A New Adaptive Mixture Distance-Based Improved Density Peaks Clustering for Gearbox Fault Diagnosis

Abstract

With the rapid development of sensors and mechanical systems, we produce an exponentially large amount of data daily. Usually, faults are prevalent in these sensory systems due to harsh operational conditions. Thus, detecting and diagnosing faults in the gearbox of mechanical systems are done by analyzing an exponentially large amount of data in the form of vibration signals and categorical features. However, the automatic fault detection method can match the increasing requirement for high-quality products in the course of intelligent manufacture. Thus, to acquire more distinguishable fault features under varied conditions, a new adaptive mixture distance-based simple and efficient density peaks clustering algorithm is proposed for handling mixed data as real-world datasets encompassing both numerical and categorical attributes. Our approach revolves around the concept of a sequence of the weighted exponential kernel using a symmetry-favored c-nearest neighbor to estimate the global parameter and the local density of each data point. Then, the initial clusters are extracted from a decision graph using an adaptive threshold parameter. The final step is to allocate the remaining data objects, if they are density reachable, to either of the initial groups. Thirteen UCI datasets and one real-world dataset from a mechanical system for gearbox defect diagnosis are employed to validate the proposed method. Five external and two internal evaluation criteria are considered to gauge how well the strategies are working. All of the findings indicate that the proposed method outperforms 13 other approaches.