Predicting road condition: four condition variables from three survey years and traffic load

The condition of a German motorway is surveyed at intervals of several years. Between two surveys lies a blind period in which maintenance measures have to be planned and tendered – on the basis of data several years old.

Road maintenance is therefore reactive: it follows measured decay. A prediction would reverse the order – measures could be planned where the condition is likely to become critical first, rather than where it was critical at the last measurement. That, exactly, was the task.

Historical condition data of the A70 in the Schweinfurt – Bamberg – Bayreuth section, surveyed in 2009, 2013 and 2017, provided by the Bavarian State Ministry of Housing, Building and Transport. Supplemented by traffic density data from 2005, 2010 and 2015.

Four condition variables were predicted: general unevenness, longitudinal evenness index (Längsebenheitswirkindex), skid resistance at 80 km/h, and road damage from cracking.

Three points in time are not much for a time series. Condition data from three surveys is the shortest series from which a progression can be modelled at all. What makes it harder: the traffic density data comes from different years than the condition data – the series do not line up.

Missing values at scale. Merging surveys taken at different points in time produced substantial data gaps. Handling them made up the bulk of the project work.

Four target variables with different behaviour. Skid resistance, unevenness and cracking do not develop alike and depend on load to different degrees. A model that treats all four the same will miss on at least one.

Merge the datasets and handle the gaps. Traffic density and condition data were merged. Variables with more than 50% missing values were excluded – a deliberate decision against filling in quantities where the imputation would be more assumption than information. The remaining gaps were imputed.

Four model approaches in direct comparison. A plain neural network, an LSTM, a CNN and the combination of LSTM and CNN were compared. The combination delivered the best scores – plausibly so, because the LSTM captures the temporal progression and the CNN the spatial structure along the route.

Result on the map, not in a table. The predictions were visualised in ArcGIS and show the severity of the road condition along the route – the form maintenance planning actually works with.

The combined LSTM-CNN model reached a coefficient of determination of 60% across the predicted condition variables.

That number needs context: 60% explained variance from three survey dates, with time-shifted traffic data and substantial gaps, is a solid result for this data situation – and considerably more than carrying the last measured value forward delivers. The value comes not from hitting the condition exactly, but from getting the order of the sections right.

Condition prediction from few measurement dates and one load variable is the task in all infrastructure maintenance: bridges, track beds, utility networks, sewers, industrial plants. The structure is always the same – rare, expensive condition surveys, continuous load, maintenance decisions between two measurements. And there too, the real difficulty is not the model but merging series that do not line up.

The bottleneck was the data situation, and that is where the most has moved. Condition data is now produced more frequently and more automatically, partly from the sensor data of vehicle fleets; load can be represented continuously instead of in five-year steps. Three points in time thereby become a real time series – and the model class we chose back then only comes into its own at that point. For handling gaps, we would today use methods that carry the uncertainty from the imputation along, instead of treating a filled-in value like a measured one. And we would state the prediction with an uncertainty interval per section: for a maintenance decision, the statement that one section will very likely become critical before another is worth more than a point value. What has stayed the same: without merging the condition and load data, there is no prediction. That is data work, not modelling.