PSI - Issue 19

Aaron Stenta et al. / Procedia Structural Integrity 19 (2019) 27–40 Stenta and Panzarella / Structural Integrity Procedia 00 (2019) 000–000

32

6

values, but because there are numerous mutually exclusive possibilities with different levels of consequence, it is chosen to have several states such as: No Failure, Small Leak, Large Leak, Rupture, et cetera. The other chance nodes, i.e. Expected Damage , Measured Damaged , Measurement Error , Current Damage , Repaired Damage , Projected Damage , and Damage Rate would have states that are numerically discretized from their minimum to maximum values. There are four primary decisions that need to be made: (i) Should an inspection be performed or not? If so, what type? (ii) Should FFS be performed to analyze inspection findings? If so, what level of analysis? (iii) Should a repair be performed based on the inspection results and/or FFS findings? If so, what type? (iv) Should the component continue to be operated? If so, for how long? The BDN in Figure 4 is used for only one unit of Run Time . It is designed to answer the simple questions above, i.e. what is the next set of decisions that I should make, given my current state of knowledge? If no action is recommended after the first unit of Run Time (say 5 years), then this process is repeated for the next 5 years and so on (with the Projected Damage at the end of one time interval used as the Expected Damage at the start of the next) until some action, such as an inspection, is recommended. This iterative time-chaining of networks is a Dynamic Bayesian Network (DBN). Note, the BDN In Figure 3 is a simplified, but working, network that assumes damage rate is constant. For more advanced industrial applications, it may be required to reduce the maximum Run Time and increase the number of DBN iterations to capture the true time-dependent behavior of the system. Although Bayesian Decision Networks (BDNs) are a very simplified graphical representation of the cause-effect relationships, along with their corresponding conditional probabilities, they can get quite complex for industrial applications. The user is not required to interact with the BDN directly. ProbFat is delivered through a simple to use web-delivery platform that provides a central database of knowledge and promotes consistency across personnel by ensuring that all users have the latest software and can access it from any mobile or desktop device. In general, ProbFat has three primary means for updating knowledge: (i) Gathering and incorporating continuous monitoring data. (ii) Uploading inspection reports and critical observations. Based on inspection findings, ProbFat can be re trained to provide post-inspection repair recommendations and plan for subsequent inspections. (iii) Modification to existing conditional probabilities in the BDN by authorized organizational users to run diagnostics or correct/resolve inconsistencies. ProbFat can be set up to integrate with existing online databases to gather these sorts of data. 4. ProbFat – Demonstration Through Example In the following sections we demonstrate ProbFat through industrial examples that are chosen to highlight the capabilities of the BDN to optimize life-cycle decisions. Hopefully the reader sees the diversity in the tool, and its ability to be easily adapted to other applications. These are not generic tools that are pre-built for all applications. We are demonstrating that they are a means of solving industrial problems more efficiently and intuitively. 4.1. Coke Drum Application – Low Cycle Fatigue A typical Bayesian Decision Network used to optimize the operation of a coke drum is shown in Figure 4. This network is a graphical representation of the financial optimization analysis performed in Panzarella (2016), which was a purely analytical approach. By using a BDN like this, the multiple cause-effect relationships are more clear and additional factors can be added to create as realistic a model as necessary. The default states and conditional probabilities used in this network are too many to list here but most can be found in the above reference. Only a few parameter values will be varied here in order to show some basic trends. For simple problems, it is possible to use human intuition alone to arrive at a nearly optimal result, but when problems become as complex as this one, it is difficult to predict what the optimal decisions will be. That’s why decision networks like this are so valuable to assist with the decisions for complex scenarios at all stages of the life cycle management of any asset. The coke drum here is just one of many possible examples. Any asset can be better managed by constructing similar networks. The concepts remain the same, just the number and meaning of the nodes change. For a coke drum, catastrophic failure is not normally a concern. In this simplified example, the optimal Drum Cycle time and the Total Number of Operating Cycles are found by balancing the rate at which Annualized Revenue with the increased Annualized Repair Cost as the coke drum ages. One of the main forms of damage that