Smart factories are composed of heterogeneous cyber-physical systems. In light of their complexity and the lack of transparency in their design, monitoring the health of these machines in real-time is made possible by the use of non-intrusive sensors. Such sensors produce mixed signals capturing component-specific signatures. Retrieving the activation statuses of the components (over the different operating modes of a machine) is essential for estimating their associated performance indicators. This is a special case of Underdetermined Blind Source Separation (UBSS), yet a sensor fusion perspective is adopted in this paper. A harmonic component detector produces observations in the Time-Frequency (TF) domain, inherently entailing noise-induced false alarms. The main contribution of this paper consists in a clutter-resilient multi-harmonic component tracking algorithm, based on the Sequential Monte-Carlo Probability Hypothesis Density (SMC-PHD) filter. Additionally, this paper presents a track association algorithm adapting the results obtained in the multi-target tracking framework for unsupervised multi-label classification. The combination of the two algorithms mitigates typical difficulties encountered in traditional UBSS problems, such as non-stationary and partially-coupled mode decomposition. The performance of the proposed technique is assessed upon synthetic data.