In a penicillin fermentation process, substrate concentration and biomass concentration greatly influence the yield of the targeted product. However, there are few on-line sensors available to measure these variables in real-time. In this paper, a compact mechanism model is employed to simulate the fed-batch process, and a particle filter is introduced to estimate the substrate and biomass states. Particle filters are favorable to handle the state estimation problems with non-linearity, time-varying dynamics, and non-Gaussian distributions. In order to improve the quality of particles, optimization strategies are applied to deal with constraint issues. Furthermore, infrequent lab analyzed state information is incorporated into the estimation procedure and used to correct PF estimate. Simulation results show that the constrained PF approach has better estimation performance than extended Kalman filter in state estimation of this penicillin fermentation batch process.