If tools exceed their usage limit, undesirable changes occur on the workpiece and in the cutting process. In order to reduce scrap and rework, the service life of tools is therefore frequently tracked in practice and a tool change is carried out at an early stage. However, the service life of tools varies, sometimes by more than 100%. If the same time for tool change is specified for all tools, many tools are replaced, although they could continue to produce good parts reliably. The free-space wear can be used to estimate more precisely when the usage limit of tools has been reached. Tools are thus used longer on average, which reduces setup times and tool costs.
Process-parallel monitoring of tool wear thus makes it possible to use each individual tool up to its usage limit. However, up to now, such monitoring has had to be taught on the monitored machine for the monitored process. This usually requires many reference processes. As part of the IIP-Ecosphere project, IFW is researching the monitoring of processes with data acquired on other machines. If machines execute the same processes, the approaches transfer models for monitoring between the machines. This eliminates the learning phase at the monitored machine. This enables economical monitoring even for smaller batch sizes, including single-part production.
The approaches are tested on three machine tools for monitoring the wear mark width during milling (Fig. 1). Two machines (A and B) provide the data for training the monitoring models and thus simulate, for example, series production or product tests at the tool manufacturer. On the third machine, the machine to be monitored, the model is then used to determine the flank wear. This makes it possible to determine the flank wear of the tool already during the first cutting process on the monitoring machine.
The extent to which machine learning methods offer added value for monitoring is also being examined. For example, deep neural networks are used for wear monitoring. These independently identify wear indicators from the data of the machines, which are meaningful across machines. In addition, federated learning is used, whereby the data remains at the individual machines and accordingly does not need to be shared. Instead, knowledge snippets are shared to improve the common model, which emerge from the data of the individual machines.
