Diagnosing and controlling the level of stress in order to reduce the risks is so important. In this study, a system for detecting five levels of stress, i.e. physical stress, semi-emotional stress, emotional stress, cognitive stress, and resting state in people based on physiological signals, is presented. In the proposed method, the Non-EEG Dataset for Assessment of Neurological Status database, which is available on the Physionet website, is used. This database contains physiological signals from twenty people. These data were collected using non-invasive wrist biosensors. A set of statistical and frequency and wavelet features are calculated for electrodermal (EDA), temperature, acceleration, heart rate (HR) and arterial oxygen level (SpO2) signals. The determined features are applied as input to the classification units to detect the stress levels. Support vector machine (SVM), k nearest neighbor (kNN), decision tree (DT), ensemble learning and neural networks are evaluated as classification methods. Experimental results show that neural networks can separate different neural states of 5 classes with 97% accuracy.