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

1. Gabatarwa & Bayyani

Wannan takarda tana magance matsala mai mahimmanci a cikin Fasahar Mount Surface (SMT) don kera Allon Kewayawa na Bugawa (PCB): gano lahani yayin matakin buga kullin solder. Hanyoyin bincike na gargajiya, waɗanda suka dogara da zato na ƙididdiga na rarraba al'ada don ƙarar kullin solder, sun kasa lokacin da lahani na bugu ya karkata bayanan bisa tsari. Cibiyar Sadarwa mai Maimaitawa da Convolutional (CRRN) da aka gabatar sabon samfurin gano abin da ba daidai ba ne wanda ke koyo kawai daga tsarin bayanan al'ada kuma yana gano abubuwan da ba su dace ba ta hanyar kuskuren sake gina. An tsara shi musamman don sarrafa yanayin lokaci da wuri na bayanan Binciken Kullin Solder (SPI), inda lahani ke bayyana a matsayin tsarin sararin samaniya waɗanda ke haɓaka akan jerin samar da PCB.

50-70%

na lahani na PCB sun samo asali ne a matakin buga solder.

Koyo na Aji ɗaya

CRRN an horar da shi keɓance akan bayanan al'ada, yana kawar da buƙatar samfuran lahani masu lakabi.

2. Hanyar Aiki: Tsarin CRRN

CRRN wani na'ura mai sarrafa kansa na musamman ne wanda ya ƙunshi manyan sassa guda uku da aka tsara don ingantaccen koyo da sake gina fasalin lokaci da wuri.

2.1 Mai Rufe Sararin Samaniya (S-Encoder)

S-Encoder yana matsawa bayanan sararin samaniya na firam ɗaya na SPI (misali, taswirar ƙarar kullin solder) zuwa ƙaramin vector na latent ta amfani da yadudduka na convolutional na yau da kullun. Yana canza shigarwar $X_t \in \mathbb{R}^{H \times W \times C}$ zuwa wakilcin fasalin sararin samaniya $h_t^s$.

2.2 Mai Rufe da Buɗe Lokaci da Wuri (ST-Encoder-Decoder)

Wannan shine zuciyar CRRN, mai alhakin ƙirƙirar ƙirar dogaro na lokaci a kan jerin fasalin sararin samaniya $\{h_1^s, h_2^s, ..., h_T^s\}$.

2.2.1 Ƙwaƙwalwar Lokaci da Wuri ta Convolutional (CSTM)

Sabon naúrar maimaitawa da aka ƙera don maye gurbin ConvLSTM na gargajiya. An tsara CSTM don ƙarin ingantaccen cire tsarin lokaci da wuri, mai yiyuwa ta hanyar gyara hanyoyin ƙofofi ko ayyukan tantanin ƙwaƙwalwa don zama mafi inganci na sigogi ko mafi dacewa don takamaiman tsarin bayanan SPI. Ana iya wakiltar sabunta yanayin a zahiri kamar haka:

$C_t, H_t = \text{CSTM}(H_{t-1}, C_{t-1}, h_t^s; \Theta)$

inda $C_t$ shine yanayin tantanin halitta, $H_t$ shine yanayin ɓoye, kuma $\Theta$ sigogi ne masu koyo.

2.2.2 Tsarin Kulawa na ST

Don magance matsalar gradient da ke ɓacewa a cikin dogayen jerin gwano, an haɗa tsarin Kulawa na ST. Yana ba da damar mai buɗewa ya mai da hankali da gaske akan jihohin ɓoye masu dacewa daga mai rufewa a ko'ina cikin sarari da lokaci, yana sauƙaƙe mafi kyawun kwararar bayanai. Ma'aunin kulawa $\alpha_{t,t'}$ don matakin mai buɗewa $t$ yana duban matakin mai rufewa $t'$ ana iya ƙididdige shi kamar haka:

$\alpha_{t,t'} = \frac{\exp(\text{score}(H_t^{dec}, H_{t'}^{enc}))}{\sum_{k}\exp(\text{score}(H_t^{dec}, H_{k}^{enc}))}$

Vector na mahallin sannan jimlar ma'auni ce: $c_t = \sum_{t'} \alpha_{t,t'} H_{t'}^{enc}$.

2.3 Mai Buɗe Sararin Samaniya (S-Decoder)

S-Decoder yana ɗaukar fitarwa daga ST-Decoder (jerin vectors na mahallin lokaci da wuri) kuma yana amfani da jujjuyawar juzu'i don sake gina ainihin jerin firam ɗin SPI $\{\hat{X}_1, \hat{X}_2, ..., \hat{X}_T\}$.

3. Cikakkun Bayanai na Fasaha & Tsarin Lissafi

Babban manufa ita ce asarar sake ginawa don jerin al'ada. Aikin asara $\mathcal{L}$ yawanci shine Kuskuren Matsakaicin Matsakaici (MSE) tsakanin ainihin jerin gwano da na sake ginawa:

$\mathcal{L} = \frac{1}{T} \sum_{t=1}^{T} \| X_t - \hat{X}_t \|_2^2$

Yayin ƙididdiga, ana ƙididdige makin abin da ba daidai ba $A_t$ don firam a lokacin $t$ bisa la'akari da kuskuren sake ginawa:

$A_t = \| X_t - \hat{X}_t \|_2^2$

Ana amfani da kofa $\tau$ zuwa $A_t$ don rarraba firam ɗin (kuma ta hanyar tsawaitawa, PCB) a matsayin al'ada ko mara kyau. Ƙarfin samfurin yana cikin rashin iyawarsa don sake gina tsarinsa daidai da ba a gani ba yayin horo (watau abubuwan da ba su dace ba).

4. Sakamakon Gwaji & Aiki

Takardar ta nuna fifikon CRRN akan samfuran gargajiya kamar Autoencoders na yau da kullun (AE), Autoencoders na Bambance-bambance (VAE), da ƙananan samfuran maimaitawa. Babban sakamakon ya haɗa da:

Mafi Girman Daidaiton Gano Abin da ba daidai ba: CRRN ya sami manyan ma'auni na aiki (misali, makin F1, AUC-ROC) akan bayanan SPI waɗanda ke ɗauke da lahani da bugu ya haifar idan aka kwatanta da ma'auni.
Ingantaccen Matsayin Lahani: Bayan gano binary, CRRN yana samar da taswirar abin da ba daidai ba ta hanyar haskaka yankuna masu babban kuskuren sake ginawa. An nuna wannan taswirar tana da ikon nuna bambanci, yana taimakawa cikin nasara wajen rarraba takamaiman nau'ikan lahani na bugu (misali, toshewar stencil, rashin daidaito).
Ƙarfi ga Dogayen Jerin Gwano: Tsarin Kulawa na ST ya tabbatar da mahimmanci don kiyaye aiki akan dogayen jerin lokutan samar da PCB, wani yanayi na gama-gari a cikin layukan SMT na ainihin duniya.

Bayanin Ginshiƙi: Taswirar aiki ta zato za ta nuna lanƙwan AUC-ROC na CRRN sama da na AE, VAE, da autoencoders na tushen LSTM, musamman a ƙananan ƙimar kuskure mai mahimmanci don aikace-aikacen masana'antu.

5. Tsarin Nazari & Nazarin Lamari

Yanayi: Layin taron PCB yana fuskantar haɗin solder na lokaci-lokaci. Tsarin ƙididdiga na SPI na gargajiya ya kasa gano tushen dalili yayin da yake alamar yawancin pad a matsayin "wuce gona da iri" saboda canjin rarraba.

Aikace-aikacen CRRN:

Lokacin Horarwa: An horar da CRRN akan bayanan taswirar ƙarar SPI na makonni da yawa daga lokutan aikin bugu da aka sani mai kyau.
Ƙididdiga & Ganowa: Yayin samarwa kai tsaye, CRRN yana sarrafa jerin PCBs. Ya yi alama da takamaiman PCB tare da babban makin abin da ba daidai ba gabaɗaya.
Nazarin Tushen Dalili: Taswirar abin da ba daidai ba da aka samar don PCB da aka yi alama yana nuna tsarin sararin samaniya mai ci gaba na babban kuskure tare da axis ɗaya na allon, ba kawai keɓaɓɓun pad ba.
Bincike: Wannan tsarin sararin samaniya yana da halayen lahani na lalacewar ruwan wukake na bugu, wanda ke shafa kullin ba daidai ba. An faɗakar da kulawa don maye gurbin ruwan wukake, yana hana ƙarin gungu marasa inganci.

Wannan lamarin yana motsa sarrafa inganci daga "ganon allunan marasa kyau" zuwa "binciken kayan aiki masu gazawa," yana ba da damar kulawa na tsinkaya.

6. Ayyukan Gaba & Hanyoyin Bincike

Daidaitawar Yanki: Aiwatar da tsarin CRRN zuwa wasu ayyukan gano abubuwan da ba su dace ba a cikin lokaci da wuri a cikin Masana'antu 4.0, kamar nazarin girgiza a cikin injinan juyawa, hoton zafi a cikin taron lantarki, ko sa ido kan bidiyo don amincin layin taro.
Haɗawa tare da Tagwaye na Digital: Saka CRRN a matsayin na'urar gano abin da ba daidai ba a cikin tagwayen dijital na layin SMT don kwaikwaiyon kwaikwaiyo da nazarin tsinkaya.
Ƙaramin Hoto ko Koyo Mai Kulawa: Haɗa CRRN don haɗa ƙananan misalan lahani masu lakabi don inganta takamaiman ganowa don sanannun lahani mai mahimmanci.
Haɓaka Bayyanawa: Haɓaka hanyoyin don sanya ma'aunin kulawa na ST da taswirorin abin da ba daidai ba su zama masu fassara ga injiniyoyin shuka, watakila ta hanyar haɗa abubuwan da aka fi mayar da hankali zuwa takamaiman sassan jiki na bugu.
Turawa Bakin: Haɓaka samfurin don turawa akan na'urori a bakin teku a cikin injin SPI don ƙananan jinkiri, ganin abin da ba daidai ba a cikin wuri.

7. Nassoshi

Yoo, Y.-H., Kim, U.-H., & Kim, J.-H. (Shekara). Cibiyar Sadarwa mai Maimaitawa da Convolutional don Gano Abubuwan da ba su dace ba a cikin Lokaci da Wuri a cikin Binciken Kullin Solder. IEEE Transactions on Cybernetics.
Hochreiter, S., & Schmidhuber, J. (1997). Dogon Ƙaramin Ƙwaƙwalwar Ajiya. Lissafin Jijiyoyi.
Vaswani, A., et al. (2017). Kulawa Duk Abinda Kake Bukata. Ci gaba a cikin Tsarin Bayanai na Jijiyoyi.
Zhao, Y., et al. (2017) Autoencoders na Spatiotemporal Stacked don Gano Abubuwan da ba su dace ba a cikin Bidiyoyi. Ƙirar Ƙira.
Rahotannin Ƙaddamarwar Kera Lantarki ta Duniya (iNEMI) kan yanayin fasahar SMT da nazarin lahani.

8. Nazarin Kwararru & Bita Mai mahimmanci

Babban Fahimta

Wannan takarda ba wani ƙarin sauyi ne kawai na jijiyoyin jijiyoyi ba; yana da takamaiman harin tiyata akan matsalar ɓarna na masana'anti na biliyoyin daloli. Marubutan sun gano daidai cewa ainihin ƙimar a cikin masana'antar wayo ba a cikin gano allon da ba shi da inganci ba—yana cikin binciken na'urar da ta yi shi, a ainihin lokacin, kafin ta samar da dubu. Ta hanyar tsara lahani na bugu a matsayin abubuwan da ba su dace ba a cikin lokaci da wuri, sun wuce ƙididdiga na kowane pad zuwa hangen tsarin tsarin. Wannan shine bambanci tsakanin makaniki yana sauraron bugun injin guda ɗaya da injiniyan sararin samaniya yana nazarin dukan rikodin bayanan jirgin.

Kwararar Hankali

Hankalin gine-ginen yana da ƙarfi kuma yana nuna darussan da aka koya daga filayen kusa. Amfani da hanyar sake ginawa (autoencoder) don koyo na aji ɗaya an kafa shi sosai a cikin wallafe-wallafen gano abin da ba daidai ba, yayin da yake kawar da kusan aikin da ba zai yiwu ba na tattara bayanan da aka lakafta don kowane yanayin gazawar bugu. Sabon abu yana cikin haɗakar: haɗa ƙwarewar sararin samaniya na CNNs (wanda aka tabbatar a cikin nazarin hoto) tare da ƙirar lokaci na cibiyoyin sadarwa masu maimaitawa, sannan a ƙarfafa shi da tsarin kulawa. Kulawar ST wani kai tsaye ne, amfani da nasarar tsarin mai canzawa a cikin NLP (kamar yadda aka gani a cikin takarda mai mahimmanci "Kulawa Duk Abinda Kake Bukata") don magance kwatankwacin masana'antu na dogon lokaci dogaro—bin lalacewar sashi na injiniya a cikin sa'o'i na samarwa.

Ƙarfi & Kurakurai

Ƙarfi: Taswirorin abin da ba daidai ba na samfurin shine babban fasalin kashe shi. Wannan yana ba da hankali mai aiki, ba kawai ƙararrawar ƙararrawa ba. Mayar da hankali kan bayanan SPI na ainihin duniya ya kafa binciken a cikin alaƙar masana'antu mai ma'ana, sabon abu mai ban sha'awa ga samfuran da aka gwada kawai akan bayanan ilimi da aka tsara kamar bambance-bambancen MNIST don gano abin da ba daidai ba. Naúrar CSTM da aka gabatar tana nuna fahimtar cewa ConvLSTM na kasuwa na iya zama wuce gona da iri ko rashin inganci don wannan takamaiman tsarin bayanai.

Kurakurai masu yuwuwa & Tambayoyi: Takardar tana da sauƙi akan farashin lissafi da jinkirin ƙididdiga. A cikin babban layin SMT mai saurin samar da allon kowane 'yan daƙiƙa kaɗan, shin CRRN zai iya ci gaba? Horon "aji ɗaya" yana ɗauka da tsaftataccen bayanan, ba tare da lahani ba, wanda babban ƙalubale ne a ainihin saitunan masana'antu—yaya ƙarfinsa ga ɗan gurɓataccen bayanan horo? Bugu da ƙari, yayin da gine-ginen ya kasance mai zurfi, al'umma za su amfana daga nazarin zubar da ciki wanda ke tabbatar da wajibcin kowane sashi (CSTM vs. ConvLSTM, tare/ba tare da Kulawar ST ba) don wannan takamaiman aikin.

Fahimta mai Aiki

Ga injiniyoyin masana'antu, wannan binciken tsari ne don canzawa daga amsawa zuwa sarrafa inganci na tsinkaya. Mataki na nan da nan shine gwada CRRN akan layin SPP guda ɗaya mai mahimmanci, yana mai da hankali kan taswirar abin da ba daidai ba don jagorantar jadawalin kulawa. Ga masu binciken AI, aikin yana tabbatar da babban yuwuwar amfani da ci-gaba na samfuran jeri-zuwa-jeri tare da kulawa ga lokaci-lokaci na masana'antu da bayanan jeri na hoto. Gaba gaba, kamar yadda aka nuna a cikin taswirar hanya na iNEMI, yana motsawa daga ganowa zuwa rubutu—shin sararin ɓoye na CRRN ba zai iya alamar wani ruwan wukake da ya lalace ba har ma ya ba da shawarar matsi mafi kyau da saurin gyara don ramawa har zuwa lokacin kulawa na gaba? Wannan shine ainihin tsalle daga mai gano wayo zuwa tsarin samarwa mai daidaita kansa.