Junghans, SebastianSebastianJunghansFlehmke, MalteMalteFlehmkeStützle, JoshuaJoshuaStützleMöller, CarstenCarstenMöllerDege, Jan HendrikJan HendrikDege2023-12-152023-12-152023-12-11Research Paper Series, SSRN (2023-11-29)https://hdl.handle.net/11420/44643In manufacturing of structural components for the aerospace industry, the application of milling tools with fixed service life time is a common practice. A tool replacement is often based on predetermined schedules rather than the actual wear. Consequently, tools may be replaced too soon or too late with respect to a defined final state of wear. To address this issue, different evaluation methods based on sensor data from the machining process have been proposed in the literature as a means of predicting tool wear and automating tool change decisions. However, decision models often do not take into account the variability of the data due to a change in the cutting parameters and tool load, thereby limiting their applicability to specific situations. In this paper, a convolutional neural network (CNN) is trained based on the pretrained GoogLeNet architecture to detect different states of tool wear in circumferential milling of Ti-6Al-4V with varying cutting parameters. Experiments have been carried out to capture sensor data, such as structure-borne noise, machine drive data, and cutting forces, in comparable states of tool wear under varying cutting parameters, for the purpose of model training and assessment. Different types of sensor data and time-series-to-image encoding techniques have been investigated in terms of model performance. The conducted experiments successfully applied gramian angular fields (GAF), markov transition fields (MTF), recurrence plots (RP), wavelet transform (WT) and short time fourier transform (STFT) to different signal regions (entry, full cut and exit). Using a trained CNN for classification, accurate states of tool wear can be predicted based on the created images. WT performed best in the investigations. Furthermore, an investigation was conducted to determine the level of deviation in actual tool wear on a sample of 10 tools that were removed from an active production line after being used for a fixed duration of 60 minutes. The level of deviation on tool wear across the 10 tools was found to be significantly high, indicating that the tools were not wearing out uniformly and that there may be a need for better tool monitoring and maintenance practices. This ML-based approach enhanced the tool replacement decisions by predicting similar states closer to the actual wear of the tools. Thus, the use of a CNN is found to be a helpful method to predict tool wear even under variable process conditions. They are able to provide more accurate and timely tool replacement decisions, reducing the risk of premature or delayed replacement, and improving overall manufacturing efficiency.enMLE@TUHHEngineering and Applied OperationsConference Paper10.2139/ssrn.4657889Conference Paper