Data-driven lead-acid battery lifetime prognostics
Abstract: To efficiently transport goods by heavy-duty trucks, it is important that vehicles have a high degree of availability and in particular avoid becoming standing by the road unable to continue the transport mission. An unplanned stop by the road does not only cost due to the delay in delivery, but can also lead to a damaged cargo. High availability can be achieved by changing components frequently, but such an approach is expensive both due to the frequent visits to a workshop and also due to the component cost. Therefore, failure prognostics and flexible maintenance has significant potential in the automotive field for both manufacturers, commercial fleet owners, and private customers.In heavy-duty trucks, one cause of unplanned stops are failures in the electrical power system, and in particular the lead-acid starter battery. The main purpose of the battery is to power the starter motor to get the diesel engine running, but it is also used to, for example, power auxiliary units such as cabin heating and kitchen equipment. Detailed physical models of battery degradation is inherently difficult and requires, in addition to battery health sensing which is not available in the given study, detailed knowledge of battery chemistry and how degradation depends on the vehicle and battery usage profiles.The main aim of the given work is to predict the lifetime of lead-acid batteries using data-driven approaches. Main contributions in the thesis are: a) the choice of the Random Survival Forest method as the model for predicting a conditional reliability function which is used as the estimator of the battery lifetime, b) variable selection for better predictability of the model and c) variance estimation for the Random Survival Forest method.When developing a data-driven prognostic model and the number of available variables is large, variable selection is an important task, since including non-informative variables in the model have a negative impact on prognosis performance. Two features of the dataset has been identified, 1) there are few informative variables, and 2) highly correlated variables in the dataset. The main contribution is a novel method for identifying important variables, taking these two properties into account, using Random Survival Forests to estimate prognostics models. The result of the proposed method is compared to existing variable selection methods, and applied to a real-world automotive dataset.Confidence bands are introduced to the RSF model giving an opportunity for an engineer to observe the confidence of the model prediction. Some aspects of the confidence bands are considered: a) their asymptotic behavior and b) usefulness in the model selection. A problem of including time related variables is addressed in the thesis with arguments why it is a good choice not to add them into the model. Metrics for performance evaluation are suggested which show that the model can be used to find and optimize cost of the battery replacement.
CLICK HERE TO DOWNLOAD THE WHOLE DISSERTATION. (in PDF format)