How our GPS filters the relevant signals from thousands of sensor data points

The real challenge in data handling begins after data acquisition: from the continuous stream of measured values, we have to filter out the signals that are truly relevant. A single ‘outlier’ does not necessarily have any significance. But what if the value appears again or persists over a longer period?

Example: GPS module Smart Water:

Example: GPS module Smart Energy:

Our goal is to generate only relevant alerts. Notifications that support daily dialysis operations without overwhelming staff. To achieve this, we need to understand which values are normal in which context, which deviations are harmless – and which indicate a genuine anomaly.

This is exactly where we are further developing our GPS: moving away from pure data collection towards a self‑learning system that identifies and names the relevant signals from thousands of sensor data points.

Why context is crucial for evaluation

Sensor data only becomes meaningful when their temporal and operational context is known:

A grid power draw of 20–40 kW is completely normal during the day when the Reverse Osmosis (RO) system and the entire centre are operating at full capacity; but unusual during standby at night, for example between 00:00 and 06:00. At such times, depending on the base load, even a consumption of 5 kW could already be noticeable and might indicate equipment that has not been switched off.

Weekdays also play a role:

Data from a dialysis‑free Sunday look very different from those of a fully booked dialysis day. Values that are inconspicuous on Sundays would trigger a warning during the working week.

Why outliers are not automatically problems

A single value that ‘stutters’ briefly is often harmless. Small fluctuations are part of any dataset. What matters is whether

• the value reappears,
• it persists for a longer period,
• or it occurs in a context (for example, a specific time window) in which it is not expected.

The distinction between signal and noise is crucial. And this is where the further development of our system comes in. Our GPS software must be able to evaluate incoming data and recognise context‑dependent normal ranges, so that alerts remain truly relevant.

What our GPS software already delivers today

We currently receive around 1,690 continuous sensor readings from various centres. We record between 35 and 65 data points for each centre. From energy consumption and conductivity to operating conditions, everything is displayed in real time.

This data is evaluated in a structured manner.

For each sensor, we have defined individual thresholds and normal ranges – always within the appropriate context; depending on the time of day, day of the week and operating mode. This allows us to check daily which anomalies are truly relevant.

Next goal: Machine Learning

By defining thresholds and context‑dependent normal ranges, we are already creating a key prerequisite today: data can be assessed and classified. This makes it possible to see which trends are unremarkable and which deviate from the norm.

In the next step, these classifications can be used to generate training data. This makes it possible to apply learning methods that recognise patterns in larger datasets and understand correlations beyond individual measurement points. This leads to even more accurate evaluations.

Long-term perspective: Predictive Maintenance

In the longer term, this approach opens up the possibility of predictive maintenance. The goal is no longer to detect anomalies only when thresholds are exceeded, but as soon as problematic developments begin to emerge.