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

Teburin Abubuwan Ciki

1. Gabatarwa & Bayyani

Wannan takarda tana magance wata kalubale mai mahimmanci a cikin Fasahar Haɗaɗɗun Surface (SMT) don kera Allon Kewayawa (PCB): gano abubuwan da ba su dace ba da lahani na bugu ke haifarwa yayin matakin buga kwalliyar solder. Hanyoyin bincike na gargajiya, kamar Binciken Kwalliyar Solder (SPI), sun dogara da ƙofofin ƙididdiga waɗanda ke ɗaukar rarraba al'ada na yawan kwalliyar solder. Wannan hanyar ta kasa lokacin da rashin aikin bugu ya karkatar da rarraba bayanai bisa tsari. Maganin da aka gabatar shine Cibiyar Sadarwa mai Maimaitawa da Convolutional (CRRN), ƙirar gano abubuwan da ba su dace ba ta aji ɗaya wacce ke koyo kawai daga tsarin bayanai na al'ada kuma ta gano abubuwan da ba su dace ba ta hanyar kuskuren sake gini. Babban ƙirƙira yana cikin ikonta na rarraba tsarin abubuwan da ba su dace ba na lokaci da wuri daga jerin bayanan SPI, wanda ya wuce sauƙaƙan saita ƙofa zuwa wakilcin da aka koya na aikin al'ada.

2. Hanyoyi & Tsarin Gine-gine

CRRN wani na'ura mai sarrafa kansa na maimaitawa da convolutional (CRAE) ne wanda aka ƙera don bayanan jerin lokaci da wuri. Tsarin gine-ginensa an keɓance shi don ɗaukar fasalin sararin samaniya (misali, siffar kwalliyar solder akan fakitin) da dogaro na ɗan lokaci (misali, tsarin a kan alluna ko fakitoci masu bi da bi).

2.1 Bayyani game da Tsarin CRRN

Cibiyar sadarwa ta ƙunshi manyan sassa uku:

1. Mai Shirya Sararin Samaniya (S-Encoder): Yana ciro fasalin sararin samaniya daga firam ɗin shigarwa guda ɗaya (misali, hoton aunawa na SPI guda ɗaya) ta amfani da yadudduka na convolutional.
2. Mai Shirya da Mai Fassara na Lokaci da Wuri (ST-Encoder-Decoder): Babban tsarin da ke sarrafa jerin gwano. Ya ƙunshi tarin tubalan Ƙwaƙwalwar Ajiyar Lokaci da Wuri ta Convolutional (CSTM) da tsarin ST-Attention don ƙirar motsin lokaci da dogaro mai nisa.
3. Mai Fassara Sararin Samaniya (S-Decoder): Yana sake gina jerin shigarwa daga wakilcin ɓoyayyen lokaci da wuri ta amfani da jujjuyawar convolutional.

An horar da ƙirar ne kawai akan jerin bayanan SPI na al'ada. Yayin ƙididdiga, babban kuskuren sake gini yana nuna karkata daga tsarin al'ada da aka koya, yana nuna alamar wani abu da ba a saba gani ba.

2.2 Ƙwaƙwalwar Ajiyar Lokaci da Wuri ta Convolutional (CSTM)

CSTM sabon naúra ne da aka ƙera don ciro tsarin lokaci da wuri cikin inganci. Yana haɗa ayyukan convolutional cikin tsarin ƙwaƙwalwar ajiya mai maimaitawa, kama da Convolutional LSTM (ConvLSTM) amma an inganta shi don takamaiman aiki. Yana sabunta yanayin tantanin halitta $C_t$ da yanayin ɓoye $H_t$ ta amfani da ƙofofin convolutional, yana ba shi damar adana alaƙar sararin samaniya a cikin lokaci: $$i_t = \sigma(W_{xi} * X_t + W_{hi} * H_{t-1} + b_i)$$ $$f_t = \sigma(W_{xf} * X_t + W_{hf} * H_{t-1} + b_f)$$ $$C_t = f_t \odot C_{t-1} + i_t \odot \tanh(W_{xc} * X_t + W_{hc} * H_{t-1} + b_c)$$ $$o_t = \sigma(W_{xo} * X_t + W_{ho} * H_{t-1} + b_o)$$ $$H_t = o_t \odot \tanh(C_t)$$ inda $*$ ke nufin convolution kuma $\odot$ yana nufin ninka kashi-kashi.

2.3 Hankali akan Lokaci da Wuri (ST-Attention)

Don magance matsalar gradient da ke ɓacewa a cikin jerin gwano masu tsayi, an ƙera tsarin ST-Attention. Yana sauƙaƙe kwararar bayanai daga ST-Encoder zuwa ST-Decoder ta ba da damar mai fassara ya "kula" da jihohin mai shiryawa masu dacewa a cikin duk matakan lokaci, ba kawai na ƙarshe ba. Wannan yana da mahimmanci don ɗaukar dogaro na dogon lokaci a cikin tsarin masana'antu, kamar jinkirin motsi a cikin aikin bugu.

3. Cikakkun Bayanai na Fasaha & Tsarin Lissafi

Manufar horo ita ce rage asarar sake gini tsakanin jerin shigarwa $X = \{x_1, x_2, ..., x_T\}$ da jerin da aka sake gina $\hat{X} = \{\hat{x}_1, \hat{x}_2, ..., \hat{x}_T\}$, yawanci ta amfani da Kuskuren Matsakaicin Murabba'i (MSE): $$\mathcal{L}_{recon} = \frac{1}{T} \sum_{t=1}^{T} \| x_t - \hat{x}_t \|^2$$ Ana saka maki don abin da ba a saba gani ba na sabon jerin gwano a matsayin wannan kuskuren sake gini. Ana amfani da ƙofa (wanda galibi ana ƙayyade shi ta hanyar gwaji akan saitin tabbatarwa na bayanan al'ada) don rarraba jerin gwano a matsayin na al'ada ko mara kyau.

4. Sakamakon Gwaji & Aiki

Takardar ta nuna fifikon CRRN akan ƙirar gargajiya kamar na'urori masu sarrafa kansa na al'ada (AE), Na'urori masu Sarrafa Kai masu Bambance-bambance (VAE), da ƙirar maimaitawa mafi sauƙi. Manyan sakamako sun haɗa da:

• Mafi Girman Daidaiton Gano Abubuwan da ba su dace ba: CRRN ta sami mafi girman ma'auni na aiki (misali, maki-F1, AUC-ROC) akan bayanan SPI na ainihi idan aka kwatanta da ma'auni.
• Ingantaccen Rarraba Abubuwan da ba su dace ba: Ƙirar tana samar da "taswirar abin da ba a saba gani ba" wacce ke gano fakitoci marasa kyau a cikin PCB, yana ba da bincike mai fassara. An tabbatar da wannan taswirar ta hanyar aikin rarrabuwa na lahani na bugu na biyu, yana nuna babban ikon nuna bambanci.
• Ƙarfi ga Jerin Gwano masu Tsayi: Tsarin ST-Attention ya ba da damar ingantaccen koyo a kan mahallin lokaci mai tsayi inda wasu ƙirar suka kasa.

Bayanin Ginshiƙi: Taswirar sandar hasashe za ta nuna CRRN ta fi AE, VAE, da LSTM-AE a cikin sharuddin Yankin Ƙarƙashin Lanƙwasa (AUC) don gano abubuwan da ba su dace ba akan bayanan SPI.

5. Tsarin Nazari & Nazarin Lamari

Aiwatar da Tsarin (Misali mara Code): Ka yi la'akari da yanayin da stencil na SPP ya fara toshewa a hankali bayan lokaci. SPI na gargajiya zai iya yiwa fakitoci alama kawai idan yawansu ya faɗi ƙasa da ƙofa mai tsayi. Duk da haka, CRRN za ta sarrafa jerin aunawar SPI na duk fakitoci. Tana koyon alaƙar al'ada tsakanin yawan fakitoci a fadin allo da kuma cikin lokaci. Toshewar a hankali tana gabatar da jinkirin motsi mai alaƙa da sararin samaniya (misali, fakitoci a wani yanki na musamman suna nuna ci gaba mai ƙasa). CSTM na CRRN yana ɗaukar wannan bambancin tsarin lokaci da wuri, kuma kuskuren sake gini yana haɓaka kafin fakitoci guda ɗaya su keta ƙofar ƙaƙƙarfan ƙofa, yana ba da damar kulawa na tsinkaya. Tsarin ST-Attention yana taimakawa haɗa abin da ba a saba gani ba na yanzu zuwa jihohin mai shiryawa daga sa'o'i da suka gabata lokacin da motsin ya fara.

6. Ayyuka na Gaba & Hanyoyin Bincike

• Gano Abubuwan da ba su dace ba ta Hanyoyi daban-daban: Haɗa CRRN tare da bayanai daga wasu na'urori masu auna firikwensin (misali, tsarin gani, na'urori masu auna matsa lamba a cikin bugu) don cikakken tagwayen dijital na masana'anta.
• Koyo game da Abubuwan da ba su dace ba tare da ƴan Misalai/ba tare da Misalai ba: Daidaita ƙirar don gane sabbin nau'ikan lahani da ba a gani ba tare da ƙananan misalan da aka yiwa lakabi, watakila ta amfani da dabarun koyo-meta.
• Turawa zuwa gefe: Inganta CRRN don ƙididdiga na ainihi akan na'urori a gefe a cikin layin samarwa don ba da damar amsa da sarrafawa nan take.
• Bayanin Ƙirƙira na Ƙarya: Yin amfani da mai fassara don samar da "gyare-gyaren" nau'ikan al'ada na shigarwar da ba ta dace ba, yana ba ma'aikata bayani mai haske game da yadda allo ya kamata ya yi kama.

7. Nassoshi

1. 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 Binciken Kwalliyar Solder. IEEE Transactions on Cybernetics.
2. Goodfellow, I., et al. (2014). Nets na Adawa na Generative. Ci gaba a cikin Tsarin Bayanai na Jijiya.
3. Vaswani, A., et al. (2017). Hankali shine Duk Abin da Kuke Bukata. Ci gaba a cikin Tsarin Bayanai na Jijiya.
4. Zhu, J.-Y., et al. (2017). Fassarar Hotuna zuwa Hotuna mara Haɗin gwiwa ta amfani da Cibiyoyin Sadarwa masu Adawa na Ci gaba da Daidaituwa. Taron Ƙasa da Ƙasa na IEEE akan Kwamfutar Gani (ICCV).
5. Rahotannin Ƙaddamar da Masana'antar Lantarki ta Duniya (iNEMI) kan yanayin fasahar SMT.

8. Nazarin Kwararru & Bita Mai mahimmanci