarwa mai Maimaitawa mai Maimaitawa (CRRN), sabon samfurin gano abin da ba na al'ada ba na aji ɗaya. CRRN tana koyo ne kawai daga bayanan aiki na al'ada kuma tana gano abubuwan da ba su dace ba ta hanyar auna kuskuren sake ginawa. Babban sabon abu a cikinta shine yin samfurin tsarin lokaci da wuri da ke cikin jerin bayanan SPI a fadin fakitin PCB da yawa.

2. Hanyar Aiki: Tsarin CRRN

CRRN wani na'urar sarrafa kansa ne na musamman da aka tsara don jerin bayanai na 2D (misali, taswirar yawan kullin solder akan lokaci). Yana rarraba tsarin sake ginawa zuwa sassan sarari da na lokaci da wuri.

3. Cikakkun Bayanai na Fasaha & Tsarin Lissafi

Za'a iya wakilta ainihin CSTM da tsarin hankali ta hanyar lissafi. Ana ba da aikin tantanin halitta na ConvLSTM na yau da kullun ta:

Inda $*$ ke nufin rufewa kuma $\odot$ yana nufin ninka kashi-kashi. CSTM yana gyara waɗannan ayyukan don ƙarin inganci a cikin kamawar tsarin lokaci da wuri. Tsarin ST-Attention yana ƙididdige vector na mahallin $c_t$ don mai bayyana a lokacin $t$ a matsayin jimillar nauyin duk jihohin ɓoyayyun mai rufewa $h_s$:

A nan, $a(\cdot)$ samfurin daidaitawa ne (misali, ƙaramin cibiyar sadarwa), kuma $\alpha_{ts}$ su ne ma'aunin hankali da ke ƙayyade mahimmanciyar jihar mai rufewa $s$ don matakin mai bayyana $t$.

4. Sakamakon Gwaji & Aiki

Takardar ta nuna fifikon CRRN akan samfuran gargajiya kamar na'urori masu sarrafa kansa (AE), na'urori masu sarrafa kansa masu bambance-bambance (VAE), da samfuran tushen ConvLSTM don gano abubuwan da ba su dace ba akan bayanan SPI. Ma'auni masu mahimmanci na aiki sun haɗa da:

• Kuskuren Sake Ginawa (MSE/MAE): Ƙananan kuskure don jerin abubuwa na al'ada, babban kuskure don jerin abubuwan da ba su dace ba, yana haifar da rabuwa bayyananne.
• Ma'auni na Gano Abubuwan da ba su dace ba: Babban Yankin Ƙarƙashin Lanƙwasa ROC (AUC-ROC), Daidaito, Tunawa, da maki-F1 wajen bambance lahani da jerin PCB na al'ada.
• Ƙarfin Bambance-bambance na Taswirar Abubuwan da ba su dace ba: Taswirar kuskuren sake ginawa ta sarari ("taswirar abin da ba na al'ada ba") da CRRN ta samar an yi amfani da ita azaman fasalin shigarwa don aikin rarrabuwar lahani na bugu na gaba. Babban daidaiton rarrabuwar da aka samu ya tabbatar da cewa taswirorin abubuwan da ba su dace ba suna sanya wuri da wakiltar ainihin tsarin lahani, ba kawai hayaniya ba.

Bayanin Ginshiƙi (An fayyace): Taswirar sandar za ta nuna CRRN ta fi samfuran tushe (AE, VAE, ConvLSTM-AE) a cikin ma'auni masu mahimmanci (AUC-ROC, F1-Score). Wani ginshiƙi na biyu zai iya nuna lanƙwasa daidaito-tunawa, tare da lanƙwasa CRRN yana rungumar kusurwar sama-dama, yana nuna ingantaccen aiki. Samfurin taswirorin abubuwan da ba su dace ba za su kwatanta yankuna masu babban kuskure da suka ta'allaka kan fakitin da lahani na musamman na bugu ya shafa kamar toshewa ko kuskuren daidaitawa.

5. Tsarin Nazari: Nazarin Shari'ar da ba ta ƙunshi Code ba

Yanayi: Layin tara PCB yana fuskantar lahani na gada na solder na tsaka-tsaki. SPI na gargajiya yana alamar fakitin bazuwar, amma ba a gano ainihin dalili ba.

Aikace-aikacen CRRN:

1. Tattara Bayanai: Jerin taswirorin yawan kullin solder daga ɗaruruwan PCB da aka sani suna da kyau ana ciyar da su cikin CRRN don horo.
2. Aiwatar da Samfurin: CRRN da aka horar yanzu yana sarrafa bayanan SPI kai tsaye a cikin jerin abubuwa (misali, kowane alluna 10).
3. Gano Abubuwan da ba su dace ba: Jerin allo yana nuna babban kuskuren sake ginawa. Taswirar abin da ba na al'ada ba na CRRN ya nuna ba kawai fakitin guda ɗaya ba, amma layin fakitin da ke kusa tare da yawan da ba na al'ada ba.
4. Binciken Asalin Dalili: Tsarin sarari (layi) yana nuna stencil da aka zazzage ko matsalar wukar likita a cikin Buga Kullin Solder (SPP), alaƙar lokaci wanda binciken fakitin kawai zai rasa. An sanar da kulawa ga takamammen kayan aikin bugu.

Wannan tsarin yana canzawa daga "gano allon mara kyau" zuwa "binciken tsarin da ya kasa," yana ba da damar kulawa mai hasashe.

6. Nazari mai mahimmanci & Ra'ayi na Kwararre

Ainihin Bayani: Wannan ba wani takarda ne kawai na cibiyar sadarwa ba; harin da aka yi niyya ne akan matsalar masana'antu na biliyoyin daloli—lalacewar kayan aiki a ɓoye. Marubutan sun gano daidai cewa ainihin ƙimar bayanan masana'antu mai hankali ba ta cikin hotunan guda ɗaya ba amma a cikin labarin lalacewa da aka faɗa a cikin jerin raka'o'in samarwa. Ta hanyar haɗa ƙwararrun sarari na CNNs tare da ƙwaƙwalwar lokaci na LSTMs da hankalin hanyoyin hankali, CRRN ya wuce rarrabuwar lahani zuwa fassarar sa hannun gazawa.

Kwararar Hankali: Hankali yana da inganci a masana'antu: 1) Bayanan al'ada suna da yawa, bayanan da ba su dace ba suna da wuya—don haka yi amfani da koyo na aji ɗaya. 2) Lahani yana da sarari (wurin gida akan allo) da girma na lokaci (yana ƙara muni)—don haka yi amfani da samfurin lokaci da wuri. 3) Dogon jerin abubuwa suna ɓoye alamun gargadi na farko—don haka ƙara hankali don haɗa dalili da sakamako a cikin lokaci. Wannan misali ne na littafin koyarwa na ƙira na gine-gine da aka motsa ta matsala, ba kawai tara samfura ba.

Ƙarfi & Kurakurai:

• Ƙarfi (Gaskiyar Gine-gine): Ƙirar na'ura (S-Encoder, ST-Module, S-Decoder) tana da kyau. Ya raba koyon fasalin sarari daga samfurin motsin lokaci, wanda mai yiwuwa yana taimakawa kwanciyar hankali na horo da fassarar. Amfani da hankali yana da gaskiya don matsalar dogon jerin abubuwa.
• Ƙarfi (Dabarar Tabbatarwa): Yin amfani da taswirar abin da ba na al'ada ba don aikin rarrabuwa na biyu yana da wayo. Ya tabbatar da cewa samfurin yana ciro fasali masu ma'ana, kamar yadda ake amfani da fasalin mai rarrabuwa a cikin CycleGAN don ayyuka na gaba, ya wuce makin kuskure na akwatin baƙi.
• Mai yuwuwar Kuskure (Yunwar Bayanai & Rikitarwa): Ko da yake aji ɗaya, samfurin yana da rikitarwa. Horar da ConvLSTM mai zurfi tare da hankali yana buƙatar jerin bayanai na al'ada masu yawa da albarkatun lissafi. Don layukan samarwa masu haɗawa mai yawa, ƙarancin ƙima, tattara isassun bayanan "al'ada" don kowane bambance-bambancen samfur na iya zama kalubale.
• Mai yuwuwar Kuskure (Gibi na Bayyani): Ko da yake taswirar abin da ba na al'ada ba tana sanya wurin kurakurai, bayyana dalilin da wannan tsarin ya dace da takamaiman lahani na bugu (misali, "wannan tsarin yana nufin kuskuren daidaitawa na Z-axis na 50μm") har yanzu yana buƙatar fassarar ɗan adam na kwararre. Samfurin yana binciken rashin lafiya amma baya suna ainihin ƙwayoyin cuta.

Bayanai masu Aiki:

1. Ga Masu Kera: Gwada wannan akan layin SPP mafi mahimmanci ko matsala. Komawar zuba jari ba kawai a cikin kama ƙarin lahani ba ne, amma a rage tsayawar aiki da ba a tsara ba da sharar stencil ta hanyar faɗakarwar hasashe. Fara da kayan aikin bayanan SPI don ɗaukar jerin lokaci.
2. Ga Masu Bincike: Mataki na gaba shine sanya wurin lahani na dalili. Shin za mu iya mayar da siginar kuskuren lokaci da wuri ba kawai zuwa wurin allo ba, amma zuwa takamammen kayan aikin jiki na bugu? Bincike kan haɗa samfuran tushen kimiyyar lissafi tare da hanyar CRRN na bayanai na iya haɗa gibin bayyani.
3. Ga Dillalan Kayan Aiki: Wannan shiri ne don tsarar gaba na tsarin SPI da AOI (Binciken Gani na atomatik). Matsa daga sayar da "tashoshin bincike" zuwa sayar da "tsarin kulawa da lafiyar tsari" tare da haɗaɗɗun samfura kamar CRRN. Gasar za ta kasance a cikin hankalin software, ba kawai ƙudurin firikwensin ba.

A ƙarshe, Yoo et al. sun ba da gudummawa mai mahimmanci wacce ke da ƙwaƙƙwaran ilimi kuma tana da alaƙa da masana'antu. Yana misalta yanayin da aka gani a cikin bincike na jagora daga cibiyoyi kamar Laboratory for Manufacturing and Productivity na MIT da al'ummar AI na Masana'antu: yin amfani da ci gaba mai zurfi na koyo ba don ayyuka na gaba ɗaya ba, amma don magance ƙayyadaddun matsalolin aiki masu ƙima tare da daidaitaccen gine-gine.

7. Ayyuka na Gaba & Jagororin Bincike

Tsarin CRRN yana da yuwuwar fiye da binciken kullin solder:

• Kera Semiconductor: Gano lahani masu ɗanɗano, masu alaƙa da sarari a cikin taswirorin wafer akan lokaci (misali, sakamakon karkatar da kayan aikin zubarwa).
• Ingancin Baturi: Nazarin hotuna masu biyo baya daga hanyoyin rufewa na lantarki don hasashen lahani na rufewa wanda ke haifar da gazawar tantanin halitta.
• Kulawa mai Hasashe don Robotic: Sa ido kan bayanan lokaci-lokaci daga na'urori masu auna ƙarfi/torque akan hannun mutum-mutumi yayin tattarawa don gano tsarin da ba na al'ada ba wanda ke nuna lalacewar inji.
• Jagororin Bincike:
  1. Samfuran Haske & Daidaitawa: Haɓaka nau'ikan CRRN waɗanda za a iya daidaita su da inganci don sabbin layukan samfura tare da ƙayyadaddun bayanai (misali, ta amfani da koyon meta ko dabarun ƴan harbi).
  2. Haɗawa tare da Tagwayen Digital: Ciyar da makin abubuwan da ba su dace ba na CRRN da taswirori cikin tagwayen digital na masana'anta don kwaikwayon tasirin lahani na bugu da ake zargi akan yawan amfanin gona na gaba da tsara kulawa a zahiri.
  3. Gano Abubuwan da ba su dace ba na Modal da yawa: Ƙaddamar da CRRN don haɗawa ba kawai bayanan ƙarar SPI ba, har ma da hotunan gani na 2D masu daidaitawa ko taswirorin tsayi na 3D daga wasu na'urori masu auna firikwensin don ƙarin ƙarfi sa hannun laifi.

8. Nassoshi

1. Yoo, Y.-H., Kim, U.-H., & Kim, J.-H. (Shekara). Convolutional Recurrent Reconstructive Network for Spatiotemporal Anomaly Detection in Solder Paste Inspection. IEEE Transactions on Cybernetics.
2. Shi, X., Chen, Z., Wang, H., Yeung, D.-Y., Wong, W.-K., & Woo, W.-c. (2015). Convolutional LSTM Network: A Machine Learning Approach for Precipitation Nowcasting. Advances in Neural Information Processing Systems (NeurIPS).
3. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., ... & Polosukhin, I. (2017). Attention is All You Need. Advances in Neural Information Processing Systems (NeurIPS).
4. Zhu, J.-Y., Park, T., Isola, P., & Efros, A. A. (2017). Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks. IEEE International Conference on Computer Vision (ICCV).
5. Ruff, L., Vandermeulen, R., Goernitz, N., Deecke, L., Siddiqui, S. A., Binder, A., ... & Kloft, M. (2018). Deep One-Class Classification. International Conference on Machine Learning (ICML).
6. Coleman, C., Damodaran, S., DeCost, B., et al. (2020). Defect Detection in Additive Manufacturing via Deep Learning. JOM, 72(3), 909–919.