Deep Learning Toolkit for Predictive Maintenance User Guide
The Deep Learning Toolkit for Predictive Maintenance is a predictive maintenance solution that uses data collected from sensors connected to a machine to predict a possible fault of the machine.
To learn about the status of the machine, i.e. whether if the machine is working as expected, this tool does not need any log of the machine operation. The only data needed is readings from different sensors (e.g. temperature, pressure, and current) encoded with the OGC- Sensors Things standard data model. This data is automatically labelled from the tool according to thresholds specific to each sensor. According to these thresholds, if all the values are in a “safe” range, the data is labelled as ok. On the other hand, if one or more sensors result outside the defined “safe” range, the data from the previous 20 minutes from that specific sensor is labelled to mark a fault on the component the sensor is attached to.
Once labelled, an LTSM neural network uses the data to perform an online training process of a model that can predict if a fault is going to happen on the machine in the following 20 minutes. If the model is correctly trained, the tool can help the operator with scheduling maintenance operation and minimise its impact on the production processes.
The operator gets live information from the UI, which is accessible from the portal. With this UI, the operator can monitor information coming from the tool, such as the failure prediction and confidence in the prediction. Sensors readings can be monitored from the UI as well to help identify the location of a fault.
To access the Deep Learning Toolkit (DLT) for Predictive Maintenance user interface, open a web browser and type the URL provided by LINKS. The login section will appear, and you will have to enter the EFPF username and password that have been enabled to access the tool. After you have logged in, the DLT interface will appear. As shown below, there are three tabs that can be accessed: Predictive Maintenance, Sensor Data and Help. Click on each tab to open the corresponding section.
Predictive Maintenance Tab
In this tab the three main elements of the Deep Learning Toolkit user interface are shown, namely two graphs and the semaphore widget.
- The first graph, named “Prediction of an imminent failure”, shows which is the prediction of the DLT about the possibility of an imminent failure of the monitored machine. In particular, a prediction value equal to 0 at a given time instant means that, in that moment, the DLT believes that there will be no imminent failures. On the other hand, if the prediction value is equal to 1, the analysis of the sensor data led the DLT to believe that there could be a failure soon.
- The confidence that the DLT has in each prediction is shown in the second graph, named “Confidence score for the prediction”. A high level of confidence, i.e. near to 1, means that the DLT is pretty sure about the prediction for that time instant, while a low level of confidence, i.e. around 0.5, means that the DLT is more uncertain and therefore the prediction is less reliable.
- The semaphore allows you to monitor the status of both the machine and the DLT at a glance. The green light means that the machine is operating normally and that the DLT is confident about its prediction. Yellow stands for low confidence in the prediction from the DLT. This can happen for various reasons and means that you should not make decisions based on the DLT’s output. Red light can intend either a failure or that the machine is not running, meaning that the DLT is not receiving data to be used to make predictions.
Sensor Data Tab
In this tab graphs about sensors values are grouped according to their position on the machine. You can pick the graph you are interested in by clicking on the menu close to the title and by selecting the desired sensor.
- The first group shows graphs containing readings coming from temperature and current sensors. The time resolution of these graphs changes according to the amount of available data and the sampling frequency of the sensors.
- Pressure sensors are shown in the second group.
If you need to recall some information about the Predictive Maintenance section or the Sensor Data section, in the Help tab you can find the detailed descriptions of these two sections.