Previsão de falhas de garantia em produtos complexos: uma abordagem híbrida entre Weibull e Crow-AMSAA

Abstract

Forecasting warranty failures is a strategic activity for companies that sell complex products, especially in the automotive sector, where the diversity of components and failure modes poses additional challenges. Accurate estimation of these events is essential to ensure that resources allocated to warranty claims are adequately provisioned, avoiding both shortages and excesses. In this context, the role of statistical modeling techniques applied to reliability stands out. This work aims to propose a hybrid approach that combines the Weibull and Crow-AMSAA models to optimize warranty failure prediction, seeking a balance between predictive accuracy and operational feasibility. The methodology considers steps applied to a set of real data from a heavy vehicle manufacturer, comprising 32,636 occurrences. Taxonomic categorization classified them as originating from 2,399 distinct failure modes, where 55.6% represent single occurrence situations that account for only 4.6% of total costs. Submitted to an Augmented Dickey-Fuller test, this data set revealed a predominance of non-stationary series (76.4% of failure modes), confirming statistical properties that vary over time. To handle this amount of information, frameworks were built to assist in preparing the data for statistical modeling. The statistical models were evaluated comparatively using an R language script, using error measures and correlation indices to assess predictive robustness through the statistical method of random resampling bootstrap. The development of a roadmap to describe the implementation steps of this hybrid strategy groups the findings of this research into a logical sequence of implementation. The proposal was validated through the empirical application of the roadmap to a set of real data, demonstrating that the proposed hybrid model presented greater predictive robustness, significantly reducing the mean and quadratic errors in relation to the Crow-AMSAA model, both in failure and cost prediction. The proposed model also demonstrated optimization of the selection of failure modes to be modeled by Weibull, suggesting a participation between 40 and 50% of the cost identified in the exploratory forecast, thus avoiding wasting time on failure modes with low financial representativeness. It is concluded that the hybrid model is preferably recommended for performing warranty cost forecasts in automotive processes, as the combination of the Weibull and Crow-AMSAA models proved effective in balancing analytical accuracy and operational scope, confirming the complementarity between the approaches.