Health monitoring provides insights about system health beyond raw sensor values, and helps you prevent severe failures and unplanned production stops by exposing the true health of your machine. When using incremental learning, the need of historic data is replaced by the power of real-time data.
With Ekkono’s edge machine learning SDK it is possible to incrementally train a health monitor based on streaming data in your device. Incremental learning eliminates the need for historical data, since the actual learning is performed inside the individual devices with real-time data collected out on the field. The learning starts with a short initial training period to determine the normal data state when the machinery is healthy. The normal data state is then used as a baseline to detect deviations or irregularities in the streaming data. Detected deviations are compounded into an overall health score, which can be used to determine the need for manual inspection of data or device.
By leveraging our incremental learning capabilities, it is possible to reduce the need for the extensive, and often expensive, data collection typically required in traditional batch training performed on historical data.
In practice there is often a lack of historical data including enough failures to support machine learning based modeling, and using simple threshold based models to determine health there is a risk of missing important indications of degradation that do not exceed the specified thresholds. Those deviations are readily detectable by our machine learning algorithms.
Learn the healthy state of the device and monitor how the real-time data deviates from this. The input sensor or array of sensors used should be related to the health of the device. Using incremental learning, the result is a unique health model for each individual device.
Any deviation from the normal state indicates that something has changed in the device. Use the health monitor to determine when manual inspection of data or the device is needed.