Improving the accuracy of the acclerometer based LSTM model in python.

I am performing time series classification using accelerometer data. I want to use a deep learning method for my project. I am classifying accelerometer data into various road conditions. I have a labeled dataset for various road conditions. I am presently using LSTM to perform the time series classification Working Code: I have a working LSTM classification code which runs perfectly on a different dataset. # The code directory name is --Description of the working project --Six Classes --Dataset used UCI HAR dataset --Result 70 percent Accuracy --- This was just to justify that the code works to check please The Problem: I am trying a different dataset with the same structure and format. The dataset and project is inside the rar file name [login to view URL] I am getting strange results like 31% accuracy and a confusion matrix in which only a single class is showing any [login to view URL] attachment name of the confusion matrix is [login to view URL] The training and testing dataset is balanced among various classes. Comparison Tests When I use the same dataset and ensemble machine learning method, I obtain 50% percent accuracy using an ensemble machine learning method. When I use the same dataset and calculate features and then use ensemble machine learning method, I obtain 78% percent accuracy with a better confusion matrix. Confusion matrix of Ensemble machine learning classifier is present in the attachment name [login to view URL] I want to understand what I am doing wrong to get strange results with LSTM. Things to try: 1. Trying hyperparameter optimization changing the parameters such as learning rate, number of hidden layers etc. 2. Come up with a feature to train deep learning neural network. The rar file having the Code having the problem is [login to view URL]