Nunin Farko na 512-Launi Shift Keying Sigina Demodulation Ta Amfani da Neural Equalization don Sadarwar Kamara ta Optical

Tsarin Abubuwan Ciki

1. Gabatarwa & Bayyani

Wannan takarda ta gabatar da nunin gwaji na farko na watsa sigina 512-Launi Shift Keying (512-CSK) don Sadarwar Kamara ta Optical (OCC). Babban nasara shine demodulation mara kuskure a nisan mita 4 ta amfani da na'urar daukar hoto ta Sony IMX530 CMOS na kasuwanci tare da ruwan tabarau na mm 50 da kuma tsarin neural network na rarrabuwa mai yawa (NN) wanda ke aiki azaman mai daidaita sigina mara layi. Wannan aikin ya tura iyakar yawan bayanai na OCC sosai, yana motsawa daga tsare-tsaren 8, 16, ko 32-CSK da aka nuna a baya zuwa yankin babban matakin daidaitawa na launuka 512 (9 bits/symbol).

Binciken ya magance kalubale na asali a cikin OCC: cudanya tsakanin launuka (inter-color crosstalk) wanda ke haifar da rashin daidaiton hankalin launi na tacewar RGB na kamara, wanda ke karkatar da taurarin CSK da aka watsa bisa sararin launi na CIE 1931. Neural equalizer da aka gabatar yana daidaita wannan karkatarwar kai tsaye daga bayanan sensor na danye, yana ƙetare buƙatar ƙirar sarrafa sigina mai sarkakiya.

2. Tsarin Fasaha

2.1 Saitin Mai Karɓa & Shiri

Tsarin mai karɓa an gina shi a kusa da tsarin kamara na Sony Semiconductor Solutions wanda ke iya fitar da bayanan RGB na danye na bit 12 ba tare da wani sarrafa bayanai ba (demosaicing, cire amo, daidaita farin). Wannan bayanan danye yana da mahimmanci don dawo da launi daidai. Ana kama sigina ta ruwan tabarau na mm 50 daga na'urar watsa jeri ta LED mai tsari 8x8 (panel cm 6.5). Ana fara canza ƙimar RGB da aka karɓa zuwa madaidaitan launi na CIE 1931 (x, y) ta amfani da matrix na canza sararin launi na yau da kullun kafin a shigar da su cikin neural equalizer.

2.2 Tsarin Neural Network Equalizer

Zuciyar tsarin demodulation shine neural network mai yawan lakabi. Manufarsa ita ce aiwatar da daidaitawa mara layi, yana taswirar madaidaitan (x, y) da aka karɓa da aka karkatar zuwa alamar da aka watsa mai yuwuwa (don 512-CSK).

• Sashen Shigarwa: raka'a 2 (madaidaitan launi x, y).
• Sashe na Boye: N_h sassa tare da N_u raka'a kowanne (cikakkun bayanai game da tsarin an nuna amma ba a ƙididdige su gaba ɗaya a cikin ɓangaren da aka ɗauko ba).
• Sashen Fitarwa: M = 9 raka'a, wanda ya dace da bit 9 na alamar 512-CSK. An horar da hanyar sadarwa don rarrabuwa mai yawan lakabi.

Hanyar sadarwa tana fitar da rarraba yuwuwar bayan $p(1|x, y)$ ga kowane bit. An ƙididdige Ma'aunin Log-Likelihood Ratio (LLR) $L_i$ daga waɗannan yuwuwar kuma daga baya na'urar warware LDPC (Low-Density Parity-Check) ta warware shi don gyaran kuskure na ƙarshe.

2.3 Taswirar Taurari 512-CSK

An sanya alamomi 512 a cikin gamut na CIE 1931 na mai watsa RGB-LED da dabara. Taswirar ta fara ne daga kusurwar da ta dace da launin shuɗi na farko $(x=0.1805, y=0.0722)$ kuma ta cika sararin da ke akwai ta "hanyar triangular." Wannan yana nuna ingantaccen algorithm na tattarawa don haɓaka nisan Euclidean tsakanin wuraren taurari a cikin gamut na launi na zahiri, wanda yake da mahimmanci don rage yawan kuskuren alama.

3. Sakamakon Gwaji & Bincike

3.1 Aikin BER vs. Girman Jerin LED

Gwajin ya bambanta adadin LED masu aiki a cikin jerin mai watsawa daga 1x1 zuwa 8x8. Wannan yana canza ƙarfin haske da yankin da sigina ke mamaye a kan sensor ɗin hoto. An kimanta halayen Bit Error Rate (BER) akan wannan mabambanci. Nasarar aiki mara kuskure tana nuna ƙarfin neural equalizer a cikin ƙarfin sigina daban-daban da aka karɓa da kuma bayanan sarari. Amfani da cikakken tsari na 8x8 yana ba da mafi kyawun aiki ta hanyar matsakaicin pixel da yawa da rage tasirin amo.

3.2 Kwatantawa da Aikin Baya

Takardar ta haɗa da taƙaitaccen hoto (Hoto 1(c)) wanda ke kwatanta wannan aikin da nunin OCC-CSK na baya. Maɓambanta masu mahimmanci sune:

• Matsayin Daidaitawa: 512-CSK ya zarce 8-CSK [1], 16-CSK [2,3], da 32-CSK [4,5] da aka ruwaito a cikin ayyukan gwaji na baya.
• Nisa: Aikin mita 4 yana da gasa, musamman idan aka yi la'akari da babban matakin daidaitawa. Yana tsakanin nunin babban mataki na ɗan gajeren zango (3-4 cm) da nunin ƙananan mataki na tsayin zango (80-100 cm).
• Dabarar: Amfani da neural network don daidaitawa mara layi kai tsaye daga bayanan sensor na danye wata sabuwar dabara ce kuma mai yuwuwar zama mafi gabaɗaya idan aka kwatanta da dabarun biyan diyya na tushen ƙira.

4. Bincike na Cibiyar & Fassarar Kwararru

Hankali na Cibiya: Wannan takarda ba kawai game da cimma adadin launuka mafi girma ba ne; yana da dabara daga ƙirar farko ta kimiyyar lissafi zuwa koyo na farko na bayanai a cikin dawo da sigina na gani. Marubutan sun yarda a fakaice cewa madaidaicin bututun karkatarwa mara layi a cikin kamara (cudanya ta tacewa, rashin layi na sensor, kayan aikin ruwan tabarau) ana iya sarrafa su da kyau ta hanyar mai kiyasin aiki na duniya (neural network) fiye da ƙirar bincike da aka samo amma ba ta cika ba. Wannan yayi daidai da sauyin da ake gani a wasu fagage kamar sadarwar mara waya, inda ake ƙara amfani da Deep Learning don daidaita tashoshi da gano alamomi a cikin tashoshi masu sarkakiya, marasa layi.

Kwararar Hankali: Hankali yana da ban sha'awa: 1) Ana buƙatar babban matakin CSK don kayan aiki. 2) Babban matakin CSK yana da matukar hankali ga karkatar launi. 3) Karkatar launi na kamara yana da sarkakiya kuma mara layi. 4) Saboda haka, yi amfani da mai biyan diyya mara layi (NN) wanda aka horar da shi har zuwa ƙarshe akan bayanan gaske. Amfani da bayanan sensor na danye wani babban nasara ne—yana ba neural network mafi girman adadin bayanan da ba a canza ba kafin kowane ISP na kamara (Mai Sarrafa Sigina na Hoto) ya gabatar da nasa, sau da yawa na keɓaɓɓu kuma ba za a iya juyawa ba, canje-canje. Wannan hanya tana tunawa da falsafar a cikin daukar hoto na zamani na lissafi, inda algorithms ke aiki akan bayanan sensor na danye don mafi girman sassauci.

Ƙarfi & Kurakurai: Babban ƙarfi shine babban tsalle a cikin ingantaccen yanayin launi, gwaji yana tabbatar da abin da a baya yankin kwaikwayo ne kawai. Neural equalizer yana da kyau kuma mai ƙarfi. Duk da haka, kuskuren—na gama gari ga yawancin takardun sadarwa na tushen ML—shine yanayin "akwatin baƙi". Takardar ba ta shiga cikin binciken tsarin NN, girman bayanan horo, ko ikon gabaɗaya zuwa kamarori daban-daban, ruwan tabarau, ko yanayin hasken muhalli ba. Shin hanyar sadarwa za ta buƙaci sake horo ga kowane sabon samfurin mai karɓa? Kamar yadda aka lura a cikin wani muhimmin bita kan koyon inji don sadarwa ta O'Shea & Hoydis, aikin mai karɓa na tushen DL ya dogara da ƙarfin ƙarfi da daidaitawa ga yanayin canji. Bugu da ƙari, nisan mita 4, yayin da yake da kyau, har yanzu yana nuna iyaka na ƙarfi/SNR. Dogaro da na'urar warware LDPC don aikin ƙarshe mara kuskure yana nuna cewa yawan kuskuren alama a fitarwar NN ba sifili ba ne, yana haifar da tambayoyi game da aikin equalizer na kadaitaka a ƙarƙashin ƙananan SNR.

Hankali Mai Aiki: Ga masu bincike, mataki na gaba a bayyane shine buɗe akwatin baƙi. Bincika tsarin NN (CNN na iya sarrafa bambance-bambancen sarari a kan sensor mafi kyau), bincika ƙaramin harbi ko koyon canja wuri don daidaitawa da sabbin kayan aiki, da haɗa equalizer tare da gyaran kuskure na gaba a cikin tsari mafi gabaɗaya, kamar turbo. Ga masana'antu, wannan aikin yana nuna alamar cewa babban yawan bayanai, VLC mara flicker ta amfani da kamarori na kasuwanci yana matsawa kusa da gaskiya. Haɗin gwiwa tare da Sony don sensor yana da mahimmanci; kasuwancin zai dogara da saka irin wannan sarrafa neural cikin sauƙi a cikin ASIC na kamara ko amfani da na'urorin haɓaka AI na na'ura da ke cikin wayoyin hannu. Ma'auni da za a kallo shine IEEE 802.15.7r1 (OCC), kuma gudummawar irin wannan na iya yin tasiri kai tsaye ga ci gabanta.

5. Cikakkun Bayanai na Fasaha & Tsarin Lissafi

Canza Sararin Launi: Ana yin canji daga ƙimar RGB da aka karɓa (daga sensor ɗin danye) zuwa madaidaitan xy na CIE 1931 ta amfani da matrix na yau da kullun da aka samo daga halayen launi na sensor dangane da mai lura da ma'aunin CIE. Takardar ta ba da takamaiman matrix da aka yi amfani da ita: $$ \begin{pmatrix} x \\ y \end{pmatrix} = \begin{pmatrix} 0.4124 & 0.3576 & 0.1805 \\ 0.2126 & 0.7152 & 0.0722 \end{pmatrix} \begin{pmatrix} R \\ G \\ B \end{pmatrix} $$ Wannan sauƙaƙan canji ne na layi. A aikace, ƙirar mafi daidai na iya buƙatar taswira mara layi ko matrix da aka keɓance ga takamaiman tacewar launi na sensor.

Fitar Neural Network zuwa LLR: NN mai yawan lakabi yana fitar da yuwuwar $p_i(1|x, y)$ cewa bit na $i$-th (daga cikin 9) shine '1'. An ƙididdige Ma'aunin Log-Likelihood Ratio (LLR) $L_i$ don wannan bit, wanda aka ciyar wa na'urar warware LDPC, kamar haka: $$ L_i = \log \left( \frac{p_i(1|x, y)}{1 - p_i(1|x, y)} \right) $$ Babban LLR mai kyau yana nuna babban amincewa cewa bit shine 1, babban ƙimar mara kyau yana nuna babban amincewa cewa shi ne 0.

6. Tsarin Bincike & Misalin Lamari

Tsari: Bututun "Mai Karɓa da aka Koya" don OCC

Wannan binciken ya misalta tsarin ƙirar "mai karɓa da aka koya" na zamani wanda za a iya amfani da shi fiye da OCC. Ana iya rarraba tsarin zuwa tubalan jeri, masu daidaitawa:

1. Samun Bayanai Masu Sanin Kayan Aiki: Kama sigina a wuri mafi wuri, mafi yawan danye a cikin sarkar sarrafawa (misali, bayanan RAW na sensor, samfuran RF I/Q).
2. Preprocessing Mai Bambanci: Aiwatar da ƙaramin, preprocessing mai mahimmanci (misali, canza sararin launi, daidaitawa) ta hanyar da za ta ba da damar kwararar gradient idan an horar da shi har zuwa ƙarshe.
3. Cibiyar Neural Network: Yi amfani da neural network (MLP, CNN, Transformer) don aiwatar da aikin demodulation/equalization na cibiya. An horar da hanyar sadarwa tare da aikin asara wanda ke rage yawan kuskuren alama ko bit kai tsaye, sau da yawa ta amfani da asarar giciye don ayyukan rarrabuwa.
4. Warwarewa na Hybrid: Haɗa fitarwa mai laushi na hanyar sadarwar jijiyoyi (yuwuwar, LLRs) tare da na'urar warware gyaran kuskure na zamani, wanda ba na jijiyoyi ba (kamar na'urar warware LDPC ko Polar code). Wannan yana haɗa sassaucin koyo tare da ingantaccen ka'idar lambobi na gargajiya.

Misalin Lamari Ba Lamba Ba: Aiwatar da Tsarin zuwa Ƙarƙashin Ruwa VLC

Yi la'akari da amfani da wannan tsari ɗaya zuwa Sadarwar Hasken Gani Ƙarƙashin Ruwa (UVLC), wanda ke fama da mummunan lahani na tashoshi kamar watsawa da rugujewar da ke haifar da rugujewa. Za a iya gina "Mai Karɓa da aka Koya" don UVLC kamar haka:

• Mataki 1: Yi amfani da mai gano hoto mai sauri ko kamara mai ɗaukar jerin ƙarfi na danye.
• Mataki 2: Yi preprocessing don ware yankin sigina mai ban sha'awa da aiwatar da daidaitawa mai ƙarfi.
• Mataki 3: Horar da 1D Convolutional Neural Network (CNN) ko Recurrent Neural Network (RNN) kamar LSTM akan wannan jerin bayanan danye. Aikin hanyar sadarwa shine daidaita tasirin tashoshi na lokaci-lokaci da kuma cire taswirar alamomi. Za a tattara bayanan horo a ƙarƙashin yanayi daban-daban na turbidity na ruwa da tashin hankali.
• Mataki 4: Hanyar sadarwa tana fitar da yanke shawara mai laushi don na'urar warware FEC, yana ba da damar sadarwa mai ƙarfi a cikin tashoshi mai ƙarfi sosai inda ƙididdigar tashoshi ta gargajiya ta kasa.

7. Aikace-aikacen Gaba & Hanyoyin Bincike

• Li-Fi na Tushen Wayar Hannu: Manufa ta ƙarshe ita ce haɗa wannan fasaha cikin wayoyin hannu don amintaccen canja wurin bayanai mai sauri tsakanin takwarorinsu ko sanya wuri a cikin gida tare da daidaiton santimita, ta amfani da kayan aikin kamara da ke akwai.
• Sadarwar Motoci V2X: Yin amfani da fitilun mota/fitilun wutsiya da kamarori don sadarwar Motoci-zuwa-Kowane Abu (V2X), yana ba da ƙarin hanyar haɗin bayanai mai ƙarfi wanda ya dace da DSRC/C-V2X na tushen RF.
• AR/VR da Metaverse Interfaces: Ba da damar hanyoyin haɗin bayanai masu ƙarancin jinkiri, babban bandeji tsakanin tabarau na AR da kayan aiki ko tsakanin na'urori don daidaitattun abubuwan gama gari.
• Hanyoyin Bincike:
  1. Tsarin da aka Koya har zuwa Ƙarshe: Bincika haɗin gwiwar daidaita siffar taurarin mai watsawa (ta hanyar neural network) da equalizer na mai karɓa, kama da manufar sadarwar "autoencoder".
  2. Ƙarfi & Daidaitawa: Haɓaka ƙirar mai karɓa na jijiyoyi waɗanda suka ƙarfi ga nau'ikan kamara daban-daban, hasken muhalli, da rufe wani ɓangare. Wannan yana da mahimmanci ga ƙoƙarin daidaitawa kamar IEEE 802.15.7.
  3. OCC Mai Girma Mai Girma: Haɗa babban matakin CSK tare da dabarun rufewa ko canza sarari ta amfani da kamarori masu saurin firam ɗin firam ko na tushen al'amura don karya shingen Gbps.
  4. Sadarwar Ma'ana: Matsawa bayan dawo da bit, ta amfani da hanyar haɗin OCC don watsa bayanan ma'ana (misali, masu gano abu, bayanan taswira) kai tsaye, yana inganta don nasarar aiki maimakon yawan kuskuren bit.

8. Nassoshi

1. H.-W. Chen et al., "8-CSK watsa bayanai sama da 4 cm," Majalisar Taro/Mujallar da ta dace, 2019.
2. C. Zhu et al., "16-CSK sama da 80 cm ta amfani da LED quadrichromatic," Majalisar Taro/Mujallar da ta dace, 2016.
3. N. Murata et al., "16-digital CSK sama da 100 cm bisa IEEE 802.15.7," Majalisar Taro/Mujallar da ta dace, 2016.
4. P. Hu et al., "Tri-LEDs na tushen 32-CSK sama da 3 cm," Majalisar Taro/Mujallar da ta dace, 2019.
5. R. Singh et al., "Tri-LEDs na tushen 32-CSK," Majalisar Taro/Mujallar da ta dace, 2014.
6. O'Shea, T., & Hoydis, J. (2017). "Gabatarwa ga Deep Learning don Sashen Jiki." IEEE Transactions on Cognitive Communications and Networking. (Tushen mai iko na waje akan ML don sadarwa)
7. Ma'aunin IEEE don Gida da Manyan Hanyoyin Sadarwa--Sashi na 15.7: Gajeren Zango na Sadarwar Wireless ta Optical. IEEE Std 802.15.7-2018. (Ma'auni mai iko na waje)
8. Hukumar Kula da Hasken Ƙasa da Ƙasa (CIE). (1931). Hukumar kula da hasken duniya ta ci gaba, 1931. Cambridge: Cambridge University Press. (Tushen mai iko na waje don kimiyyar launi)
9. Kamfanin Sony Semiconductor Solutions. IMX530 Sensor Datasheet. (Tushen kayan aiki mai iko na waje)
10. Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep Learning. MIT Press. (Tushen mai iko na waje akan hanyoyin sadarwar jijiyoyi)