Since many years smartphones are utilised for human activity recognition (HAR), important healthcare recommendations and telemedicine. Deep learning (DL) and machine learning techniques are commonly employed in studies of statistical models of human behaviour. However, the performance of present HAR platforms is constrained by complex physical activities. In this study, we developed the Ada-HAR human activity identification and real-time monitoring system, which is able to recognise more human motions in erratic situations. The chosen hierarchical clustering and classification methods are able automatically identify and categorise 12 activities (five dynamics, six statics, and a series of transitions). Finally, actual tests were run to verify the effectiveness and reliability of the suggested methods. The results demonstrate that the DL-based classifier achieves a higher identification rate (95.15% for waist and 92.20% for pocket) in comparison to the techniques discussed in the literature. Finally, the Ada-HAR system can track human behaviour in real-time regardless of how the smartphone is pointed. Here sensor activity dataset is considered which consists of user’s activity log. By applying algorithms named LSTM, Adamax, Adagrad, SGD, indentifiesExi user’s activity based on the movement of the mobile. Existing Algorithms named LSTM, K-NN, DT, ANN, SVM, NB applied on different dataset produced better result and accuracy. RNN is an updated version of LSTM. RNN stores only the current activities and fails to store previous conclusions. Fortunately, LSTM stores existing and completed work in any dynamic situations. The LSTM works with random weights which led to local optimum and high time complexity. This paper has addressed the above issue by the usage of optimization algorithms in weight updation of LSTM. This increase the accuracy of the model. Here Adamax, Adagram, Stochastic gradient descent are used for weight updation.