An Adaptive Fuzzy Recurrent Stochastic Configuration Network for Food Quality Prediction with Limited Samples

Accurate food quality prediction remains challenging when labeled samples are limited, and the relationships between measurable characteristics and target quality attributes are nonlinear and heterogeneous. To address this problem, this study proposes an adaptive fuzzy recurrent stochastic configuration network that integrates adaptive fuzzy rule-number selection, rule-coupled recurrent subreservoirs, global-residual-guided stochastic configuration, and regularized analytical output-weight learning. The proposed framework was evaluated on the Mackey–Glass benchmark and two practical food quality prediction tasks involving litchi soluble sugar content and wine quality. On the Mackey–Glass benchmark, it achieved a root mean square error of $4.47\times {10}^{-4}$, representing a $37.8\%$ reduction relative to the conventional fuzzy recurrent stochastic configuration network. For litchi sugar prediction, it achieved a root mean square error of 0.7200, a mean absolute error of 0.5660, a mean absolute percentage error of 2.632%, and a coefficient of determination of 0.9352, with the root mean square error reduced by approximately $10.2\%$ relative to the conventional fuzzy recurrent stochastic configuration network. For white and red wine, the corresponding root mean square errors were 0.1829 and 0.1655, with coefficients of determination of 0.9513 and 0.9677, respectively. These results indicate that the proposed framework provides favorable predictive performance under the evaluated limited-sample settings and predefined data partitions, supporting its applicability to nonlinear food quality prediction with limited labeled data.