Fuzzy C-Means (FCM) is a clustering algorithm based on partition of the universe. However, the partition generated by an equivalence relation is strict in practical application and exhibits relatively poor fault-tolerant mechanism. In this paper, a novel binary relation based on improved FCM with the principle of refined justifiable granularity is presented. Different expressions of the proposed binary relation under different values of weight parameter are discussed, and the changes of the properties of the binary relation under different parameter values are provided. By measuring the significance of attributes in the feature space, a feature selection method, called forward heuristic feature selection (FHFS), is designed to construct the low-dimension feature space based on maximizing the original data and information retention through the defined degrees of aggregation and dispersion. It is shown how the results of feature selection and classification performance vary when the values of the weight factor locate in different ranges. To illustrate the superiority and effectiveness of the proposed FHFS algorithm, nine high-dimensional datasets and eight image datasets from UCI repository are used and compared with other feature selection methods, respectively. The results of experimental evaluation and the significance test show that the proposed learning mechanism is a superior algorithm.
1 January 2023 in IEEE Transactions on Instrumentation and Measurement
Gas sensor drift, which has the characteristics of randomness and nonlinearity, is an inevitable problem in electronic nose (E-nose) systems. In this study, a domain-adaptive deep neural network (DNN) framework, called M 2 FL-CCC, is proposed to suppress sensor drift and improve E-nose performance. This framework mainly contains two key parts: multibranch multilayer feature learning (M 2 FL) and a comprehensive classification criterion (CCC). In terms of the network structure, a multibranch multilayer structure is designed for customized and joint sensor feature extraction. To fuse the full features of different levels in the network, a joint training strategy is leveraged for multilayer classifiers. Regarding the classification strategy design, a CCC is proposed to fuse the prediction results of the base classifiers and the separation degree between a specific target sample and the source samples. In addition, to optimize the training process, we adopt an improved additional margin softmax classifier with nonlinear dynamic parameter adjustment. Experiments are conducted on public E-nose drift data, and the results show that the M 2 FL-CCC framework is superior to other compared methods.