I am a diagnostics engineer with an established Condition monitroing company. We produce the hardware (Piezo accelerometers and the permanent online data collectors) The signals we retrieve are of the highest quality.
In this CM sector things moved rather slow over the past decade, till this day most of the analysis is done with our eyes, inspecting the spectra for known frequency components or patterns and tracking them. Using static alarming thresholds. Except form the fact that this is very innefficcient it is also very susceptible to human error.
A first major improvement was applying simple statistics.
The basic signal before FFT, the TWF is what every sensor produces and is the start for analysis currently. Applying the signal conditioning and FFT using matlab we extract per frequency component the regression over time and trend. (plus weighting and some harmonic pattern finders)
Getting this into an automated code is currently my job, labeling frequency components and storing them with the data as well. As you might have noticed this is far from AI. The main reason is to go step by step reproducing the way we analyse ourselves more or less, having still full knowledge of the process. Another reason is mostly the lack of interest by the management over the past years.
The plan is to go from analysing all data to analysing only devient data. This is where the statistics come in handy.
Let’s have a look at the following image (on the left):
The benefit for the analyst (or the one paying his salary) is big, as we are currently only at the end of step 2 I can’t say what the time saving’s of the next steps are.
I made a dashboard visualisation displaying the status of the machines:
see the image above on the right
Using mqtt for real time alarming and bokeh and datashader for historical data plotting the goal is to display the results through the same web-platform.
I am looking for some people that share interest in this project, open minded to open source on the project.