Design paradigms for solving sensor detection reliability problems
There are two main paradigms for solving classification and detection problems in sensor data; model-driven and data-driven. This article explores how digital signal processing using low power micro-computer units (MCUs) connected to an analog sensor was used to produce significantly improved quality of signal data. The problem to solve was one of false no detection.
Model-driven is the way everybody learnt to do it in Engineering School. Being a graduate of Imperial College, London - this is how I learned. Start with a solid idea of how the physical system works. Consider the states or events you want to detect, generate a hypothesis about what aspects of that might be detectable from the outside and what the target signal will look like. Collect samples in the lab and try to confirm a correlation between what you record and what you are trying to detect. Then, engineer a detector to record those hard-won features reliably in the real world.
Data-driven is a new way of thinking, enabled by machine learning. This wasn't taught when I was at studying engineering. Select an algorithm that can spot connections and correlations that you may not even know to suspect. Turn it loose on the data. Magic follows. But only if you do it right. The key here is understanding which algorithm to use with the data and outcomes you need to achieve. This is where some data science understanding comes in handy.
Both of these approaches have their pluses and minuses.
Model-driven approaches:
+ Limit complexity
+ Are powerful because they rely on a deep understanding of the system or process, and can benefit from scientifically established relationships.
- Can’t accommodate infinite complexity and generally must be simplified.
- Have trouble accounting for noisy data and unincluded variables.
- Are expensive and take time, it is inherently trial and error.
Data-driven approaches:
+ Can handle noisy and incomplete data.
+ Can handle complexity and indeed provide the means to measure correlation between inputs and predictions.
+ Fit well with Agile/Iterative development processes.
- Based on machine learning may require a lot of data to get good results.
- Are difficult to support on small memory low power micro-controllers.
There is no doubt that modelling a system takes time. It is inherently a trial-and-error approach, rooted in the old scientific method of theory-based hypothesis formation and experiment-based testing. Finding a suitable model and refining it until it produces the desired results is often a lengthy process. This is where data-driven can help accelerate the process, allowing a sensor design to be modelled and refined using real data collected to drive an iterative improvement in performance that can be quantified. The result has been a digital signal processing algorithm developed from a data-driven approach using machine learning techniques and tools and then squeezed onto a 4K bytes memory micro controller. This extends and enhances the analog sensor's output and produces a simple and reliable signal!
