The core of AI Powered Maintenance: AI supports and the Decisions are made by humans

“Predictive maintenance” is a buzzword often used these days. The concept behind is not clearly defined, but most agree that predicting the exact time of failure of parts is the core element. Consequently, the goal of “predictive maintenance strategies” is to trigger the maintenance department to change parts as close as possible to the predicted time of failure.

The typical industrial environment is not made for predictive maintenance strategies

Predicting failures on mechanical parts have been around for centuries (e.g. turbines), but the situation is vastly different to current industrial use cases. In the turbine case, the data science part was already considered during the engineering process. Sensors have already been integrated to measure certain parameters and the same type of turbine operated under the nearly same conditions provides a huge dataset to analyze and predict failures. For industrial machines, the situation could not be different. Most machines do not have sensors, or do not have the right sensors in the right place. The same machines operates under totally different environmental set ups. Different operators, different temperature, different humidity, different parts etc. extremely open up the space for potential failures and data science concepts struggle to cover a general “use case”. In addition to that, some machines may require to be operated on 3 shifts, whereas others may only be operated for a few hours per week. Consequently, machines do mostly also have different maintenance strategies which may also be subject to change over time.

Predicting the failure is always wrong

Let’s face the truth: No data science concept can exactly predict the time of failure. It may narrow it down, but the reality is always different. There is only one way to determine the exact point in time: Run the machine until something fails. And here comes Senzoro’s AI powered maintenance approach into play: As we have built the biggest industrial ultrasound database, we know very well, how parts sound like before they fail and this information is consequently given to the maintenance department which now has to decide to either

(1) Run the machine to failure (and maybe check/order the spare part)

(2) Schedule the part for change and choose the urgency based on the machine priority in the process chain

The decision can only be made by humans and our AI learns from every event. With every change and/or failure event, our AI will become smarter and closer to the real point of failure, but even on the first day, our AI can already tell if the part is likely to fail or not.