Artificial Intelligence (AI) and Markov Process Based Data Mining on Predicting Bridge Operating Conditions

Details

• Creators:
  Fang, F Clara ; JimenezGil, Daniel ; Xu, Peter ; Kazemzadeh, Mohammadrahim
• Corporate Creators:
  University of Hartford. Department of Civil and Environmental Engineering
• Corporate Contributors:
  Connecticut Department of Transportation. Bureau of Policy and Planning ; United States. Department of Transportation. Federal Highway Administration
• Subject/TRT Terms:
  Artificial Intelligence Bridge Management Systems Data Mining Forecasting Inspection Markov Processes Neural Networks
• Publication/ Report Number:
  CT-2322-F-24-1
• Resource Type:
  Tech Report
• Geographical Coverage:
  Connecticut ; United States
• Edition:
  Final Report, June 8, 2022 – July 31, 2024
• Corporate Publisher:
  Connecticut. Department of Transportation
• Abstract:
  To facilitate bridge management decision-making and implement preventive maintenance strategies, we endeavor to develop a machine learning model to predict the future condition rating of bridges using historical inspection data, with a funding contract with Connecticut State’s Department of Transportation. We collect the data from the USA's National Bridge Inventory (NBI) about the nation's bridges, including their location, design, condition, and usage. We process, reformat and transform the collected datasets (1992-2021) so that the machine learning model can be trained to predict the bridge conditions over the future 100 years. During the data processing, we tackle the challenges associated with predicting bridge condition ratings, such as dealing with heterogeneous data sources, handling complex high-dimensional data, and incorporating mixed types of data features. Also, we address the need for feature selection and long-term prediction capabilities. To meet with these challenges, we propose a comprehensive framework based on deep neural networks, which enables us to effectively predict future 100-year bridge condition ratings based on the historical data during 1992-2021. Our approach utilizes all inspection features used for bridge condition rating. We employ an autoencoder neural network to compress the high-dimensional categorical features, and a long short-term memory (LSTM) recurrent neural network (RNN) to compress the high-dimensional sequential data into lower-dimensional latent features. The latent features obtained from the pre-trained autoencoder, and LSTM networks serve as inputs to a multilayer perceptron (MLP) network for the final bridge condition rating prediction. We adopt a two-stage training process, wherein the autoencoder, LSTM, and MLP networks were trained and tuned separately. The final pre-trained model is then applied to perform a comprehensive study of predicting the condition rating of all bridges in Connecticut. All the results predicted are organized in the deterioration trends of all bridges, the deterioration curves of selected individual bridges and the bridges with potential risks.
• Format:
  PDF
• Funding:
  SPR-2322
• Collection(s):
  US Transportation Collection
• Main Document Checksum:
  urn:sha-512:62ae31f39b8a0a3ec5e169200a7b7bb7da25703a9b84b6591b9243edf5d02fd9a8473146415f03e5c6056766cc4b0f94727f9409b7a8a9357883b2f23dd49140
• Download URL:
  https://rosap.ntl.bts.gov/view/dot/83075/dot_83075_DS1.pdf
• File Type:
  [PDF - 14.88 MB ]