Validation Of Smart Sensors For AI-Based Condition Monitoring In Industry

By André Schneider and Martin Lehmann

Industry has challenging requirements for the quality and reliability of the distributed smart sensors used in condition monitoring. As a result of digitalization and the increasing use of AI-based monitoring solutions, we can expect an increase in demand for systematic, efficient design and validation processes. In the future, a product’s journey from prototype to volume production must be accompanied by innovative test and validation strategies for qualifying each sensor system under conditions that are as close as possible to those of the future operating environment. The main goal here is to minimize both the need for future rework (which tends to be expensive), and the risk of failure.

In the context of industrial digitalization, sensors are essential components that deliver crucial data for optimizing production processes and the systems that monitor machinery. The wide availability of low-cost wireless communication and energy-efficient microelectronics is paving the way for inexpensive smart sensor systems that can be easily integrated into machine components. This means that the sensors themselves are now complex cyber-physical systems that, when used in an industrial environment, must meet rigorous requirements with regard to reliability, energy consumption, and robust wireless communication. What’s more, the growing share of software-implemented functions coupled with the integration of AI means that the features of these systems can evolve during their lifetime.

It therefore makes sense to use established processes for developing mechatronic, data-analysis, and monitoring systems as the basis for defining a method for the fast and efficient development of smart sensor systems. The first key steps are to collect ground-truth reference data directly at the machinery using high-precision sensors and analyze this data – at first independent of the target hardware for the smart sensor system. The next step is to evaluate potential hardware solutions, both in terms of how to integrate the data analysis as well as how to collect, convert, and process the data.

To test this approach, the authors used an IoT device designed to monitor the condition of an ultrapure water supply for semiconductor production. The aim of this sample application was early detection of valve damage through continuous analysis of vibrations. The individual steps were to collect reference data as part of a measurement campaign, to define test programs for the sensor system, and to test the overall solution. This involved compiling documentation for all the data collected and for the results of all measurements and tests. Overall, the authors showed that relocating validation steps to a suitably equipped laboratory environment helped reduce or eliminate the need for expensive and time-consuming tests in the real application environment.

Reference measurements generally require high-end measurement technology. Data aggregation was performed with the SKALI.KIT measurement system developed at Fraunhofer IIS/EAS. In just four weeks, six terabytes of raw data were collected and analyzed to find the patterns that would ultimately characterize the individual operating conditions of the machinery. The next step was to simulate the real situation using valves in the lab. A shaker reproduced the measured vibrations, which were captured again – this time using the exact hardware that the user could be expected to cost-effectively install throughout their facility. An edge device continuously aggregated data on vibrations with frequencies of up to 10 kHz. There followed a comparison of characteristic features, including frequency spectra and statistical KPIs. The final step was to perform a quantitative analysis to determine whether the chosen combination of sensors and edge device met all the requirements for AI-based condition monitoring of the ultrapure water supply.

The method described and illustrated here also lends itself to applications beyond validation. It can help develop algorithms for efficient and robust feature extraction based on reference and real raw data, and it can help train and validate potentially suitable AI model architectures. These pretrained models can then be integrated directly into the edge device hardware without having to alter the physical setup. In the future, the system’s energy consumption will be measured and evaluated in the lab as well. Overall, the method offers an effective and targeted way of establishing AI solutions within an industrial environment.

Martin Lehmann is a scientific researcher in the Smart Multisensor Systems group at Fraunhofer IIS/EAS.