Acoustic emission monitoring of blending in continuous direct compression

Abstract

Continuous direct compression (CDC) is increasingly used in pharmaceutical manufacturing for the continuous production of oral solid dosage forms. Within a CDC line, continuous blending is an important processing step because the amount of material retained in the blender affects material transport and residence behaviour. This thesis investigates whether acoustic measurements recorded during continuous blending contain information that can be related to residence mass. Acoustic signals covering different frequency ranges were recorded during six blender experiments performed at different rotational speeds and throughputs. Time-frequency analysis, singular value decomposition, and multivariate modelling were used to characterise the signals and assess the influence of operating conditions. Partial least squares discriminant analysis achieved 97.8% classification accuracy using frequency band features and 100% using a higher-resolution spectral representation. Multioutput partial least squares regression further showed that both rotational speed and throughput were strongly represented in the acoustic measurements. Residence mass was modelled using gravimetric measurements from each experimental run and evaluated using leave-one-run-out validation. High-dimensional spectral models produced strong training fits but poor prediction of omitted runs. Reducing the acoustic feature set and one PLS latent substantially improved cross-run generalisation. The best acoustic-only model achieved a training R2 of 0.940, a LORO R2 of 0.854, and a LORO RMSE of 266 g. Engineered process variables showed strong individual correlations with residence mass but did not improve LORO performance. The modelling procedure was also applied to separately acquired recordings using a different high-frequency measurement setup. Resampling reduced a systematic prediction offset, while a reference gravimetric measurement was required for bias correction. Overall, the results show that acoustic measurements contain information associated with both blender operating conditions and residence mass, while also highlighting the importance of model complexity, run-level validation, block selection, and measurement consistency when applying the approach to continuous pharmaceutical blending.