This paper presents a new approach to the on-line tracking of pulverized coal and biomass fuels through flame spectrum analysis. A flame detector containing four photodiodes is used to derive multiple signals covering a wide spectrum of the flame from visible, near-infrared and mid-infrared spectral bands as well as a part of far-infrared band. Different features are extracted in time and frequency domains to identify the dynamic "fingerprints" of the flame. Fuzzy logic inference techniques are employed to combine typical features together and infer the type of fuel being burnt. Four types of pulverized coal and five types of biomass are burnt on a laboratory-scale combustion test rig. Results obtained demonstrate that this approach is capable of tracking the type of fuel under steady combustion conditions.