PSI - Issue 19

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

35

9

4.2. Fatigue Crack Inspection The coke drum is an example of an asset where increased maintenance and repair cost is more concerning than a catastrophic failure. It is assumed that after cracks initiate, there is sufficient time to inspect and repair them during regular, scheduled shutdowns before any failure occurs, so that the risk of failure is very low. For other assets where the consequence of failure is much higher and where a crack might not be as easily detected, a modified decision network is required. When building complex decision networks, it is better to focus on subnetworks that can be more easily designed and understood and then to assemble these into larger networks. A subnetwork for fatigue crack initiation/growth is shown in Figure 6. This network includes the cost of failure, should it occur. Here, no inspection or repair decisions are considered, but these are included in the next example for comparison.

Fig. 6. Crack Growth Rate with No Inspection or Repair Bayesian Decision Network.

A decision node that represents a typical design choice is also included with two possible states: (i) a higher cost more crack resistant material, or (ii) a lower cost less-crack resistant one. The more crack resistant material is able to withstand a larger number of cycles until a crack initiates and the crack grows more slowly once it does. Obviously, this is a great simplification of the real scenario, but it illustrates the typical tradeoff decisions that need to be made. Should an asset be built out of a more expensive but crack-resistant material that could save money in the long run (by reducing expected inspection, repair and failure costs) or should a cheaper material be used? Using a probabilistic decision network, this question can be answered with whatever information is available, no matter how good or bad that information is, because uncertainty in any parameter is directly quantified by using probability distributions for its possible values. Since only relative costs are the driving factor for determining optimal decisions, this network uses costs and benefits that are all chosen as relative magnitudes. For example, the Failure Cost is assumed to be 100 times the Installation Cost , the Design Cost of a crack-resistant material is assumed to be 100 times that of a less resistant material, and the Revenue Per Cycle of operation is assumed to be 1/100 the Installation Cost . A normal probability distribution is used to represent the Cycle of Crack Initiation with a mean value of 200 cycles for a crack-resistant material and 150 cycles for a less crack resistant material. The same standard deviation of 20 cycles is assumed for both. These are just representative values for illustration that could come from any desired fatigue model (further dependencies are not shown here for the sake of simplicity). Once a crack initiates, it