Deep Learning (DL) for monitoring slowly evolving degradation processes typically involves overcoming data drift, complexity, and unavailability issues resulting from dynamic and harsh conditions, and rarity of labeled failure patterns, respectively. While degradation patterns are mostly hidden in such complex data, observation-based DL lean towards producing uncertain predictions and/or overfit the model during training process. This problem is usually caused by the insignificance of certain data representations. Therefore, and particularly due to the sequential nature of data in such a degradation process, it is necessary to consider neighboring observations to judge the accuracy of its representation or improving it. In this context, instead of traditional observation-based learning philosophy, this paper presents data-driven sequential mapping, while health indices can also be represented as a vector of sequential data and not as a single regressor output changing the model’s architecture. Using a dataset generated from a mathematical model mimicking bearing degradation life cycles and responding to the aforementioned three main challenges, a comparative study is built on investigating observation-based and sequence-based learning paths. According to a well-defined visual and numerical evaluation criterion, a sequence-based methodology reflects a better understanding of data representations through parameter tuning reaching better approximation and generalization. Such results support the necessity to such learning mechanism, especially for sequential data, dealing with some sort of correlation, and degrade controversially.