Utafiti huu unachunguza jinsi ya kutenganisha gas-liquid two-phase flow katika evaporative cooling system kuwa slug flow, unstable slug flow na churn flow kwa kutumia images. Watafiti walitumia samples 300 kwa kila moja ya flow classes tatu, jumla samples 900; walilinganisha spatial tensors zinazotegemea single image frame na spatio-temporal tensors zinazoundwa kwa kuweka juu ya nyingine RGB frames tano zilizochaguliwa kwa equal time intervals. Full model inayounganisha ResNet-18, SwinTransformer-B na EfficientNet-B0 ndani ya mixture-of-experts system na kupitisha outputs kupitia kile utafiti unachokiita Mamba attention mechanism ilifikia overall recognition accuracy ya %83,9 kwenye test set. Hata hivyo, utafiti haujapitia peer review; training-test split ratio, camera frame rate, model-training hyperparameters, independent external-test results na mean pamoja na uncertainty values kutoka repeated experiments hazijaripotiwa.

Full model ilitoa accuracy iliyo higher kwa 8,9 percentage points kuliko %75,0 accuracy ya SwinTransformer-B model inayotumia single image frame pekee. Waandishi wanaeleza tofauti hii kama “%8,9 improvement”; mathematical difference kati ya accuracy values mbili zinazolinganishwa ni 8,9 percentage points, si relative percentage increase. Strongest class result ya full model ilikuwa churn flow recall ya %93,1, huku unstable slug flow recall ikibaki %68,2. Kwa hiyo, increase ya overall accuracy haikusambazwa sawa katika flow classes zote.

Swali kuu la utafiti ni nini?

Katika gas-liquid two-phase flow, shape, distribution, continuity na time-dependent motion ya gas bubbles vinaweza kuathiri moja kwa moja flow resistance na heat transfer. Tatizo kuu la utafiti ni kwamba traditional flow-regime identification methods mara nyingi hutathmini spatial appearance ya flow na temporal evolution yake kando.

Methods zinazotegemea time signals kama differential pressure au conductivity zinaweza kufuatilia temporal variation ya flow, lakini huwakilisha geometric structure ya gas na liquid phases ndani ya pipe kwa kiwango kidogo. Visual methods zinazotegemea single image frame zinaweza kuonyesha bubble shape na texture; lakini hazitumii moja kwa moja jinsi flows mbili zinazofanana zinavyotofautiana baada ya frames chache.

Tofauti hii ni muhimu hasa kati ya slug flow na unstable slug flow. Unstable slug flow imefafanuliwa kama transition stage kutoka regular slug flow kwenda regimes nyingine. Katika single image frame, textures za classes hizi mbili zinaweza kufanana sana. Main hypothesis ya utafiti ni kwamba kuunganisha multiple frames katika tensor moja kutaruhusu model kujibu si tu “flow inaonekanaje sasa?” bali pia “flow inabadilika vipi ndani ya short time interval?”.

Experimental setup iliundwaje?

Samples zilipatikana kutoka physical experimental setup yenye self-circulating evaporative cooling system na image-acquisition system. Main components za cooling loop ni:

• Heater,
• Vertical gas outlet pipe,
• Condenser,
• Liquid return pipe,
• Lighting panel,
• Imaging camera.

Inner chamber ya heater ilikuwa imezama kikamilifu katika coolant. Liquid ikipokea energy kutoka heater, sehemu yake huvaporize na kuunda gas-liquid mixture. Density difference husababisha mixture kupanda kwenye vertical gas outlet pipe. Gas phase hutoa heat kwenye condenser, huliqueify tena na kurudi kwenye heater kupitia liquid return pipe kwa gravity.

Vertical height kati ya top point ya gas outlet section na bottom ya heater imetolewa kama H = 110 cm. Initial liquid levels tofauti zilitumika katika experiment:

• h₁ = 100 cm,
• h₂ = 75 cm,
• h₃ = 45 cm,
• h₄ = 20 cm.

Initial levels tofauti ziliunganishwa na heat loads tofauti ili kuunda flow regimes tofauti katika gas outlet pipe. Katika Kielelezo 2, horizontal axis ina applied heat load katika watts na vertical axis ina observed flow class. Graphs zinaonyesha kwamba flow regime haitegemei heat load pekee, bali pia pipe diameter na initial liquid level. Heat load inapoongezeka, katika operating points nyingi kuna transition kutoka regular slug flow kwenda unstable slug flow na katika baadhi ya conditions kwenda churn flow; hata hivyo, transition thresholds si sawa kwa kila liquid level na pipe diameter.

Kwa nini imaging conditions ni muhimu?

Watafiti walilinganisha front, side na back illumination arrangements. Katika front illumination, bubble shadows zilikuwa juu ya bubbles zenyewe na gas-liquid contrast ikawa dhaifu; katika side illumination, shadows ziliundwa kwa angle relative na camera direction. Back illumination ilichaguliwa kwa experiment kwa sababu ilifanya gas-liquid interface ionekane wazi zaidi.

Inner diameters za transparent pipes zilikuwa 8 mm na 16 mm. Images zilichukuliwa kutoka selected observation section katika vertical gas outlet pipe. PDF inataja device kama “ordinary camera”, yaani standard camera; brand, model, resolution, exposure time na frames per second hazijatolewa. Upungufu huu unazuia kubainisha five selected frames katika temporal tensor zinawakilisha physical time interval ya sekunde ngapi.

Flow regimes tatu zinatofautianaje kwenye images?

Kielelezo 3 kinaonyesha consecutive image frames za kila flow class. Katika slug flow images, elongated gas regions zenye discernible boundaries zinaonekana. Katika unstable slug flow, regions hizi huanza kuvunjika na kubadilisha shape. Katika churn flow images, badala ya continuous regular gas slug, kuna denser, irregular na rapidly changing mixture texture.

Single frame haiwezi daima kuonyesha wazi tofauti kati ya slug flow na unstable slug flow. Main reason ya kutumia frames tano katika utafiti ni kuingiza boundary breakup na weakening ya periodicity kwa muda ndani ya model.

Spatio-temporal input tensor iliundwaje?

Samples 300 zilikusanywa kwa kila flow regime, jumla samples 900. Kila sample ina consecutive image frames 30. Watafiti waliandaa input schemes mbili tofauti:

RGB image moja ina color channels tatu. Stacking RGB frames tano huongeza channel count hadi 5 × 3 = 15. Hivyo model huchakata image information kutoka time points tano tofauti ndani ya sample ileile kwa pamoja.

Utafiti hautaji ni five frame indices gani zilichaguliwa kutoka 30-frame sequence au physical time interval kati ya frames. Kauli “equal time intervals” inaeleza selection pattern kwa ujumla, lakini haitoi detail ya kutosha kwa full method reproduction.

Temporal information ilibadilishaje accuracy ya classical models?

Single-frame spatial tensor na five-frame spatio-temporal tensor zililinganishwa katika ResNet-18, SwinTransformer-B na EfficientNet-B0 models.

Overall accuracy iliongezeka katika models zote tatu. Largest increase ilionekana katika SwinTransformer-B; accuracy ilipanda kutoka %75,0 hadi %81,1. Hata hivyo, overall accuracy pekee haimaanishi kwamba kila class imeboreshwa.

Confusion matrices zinaonyesha classes zipi zilipata au kupoteza?

Diagonal values za row-normalized confusion matrices zinaonyesha correct-recognition rate ya kila true class, yaani class recall.

Five-frame tensor iliboresha separation ya slug flow na unstable slug flow kwa ujumla. Kwa mfano, katika SwinTransformer-B, slug flow recall iliongezeka kutoka %75,4 hadi %86,4 na unstable slug flow recall kutoka %58,7 hadi %77,8.

Kwa upande mwingine, churn flow recall ilipungua katika models zote tatu:

• Katika ResNet-18 kutoka %93,4 hadi %83,6,
• Katika SwinTransformer-B kutoka %90,2 hadi %79,1,
• Katika EfficientNet-B0 kutoka %93,4 hadi %80,6.

Waandishi wanatafsiri result hii kama “compression ya churn-flow feature space” na “feature-distribution shift”. Kwa maneno mengine, temporal information inapopanga upya decision boundary ili kutenganisha slug classes zinazofanana, representation space iliyotengwa kwa churn flow yenye visual difference kubwa inaweza kubanwa. Mechanism hii si physical quantity iliyopimwa directly; ni model interpretation inayotokana na confusion-matrix behavior.

Mixture-of-experts model inafanyaje kazi?

Katika proposed classifier, neural networks tatu tofauti zilitumika kama “experts”:

• ResNet-18: Convolutional network inayotumia residual connections ili kupunguza vanishing au exploding gradient problems katika deep networks.
• SwinTransformer-B: Visual Transformer architecture inayomodel long-range image relations kwa shifted-window self-attention.
• EfficientNet-B0: Convolutional architecture inayoscale network depth, width na image resolution kwa pamoja.

Models hizi zinaapproach input ileile kwa different representation principles. Mixture-of-experts (mixture of experts, MoE) system hai-average expert outputs uniformly; badala yake, huzipa weights kupitia gating network kwa kila sample. Lengo ni kubainisha dynamically ni expert gani ni useful zaidi kwa specific flow sample.

Kielelezo 6 kinaonyesha B₁, B₂ na B_n kama base models; V₁, V₂ na V_n kama features zinazotolewa na models hizi; G kama gating network; V kama weighted fused feature; na ATT kama Mamba-based processing section.

Architecture ya gating network ikoje?

Kielelezo 7 kinaonyesha gating network kwa sequence ifuatayo:

1. Input tensor: (900, 15, 224, 224),
2. 3 × 3 Conv1, channel transformation 15 → 6 na 2 × 2 MaxPool,
3. 3 × 3 Conv2, channel transformation 6 → 16 na 2 × 2 MaxPool,
4. 3 × 3 Conv3, channel transformation 16 → 32 na 2 × 2 MaxPool,
5. Adaptive average pooling, iliyoonyeshwa katika figure kama “(32,32), 1 × 1”,
6. Flattening,
7. 32 → 120 fully connected layer na ReLU,
8. 120 → 3 fully connected layer na ReLU,
9. Softmax output inayozalisha weights kwa experts tatu.

Katika schematic, kuandika samples 900 za entire dataset kama first dimension na kuonyesha baadhi ya intermediate dimensions kama “(32,32)” haijaelezwa kikamilifu kwa standard mini-batch notation. General function ya network inaeleweka, lakini layer output dimensions na batch usage zinahitaji maelezo ya ziada kwa code-level reproduction.

Mamba-based mechanism inatumia mathematical operations gani?

Kile utafiti unachokiita “Mamba attention mechanism” huchakata fused features upya kupitia selective state-space model. Input inaweza kufafanuliwa kama:

\[ X \in \mathbb{R}^{B_{\mathrm{batch}} \times L \times d} \]

Hapa:

• B_batch ni mini-batch size,
• L ni sequence length,
• d ni feature dimension.

PDF inatumia symbol B kwa batch size na pia kwa input-projection matrix. Ili kuepuka confusion, batch size imeonyeshwa hapa kama B_batch; matrix symbols katika original formula hazijabadilishwa.

Dynamic parameters huzalishwa kwa linear transformations:

\[ \begin{cases} \Delta = W_{\Delta}X \\ \mathbf{B} = W_{B}X \\ \mathbf{C} = W_{C}X \end{cases} \]

• Δ ni time-step adjustment parameter inayobainisha state inasasishwa kwa kasi gani.
• B ni input projection inayobainisha current input inahamishwa kwa kiwango gani kwenda hidden state.
• C ni output projection inayobainisha features zipi kutoka hidden state zinahamishwa kwenda output.
• W_Δ, W_B na W_C ni learnable weight matrices.

Text inaeleza kwamba Δ value imewekewa positive range kwa kutumia sigmoid function; hata hivyo, sigmoid operation haijaonyeshwa wazi katika written form ya Equation 1. Kwa hiyo actual transformation iliyotumika inahitaji kuthibitishwa kwa code au supplementary method description.

Document inatoa dimension ya Δ kama [ B_{\mathrm{batch}} \times L \times d ], input projection B kama [ B_{\mathrm{batch}} \times L \times n ] na W_B kama [ d \times d ]. Ikiwa hidden-state dimension n na feature dimension d si sawa, haijaelezwa jinsi W_B transformation inazalisha n dimension. Hii pia ni dimensional ambiguity katika formula notation.

Discretized form ya state-space equation imetolewa kama:

\[ \begin{cases} \widetilde{A}_{t} = e^{\Delta_{t}A} \\ \widetilde{B}_{t} = (\Delta_{t}A)^{-1}(e^{\Delta_{t}A}-I)\cdot\Delta_{t}B_{t} \\ h_{t} = \widetilde{A}_{t}h_{t-1}+\widetilde{B}_{t}x_{t} \\ y_{t} = C_{t}h_{t} \end{cases} \]

• A ni learnable state matrix yenye dimension [ n × n ].
• h_t ni n-dimensional hidden state wakati t.
• x_t ni d-dimensional input vector wakati t.
• Ã_t na B̃_t ni time-dependent discrete counterparts za continuous state-space system.
• I ni identity matrix.

Kulingana na utafiti, large Δ_t value husababisha e^ΔtA term kudecay haraka na past information kusahaulika haraka zaidi; small Δ_t huhifadhi longer-term dependencies. Time step hapa haijatolewa kama duration ya sekunde iliyopimwa katika experiment; ni learnable sequence-update parameter ndani ya network.

Ili kuunganisha state-space output inayobeba past information na current input, skip connection imetumika:

\[ y_{t} = C_{t}h_{t}+Dx_{t} \]

Katika formula hii, C_th_t inawakilisha feature iliyocode kutoka past state, huku Dx_t ikiwa skip path inayosisitiza current input directly. Learnable D matrix hurekebisha relative contribution ya components mbili.

Baada ya Mamba layer, residual connection, LayerNorm na feed-forward network zilitumika. Residual connection inalenga gradient flow na information preservation, LayerNorm inasawazisha feature scales, na feed-forward network inaongeza nonlinear feature interaction kwenye linear state-space operations.

Ablation experiments zinaonyesha kila component imeleta faida gani?

Average accuracy ya classical classifiers tatu zinazotumia spatio-temporal tensor ni %79,8. Model-3 imezidi average hii kwa 4,1 percentage points.

Katika Model-1, biases za experts tofauti kufidia kwa kiasi zilipandisha churn flow recall juu ya experts wengi mmoja mmoja. Katika Model-2, gating network iliweight expert outputs kulingana na sample na kuongeza overall accuracy kutoka %81,7 hadi %82,8.

Katika Model-3, kubadilisha Transformer-based section na Mamba kuliongeza overall accuracy hadi %83,9 na churn flow recall hadi %93,1. Kwa upande mwingine, slug flow recall ilibaki %84,4 bila kubadilika, huku unstable slug flow recall ikipungua kutoka %71,4 hadi %68,2.

Result hii inaonyesha kwamba Mamba component haikuboresha classes zote kwa wakati mmoja. Ingawa conclusion section ya utafiti inaeleza kwamba Mamba inakamata important transient dependencies za transition states, visible gain katika class-level ablation data imekusanyika zaidi katika churn flow. Unstable slug flow ilibaki class yenye lowest recall katika full model ya utafiti.

Ni conclusions zipi zinazoungwa mkono na utafiti?

• Katika experimental dataset hii, kutumia image frames tano pamoja kuliongeza overall accuracy ya classical models zote tatu ikilinganishwa na single-frame input.
• Temporal information iliboresha separation ya slug flow na unstable slug flow zenye similar visual textures, hasa katika SwinTransformer-B model.
• Kuongeza input dimension pekee hakukuboresha classes zote sawa na kulipunguza churn flow recall.
• Kuweka weights kwa experts tofauti kupitia gating network kuliruhusu matumizi ya more balanced strengths za individual experts.
• Mamba-based full model ilitoa highest overall accuracy na highest churn flow recall katika dataset hii.

Utafiti haujathibitisha nini?

• Hauonyeshi kwamba model itafikia accuracy ileile katika coolants nyingine, pipe geometries tofauti au laboratories tofauti.
• Real-time industrial au nuclear-facility application haijatekelezwa.
• Model latency, memory requirement, energy consumption au frames-per-second processing rate hazijaripotiwa.
• Regimes kama bubbly, annular, stratified au mist flow nje ya classes tatu hazijajaribiwa.
• Generalization haijavalidate kwa independent external dataset.
• Robustness dhidi ya image quality, illumination changes, camera angle, pipe fouling au image noise haijatathminiwa.
• Hakuna confidence interval, standard deviation, p value au other statistical-significance analysis iliyotolewa kwa reported accuracy differences.
• High classification accuracy haimaanishi moja kwa moja higher cooling efficiency, lower energy consumption au plant safety.

Nguvu na main limitations za utafiti ni zipi?

Nguvu za utafiti ni kukusanya images kutoka physical evaporative-cooling setup, kulinganisha single-frame na multi-frame inputs kwenye same base models, kutoa class-level confusion matrices, na kutenganisha MoE na Mamba components kwa ablation experiments. Kuonyesha recall ya kila flow regime pamoja na total accuracy kunaweka wazi model imepata katika classes zipi na imerudi nyuma katika zipi.

Main limitations ni:

• Utafiti ni preprint ambayo haijapitia peer review.
• Haijaelezwa samples 900 ziligawanywa kwa ratio gani kati ya training na test sets.
• Haijaelezwa samples zilizotokana na same experimental condition au same 30-frame sequence ziligawanywa vipi kati ya training na test. Kwa hiyo risk ya condition- au sequence-level data leakage haiwezi kutathminiwa.
• Random training-test split ilisemwa kuwa imerudiwa multiple times, lakini repeat count, mean accuracy, standard deviation na confidence interval hazijatolewa. Results za article zimeelezwa wazi kuwa zimetoka kwenye “representative single run”.
• Camera model, frame rate, exposure, image-capture duration na temporal spacing ya frames tano hazijatolewa.
• Optimization algorithm, learning rate, epoch count, batch size, loss function, data-augmentation method, early-stopping criterion na hardware information hazijatolewa.
• Flow regimes tatu, pipe diameters mbili na experimental platform moja tu zimechunguzwa.
• Kuna inconsistency ndani ya PDF kati ya final ablation values na values katika baadhi ya supplementary intermediate graph pages.
• Sigmoid operation na baadhi ya matrix dimensions hazijaelezwa kikamilifu katika formula notation.

Utafiti una maana gani kwa past, present na future?

Flow-regime recognition research hapo awali imetegemea signal types tofauti kama pressure fluctuations, conductivity, chaos analysis, wavelet transform, entropy na image texture. Utafiti huu, badala ya kutatua moja kwa moja separation kati ya time-signal na image analysis, unapendekeza kuunganisha sequential images katika single multi-channel tensor.

Current contribution yake ni kuonyesha kwamba temporal information kuongeza overall accuracy pekee haitoshi. Five-frame input inaweza kusaidia kutenganisha two easily confused classes huku ikidhoofisha representation ya class tofauti zaidi visually. Kwa hiyo input-level fusion na classifier-level resource allocation zinapaswa kuundwa pamoja.

Katika future, flow regimes zaidi, fluids tofauti, pipe diameters tofauti na independent experimental setups zinapaswa kutumika. Pia real-time operation ya model, computational cost, robustness kwa illumination tofauti na generalization capability across experimental conditions zinahitaji kutathminiwa. Current results zinatoa promising laboratory classification approach; hazitoi industrial-scale validated control au safety system.

Mbinu na Matokeo ya Utafiti

Muhtasari wa kiufundi wa method

Main numerical findings

Scientific message ya figures na tables

• Kielelezo 1: Kinaonyesha physical layout ya heater, gas outlet pipe, condenser, return pipe, light panel na observation section.
• Kielelezo 2: Kinaonyesha jinsi flow class inavyobadilika kadiri heat load, initial liquid level na pipe diameter zinavyobadilika.
• Kielelezo 3: Kinaonyesha morphological na temporal differences za flow regimes tatu katika consecutive images.
• Jedwali 1: Linatoa defining characteristics za slug, unstable slug na churn flow.
• Jedwali 2: Linalinganisha structure ya single-frame na five-frame tensors.
• Kielelezo 4: Kinaonyesha five-frame input iliongeza overall accuracy katika classical classifiers zote tatu.
• Kielelezo 5: Kinaonyesha temporal information iliboresha slug classes huku ikipunguza churn-flow recall.
• Kielelezo 6: Kinaonyesha general architecture yenye expert models, gating network, feature fusion, Mamba section na classification head.
• Kielelezo 7: Kinaonyesha convolutional gating network inayozalisha Softmax weights kwa experts tatu.
• Kielelezo 8: Kinaonyesha sequence ya Mamba layer, residual connection, LayerNorm na feed-forward network.
• Jedwali 3 na Kielelezo 9: Vinaonyesha separate contributions za direct fusion, dynamic gating na Mamba component kwenye overall accuracy na class recall.

Maelezo ya Chanzo na Mbinu

Kichwa cha asili cha utafiti: Research on Gas-Liquid Two-Phase Flow Pattern Identification Method Based on Spatio-Temporal Feature Fusion

Waandishi na mpangilio wao: Jiaqi Rao; Bin Xiong

Co-first author: PDF haina information ya equal contribution au co-first authorship.

Corresponding author: Bin Xiong

Institutional affiliations:

1. State Key Laboratory of High Density Electromagnetic Power and Systems, Institute of Electrical Engineering, Chinese Academy of Sciences, Haidian District, Beijing 100190, China
2. University of Chinese Academy of Sciences, Shijingshan District, Beijing 100049, China

Author contributions: Jiaqi Rao alifanya conceptualization, methodology, software, formal analysis, investigation, data curation, first-draft writing, visualization na validation. Bin Xiong alitoa funding acquisition, review na editing pamoja na supporting visualization contributions.

DOI: 10.2139/ssrn.6946753

Publication au preprint platform: SSRN

Journal au conference: Peer-reviewed journal au conference publication haijathibitishwa kwa version hii.

Original publisher au platform service: Utafiti uko kwenye SSRN preprint platform; SSRN service inaendeshwa na Elsevier. Peer-reviewed journal publisher haijathibitishwa.

Publication year: 2026

Source type: Preprint research article inayounganisha experimental image-data collection na deep-learning classification

Peer-review status: Utafiti huu ni preprint na haujapitia peer review.

Official link:Ukurasa wa utafiti wa SSRN

Funding: Utafiti unaripoti support ya National Natural Science Foundation of China kwa grant number 51677178. Watafiti pia wanashukuru relevant division ya Institute of Electrical Engineering, Chinese Academy of Sciences kwa experimental facilities na sample-collection support.

Makala hii ya Kituruki imeandaliwa kwa kuchunguza main text, formulas, tables, graphs na schematics za uploaded PDF pamoja na attached visual pages. Hakuna new scientific finding, field success, industrial-application result au real-time performance claim ambayo haipo katika PDF iliyoongezwa.

Main limitations za utafiti ni kutokuwepo kwa peer review, kuwa limited kwa experimental platform moja na flow classes tatu, kutokuelezwa kwa training-test ratio, kutotolewa kwa camera na frame-rate information, kutokamilika kwa training hyperparameters, kutokuwepo kwa independent external validation, repeated experiments kuwasilishwa kwa representative single run pekee na kutoripotiwa kwa statistical uncertainty ya results.

Kuna inconsistency ndani ya PDF kati ya ablation accuracies za %81,7, %82,8 na %83,9 katika final Figure 9 na conclusion text, na values za %81,1, %81,7 na %83,3 zinazoonekana katika baadhi ya supplementary intermediate graph pages. Makala hii imetumia %81,7, %82,8 na %83,9 kwa sababu zinaendana na main text, final figure na results section. Kabla ya publication, inapendekezwa numerical difference hii ithibitishwe kupitia researchers au final experimental records.

Utafiti huu ni preprint ambayo haijapitia peer review; results zinapaswa kusomwa kwa kuzingatia limitation hii. Findings zinaonyesha classification performance kwa specific image data zilizokusanywa katika laboratory setup, na hazitoi evidence ya industrial success iliyothibitishwa katika real field conditions.