Predictive services

Quick Check

Initial situation

With the help of data from continuous monitoring of machine tools, the company TRUMPF GmbH promises to

TRUMPF Co. KG (hereinafter TRUMPF) to detect error states early and thus prevent machine failures. TRUMPF plans to offer early error detection to its customers as a predictive service. The Quickcheck was intended to test the feasibility of a predictive service for laser cutting heads. The performance of these can be influenced, for example, by dust particles, which can be triggered by various events.

Solution idea

Methods from the field of machine learning are used to detect possible impairments and errors in the cutting process at an early stage. Various Trumpf data sources were used for this purpose, in particular log files and service reports, which are also available later directly on the machine.

Figure 1: Occurrence of different events over time (blue & green) until a fault occurs (orange) for four different machines. For each machine, both the frequency of the events (y-axis) and the time period (x-axis) as well as the temporal distribution of the events differed significantly. Own representation of the IAO — Figure 1: Occurrence of different events over time (blue & green) until the occurrence of a fault (orange) for four different machines. For each machine, both the frequency of the events (y-axis) and the time period (x-axis) as well as the temporal distribution of the events differed significantly. Own representation of the IAO.

Benefit

With the help of the Predictive Service, a failure of the laser cutting head of the machine tool can be detected at an early stage. With a reliable prediction, the service can
customers, e.g. as a monitoring solution. On the one hand, this prevents machine downtimes. It also reduces the number of service calls for TRUMPF technicians.

Implementation of the AI application

Trumpf supplied data from various sources, which was merged into a large data set. Using various parameters, an attempt was to be made to predict the remaining runtime until the error message using methods from the field of supervised learning. During the data preparation required for this, however, it turned out that the events that occurred before an error were very different (see Figure 1). For this reason, the feasibility of the predictive service for the use case could not be confirmed during the quick check with the available data.