providing decision support for transport infrastructure maintenance planning: through application of multi-criteria and machine learning methods
Zaharah Allah Bukhsh is a PhD student in the department of Construction Management & Engineering. Her supervisor is prof.dr.ir. A.G. Doree from the faculty of Engineering Technology (ET).
Functional and serviceable transport infrastructure presents one of the essential predispositions for the economic growth of a country. The importance of maintaining transport infrastructure is increasingly recognized as we witness the aging of infrastructure, an increase in the frequency of extreme weather events, expanding performance demands, and shrinking financial resources. Under these circumstances, transportation agencies are facing competing demands to optimally spend the limited budget and satisfy various performance requirements related to reliability of assets, safety of users, availability of the network and impact on the environment. The multiple performance requirements of infrastructure give rise to several decision-making dilemmas.
Typically, infrastructure managers analyze the condition data, estimate the future performance of assets, and decide on the maintenance actions implicitly based on their technical knowledge, past experiences, and judgments. However, due to varying personnel knowledge, cognition capacity and experiences, the intuition-based decision-making suffers from inconsistency (e.g., different decisions for the same scenarios), and distortion (e.g., over and under emphasis of specific attributes, such as cost and condition states). This results in the decisions which are difficult to follow, justify, and reproduce in the future. The implicit decision-making by asset managers can also be attributed to several related factors such as the single-objective optimization methods, poor integration of qualitative data and preferences of experts, distributed asset management systems, and data accessibility challenges.
Aligned within the focus of two European projects, namely DESTination RAIL and COST ACTION TU1406, the objective of this research is to improve the decision-making process of maintenance planning by developing applied decision support methods and predictive models to aid transport infrastructure managers. The developed data-driven decision support methods firstly enabled the optimal maintenance planning of assets over the multi-year period, and secondly used the data from asset management systems for predictive modeling of unseen future events. The proposed approaches explicate the implicit reasoning of experts and pave a way forwards towards evidence-based asset maintenance solutions. The methodological developments of the thesis are highlighted below:
· A multi-attribute utility theory (MAUT) method is developed to accommodate multiple objectives of maintenance for all the assets of the network. The proposed methodology also introduces a procedure to quantify the objectives in the form of performance indicators. Additionally, the model transforms the subjective preferences of a decision-maker into objective values in the form of utility functions, and performs trade-offs among multiple performance attributes. The resulting prioritization of assets directs the investment decisions of maintenance for assets of the network. (Chapter 2 and 3).
· The MAUT method is further extended into a holistic computational framework that aims to find the best time to maintain an asset under the budget and performance constraints. The framework seeks to develop an optimal multi-year maintenance plan by synthesizing the type of maintenance treatment; estimating the future performance of assets by Markov chain processes, and utilizing the genetic algorithm for optimization. The proposed approach enables asset managers to simulate various maintenance planning scenarios under different budget and performance requirements. (Chapter 4).
· With the objective to make the maintenance planning procedure smarter by using the asset management data, the predictive models are developed using tree-based (machine learning) classification techniques. The classifiers accurately identify correct maintenance notifications and treatment type through modeling the historical data of unplanned maintenance triggers of railway switches. Besides, the feature importance analysis of predictive models shows the essential data attributes and also reveals intrinsic decision logic. (Chapter 5)
· The large amount of historical data deemed as big data is processed for the predictive maintenance of road bridges by developing a deep neural networks with entity embeddings. These models can extract insights and have shown to learn complex non-linear features from the inspections and damages data of the bridges. In addition to discrete models, a unified model utilizing the multi-task neural network is developed which jointly learns to solve multiple tasks through utilizing shared embedding and task-specific layers. The introduced deep models can be used for the transfer learning to gain performance improvements on tasks in a low-data regime. (Chapter 6)
This paper-based thesis addresses the challenges of maintenance planning by proposing multi-criteria methods and machine learning models. The proposed multi-criteria methods reduce the preferences of experts into objective data, establish the ranking of discrete assets and create multi-year maintenance plans to facilitate asset managers in deciding which assets to maintain, when to maintain them and what are the consequences of delaying maintenance in terms of budget and performance of assets. The developed predictive models learn from the historical asset management data and facilitate in maintenance planning through predicting the (future) condition states, risk levels, need of maintenance for assets.
This research has made progress towards more consistent, explicit, and evidence-based maintenance planning approaches, which makes the decision processes concrete, transparent, and reproducible. The suggested methods specifically concentrated on providing support to infrastructure managers; therefore, the usefulness of the proposed approaches are validated on the real datasets of highway bridges and railway switches. Moreover, where it was possible, the digital tool and code are provided to motivate the implementation of the methods in practice. Finally, these methods eliminate the gap between the appropriate use of historical data and implicit judgment-driven decision-making of experts and pave a way forward towards data-driven resources efficient asset management practices.
