This paper proposes a three-phase metaheuristic-based approach for induction machine bearing failure detection and diagnosis. It consists of extracting and processing different failure types features to set up a knowledge base, which contains different failure types. The first phase consists in pre-processing the measured signals by aggregating them and preparing the data in exploitable formats for the clustering. The second phase ensures the induction machine operating mode diagnosis. A measured signals clustering is performed to build classes where each one represents a health state. A variable neighborhood search (VNS) metaheuristic is designed for data clustering. Moreover, VNS is hybridized with a classical mechanics-inspired optimization (CMO) metaheuristic to balance global exploration and local exploitation during the evolutionary process. The third phase consists of two-step failure detection, setting up a knowledge base containing different failure types, and defining a representative model for each failure type. In the learning step, different class features are extracted and inserted in the knowledge base to be used during the decision step. The proposed metaheuristic-based failure detection diagnosis approach is evaluated using PRONOSTIA and CWR bearing data experimental platforms vibration and temperature measurements. The achieved results clearly demonstrate the failure detection and diagnosis, efficiency, and effectiveness of the proposed metaheuristic approach.