Human activity recognition techniques have achieved significant advancements in recent years. However, the performance of the generalization model may be hampered by the methods' heavy reliance on human feature extraction. Deep learning methods are becoming more and more effective, which has led to a lot of interest in employing these approaches to understand human behaviors in mobile and wearable computing settings. In place of the conventional hyperbolic tangent (tanh) activation function for human activity recognition, which can be applied in a variety of applications, in this study, the main part of LSTM neural networks is developed by employing 26 state functions to suggest Deep Learning Long Short-Term Memory (DLLSTM) classifiers. In LSTM network units, the sigmoid and tanh functions are often used as activation functions. The vanishing gradient issue that RNNs encounter can be effectively solved by LSTM networks. The effectiveness of the suggested DLLSTM classifiers for classification tasks was investigated using three different deep learning optimization techniques. The simulation results show that the suggested classifiers, which utilize the Modified -Elliott, Gaussian, and wave as DLLSTM classifiers, outperform the tanh classifier by getting a perfect accuracy rate of 99.92%, 99.5%, and 99.95% as opposed to their 96.4%, respectively.