Autonomous rehabilitation training for assisted patients with injured upper-limbs promotes the regenerative communication between muscle signals and brain consciousness. Surface electromyographic (sEMG) is a type of electrical signals of neuromuscular activity recorded by electrodes on the surface of the human body, which is widely applied for detecting gestures and stimuli reactions. Experimental results proved the importance of the sEMG signals for extracting such reactions, in which, the segmentation and classification of the sEMG are vital tasks. The objective of the present work is to segment and classify the sEMG signals of patients to assist the design of clinical rehabilitation devices based on the classification of sEMG signals. In the pre-processing stage, a dual-tone multi-frequency signaling is designed for signal coding; subsequently, the pre-processed sEMG signal is transformed by the Fast Fourier Transfer. Afterward, a time-series frequency analysis is performed by applying Hidden Markov Models. A basic traditional long short-term memory (LSTM) model is addressed for waveform-based classification to be compared to the proposed improved deep BP (back-propagation)–LSTM for sEMG signal classification. Seventeen performance features are selected for evaluating the proposed multi-classification, deep learning model for classifying six actions, namely moving gesture of grip, slowly moving, flexor, straight lift, stretch, and up-high lift; which were proposed by rehabilitation physician. The experiment results indicated the superiority of the proposed method compared to other well-known classifiers, such as the neural network, support vector machine, decision trees, Bayes inference, and recurrent neural network. The proposed deep BP–LSTM network achieved 92% accuracy, 89% specificity, 91% precision, and 96% F1-score, in the multi-classification of the sEMG signals.