Towards a model based asset deterioration framework represented by probabilistic relational models

Abstract

Most asset deterioration tools are designed for a specific application, as a consequence, a small change of the specification may result in a complete change of the tool. Inspired by the model-based approach of separating problem specification from analysis technique, we propose a model-based asset deterioration assessment framework using probabilistic relational models. The probabilistic relational models express abstract probabilistic dependency covers a range of deterioration modelling assumptions. An expert in the domain of asset deterioration can then use his knowledge of the factors that affect deterioration to customise the abstract models to a specific application, without requiring a detailed understanding the underlying computational framework. We illustrate the use of the framework with multiple variants of deterioration models. can be adapted when there is insufficient data, both with expert knowledge (Frangopol et al., 2004, Zhang and Marsh, 2018), or by learning from similar groups (Memarzadeh et al., 2016, Zhang and Marsh, 2018). In the work of Zhang and Marsh (2018), six generic BN models for asset deterioration were developed, which both provides us the possibility of adopt different deterioration models, but also enables us to include alternative data and unused expert knowledge. These model variants cannot yet be presented to an asset deterioration domain expert in a unified framework: adapting the underlying concepts to a particular context requires a deep understanding of their implementation as BNs. In model-based machine learning (MBML) (see Bishop (2013) and Ghahramani (2015)), models and problem specifications are defined in a compact language, while inference or machine learning algorithm codes are generated automatically. Bayesian networks are such a language, though they lack structure. More recently, probabilistic programming languages such as Figaro (Pfeffer, 2009), has been developed which could also be used in our framework. Model-based approaches, both in MBML and MBSA, often use the object-oriented paradigm to provide a library of generalized models for reuse. This is not provided by traditional BNs, with a fixed set of variables and relationships. This issue has been widely researched for BNs, with proposals including idioms in Neil et al. (2000) and fragments in Laskey and Mahoney (2000). Probabilistic relational models (PRM), developed by Koller (1999) combines relational structure with probabilistic graphical models (i.e. BNs). A PRM combines probabilistic dependencies with a relational schema that describes the entities in the problem domain. This representation provides a separation between model library and structure relationships. Therefore, we propose to develop a model-based framework for asset deterioration assessment in the spirit of MBSA. The framework separates reusable low-level models from modelling choices and asset descriptions. The framework is encoded with a PRM representation of a hierarchical Bayesian network, with a range of generalised models for asset deterioration each represented by its probabilistic dependencies, and the problem specification of the target domain is represented as the relational schema. 3 MODEL-BASED ASSET DETERIORATION ASSESSMENT FRAMEWORK 3.1 Asset Deterioration Model using Hierarchical BNs 3.1.1 A Simple Deterioration Model Figure 1. A simple deterioration model. For a system that is either working or failed, given historical data on the times that it remained in working condition, we can estimate the distribution of time for its transition to the failed state and so predict its likelihood of failing. A basic deterioration BN model, from Zhang and Marsh (2016), is shown in Figure 1. This is a hierarchical BN model that both learns from data and can be used for decision support. However, this specific model can only be used to describe a type of asset with two-state and deterioration that follows a one-parameter distribution. This is not usually the case in asset deterioration, for example a 4 point grading system is used to describe bridge condition, and a two-parameter Weibull distribution is used to fit the bridge transition distribution in Sobanjo (2011). So instead, the model has to be adapted: the variables are similar but the number of them and links between them must change. Figure 2. Effects of prior knowledge and data quantity in distri-