The invention discloses a stress-strain prediction method based on machine learning, and belongs to the technical field of detection and prediction. The prediction method comprises the steps of takingthe machine learning as a medium, processing the stress-strain experiment data as input and output of a learning model, selecting a proper algorithm and the corresponding training parameters, training, so that a prediction network is obtained. During the prediction process, the computer is operated to control the loading force of the loading device each time, a demodulator is used for collectingthe measurement data, a data analysis software is used for processing the experimental data, an appropriate learning model is established, the model is trained, and therefore accurate prediction of the strain field of the tested system is achieved. The method is suitable for the stress-strain field prediction of any optical fiber strain sensor detection system, avoids the errors caused by consideration of determination of loading force and pre-tightening force, simplification of the elastic modulus range of a test piece and a complex model and the like, greatly improves the calibration precision, and is convenient and rapid to operate and easy to popularize.