Deep transfer learning leveraging the prediction of atmospheric corrosion monitor sensors

Atmospheric corrosion sensors (ACM sensors), based on the principle of measuring galvanic corrosion between two electrodes with different electrochemical potentials, have proven to be an efficient way of assessing the corrosiveness of the environment from which the expected corrosion performance of a metal can be derived.
Nonetheless, the data generated by ACM sensors is complex, and it is challenging to determine interpretable correlations between the different variables measured using only analytical corrosion expertise. Recently, studies1,2 have shown that machine learning models are able to unravel unexplored patterns between the environment and the current output, thereby unlocking the potential to develop predictive atmospheric corrosion models. However, one limiting factor is the large amount of data from long testing times required to make proper predictions in the environment where the sensor was placed.
Our research argues that the specific subdomain of machine learning called transfer learning (TL) can solve this issue. In essence, TL attempts to transfer knowledge from a previously developed machine learning model (the source) to another similar task (the target). As such, it provides a way to improve the modeling performance in domains where data is scarce. When predicting the sensor's output for a year, fine tuning of previously trained source model is required. This source model can be trained on data derived from a sensor installed in a controlled lab environment or a different geographical setting. The results obtained so far suggest that when only 1 to 3 weeks' worth of hourly ACM sensor data is available in a particular outdoor setting, a two-fold increase is expected in the target model accuracy of 1 year.