Emotions play an important role in human psych science and Cognitive. It develops from everyday experiences from the people you communicate with and the environment in which they communicate. In this paper, we propose an efficient Cognitive signal analysis using the Conglomerate machine learning technique for human emotions classification. The main contributions of the proposed technique are, A scheduled optimal wavelet packet transform (SOWPT) used to obtain characteristic features (i.e., feature extraction), which provides both time scale and time-invariant property features. It is also used to decompose the ECG signals into the sample of sub-signals. The fish swarm-based feature selection (FSFS) algorithm is utilized to select optimal features among multiples, which reduces the number of features in the ECG signal. Taguchi-based discernibility matrix used to calculate an optimal best feature in EEG signal, which enhances classification correctness by selective manner. The extracted features from both ECG/EEG signals are classified in the final stage i.e., classification, we use numerical shape induced artificial neural network (NS-ANN) for signal categorization. Moreover, we develop a Grain Size Optimal Deep Neural Network (GS-DNN) for emotions classification with the help of classified Cognitive signals. To evaluate the performance of the proposed technique through different real-time data-sets using our proposed method. The results can compare with the existing state-of-art techniques in terms of classification accuracy, precession, F-measure, and Recall.