ICCES UCSD

Simple Neural Network Application for Traffic Monitoring : 

Simple Neural Network Application for Traffic Monitoring Michael Fraser Linjun Yan Xianfei He Ahmed Elgamal Joel P. Conte Tony Fountain

Civil Infrastructure: 

Civil Infrastructure

Slide5: 

Physics-based model validation Finite element model updating System identification Decision Support: Emergency response Preventive maintenance Retrofit/strengthening Reliability/risk analysis Structural health assessment Knowledge-Based Integration Advanced Query Processing Grid Storage (Curated Database) File systems, Database Systems High speed networking Storage hardware Networked Storage (SAN) Sensor nets (real-time data, video streams) DATA INFORMATION KNOWLEDGE DECISION

UCSD Composite Bridge Decks Testbed: 

UCSD Composite Bridge Decks Testbed

Construction Details: 

Construction Details

Current Monitoring System: 

Current Monitoring System

Web Portal (http://healthmonitoring.ucsd.edu): 

Web Portal (http://healthmonitoring.ucsd.edu) Peak strain: 2.5974E-4 – Peak Time: 15:43:24 15:43:16 15:43:42 April 7, 2003

Slide11: 

Typical Campus Traffic Cars, SUV’s, Maintenance Trucks

Slide12: 

Typical Campus Traffic Delivery Trucks

Slide13: 

Typical Campus Traffic Campus Buses

Slide14: 

Typical Campus Traffic Construction Equipment

Query Database of Recorded Strains by Vehicle Type : 

Query Database of Recorded Strains by Vehicle Type

Scatter Plots of Peak Strain: 

Scatter Plots of Peak Strain

Peak Strain Distributions: 

Peak Strain Distributions All of the data is available for download through the webportal

Neural Network Application for Traffic Monitoring: 

Neural Network Application for Traffic Monitoring Sample Application

Finite Element Modeling: 

Finite Element Modeling FE Model composed of sixty beam-column elements The bridge deck response, under simulated traffic loading, was analyzed using the computational framework OpenSees

Slide20: 

Response of bridge-deck system (at quarter span) under simulated bus loading.

Slide21: 

Prior to applying the neural network for estimating vehicle properties, Principal Components Analysis (PCA) was employed for feature extraction. For each pair of time histories corresponding to one particular event, the number of features was reduced from 20,000 (each time step in each of the two strain time histories is a feature), to the first ten principal components. These ten PCA features were used as input in the neural network.

Slide22: 

The neural network was then defined with the 10 input units, 15 hidden units (preliminary studies showed 15 to be sufficient), and a single output unit (corresponding to normalized speed or wheelbase). The backpropagation learning algorithm was used to train the network, and adjust the weights by minimizing the error between network outputs and targets (corresponding desired values for the outputs).

Error for Speed Estimation: 

Error for Speed Estimation Predicted

Error for Wheelbase Estimation: 

Error for Wheelbase Estimation Percent of Test Data Correctly Predicted

Slide25: 

Please join us! SDSC Chaitan Baru Mike Bailey Texas A&M Norris Stubbs Caltrans Cliff Roblee Charles Sikorsky LANL Chuck Farrar UCSD, CSE Ingolf Kruger Ramesh Rao Larry Smarr Stanford Kincho Law James Peng UC Berkeley Gregory Fenves Frank McKenna Kris Pister Structural Engineering Zhaohui Yang Jinchi Lu Michael Todd Vistasp Karbhari Frieder Seible Blue Road Research Eric Udd Project Team Joel P. Conte, UCSD, SE Ahmed Elgamal, UCSD Magda El Zarki, UC Irvine Tony Fountain, SDSC Sami Masri, USC Mohan Trivedi, UCSD, ECE Amarnath Gupta, SDSC Remy Chang, UCSD, ECE Corneliu ‘Neil” Cotfana, SDSC Michael Fraser, UCSD Tarak Ghandi, UCSD, ECE Xianfei Daniel He, UCSD M. Erdem Kurul, SDSC Vipin Mehta, UC Irvine, ECE Dung Nguyen, UCSD, SE Kendra Oliver, UCSD, SE Minh Phan, UCSD, SE Peter Shin, SDSC Mazen Wahbeh, USC Linjun Yan, UCSD, SE