A design concept for high-performance components involves the combination of different materials in hybrid workpieces. Different material properties and chemical compositions influence the machining quality of hybrid workpieces. To achieve a constant workpiece and process quality, it is necessary to adjust the process parameters to the individual material. Thus, it is mandatory to classify the material during machining for the relevant range of process parameters. This paper examines teaching strategies for neural networks to determine the machined material in process by a small amount of cross points. For this purpose, different training sets are compared. Process parameters with different cutting speeds, feeds and with constant and varying depth of cut are examined. In addition, the signal sources necessary for robust material classification are compared and investigated. The investigation is performed for the cylindrical turning of friction welded EN AW-6082/20MnCr5 shafts. The study shows that an F1 score of 0.99 is achieved at a constant cutting depth, provided that only the corner points of the process window and the machine control signals are used for training. With an additional variation of the cutting depth, the classification rate is significantly improved by the use of external sensors such as the acceleration sensor.