Machines operating in a production environment inevitably undergo changes throughout their usage life span, especially due to environmental influences, modifications or wear and tear. This introduces complexities in leveraging machine learning to monitor their condition for maintenance purposes. Depending on the significance of the changes, the performance of the condition monitoring system could degrade drastically. Especially, if changes occur frequently, retraining the system from scratch is not feasible. Therefore, we investigated active learning and self-training as adaptation strategies to enhance the robustness of the condition monitoring process. We evaluated both strategies using a demonstrator consisting of several electric motors and a single vibration sensor. The classification task is to identify, which motors are running based on their superimposed vibration data. We emulated different changes occurring in typical production environments and rated them as small, medium and large scale changes. For a comparative analysis of both strategies, we trained a benchmark model based on a convolutional neural network and evaluated the performance of active learning and self-training for the different changes in relation to this reference model. The results indicate that employing both active learning and self-training in a production environment to adapt the model can enhance its robustness. Self-training is the preferred option for small changes, as it adapts the model without the need for user interaction. For medium and large scale changes, on the other hand, self-training can fail, while active learning is a sensible strategy, despite the operator having to label some of the data.