Reinforced Two-Stream Fuzzy Neural Networks Architecture Realized With the Aid of One-Dimensional/Two-Dimensional Data Features

A novel structure of reinforced two-stream fuzzy neural networks (TSFNNs) realized with the aid of fuzzy logic and transfer learning method is presented. This architecture consists of a TSFNN and a fusion strategy. TSFNN architecture consists of two combined networks of both fuzzy rules-based radial basis function neural networks (FRBFNN) and convolutional neural networks (CNNs). In the TSFNN architecture, one stream employs the deep CNN to extract the spatial information of images and effectively learn the high-level features and another stream uses the FRBFNN to analyze the distribution of data points over the input space and learn to capture complex relationships in data. In the fusion strategy, the outputs of two streams are concatenated by a softmax function, which normalizes the output to a probability distribution. A transfer learning method is considered to reconstruct new data representation as the inputs of CNN to mine potential spatial features of data. Moreover, L-2-norm regularization is used to alleviate the possible overfitting and enhance the generalization ability. The proposed method not only inherits the advantages of FRBFNN and CNN such as global feature extraction ability, good local approximating performance, ability of handling uncertainty by fuzzy logic but also improves the classification performance under the synergy between two-stream architecture and the fusion strategy. Experimental results obtained for a diversity of datasets as well as partial discharge datasets be using in the real life of fault diagnosis and black plastic wastes datasets for recycling confirm the effectiveness of the proposed TSFNN. A comprehensive comparative analysis is covered. This design can simultaneously capture different level information of inputs and easing the insufficient problem of extracting features from a single steam. Especially, we show that the synergistic effect of FRBFNN, CNN, enabling deep learning for generic classification tasks and multipoint crossover, and L-2-norm regularization can effectively improve the performance of the TSFNNs.