Download as pdf or txt
Download as pdf or txt
You are on page 1of 1

다층 퍼셉트론 신경망을 활용한 화재 발생 건물의 구조 안전성 예측

Prediction Structural Safety of Buildings Under Fire Using MultiLayer Perceptron Neural Networks
서정우 (연세대학교 건축공학과)

Purpose of the Study

• Safe Routing based on MLP

Safe Unsafe
Unable to Self-Evacuate

Real - Time
Beam , Girder
AI Guided Selection
Radiation Convection

CO , CO₂
Rescue Ladders Rescue Worker
Unavailable
Unsafe Safe

Time Safe
Yield Strength( 𝒌𝒚 )
Location
Elastic Modulus( 𝒌𝑬 )
Member
Unsafe

• Fire Dynamic Simulation • Heat Analysis • MultiLayer Perceptron Architecture (MLP) • Safety Check

Research Procedures and Methods

1. Target Building 2. Fire Dynamic Simulation 3. Data Analysis 4. MultiLayer Perceptron 5. Safety Check

1. Target Building – Taipei 101 32th Floor 3. Data Analysis

3-1. Heat Analysis for Internal Temperature


Validation
∆𝒕 𝒉𝑨 𝑻∞ − 𝑻𝒔 + 𝝈𝜺𝑨 𝑻𝟒∞ − 𝑻𝟒𝒔
𝑻𝒔,𝒏𝒆𝒘 = 𝑻𝒔,𝒐𝒍𝒅 + Time(s) Ex_Temp(K) In_Temp(K)
𝒎𝒄 11 312 293.96
𝑻∞ : External Temperature (𝑲) 𝑻𝒔 : Initial temperature (𝑲) 60 1250 697.78
11s 60s 100s
𝝈 : Stefan Boltzmann Constant (𝐖/𝐦𝟐 ·𝑲𝟒 ) 𝒉 : Convective heat transfer coefficient (𝐖/𝐦𝟐 ·𝑲) 100 1135 936.08
𝒎 : Mass (𝒌𝒈) 𝑨 : Cross-Section Area (𝒎𝟐 )
▪ Taipei 101 ▪ 32th Structural Floor Plan ▪ 32th Floor Plan 𝜺 : Emissivity of Gases (CO,CO₂) 𝒄 : Specific Heat ( 𝑱/𝒌𝒈·𝑲 )

• Location: Taipei, Taiwan (Republic of China) • Main Steels for Building Structure : SN490B
SN490B – Properties at Room Temperature
• Designed by C.Y. Lee & Partners
• Elastic Modulus : 200 - 215 GPa 3-2. Modeling Reduction Factor in yield Strength (𝒌𝒚 ) and Elastic Modulus (𝒌𝑬 )
• Height: 508 m • Yield Strength : 295 – 445 MPa
• Specific heat capacity : 465 J/ ( kg·K ) Member’s Temp (T) 𝒌𝒚 𝒌𝑬
• Use of 37th Floor : Office
25 1.00 1.00
Reference 100 0.92 0.89
• Taipei 101 : CC BY-SA 3.0, https://commons.wikimedia.org/w/index.php?curid=371576 Make
• 32th Structural Floor Plan : Shieh, Shaw-Song, Ching-Chang Chang, and Jiun-Hong Jong. "Structural design of composite super-columns for the Taipei 101 Tower.“ 200 0.84 0.81 Non-Linear
Training & Test Data
Regression Analysis
300 0.72 0.77 (8:2)
2. Fire Dynamic Simulation (FDS) 400 0.67 0.78
500 0.58 0.58

• PyroSim is Fire simulation software with a graphical interface for NIST’s FDS. 600 0.39 0.35
.csv File
• Supports Fire scenario modeling, 3D result visualization. 700 0.17 0.24

• Used by Fire safety engineers for building analysis. 800 0.09 0.05
• 𝒌𝒚 = −𝟕. 𝟒𝟓𝟖𝟖 × 𝟏𝟎−𝟕 × 𝑻𝟐 − 𝟓. 𝟔𝟕𝟐𝟔 × 𝟏𝟎−𝟒 × 𝑻 + 𝟎. 𝟗𝟗𝟓𝟒 • Input Data: Time, Location, Member
• 𝒌𝒚 = 𝑭𝒚 (𝑻)/𝑭𝒚 (𝟐𝟓°𝑪) • 𝒌𝑬 = 𝑬(𝑻)/𝑬(𝟐𝟓°𝑪) • 𝒌𝑬 = −𝟏. 𝟏𝟒𝟖𝟔 × 𝟏𝟎−𝟔 × 𝑻𝟐 − 𝟐. 𝟐𝟔𝟕𝟐 × 𝟏𝟎−𝟒 × 𝑻 + 𝟎. 𝟗𝟓𝟗𝟓 • Output Data: 𝑘𝑦 , 𝑘𝐸
Model Fire Scenario Data Ref: Chung, Hsin-Yang, et al. "Application of fire-resistant steel to beam-to-column moment connections at elevated temperatures." Journal of Constructional Steel Research 66.2 (2010): 289-303.

▪ Fire occurrence locations : 8 locations


▪ Fire occurrence time : 0 ~ 3600 seconds
▪ Type of fire : Electrical spark 4. MultiLayer Perceptron Neural Network
▪ Combustible materials : Wood
▪ Fire spread rate : 1.0 m/s ▪ MLP ▪ MLP Performance
[ Perspective View ] ▪ Fire Emissions Gas : CO,CO₂ Input Layer Hidden Layer Output Layer
▪ Assumptions :
- Sprinkler malfunction and initial fire suppression failure Time
- The area of the fire room was averaged at 20 square meters ෢𝒚
𝒌 𝒌𝒚
- A fire has broken out in one of the offices
➢ Fire Load : 𝟕𝟎𝟎 𝑴𝑱/𝒎𝟐 Location MSE Loss
[ Section View ] ➢ HRRPUA : 𝟏𝟎𝟎𝟎 𝐤𝐖/𝒎𝟐 (Heat Release Rate Per Unit Area) ෢𝑬
𝒌 𝒌𝑬
Member
Simulation Prediction 𝒌𝒚 , 𝒌𝑬 Target 𝒌𝒚 , 𝒌𝑬

Update weight Update weight


Update weight
External Temperature And bias And bias
σ𝑵 𝒚𝒊 𝟐
𝒊=𝟏 𝒚𝒊 −ෝ
▪ 𝐑𝐌𝐒𝐄 = = 𝟎. 𝟏𝟔𝟗𝟏
Optimizer - Adam 𝑵
σ𝑵 𝒚𝒊 𝟐
𝒊=𝟏 𝒚𝒊 −ෝ
Back propagation - Gradient ▪ 𝑹𝟐 𝑺𝒄𝒐𝒓𝒆 = 𝟏 − σ𝑵 ഥ 𝟐
= 𝟎. 𝟗𝟕𝟏𝟏
𝒊=𝟏 𝒚𝒊 − 𝒚

5. Safety Check Example of Real-Time AI Guided Selection


[ 0s ] [ 72s ] [ 163s ] [ 666s ] [ 2728s ]
Situation Result
Heat Release Rate Per Unit Volume Unsafe Fire Non-Ambulatory Non-Ambulatory
𝒌𝒚 × 𝒌𝑬 < 𝟎. 𝟓𝟓

Emergency
stairs
Warning Real-Time

𝟎. 𝟓𝟓 < 𝒌𝒚 × 𝒌𝑬 < 𝟎. 𝟔𝟖

[ 0s ] [ 70s ] [ 139s ] [ 455s ] [ 815s ] Unsafe Route Safe Route


Safe MLP
Visualization of
Time : 1200s Structural Member’s AI Guided Selection
𝟎. 𝟔𝟖 < 𝒌𝒚 × 𝒌𝑬 Output
Degradation State

REF: Eurocode 3: Design of steel structures - Part 1-2: General rules - Structural fire design

Conclusion and Further Study

➢ Conclusion: An MLP Neural Network was trained to predict structural safety by learning the Reduction Factor in yield strength (𝒌𝒚 ) and Elastic modulus (𝒌𝑬 ) as temperature increases.
➢ Further Study: By including simulations of more location cases and heat analysis on CFT columns, Can help train the MLP for more specific structural safety predictions

You might also like