A CNN-RBPNN Model with Feature Knowledge Embedding and its Application to Time-Varying Signal Classification

Abstract

A novel technique, combining the feature extraction mechanisms of a convolutional neural network (CNN) with the classification method of a radial basis probability neural network (RBPNN), is proposed for small sample set modeling and feature knowledge embedding in multi-channel time-varying signal classification. This CNN-RBPNN consists of a signal input layer, signal feature parallel extraction and integration units, and an RBPNN classifier. Each channel signal in a feature extraction unit corresponds to a 1D CNN. The extracted features are represented as feature vectors, and these vectors constitute a comprehensive feature matrix. The RBPNN classifier was designed using signal feature embedding mechanism based on radial basis kernels and the property of combining pattern subclasses into pattern classes to form complex class boundaries. A dynamic clustering algorithm was used to divide each pattern class sample into several subclasses. Typical signal samples in each pattern subclass were designated as kernel centers, in order to achieve signal categories features embedding. This process was also used to determine the number of nodes in the RBPN layer. The RBPN layer outputs were selectively summed in the pattern layer according to kernel center category, which can generate irregular class boundaries, reducing the overlap among different pattern class boundary. The proposed CNN-RBPNN replaces the full-connection layer and classifier unit of conventional CNN with RBPNN, which can extract and represent signals distribution features and structural properties, implement structural and data constraints. This can reduce the structural risks of small sample set modeling. In this study, the properties of CNN-RBPNN are analyzed and an integrated learning algorithm is proposed. An experiment was conducted using 12-lead ECG signals in a seven-classification in the case of small sample set. Results demonstrated that, the correct recognition rate is 5.7% higher than other methods in the experiment, the performance evaluation index also showed significant improvement.