Classification of time-frequency images based on locality-constrained linear coding optimization model for rotating machinery fault diagnosis

Time-frequency images have been widely employed in vibration signal analysis in the field of rotating machinery fault diagnosis. Generally, global features such as gray statistic features and texture features, are extracted from time-frequency images for fault classification, but the effect is not satisfactory. The locality-constrained linear coding model based on local features has been successfully employed in image classification. This paper contributes to this ongoing investigation by developing a locality-constrained linear coding optimization model applied in time-frequency images classification for rotating machinery fault diagnosis. The classification accuracy and generalization of the locality-constrained linear coding model depend on the selection of pooling methods, values of pooling parameters, and penalization coefficient. In the optimization model, misclassification rate is chosen as an objective function, and an improved particle swarm optimization algorithm is adopted to optimize the pooling parameters and penalization coefficient. The improved particle swarm optimization algorithm is utilized to produce optimal solutions at the training stage, and then these solutions will be evaluated at the testing stage. The promise of the novel model is illustrated by performing our procedure on vibration signals from a rolling bearing with 16 health states. Experimental results have demonstrated that the proposed approach could obviously increase the classification accuracy.