Reducing spinal loads using exoskeletons has become one of the optimal solution in reducing compression of the lumbar spine. Medical research has proved that the reduction compression of the lumbar spine is a key risk factor for musculoskeletal injuries. In this paper we present a deep learning based approach which is aimed at increasing the universality of lower back support for the exoskeletons with automatic control strategy. Our approach is aimed at solving the problem of recognizing human intentions in a lower-back exoskeleton using deep learning. To train and evaluate our approach deep learning model, we collected dataset using from wearable sensors, such as IMU. Our deep learning model is a Long short-term memory neural network which forecasts next values of 6 angles. The mean squared error and coefficient of determination are used for evaluation of the model. Using mean squared error and coefficient of determination we evaluated our model on dataset comprised of 700 samples and achieved performance of 0.3 and 0.99 for MSE and , respectively.