Generalization of LSTM CNN ensemble profiling method with time-series data normalization and regularization
Disni Rathnayake, Pasindu Bawantha Perera, Heshan Eranga, Manjusri Ishwara · 2021
This study concentrated on generalizing an anomaly detection method of time series data using ensemble LSTM CNN network with time series data normalization and regularization. Considering the relevant conditions must meet for time series normalization, an algorithm was proposed for time series normalization. Checking the stationarity and normality of time series data is fundamentally included in the proposed algorithm. Afterward, different types of time series data are visualized with different normalization methods, and the impact of each of these methods is discussed. Normalization techniques like Min - Max, Sigmoid and Tanh are used. When sigmoid normalization used with a dataset where the original data is almost in the range of zero to approximately one was able to normalize well. For the data which is not in the range of [0],[1] this method cannot be used since it tends to overlap data that is not available in original data. The study reveals that a smooth, non-linear sigmoid function performs a better transformation for many anomalous time series data as a normalization factor. The prediction errors of the LSTMCNNkeras (LSTM CNN ensemble neural network was implemented using Keras is called LSTMCNNkeras.) model are discussed and compared with and without proposed normalization approach. Also, the gross effect of both normalization and regularization steps to the prediction errors of the LSTMCNNkeras model is discussed. For the time series data where the range lies somewhere between 0 to 1, produced better predictions in LSTMCNNkeras model along with Sigmoid normalization and Dropout regularization techniques. The gross effect of Layer weight regularization with Tanh normalization was able to produce a foreseeable growth in the performance of LSTMCNNkeras model with more accurate predictions for many time series data.