Abstract Within the oil and gas industry, measurement-while-drilling (MWD) tools have several sensors to provide telemetry data that allow performing either live or post-job health diagnostics of the tool components and effectively replace/repair them to avoid unnecessary nonproductive time. Nevertheless, health checks are often purely telemetry-based, not using past job information to quantify cumulative component degradation, which increases the probability of failure during subsequent jobs. This work proposes a methodology that (1) introduces historical data to improve on failure detection and (2) benchmarks it against single-job-based models over the same features for a particular MWD tool. We find that by stacking historical features (as multivariate time series), failure prediction using one-dimensional convolutional neural networks provides better estimates of the probability of failure. We apply our methodology to five of the most common electronic component failure detection models trained over the time series the telemetry channels of 38 coiled tubing drilling MWD assets, totaling 752 jobs, and present the comparison between single-job-based models (with an average F1-score 0.66) and historical data-based models (with an average F1-score of 0.75).