How do you handle missing, noisy, or imbalanced data for predictive maintenance?

Consistent readings and input of data measurements. The more often you perform this routine will affect how reliable your Predictive program is. To make this successful historical trending is factored in. When you don’t have that data or suffer lack of current trends you can circumvent and build it in through consistent readings and follow up it up with the input of measurements, hour or temp readings, etc. In time and with transition companies can use various programs and databases to monitor equipment readings or runtime. One of which would be inspection routes performed and entered by the tech in the field by mobile device. When synced to the cloud this offers real time data and a point of analytics.

Reliable predictive maintenance programs depend on consistent data input and frequent measurements. The regularity of this process significantly influences the program's reliability, and historical trends play a crucial role in its success. When historical data is unavailable or there's a lack of current trends, you can overcome this challenge by consistently collecting readings and supplementing them with input such as hour or temperature readings. Over time, companies can leverage various programs and databases to monitor equipment readings or runtime. For instance, inspection routes conducted and recorded by field technicians through mobile devices can contribute valuable real-time data when synchronized with cloud platforms.

In all the predictive maintenance projects I've been part of, missing data wasn't a big problem, thanks to our use of portable devices capturing data samples like vibrations or temperatures at regular intervals Regarding online systems, we relied on multiplexed systems that often surpassed the reliability engineers' expectations in terms of cost-effectiveness. These multiplexed setups inherently involve some data loss. To estimate missing entries over time spans, we usually turn to interpolation or regression methods, which maintain the integrity of diagnostics and the accuracy of the resulting maintenance recommendations, which are critical aspects of predictive maintenance.

I remember dealing with significant noisy data only once, caused by electromagnetic interference at an old factory. The interference stemmed from frequency inverters installed near AC motor systems but connected using outdated DC motor wiring. This setup introduced noticeable errors in the sensor signals. The solution was straightforward—shield the cables electromagnetically using properly grounded triaxial cables, and in severe cases, route these cables through grounded conduits. Another typical approach to reduce noise, especially high-frequency noise indicative of early-stage bearing failures, is signal enveloping, which involves using filters, rectification, and signal enveloping to clear out the noise.

Data imbalance is a critical issue when employing machine learning with typical sensor data from vibrations, temperatures, and some process variables. The main challenges with imbalanced data include bias, where the model predominantly learns from the "everything is fine" scenario, overlooking potential failures. It also tends to overfit, focusing too much on the nuances of functioning data at the expense of crucial failure detection. To counter this, we use resampling techniques such as oversampling the minority class. We might also adjust the learning algorithms to prioritize fault detection through cost-sensitive learning or ensemble methods.

A frequent oversight I've seen among young vibration analysts is ignoring the dynamic range of piezoelectric vibration sensors. Another important point is data reduction. At the end of the day, if everything was fine, you don't need to store repetitive stable or non-changing data. So, you can create data reduction rules.