Recommendations

After working with the road maintenance data from the City of Copenhagen, we gathered valuable insights. In previous sections we discussed our findings and results of the models that we applied our data on. Furthermore, in chapter 3 we presented the current organizational structure, systems and road maintenance processes. Based on our experience that we gained in this project, we developed the following five recommendations for the City of Copenhagen:

Recommendation 1: Establish a data structure and data strategy: The systems RoSy, Giv et Praj and PUMA have one common ground: They include useful data about road maintenance and damages. As we have shown in this thesis, combining data from all systems leads to better results and integrating the data in a data warehouse enables a more accurate data driven decision making. Currently, only a few people understand the data from different data sources, while no one has a complete overview.

We recommend to enable a holistic view, where different teams can combine their knowl-edge to produce better overall results. Also data preprocessing becomes no longer necessary.

A standardization of values for different features like district, road class, road status, road type or the coordinate systems of the geo data would highly reduce manual effort. Therefore, data pipelines can be automated to save time and improve the process.

In order to build predictive models, the way that the data is currently generated has to be

reconsidered. Further features can be collected, however continuous and objective collection is the most important key factor as fewer assumptions lead to more accurate predictions.

Recommendation 2: Improve data quality: In our prediction models, we had to make a lot of assumptions to fill missing data values in order to use the data as input features (for example traffic data, completion dates or road type). In other cases, we left out entire features because the data was not complete and could not be replaced with assumptions (e.g.

thickness of pavement layers). As we have mentioned previously, the less assumptions that have to be made, the more accurate the resulting predictions. We recommend to improve the data quality by making it mandatory to register only complete data entries for upcoming assessments or by automating the data collection (see Recommendation 5). Furthermore we recommend to collect additional data, including the data about restoration activities (e.g. date of maintenance/restoration activities, costs, contractor information).

Recommendation 3: Verify the accuracy of the current RoSy lifetime prediction: We showed that it is possible to create reasonable PCI prediction results where we can understand the importance of the individual features. For the predictions we take the remaining lifetime prediction from RoSy as the ground truth. Since we do not have actual data when and how the streets have been repaired, all the results are based on the assumption that this is the actual value when a street has reached its lifetime. In case that the current lifetime prediction by RoSy cannot be verified as accurate enough, our model can be simply retrained on the actual values and potentially produce better results. In contrast to RoSy, our model can also include additional variables for its predictions.

We recommend to verify the accuracy of the remaining lifetime prediction that RoSy calculates for each assessed road segment. This is crucial to verify the accuracy of our assumption, which our PCI prediction is based on. Furthermore, our model can determine the feature importance of variables with the actual ground truth, meaning potential root causes of road deterioration. By understanding the root causes, resources can be allocated better to perform focused maintenance activities to extend the maximum of a roads lifetime while reducing costs.

Recommendation 4: Rethink role of Giv et Praj: During our analysis, we found it hard to include the data from Giv et Praj for various reasons. Firstly, about 25% of the reported

potholes have not been close to a street which is part of the RoSy system. The data is generated mainly by citizens without proper expertise to make an adequate judgement, in contrast to the road engineers who inspect the streets for the RoSy or PUMA data collection. The same reported damage could potentially be reported more than once by different people, or even by the same person who may be highly disturbed by the damage. Also, only the roads are taken into account where citizens who use the service live close. It is unclear for roads where no damages are reported if it is due to the fact that there are no damages or because no one uses the service to report the damage in that area. Furthermore, the current process where reported damages are manually inspected by PUMA road engineers produces a high time effort and leads to potential doubling of the damage so that it is reported in PUMA and Giv et Praj.

However, Giv et Praj can reflect the common mood from citizens of streets that are perceived as highly damaged. Therefore, we recommend to include the data in the decision-making for the renovation process. To reduce the current manual efforts we recommend to use an image recognition solution to automatically classify the reported damages and register them as tasks in PUMA.

Recommendation 5: Automate the data collection and damage classification: We have discussed the flaws of the currently available data from the present processes in previous sections. Our last recommendation for the City of Copenhagen is the most radical and turned towards a future road maintenance process. In our literature and case analysis we have shown examples of different technologies that automated the data collection and road damage classification, which we have drafted into two different future scenarios.

We recommend for future road maintenance activities at the City of Copenhagen to remove the manual inspection for road damages and replace them with an automated approach that can be conducted more objectively and in more frequent intervals. The technologies that we used in our two different future scenarios can serve as inspiration. Gathered data from automated processes should be conceptualized in a form that can be used for predictive maintenance activities.