Vibration analyses are a classic for measurement engineers. However, if you consider the complexity of these analyses and the required hardware investment, you will be amazed.So-called expert systems, which cost at least 100,000 € for 48 channels, are typically used as hardware. Fully equipped incl. software up to 250.000 €.
At PANDA, we offer this kind of system at a much more affordable price, starting at around 60€ per channel, which means that the same result can be achieved at around 4% of the standard price, and we also develop the complete Open Source data recording software.One of the reasons why the expensive hardware is called an expert system is that the time needed to understand and learn how to operate it is lost in the cost of the purchase, so that you become an expert in measuring technology at the same time, even though you only wanted to record a few data.
Considering that you need data for every AI application and with a Raspberry Pi available from 40€, it is actually no longer necessary to invest so much in hardware nowadays.One thing most people underestimate, however, is that data from sensors is not digital from the very beginning but is converted at great expense. In principle, electrons are counted as accurately and quickly as possible. A lot can go wrong in such a highly sensible manufacturing chain. The electronics in such a measuring system are therefore full of sophisticated circuit gimmicks before they reach the actual analogue-to-digital converter.
Once this has been achieved, the selected chip determines how exactly the values and how many values per second can be recorded. While amplitude resolution is usually not the decisive factor for oscillations, time resolution has the problem of the Nyquist-Shannon sampling theorem. Without going into detail here, the sampling rate should be at least twice as high as the maximum frequencies occurring in the signal being converted. To meet this requirement, laboratory measurement techniques use filters to filter higher signal components from the signal before the actual conversion, in order to get as close as possible to the theoretical Nyquist limit. This calls for very steep, active filters, which are difficult to design and generally cause noise. Furthermore, the sampling should be uniform, with so-called equidistant sampling. With professional measurement technology, however, this is always the case in contrast to low-cost consumer solutions. Often consumer technology does not use filters at all or only rudimentary switch-capacitor circuits with further disadvantages.
Going back to Nyquist, it can be really bad. You can’t digitally filter the signal in the computer if you don’t perform analogue filtering correctly. Even if there is a simple switch-capacitor analogue filter, instead of the minimum twofold distance for subsampling, the 10fold distance to the sampling rate is needed, since mirror effects can no longer tell which frequencies have actually occurred in the analogue world. In an FFT, for example, peaks occur somewhere else than where they really should be. You can no longer trust the results of a vibration analysis and have no chance to notice the error in the data itself. This is also strange if you sample at 20,000 Hz and only 1000 Hz can be evaluated sensibly. Unfortunately, this is not a theoretical factor, but occurs more frequently in practice than expected. Extreme measures are then used, such as a moving average filter.Without going into detail, by the way, you should never do this if you want to evaluate vibrations.
In our Timeswipe boards we use 16 bit amplitude resolution, which corresponds to 65,000 different measured values, and sample 48,000 values per second per channel. With a steep and low-noise 5-pol filter, frequencies up to approx. 20,000 Hz can be evaluated in the signal. Higher frequencies in the signal are correctly filtered out before reaching the converter. As a result, you don’t have to worry about Nyquist at all, because a lower desired recording rate is achieved by digital post-filtering in the driver.
Let’s take a look at how a measurement for a vibration analysis with PANDA technology in the vehicle is feasible.
The sensors, which cost approx. 300€ each and resolve over 20,000 Hz to match the boards, are attached to the different positions in the vehicle using a two-component adhesive. The cables are carefully routed into the trunk with cable ties, where they are connected to several Timeswipe devices, which are supplied by USB power banks and send the values via WLAN to a laptop or iPad for data recording.PANDA’s open source approach allows the values not only to be stored, but also to be used for real-time AI applications in the vehicle, such as intelligent condition monitoring. This is where our Metric Algorithms or neural networks come to use.
Including accessories, we are at approximately 7,000€ for 16 measuring points. This is still a lot of money, but it is a factor of 20 cheaper than an expert system.
We at Panda are working on reducing the cost of this by another zero in the future by removing the expert status of the sensors as well.