Segmented control strategy based on rheological critical threshold coupled with BP neural network for improving the uniformity of microwave vacuum drying of herbal extracts

High-viscosity, heat-sensitive herbal extracts experience uneven drying and localized overheating during microwave vacuum drying (MVD), which can degrade their active components. Constant-power models based on static porous media overlook the fluid-to-solid rheological transition, thereby limiting precise control. Using Scutellaria baicalensis extract as a model, we identified a viscosity threshold of 10,000 mPa·s for the transition from fluid to semi-fluid states. The degradation of baicalein at 300 W (6.18%) exceeded that at 1000 W (4.50%), indicating that the constant-power mode fails to mitigate both localized overheating and low-temperature thermal degradation. Consequently, we divided the drying process into fluid, semi-fluid, and solid stages and proposed a segmented power strategy (1000 W → 500 W → 300 W). A backpropagation neural network (BPNN) was employed to predict switching times based on moisture, viscosity, and density. The comprehensive score improved from 97.60 (RSM) to 97.98 (BPNN constant power) and 98.27 (staged control), while the relative standard deviation (RSD) decreased by 43.2% (from 0.387% to 0.22%). A physical fingerprint analysis identified pH, surface tension, and conductivity as key indicators. The BPNN moisture ratio prediction error (0.111%) was significantly lower than that of the Wang model (0.843%), confirming the strong nonlinearity of MVD and the necessity for dynamic modeling. This strategy effectively balances efficiency and active component retention, providing a viable engineering solution.