PREDICTIVE AI its power and use in the industry thus far , while also noting it has a fair way to go in being used to its full potential .
“ In manufacturing , audio information and audio learning models can partly replace the trained engineer who knows when a machine does not sound healthy ,” he further explains . “ Of course , the same predictive techniques can be used to gather data from several sources to inform condition maintenance . This requires less predictive power because it ’ s more concerned with accurately understanding the overall condition of an asset based on multiple inputs . Sensors can give false readings , but techniques such as cohort analysis allow these anomalies to be identified by comparing each sensor with others doing a similar role .”
Part of this , and what can be argued as a minor teething problem , is how the training data is sourced and subsequently implemented . As Wilson explains , traditional predictive AI relies on having enough of the right telemetry data , clean information about fault occurrence and relevant context . If the incidence of faults is rare , this makes it harder to train AI models , but techniques such as random forest can be used to mitigate this issue .
Offering a solution , Wilson details : “ More recently , drone surveys of visible assets have started to complement this approach , while generative AI can produce additional training data that enables computer vision models to be trained to spot faults .” energydigital . com 89