Predictive Maintenance – Forecasting the road condition of the highway A70
a Supper & Supper Use Case
For a road builder it is essential to plan measures for the maintenance of the infrastructure based on the respective roadway, building condition and their development. Predictive analysis can lead to monitoring of public infrastructure hence cost efficiency in maintenance. On the other hand, the overall development of public infrastructure can be better monitored.
Objective: Develop an artificial intelligence solution to forecast highway conditions using neural networks on 4 characteristics – general road roughness, lengthwise roughness index, grip of the road at 80 km/h driving speed and the damage level of the highway due to cracks.
The Bavarian State Ministry of Housing, Building and Transport has provided historical A70 highway (Schweinfurt, Bamberg, Bayreuth) condition assessment data from the years 2009, 2013 and 2017. In addition, the traffic density data of the A70 highway was also provided for the years 2005, 2010 and 2015.
The data provided were collected at three different points in time, which resulted in a significant pro-portion of missing values in the aggregation.
First, the traffic density data was merged with the high way condition data, variables with <50% missing values were excluded from the downstream modelling process. To meet the requirement of neural networks, that missing values should be imputed before the modelling. 4 predictive performance models were compared: Neural Network, LSTM (Long term short memory), CNN (Convolutional Neural Network) and the stacking model of LSTM + CNN. The R-square of the stacked LSTM+CNN model was the highest and shows the best model performance.
The combination of LSTM + CNN model was successfully used to predict the road condition variables, with 60% accuracy. his prediction was visualized in ArcGIS showing the degree of the road condition.
Category
→ GEO AI
→ Infrastructure
→ Predictive Maintenance
→ Spatial temporal Analysis
Technologies
Deep Learning
LSTM-CNN
