| dc.description.abstract | This study seeks to minimize the likelihood of casing failure using data analysis and machine learning algorithms. The study focuses on the design of a data-driven workflow that can address several challenging aspects that pertain to casing failure including: (1) identification of potential risk factors amongst different exposures through the adoption of risk analysis techniques (case-control study design), (2) evaluation of the type and magnitude of the impact for each risk factor, determined through the application of different association measurements, (3) identification of the levels within each potential risk factor that impose the highest risk on casing failure, (4) acknowledgement of the depths most susceptible to casing failure implemented through non-parametric and semi-parametric survival analysis techniques, (5) prediction of the overall probability of casing failure given the information for pre-defined risk factors via the application of multiple classification learning algorithms , and finally (6) have a scheme for mitigating casing failure. To account for the rigidity of machine learning algorithms and to allow for more control from the user’s end, risk assessment techniques were implemented, particularly, semi-quantitative probability-impact risk assessment matrices (PI-RAMs). These matrices were based on results obtained from frequency and survival analysis. PI-RAMs were used as feedback means to the initial predictions obtained from conventional ML algorithms to not only provide a better intuition of the overall risk, but also the contribution of each risk factor.
One major limitation to the proposed framework is the compliance of field engineers with the recommendations, due to cost-related concerns. This motivated the focus on yet another perspective of mitigating failure, that is the enhancement of casing integrity evaluation techniques, specifically fatigue failure induced from thermal stress. A steam injection simulated case has been used for the design and validation of the proposed data-driven model. A correct estimation of fatigue life of different casing parts must be made. The accuracy of fatigue life estimation is contingent on the accuracy of local strain estimations. Two classes of factors impact local strains: 1) active/direct (temperature changes, casing material, etc.) and (2) passive/indirect factors (such as cement cracks or leaks). Although physics-based models can account for direct factors, they still fail to account for indirect factors, leading to false conclusions on casing fatigue life and abusive consumption of casing parts beyond their capabilities.
The proposed data-driven estimator takes as input all direct and indirect factors, and outputs the corresponding local strains that reflect those effects. Then, using the casing material properties, along with estimated strains as input for Manson’s Eq. and estimating the fatigue life of those casing parts. Based on estimated fatigue life, the model can give recommendations on changing casing parts that are abused throughout any process (such as steam injection, or hydraulic fracturing). This would, ultimately, prevent or reduce the chances of the occurrence of casing failure. Using the proposed model, engineers can have a better understanding of the casing durability and ability to withstand the downhole conditions and practices. | |
