Minimize downtime and increase machine efficiency through predictive maintenance, part 2

In the following, we will now go through the individual steps for setting up a predictive maintenance system.


Collect Data

The first step is to collect a large amount of sensor data that represents normal and faulty operation. Modern machines and systems are often already equipped with a large number of sensors for direct control. By recording the sensor data accordingly, you can now use it for predictive maintenance.

You also need to ensure that this data is collected under different operating conditions. For example, machining different materials with a CNC milling machine requires different speeds.


Preprocess Data

Once you have collected the sensor data, the next step is to remove outliers and clean the data by filtering out the noise.


Identify Condition Indicators

Sometimes further preprocessing is necessary to obtain additional information that is not apparent in the original form of the data. Conversions, for example, can help us extract some useful features, also called condition indicators.

In the graph, you can see that the peaks of the frequency data shift to the left as the power of the pump decreases, so the peak frequencies can serve as condition indicators. So far, you have extracted some features from your data that will help you understand the normal and faulty operation of the pump.


Train Model

However, at this point you don't know for sure which parts need to be repaired or how much time is left before failure. In the next step, you can use the extracted features to train machine learning models. The models can then be used to identify anomalies in the data.

You can train a classifier to detect different types of faults, giving you insight into which parts of the machine need your attention.

Or you can also predict the trend that the machine will follow as it transitions between these two states.

Finding a model that captures the relationship between the extracted features and the progression of deterioration in, for example, pump performance, will help you estimate how much time remains before failure and when you should schedule maintenance.


Deploy and Integrate

After you develop a procedure, it can be deployed to the cloud or to your edge device.

Alternatively, if you have large amounts of data and the amount of data that can be transferred is limited, you can perform the preprocessing and feature extraction steps on your edge device and then transfer only the extracted features to your predictive model running in the cloud.

In this article, we've described why predictive maintenance is important and the steps you need to take to develop a process that can detect problems in your machines and give you advance notice of a future failure.

In productive operation, there is often not enough data on faulty conditions. In the next article, we will explain which procedures can nevertheless be used to detect abnormal conditions at an early stage.