An encoder-only transformer architecture with the traditional 1-D convolutional projections for queries, keys, and values instead of classical fully-connected layers is studied in this paper, as well as substitution of sinusoidal absolute positional encoding with the learnable vector for fixed-size context length. In order to explore alternatives to the standard softmax attention mechanism, this method implements and compares a learnable sketched kernel-based attention function with the conventional softmax function. The model is trained and evaluated on the NASA C-MAPSS FD002 dataset for remaining useful life (RUL) prediction. When using the sketched kernel-based attention, a marked improvement in prediction accuracy was noted. These results demonstrate that sketching techniques can provide an effective trade-off between computational efficiency and modeling fidelity in transformer-based prognostics given longer sequences and operational conditions. The source code is available at: https://github.com/nikitakolovorotnyy/RUL1DconvTransformer .