Future Prediction

This subsection investigates the prediction performance for future inspection. The training data for this subsection is same as the previous subsection, but we use data from year 2010 to 2017 to validate the performance of different approaches in predicting deck conditions over different future time length: from one year (year 2011) to seven years (year 2017).

The results are shown in Figure 5.14 measured by the average RPS of all states (the accuracy rates between methods are all quite close but with the same trending as RPS, hence not displayed). All methods have an increasing RPS over time; that is, the performance gets worse with the increase in prediction time. Our approach HierBN prevail other methods by having the lowest scores across these seven years. By comparing to the regular BN, we can also see the drastic improvement bring by separating the population into groups and learning between them. MCLR_G has a very close performance with the proposed approach. However, compare to MCLR, the benefit of grouping in MCLR_G is not significant in short future prediction but slowly increase over the years.

These experiments show we can provide a reasonably good prediction for a 2-year inspection (2012) with an RPS of 0.021 and an average accuracy rate of 87.6%, and a 4-year

Figure 5.14 RPS (lower better) of different approaches with the increase in prediction time.

inspection (2014) with an RPS of 0.047 and an average accuracy rate of 77.5% in HierBN. This counts for one and two biennially inspections in a regular inspection scheme respectively, which is valuable information for future inspection planning and resources scheduling.

5.5

Summary

This chapter validates the deterioration prediction models developed in Chapter 4 using synthetic data and real-world data to fulfil Objective IV.

We first produce a synthetic dataset to show that, with enough data, it is possible to learn the parameters of a deterioration distribution. However, when there is less data relevant knowledge about the parameters in the form of Bayesian priors helps to improve the performance. We also simulate data from a mix of deterioration distributions and show that we can provide a better prediction if we can separate assets into groups, where each group follows one of the distribution. But the challenge is to know how to classify assets into groups in the first place.

We tackle this challenge using the features associated with each asset and explain the processes using the NBI database case study. First, we introduce the background of the NBI database and its challenge of having limited deterioration data in some cases. To separate assets into groups, the features that have the most impact on the deterioration time should be identified. We perform dimensionality reduction to identify the key features for

predicting asset deterioration using a modified random forest. The values of features are used to separate assets into groups. We use a hierarchical BN developed in Chapter 4 to learn both the hyperparameters of the overall population and the local parameters within each group. The model can also learn the weights of each feature for influencing the deterioration rate. The transitions distributions learned in this way are further assembled to provide individual multi-state prediction given the state of an asset observed in the latest inspection and inspection time.

Lastly, we measure the performance of the developed deterioration prediction models using the NBI data. We first evaluate how many features to consider in the models. The example shows that the performance improves at a rate up to three features; after this, more features can only improve the performance slightly. We also compare the performance of our models with other available approaches for correctly predicting the condition of a multi-state system. Our results show that our proposed models excel at most predictions, especially for cases where there is little data. We also show that as the prediction time increases, the accuracy of the prediction drops. The proposed models can provide reasonably good accuracy over 1 or 2- biennially inspections in a regular inspection scheme, which could be useful for inspection planning.

Inspection and Maintenance Decisions

Support

The challenges in making inspection and maintenance decisions are to inspect a suitable asset and perform an appropriate repair action at a suitable time. Understanding asset states from its deterioration and the effectiveness of repair actions are the foundation to recommend these decisions. A range of techniques was proposed to support these decisions, and some of them have been implemented in industry. But as discussed in Section 2.3.1, classical deterioration models cannot properly handle uncertainty and situation with little deterioration data. This problem is common for critical infrastructure like bridges, where deterioration data are often uncertain, and data are rare (Section 2.2.1). Another challenge is from the complexity of the modelling with various assumptions (Section 2.3.2 and 2.3.3) and further, to reason inspection and maintenance decisions (Section 2.4) from the models, where classical approaches often lead to an unmanageable model size that becomes difficult to perform analyses.

Validated in Chapter 5, models build with a Bayesian framework offer the flexibility in handling uncertainty and can learn deterioration from data and expert knowledge even in the case with small data amount. Pointed out in Section 4.3, we can also use the models to represent various system configurations and extend them for complex and large-scale problems (later showed in Chapter 7). This study does not have access to conduct expert knowledge elicitation. Instead, this chapter shows how these models can be built using example knowledge from published documentation to support a variety of decisions, from inspection decision, repair decision, to maintenance planning using real-world case studies (Objective V).

The remainder of this chapter is as follows: Section 6.1 presents the use of deterioration models in evaluating asset conditions. It is illustrated by a GB bridge example and a US bridge example, where assets are assembled by multiple components with different

system configurations. The US example is continuously studied in the following sections: for inspection decisions in Section 6.2, repair decisions in Section 6.3 and for strategic maintenance planning in Section 6.4. Section 6.5 concludes this chapter.

6.1

Condition Prediction and Structural Evaluation

With the deterioration prediction models from Chapter 4, we can predict the condition of components or assets, and their performance is validated in Chapter 5. This section introduces two applications of them, where the predicted conditions of two assets are determined by the condition of their components but with different system configurations.