Wakati ukarabati, ubomoaji au matumizi tena ya vifaa vya jengo lililopo yanapopangwa, uso unaoonekana wa kuta pekee hautoi taarifa za kutosha. Nyuma ya rangi au drywall kunaweza kuwa na studs za chuma au mbao, insulation ya fiberglass, saruji, matofali, concrete blocks, conduits za umeme, sockets, ducts na tabaka mbalimbali. Kutokujulikana kwa vipengele hivi vilivyofichwa hufanya iwe vigumu kuandaa building information models, kutengeneza material passports, kuhesabu embodied carbon na kufanya maamuzi ya reuse.

Katika utafiti huu, workflow ya multimodal vision-language model ilitengenezwa ili kutabiri vipengele vilivyofichwa ndani ya ukuta bila kuuvunja. Mfumo hutathmini pamoja data za maandishi kutoka GIS na building records, floor plans, picha za kawaida za RGB na thermal camera images. Matokeo ya model si class label rahisi kama “ukuta huu una metal studs”. Badala yake, mfumo hutengeneza structured wall assembly graph inayojumuisha mpangilio wa tabaka za ukuta, materials, idadi ya studs, insulation filling na baadhi ya mechanical-electrical-plumbing components.

Workflow kwanza hutengeneza compact prediction yenye idadi ndogo ya main layers, kisha huipanua kuwa detailed nodes, huongeza MEP components kama electrical conduit na outlet, na hatimaye hutumia ensemble stage inayounganisha outputs nyingi za model ili kutoa prediction ya mwisho. Muundo huu wa JSON unaweza baadaye kubadilishwa kuwa 3D model, BIM wall type, material inventory au embodied carbon calculation.

Mbinu ilitathminiwa kwanza kwenye experimental wall mock-up ambayo muundo wake wa ndani ulijulikana kwa uhakika, kisha kwenye jumla ya 22 real walls katika commercial office buildings nne huko Boston, San Francisco, Toronto na Birmingham. Kwa real walls, graph-based accuracy ya jumla ilikuwa %54 wakati GIS pekee ilipotumika; lakini baada ya kuongezwa floor plan, RGB photo na thermal image iliongezeka hadi %76. Kwa enrichment hiyo hiyo ya data, embodied carbon relative error ilishuka kutoka %104 hadi %26.

Hata hivyo, mbinu haioni ndani ya ukuta kimwili. Model hufanya probabilistic inference kutokana na documents zilizopo, visible clues, thermal patterns na common construction knowledge iliyojifunza awali. Reference labels katika real-wall dataset pia zilitayarishwa kwa kutumia Revit models, drawings, thermal scans, expert interpretations na physical observations badala ya destructive verification. Kwa hiyo, matokeo yanapaswa kuonekana si kama utambuzi wa uhakika na unaodhaminiwa wa yaliyomo ndani ya ukuta, bali kama experimental predictions zinazoweza kusaidia maamuzi kabla ya destructive inspection.

Tatizo kuu la utafiti ni nini?

Katika jengo jipya, materials zinazounda kila ukuta mara nyingi zimeandikwa katika design drawings na BIM files. Lakini katika majengo ya zamani, documents hizi zinaweza kutokuwepo, kuwa incomplete au kutowakilisha renovations zilizofanywa kadri muda ulivyopita. Tofauti kubwa zinaweza kutokea kati ya initial design na hali halisi ya sasa.

Kwa mfano, interior partition inaweza kuonekana nje kama painted drywall tu. Lakini ndani ya ukuta kunaweza kuwa na:

• Wood au metal stud system,
• Fiberglass au insulation nyingine,
• Electrical conduits na outlets,
• Water pipes au air ducts,
• Horizontal metal channels,
• Concrete, brick au concrete block layers

Vipengele hivi haviwezi kuonekana moja kwa moja kutoka nje.

Njia ya jadi ya kujua kwa uhakika muundo wa ndani wa ukuta ni kufungua surface, kuchimba, kutumia borescope au kuondoa sehemu ya ukuta. Mbinu hizi zinahitaji muda, ruhusa, labor na gharama; na katika jengo linalotumika, mara nyingi haiwezekani kuzitumia kwa kila ukuta.

Hata hivyo, kwa building renovation, selective demolition na reuse decisions, ni muhimu kujua materials zilizomo ndani ya ukuta. Ikiwa idadi ya metal studs haijulikani, kiasi cha steel kinachoweza kurejeshwa hakiwezi kukadiriwa kwa uhakika; ikiwa aina ya insulation haijulikani, reuse potential huwa haijulikani; na ikiwa layer thicknesses hazijulikani, embodied carbon calculation haiwezi kukadiriwa kwa kuaminika.

Kwa nini ni muhimu kwa material passport na circular construction economy?

Material passport ni digital record inayofafanua ni materials zipi zipo katika jengo, kiasi chake, mahali zilipo, sifa zake na thamani inayowezekana ya reuse. Ili record kama hii iandaliwe, si surfaces zinazoonekana pekee, bali hidden building components pia zinahitaji kujulikana.

Katika circular construction approach inayotumiwa na utafiti, hatua ya kwanza ni “identification”. Kabla ya kuamua kama material inaweza kuondolewa, kutumika tena, kurejelewa au kutengwa kama waste, ni lazima ijulikane ikiwa kweli ipo kwenye jengo na katika physical context gani.

Mbinu inayopendekezwa inalenga kukadiria yaliyomo ndani ya kuta bila destructive testing ili kusaidia michakato ifuatayo:

• Kuandaa material passports kwa majengo yaliyopo,
• Kuhesabu embodied carbon,
• Kutambua kuta zinazoweza kuhifadhiwa wakati wa renovation,
• Kuandaa selective demolition plans,
• Kukadiria metal, wood na insulation materials zinazoweza kutumika tena,
• Kuhamisha hali iliyopo kwenda BIM model.

Utafiti huu unatofautianaje na approaches za awali?

Sehemu kubwa ya computer vision methods zilizopo inalenga kutambua materials zinazoonekana kwenye facade, kutenganisha building elements kutoka point cloud au kubadilisha scanned surfaces kuwa BIM objects. Systems hizi zinaweza kutambua visible components kama brick, concrete, window au door; lakini hazionyeshi moja kwa moja layers zilizofichwa ndani ya finished wall.

Narrow-scope machine learning systems kwa kawaida huhitaji fixed input format. Kwa mfano, model iliyofundishwa kwa thermal images za resolution maalum inaweza kufanya vibaya katika jengo lisilo na floor plan au linalotumia camera tofauti. Pia ni vigumu sana kutengeneza large and reliable labeled dataset ya hidden wall assemblies, kwa sababu kila label inahitaji physical verification ya yaliyomo ndani ya ukuta.

Utafiti huu hautengenezi foundation learning algorithm mpya. Watafiti wanaielezea wazi kama “system-level framework”. Mchango mkuu ni kutumia pre-trained general-purpose vision-language models kwa carefully designed prompts na kubadilisha output kuwa wall assembly graph inayoweza kubadilishwa kuwa BIM.

Ni data gani huingizwa kwenye mfumo?

Workflow hukubali aina nne kuu za data. Si lazima zote zipatikane kwa kila ukuta.

Gharama ya kukusanya data pia huongezeka kwa mpangilio huu. GIS record mara nyingi inaweza kupatikana kutoka public sources, ilhali thermal imaging inahitaji special camera, suitable temperature difference na site visit. Kwa hiyo, system imeundwa pia kufanya kazi na incomplete data.

Wall assembly graph ni nini?

Kwa kila ukuta, wall assembly graph inayowakilishwa na \(S_i\) hutengenezwa. Muundo huu ni rooted tree iliyopangwa katika JSON format. Root key ni “layers”. Tabaka za ukuta huonyeshwa kwa mpangilio kutoka outer surface ya upande mmoja hadi outer surface ya upande wa pili:

\[ L=[l_1,l_2,\ldots,l_n] \]

Hapa \(l_1\) ni outermost layer upande mmoja wa ukuta; na \(l_n\) ni last layer upande wa pili.

Kila layer ina angalau sifa mbili kuu:

• material: Jina la material kama drywall, metal stud frame au fiberglass insulation,
• confidence: Confidence ya model katika prediction hiyo, iliyoelezwa kati ya 0 na 1.

Ikiwa layer ina component zaidi ya moja, additional information hutumika. Kwa mfano, ikiwa kuna fiberglass insulation ndani ya metal stud frame, idadi ya studs na cavity filling huwakilishwa ndani ya hierarchical node hiyo hiyo.

Mfano uliorahisishwa uliotolewa katika utafiti ni:

{
  "layers": [
    {
      "material": "Drywall",
      "confidence": 0.95
    },
    {
      "material": "Metal Stud Frame",
      "confidence": 0.85,
      "number": 4,
      "filling": {
        "material": "Fiberglass Insulation"
      }
    },
    {
      "material": "Drywall",
      "confidence": 0.95
    }
  ]
}

Mfano huu unaelezea typical interior partition yenye drywall pande zote mbili, metal studs nne na fiberglass insulation kati ya studs.

Faida ya graph approach ni kwamba model hailazimiki kutabiri moja kwa moja exact 3D coordinates za components zote. Kwanza materials na hierarchical relationships kati yao hufafanuliwa; baadaye 3D geometry inaweza kutengenezwa kwa kutumia actual dimensions za ukuta.

Four-stage workflow inafanyaje kazi?

Hatua ya 1: Compact graph prediction

Katika hatua ya kwanza, vision-language model kwa kawaida hutengeneza simplified wall prediction yenye main layers tatu hadi tano. Lengo hapa si kutatua kila detail, bali kutambua basic wall type.

Katika hatua hii model hujaribu kujibu maswali kama:

• Ukuta ni interior partition au exterior wall?
• Outer layer ni drywall, brick, concrete au concrete block?
• Kuna stud frame?
• Studs ni wood au metal?
• Kuna takribani studs ngapi?
• Kuna insulation kati ya studs?

Hatua ya 2: Graph expansion

Katika prediction ya kwanza, “frame yenye metal studs nne na fiberglass filling” inaweza kuwa compound node moja. Katika hatua ya pili, Python script isiyotumia AI hugawanya node hii kuwa detailed components.

Kwa njia hii, kila metal stud, insulation region na horizontal channel huwa separate node. Expansion hii inahitajika ili katika hatua inayofuata electrical na plumbing components ziweze kuunganishwa kwenye specific cells.

Hatua ya 3: MEP component prediction

MEP inawakilisha mechanical, electrical na plumbing components. Katika hatua hii model huchunguza hasa insulation regions na kutabiri uwepo wa vitu vifuatavyo:

• Electrical conduits,
• Outlets,
• Water pipes,
• Air ducts.

Prompt iliyotumiwa na watafiti inaomba components hizi ziongezwe tu ikiwa kuna strong visual evidence. Hata hivyo, matokeo yalionyesha kwamba model ilikuwa bora zaidi katika kutabiri uwepo wa component kuliko kubainisha component hiyo iko kati ya studs zipi.

Hatua ya 4: Ensemble strategy

Model inapokimbizwa mara nyingi kwa inputs zilezile, inaweza kutoa graphs tofauti. Hii inatokana na stochastic variability katika outputs za vision-language models.

Kwa kuwa graphs zinaweza kuwa na idadi tofauti ya nodes na hierarchies tofauti, simple majority voting haikutumika. Badala yake, model huchunguza multiple candidate graphs kwa pamoja na kutafsiri common components na outlier predictions kupitia generative merging process ili kutoa final graph.

Kwa mfano, ikiwa predictions nne kati ya tano zina metal stud lakini idadi ya studs inatofautiana kama tatu, nne na tano, system haifanyi text matching pekee; hutafsiri candidates wote structurally na kutoa final graph.

Ni vision-language model gani ilitumika?

Utafiti unasema majaribio yote yalifanywa kwa GPT-5.4 vision-language model kupitia OpenAI API. Default sampling settings zilitumika; temperature, top-p, seed sampling au random seed hazikuwekwa wazi.

Kila wall na data combination ilipitia four-stage workflow kwa kujitegemea mara tano. Mean na standard deviation za repetitions hizi ziliripotiwa katika matokeo.

Katika additional comparison, GPT-5.2 na GPT-5.4 results zilitolewa tofauti. Comparison hii inaelezea kwa nini kuna accuracy sequences mbili tofauti katika muhtasari na sehemu fulani za discussion:

Katika baadhi ya sehemu za maandishi accuracy inaelezwa kuongezeka kutoka %47 hadi %74, na katika nyingine kutoka %54 hadi %76. Tables zinaonyesha kwamba tofauti hii inatokana na model versions mbili tofauti.

Jaribio 1: Wall mock-up yenye contents zinazojulikana kwa uhakika

Katika jaribio la kwanza, physical wall mock-up ilijengwa kuwakilisha typical partition wall katika hypothetical office building huko Boston iliyojengwa mwanzoni mwa miaka ya 2000.

Ukuta ulikuwa na components zifuatazo:

• Drywall pande zote mbili,
• Vertical metal studs nne,
• Fiberglass insulation kati ya studs,
• Electrical conduits na outlets,
• Additional hidden components kama pipe na air duct.

Faida ya jaribio hili ni kwamba watafiti walijua kikamilifu contents halisi za ukuta. Kwa hiyo, predicted graph iliweza kulinganishwa moja kwa moja na ground-truth graph.

GIS data iliyotumiwa kwa mock-up haikutoka kwenye real public record; ilitengenezwa synthetically kulingana na experimental scenario. Floor plan ilitoka kwenye drawing halisi ya Boston building iliyowakilishwa na mock-up. Kwa hiyo, watafiti wanaelezea jaribio la kwanza si kama evidence ya real-world performance, bali kama controlled ablation na calibration study.

Matatizo yaliyokumbana nayo katika thermal imaging

Kwa kuwa wall mock-up ilikuwa katika independent controlled indoor environment, ilikuwa vigumu kuunda realistic temperature difference kati ya pande mbili za ukuta. Heating pads zilitumika kujaribu kuunda temperature gradient, lakini fiberglass insulation ilizuia heat transfer kwa kiasi kikubwa.

Thermal image iliyopigwa kupitia finished drywall ilikuwa na resolution ya 382 × 288 pixels pekee na details zilikuwa dhaifu. Ili kupata thermal image iliyo wazi zaidi, additional image ilichukuliwa kwa kusogeza camera karibu na studs na insulation kabla ya last drywall layer kufungwa.

Hili ni limitation muhimu la jaribio. High-resolution thermal input iliyotumika katika mock-up experiment haiwezi kupatikana kila wakati katika finished real wall.

Katika thermal images:

• Studs na baadhi ya electrical conduits ziliweza kutambulika kwa kiasi,
• Air duct na pipe zilizokuwa nyuma ya insulation hazikuonekana kwa kiasi kikubwa.

Hii inaonyesha kwamba thermal camera si “X-ray” inayoweza kuona depth yote ya ukuta. Camera hutambua indirect effects kwenye surface temperature pekee.

Matokeo ya mock-up experiment

Katika controlled wall mock-up, baadhi ya combinations za data types mbili pekee zilifikia accuracy zaidi ya %80. GIS na RGB photo zikitumika pamoja, graph accuracy ya %93 ilipatikana. Baadhi ya three-modal combinations zilibaki katika %82–88.

Matokeo haya yanaonyesha jambo muhimu: data nyingi zaidi hazimaanishi matokeo bora kila wakati. Ikiwa data type mpya inapingana na clues zilizopo, inaweza kufanya interpretation ya model kuwa ngumu zaidi.

Kwa mock-up, all data types na suitable ensemble zilipotumika, baadhi ya runs ziliripoti:

• Graph-based accuracy 1,00,
• Embodied carbon relative error %0

Hata hivyo, matokeo haya hayamaanishi kwamba system ni %100 sahihi katika real walls zote. Wall mock-up pia ilikuwa calibration example ambapo prompts na parameters zilitengenezwa.

Jaribio 2: 22 real walls katika miji minne

Katika jaribio la pili, jumla ya 22 walls kutoka commercial office buildings nne zilichunguzwa:

Dataset ilijumuisha interior na exterior walls. Hata hivyo, examples zote zilitoka commercial office au mixed-use office buildings. Residential buildings, historic masonry, mass timber structures au non-Western local construction techniques hazikuthibitishwa katika dataset hii.

“True” contents za real walls zilibainishwaje?

Real building walls hazikuvunjwa kwa verification. Reference labels zilitayarishwa kwa kuchanganya sources zifuatazo:

• Existing Revit na BIM models,
• Original na renovation-era floor plans,
• Thermal scans zilizotafsiriwa na wataalamu,
• Physical inspections za architecture professionals wanaojua construction history ya majengo.

Multi-source approach hii inatoa reference imara chini ya access constraints. Hata hivyo, watafiti pia wanakubali kwamba kwa kuwa physical opening haikufanywa, label uncertainty haiwezi kuondolewa kabisa.

Graph-based accuracy ilihesabiwaje?

Predicted graph \(G_P\) na reference graph \(G_{GT}\) zilibadilishwa kuwa rooted trees zenye material na orientation information. Tree edit distance kati ya miundo miwili ilihesabiwa kwa Zhang–Shasha algorithm:

\[ D(G_P,G_{GT})= \min_{(e_1,\ldots,e_k)\in\mathcal{P}(G_P,G_{GT})} \sum_{i=1}^{k}c(e_i) \]

Hapa:

• \(\mathcal{P}(G_P,G_{GT})\), ni seti ya possible edit paths zote zinazobadilisha predicted graph kuwa reference graph,
• \(e_i\), ni insertion, deletion au substitution operation,
• \(c(e_i)\), ni cost ya operation husika.

Node insertion na deletion cost iliwekwa 1,0. Material ikiwa wrong, substitution cost ilikuwa 1,0; ikiwa orientation pekee ndiyo wrong, cost ilikuwa 0,1.

Ili kulinganisha graphs za ukubwa tofauti, distance ilinormalishwa kwa upper bound ifuatayo:

\[ D_{\max}=|T_{GT}|+|T_P| \]

Hapa \(|T|\) ni total node count katika tree husika.

Final graph-based accuracy ilihesabiwa kama ifuatavyo:

\[ GBA(G_P,G_{GT})= 1-\frac{D(G_P,G_{GT})}{D_{\max}} \]

Thamani inapokaribia 1, predicted graph inakaribia reference structure. Kwa kuwa reversing wall layer order inaweza kutoa physically equivalent assembly, prediction ilitathminiwa katika original orientation na pia kwa reversing second-level children; accuracy kubwa zaidi ndiyo iliyoripotiwa.

Embodied carbon relative error ilihesabiwaje?

Kila predicted component ilipewa embodied carbon coefficient. Kwa mfano, carbon contribution ya drywall ilihesabiwa kwa relation ifuatayo:

\[ EC=D_{EC}\times H\times L \]

Hapa:

• \(D_{EC}\), ni embodied carbon intensity ya drywall kwa unit area,
• \(H\), ni wall height,
• \(L\), ni wall length.

Kwa walls zote, default height ilikuwa 10 feet, takribani 3,05 metres. Katika stud walls, length ilihesabiwa kwa kutumia reference stud count na assumption ya 16-inch stud spacing.

Carbon contribution ya studs ilihesabiwa kama ifuatavyo:

\[ EC=D_{stud}\times N \]

Hapa \(D_{stud}\) ni carbon coefficient ya stud moja; \(N\) ni idadi ya studs.

Relative error kati ya predicted total embodied carbon na reference value ilihesabiwa kama:

\[ ECRE(EC_P,EC_{GT})= \frac{|EC_P-EC_{GT}|}{|EC_{GT}|} \]

Hapa \(EC_P\) ni predicted embodied carbon; \(EC_{GT}\) ni reference embodied carbon. Thamani ya chini ina maana ya result bora.

Assumptions zilizotumika katika carbon calculation

Utafiti ulitumia coefficients tofauti kwa drywall, fiberglass, concrete, concrete block, brick, metal cladding, metal na wood studs, ducts, pipes na outlets. Sehemu kubwa ya values hizi ilitokana na average cradle-to-gate values katika EC3 database zinazojumuisha stages A1–A3.

Example coefficients ni:

Hizi si values zilizochukuliwa kutoka individual environmental product declarations za real walls zilizochunguzwa. Ni average coefficients zilizochaguliwa kwa standardized comparison. Kwa hiyo ECRE metric si exact life-cycle carbon calculation ya real building, bali experimental comparison inayonyesha wall component prediction errors zinavyoweza kupotosha carbon calculation.

Matokeo ya jumla katika real walls

Matokeo yanaonyesha kwamba jump kubwa zaidi ilitokea RGB photo ilipoongezwa. Floor plan iliongeza accuracy kutoka 0,54 hadi 0,57 tu, wakati RGB iliipeleka hadi 0,71. Thermal image iliongeza total accuracy hadi 0,76.

Katika embodied carbon error, transition kutoka GIS kwenda floor plan ilileta improvement kubwa zaidi. Error ilishuka kutoka %104 hadi %44. RGB ikaishusha hadi %28, na thermal image hadi %26.

Relative error ya zaidi ya %100 kwa GIS-only inaonyesha kwamba model inaweza kutabiri wrong wall type au kiasi kikubwa kupita kiasi cha high-carbon material.

Matokeo kwa kila jiji

Boston examples zilitoa average result ya juu zaidi. Watafiti hawahusishi city differences moja kwa moja na geography. Sababu kuu inafikiriwa kuwa baadhi ya buildings zinafanana zaidi au kidogo na common construction patterns zilizojifunzwa awali na model.

Kwa mfano, model ilikosa metal stud layer katika exterior wall moja ya San Francisco. Katika baadhi ya Birmingham partition walls zilizokuwa na drywall na concrete block, model ilihusisha stud frame kimakosa. Kwa upande mwingine, katika baadhi ya unusual exterior walls huko Birmingham, iliweza kutambua brick na insulation components kwa usahihi kutokana na visual evidence.

Utendaji katika interior na exterior walls

Interior partition walls mara nyingi zina repeatable assemblies kama drywall, studs na insulation, hivyo model ilifanya vizuri zaidi katika walls hizi. Exterior walls zinaweza kuwa na combinations mbalimbali za brick, concrete, cladding, air cavity, insulation na structural layers, hivyo zikaonyesha error kubwa zaidi.

Thermal image ilichangia kiasi gani?

Contribution ya thermal image inategemea image quality na temperature difference kati ya pande mbili za ukuta. Utafiti unasema sustained temperature difference ya angalau takribani 5–10 °C kupitia ukuta ni muhimu ili studs ziweze kuonekana vizuri.

Difference hii inaweza kutokea katika hali zifuatazo:

• Kati ya indoor na outdoor environments wakati wa winter,
• Exterior wall inapopashwa na jua,
• Asubuhi mapema baada ya heating usiku,
• Kwa temporary controlled heating.

Katika interior partitions, natural temperature difference mara nyingi ni ndogo. Katika hali hiyo thermal camera inaweza kutambua tu very small surface temperature differences karibu na studs.

Thermal image enhancement experiment

Watafiti walichakata thermal images kwa njia tatu tofauti na kulinganisha performance:

• Kama zilivyotoka kwenye camera software,
• Kwa common exposure, offset na gamma settings kwa kila city,
• Kwa manual adjustment ya kila image kwa Adobe Photoshop 2025.

Thermal image transformation iliwasilishwa kwa equations zifuatazo:

\[ x=img\cdot 2^{EXPOSURE}+OFFSET \]

\[ x=[\min(\max(x,0),1)]^{1/GAMMA} \]

\[ x=\min(\max(x,0),1) \]

Hapa:

• \(img\), ni initial thermal image array,
• \(EXPOSURE\), ni exposure correction,
• \(OFFSET\), ni brightness shift,
• \(GAMMA\), ni gamma value inayodhibiti tonal distribution,
• \(x\), ni processed image value.

Katika utafiti \(OFFSET=-0,23\) na \(GAMMA=0,53\) zilitumika. Exposure value ilikuwa 0,6 kwa San Francisco na Birmingham; 1,5 kwa Boston na Toronto.

Manual enhancement iliongeza accuracy; lakini city-specific automatic processing iliongeza carbon error kutoka %27 hadi %30. Hii inaonyesha kwamba image enhancement haileti faida kwa direction ileile katika kila metric na kila example.

Error analysis ilionyesha nini?

Watafiti waligawa errors katika categories sita:

1. Wrong identification ya outer boundary layer,
2. Wrong prediction ya structural au stud system,
3. Wrong stud count prediction,
4. Wrong identification ya stud kuwa metal au wood,
5. Wrong prediction ya uwepo wa MEP components,
6. Kushindwa kuweka MEP component katika correct stud bay.

Additional data types zilipunguza hasa structural-system na stud-type errors. Model iliyotumia GIS pekee ilitegemea general assumptions, kwa mfano prevalence ya metal studs katika commercial buildings baada ya 1980. Thermal images zilipoonyesha real stud pattern kwa kiasi, assumption errors kama hizi zilipungua.

MEP placement error haikupungua sana kadri data types zilivyoongezwa. Model inaweza kuelewa kwamba outlet au electrical conduit ipo, lakini haikuweza kubainisha kwa kuaminika iko kati ya studs zipi. Watafiti walihusisha hili na limited fine-grained spatial reasoning ya current VLMs.

Je, model kweli inafanya inference kutoka images, au inategemea construction patterns ilizojifunza awali?

Hili ni mojawapo ya maswali muhimu ya kisayansi katika utafiti. Vision-language models zimefundishwa awali kwa internet-scale text na images, hivyo zinabeba general knowledge kuhusu common construction systems. Kwa mfano, zikipewa context ya commercial office interior partition, model inaweza kuelekea kwenye assumption ya metal studs, fiberglass na drywall hata bila strong visual evidence.

Watafiti wanaona ongezeko la accuracy kutoka GIS-only hadi all-data condition kama evidence ya real multimodal inference. Kutambua kwa usahihi baadhi ya unusual brick na insulation configurations huko Birmingham kutoka visual data pia kunaonyesha kwamba model hairudii default template pekee.

Kwa upande mwingine, wrong prediction ya stud system katika Birmingham concrete-block walls inaonyesha kwamba common construction assumptions zilizojifunza katika pretraining wakati mwingine zinaweza kushinda real evidence.

Utafiti hauwezi kutenganisha kwa uhakika ni sehemu gani ya success ya model inatokana na multimodal evidence na ni sehemu gani inatokana na construction patterns za pretraining. Tofauti hii imeachwa kama open research question.

Kwa nini radio-frequency wall scanner haikujumuishwa katika main experiment?

Watafiti pia walijaribu Walabot device na radio-frequency measurement setup yenye custom filtering. Device hii inaweza kutambua wood na metal studs pamoja na baadhi ya pipes.

Hata hivyo, measurements zilikuwa noisy, zilihitaji extensive post-processing na ziliongeza information kidogo juu ya thermal images. Kwa sababu hiyo, radio-frequency data haikujumuishwa katika main performance tables.

Ubadilishaji kwenda BIM na 3D model

Proof-of-concept script ilitengenezwa kubadilisha JSON wall assembly graph kuwa 3D mesh model. Kila layer node inabadilishwa kuwa geometric material layer, na MEP subnodes kuwa wall-attached components.

Watafiti wanasema kiufundi inawezekana kuunganisha hii na Revit na IFC-based workflows:

• Layer nodes zinaweza kuhamishwa kuwa Revit wall assembly types,
• Outlet, pipe na duct nodes kuwa wall-hosted family instances,
• Material list kwenda embodied carbon calculators.

Hata hivyo, haijaonyeshwa kwamba system ina-support BIM families na IFC schemas zote automatically. Full-scale na error-free BIM integration itahitaji additional engineering.

Athari gani inaweza kuwa nayo katika maisha ya kila siku na sekta?

Ikiwa system ya aina hii itakuwa reliable, inaweza kuwa important pre-screening tool hasa kwa large building portfolios. Experienced expert anaweza kutathmini wall moja kwa existing drawings na Revit model. Potential kubwa iko katika kutengeneza initial predictions kwa hundreds of walls ndani ya jengo automatically.

Possible use cases ni:

• Kutambua walls zinazoweza kuhifadhiwa kabla ya renovation,
• Kukadiria recoverable materials kabla ya demolition,
• Kuandaa embodied carbon inventory,
• Pre-assessment ya partitions ambazo haijulikani kama zina sound insulation,
• Kukadiria hidden mechanical systems juu ya ceilings,
• Kuunda whole-building material passport,
• Kuweka priority kwa critical locations za borescope au physical opening.

Workflow hii inaonekana realistic zaidi kama risk-based decision support inayosaidia kuamua wapi destructive verification inahitajika, badala ya kuchukua nafasi ya destructive testing bila masharti.

Nguvu za utafiti

• Kutoa flexible structure inayoweza kufanya kazi na variable number of data types,
• Kutoa hierarchical na BIM-convertible graph badala ya class label pekee,
• Kufanya controlled validation kwa physical wall mock-up yenye contents zinazojulikana kwa uhakika,
• Kufanya external validation attempt kwa real walls katika miji minne,
• Kupima structural accuracy pamoja na embodied carbon error,
• Kutathmini model variability kwa kuendesha kila wall na data combination mara tano,
• Kuchunguza impact ya thermal image quality kwa separate ablation study,
• Kugawa errors katika meaningful categories kama boundary layer, structural system, stud na MEP,
• Kujadili wazi possible impact ya pre-trained model bias.

Mapungufu ya utafiti

• Utafiti ni preprint ambayo haijapitia peer review.
• Real-world dataset ina walls 22 pekee.
• Examples zimepunguzwa kwa commercial office buildings.
• Residential buildings, historic structures, masonry walls, mass timber na non-Western local construction systems hazijathibitishwa.
• Reference contents za real walls hazikuthibitishwa kwa destructive testing.
• Success ya thermal imaging inategemea sana temperature difference, camera, resolution na image processing.
• Katika mock-up experiment, high-resolution thermal image ilichukuliwa kabla ya final drywall kuwekwa.
• Model inaweza kupata ugumu kutafsiri large, uncropped floor plans.
• Current VLMs haziwezi naturally process 3D point clouds.
• Conflicting data types wakati mwingine zinaweza kupunguza accuracy badala ya kuiongeza.
• Placement ya MEP components katika correct stud bay ilibaki dhaifu.
• Prelearned common construction patterns zinaweza kusababisha wrong assumptions kwa unusual walls.
• Carbon calculations zilitumia average coefficients na fixed wall-dimension assumptions badala ya actual product EPDs.
• Full IFC na Revit compatibility ya generated JSON graph haijathibitishwa kikamilifu.
• Workflow inategemea model version na prompt design; result ileile haiwezi kudhaminiwa kwa VLM nyingine.

Utafiti unasema nini, na hausisemi nini?

Ni maboresho gani yanapendekezwa kwa siku zijazo?

Watafiti wanapendekeza kazi zifuatazo:

• Physical verification ya baadhi ya walls kwa borescope au controlled opening,
• Evaluation metrics zinazotoa penalty kubwa kwa structurally important errors,
• Specialist submodels kwa thermal image, floor plan na RGB analysis,
• Fine-tuning kwenye datasets kubwa na physically verified,
• Kutumia retrieval-augmented generation kwa building databases,
• Kuscale kutoka single wall kwenda whole-floor au whole-building inventory,
• Automatic wall-dimension extraction kutoka point cloud na floor plan,
• Kutengeneza reuse-oriented design kwa materials zilizotabiriwa.

Mbinu na Matokeo ya Utafiti

Muhtasari wa kiufundi wa mbinu

Experimental setup

Data collection devices

• iPhone 13 Pro kwa RGB images,
• FLIR Boson na Micro-Epsilon TIM QVGA cameras kwa thermal images,
• Walabot-based system kwa experimental radio-frequency scanning.

Matokeo makuu ya performance

Maana ya figures na tables

Kielelezo 1, kinaonyesha jinsi GIS, floor plan, RGB na thermal inputs zinavyobadilishwa kutoka compact graph kwenda detailed wall assembly na ensemble prediction. JSON-based graph representation iko katikati ya workflow.

Kielelezo 2, kinaonyesha side-by-side construction stages za physical wall mock-up, finished surface, digital model, floor plan, RGB photo na thermal images. Figure inaonyesha kwamba thermal image inaweza kuonyesha baadhi ya studs huku components nyuma ya insulation zikibaki zimefichwa.

Jedwali 1, linalinganisha all possible data combinations katika mock-up experiment. Ukweli kwamba baadhi ya two-modal combinations zinaweza kushinda packages zenye data nyingi zaidi unaonyesha kwamba conflicting inputs zinaweza kupunguza model performance.

Jedwali 3, linalinganisha floor plan, RGB, thermal image, predicted graph, reference graph na 3D models kwa sample walls kutoka Boston, San Francisco, Birmingham na Toronto. Examples zinaonyesha kwamba model ileile inaweza kuwa perfect au high-accuracy katika baadhi ya walls na kushuka hadi accuracy ya 0,30 katika baadhi ya exterior walls.

Jedwali 4, linatoa main quantitative result ya utafiti. Kadri data types zinavyoongezeka, overall accuracy inapanda kutoka 0,54 hadi 0,76; carbon error inashuka kutoka %104 hadi %26.

Kielelezo 4, kinalinganisha error categories sita. Structural system na stud-type recognition zinaboreka sana kwa additional data, huku MEP placement ikibaki ngumu.

Jedwali 7, linaonyesha kwamba thermal image enhancement inaweza kutoa small but measurable accuracy gain. Result bora ilipatikana wakati kila image iliboreshwa manually.

Jedwali 9, linaonyesha kwamba workflow ileile ilifanya vizuri zaidi kwa newer VLM version bila additional training. Finding hii inaunga mkono dai kwamba method inaweza kunufaika na model improvements, lakini haihakikishi kwamba performance itaongezeka kwa kila future model version.

Maelezo ya Chanzo na Mbinu

Maudhui haya yanatokana na utafiti wa Nikita Klimenko, Martin JJ. Bucher, Fopefoluwa Bademosi, Lorenzo Villaggi, Dale Zhao na David Benjamin wenye kichwa “Seeing Inside Walls for Retrofit and Reuse: Multimodal Reasoning with Vision-Language Models for Hidden Wall Assemblies”.

Utafiti huu ni preprint research paper yenye electronic copy iliyochapishwa kupitia SSRN na inayosemekana kuwasilishwa kwa Elsevier. Waraka unasema wazi “This preprint research paper has not been peer reviewed”. Kwa hiyo, utafiti haujapitia peer review. PDF haina DOI inayoweza kuthibitishwa wala final journal acceptance information.

Utafiti unajumuisha experimental software workflow ya vision-language model, physical wall mock-up, 22 real walls katika commercial buildings nne, structured graph prediction, thermal na RGB imaging, BIM conversion na embodied carbon calculation.

Reference assemblies za real walls hazikuthibitishwa kwa destructive testing. Reference labels zilitayarishwa kwa Revit models, BIM files, architectural drawings, thermal scans na expert assessments. Kwa hiyo, utafiti hautumii exact physical verification ya wall contents, bali multi-source reference estimate inayoweza kupatikana kwa data zilizopo.

Accuracy iliyoripotiwa katika makala hupima similarity ya graph structure na reference graph. Thamani hii haimaanishi kwamba model imegundua kila material au MEP component kwa percentage hiyo hiyo. Embodied carbon relative error pia imehesabiwa kwa assumptions kama standard material coefficients, wall height ya 10 feet na stud spacing ya 16 inches; si verified life-cycle analysis ya real building.

VLM versions, prompts na results zilizotumika katika utafiti zimewasilishwa kama zilivyoripotiwa katika PDF. Accuracy sequences za %47–%74 na %54–%76 katika utafiti zinahusishwa na GPT-5.2 na GPT-5.4 comparisons mtawalia.

Waandishi walisema walitumia Claude Sonnet 4.5 wakati wa kuandaa makala kwa paragraph structuring, grammar correction, flow improvement na kufupisha baadhi ya sections; na kwamba walikagua final content wenyewe na kuwajibika nayo.

Maudhui haya yameandaliwa kwa msingi wa utafiti uliopo kwenye PDF iliyotolewa pekee. Hakuna madai ambayo hayapo kwenye PDF yaliyoongezwa kuhusu exact wall detection, safety guarantee, automatic BIM accuracy, commercial field success au generalizability kwa all building types.