Stress detection has a huge potential for disease prevention and management, and to improve the quality of life of people. Also, work safety can be improved if stress is timely and reliably detected. The availability of low-cost consumer wearable devices that monitor vital-signs, gives access to stress detection schemes. Heart rate variability (HRV), a stress-related vital-sign, was derived from wearable device data to reliably determine stress-levels. In order to build and train a deployable stress-detector, we collected labeled HRV data in controlled environments, where subjects were exposed to physical, psychological and combined stress. We then applied machine learning to separate and identify the different stress types and understand the relationship with HRV data. The resulting C5 decision tree model is capable of identifying the stress type with 88% accuracy, in a 1-minute time window. For the first time physical and psychological stress can be distinguished with a 1-minute time resolution from smoke-divers, firefighters, who enter high-risk environments to rescue people, and experience intense physical and psychological stress. To improve our model, we created an integrated system to acquire expert labels in real-time from firefighters during their training in a Rescue Maze. A next goal is to transfer the algorithms into generic systems for monitoring and coaching high-risk professionals to improve their stress resilience during training and reduce their risk in the field.