PCSC-net: a channel × frequency representation learning framework tailored to electrical impedance spectroscopy for robust food quality assessment

Electrical impedance spectroscopy (EIS) has shown considerable promise for rapid and non-destructive food quality assessment by providing multidimensional electrical responses. However, conventional one-dimensional modeling of EIS data tends to entangle channel-wise and frequency-wise information, making it difficult to capture discriminative component-specific frequency-response patterns and limiting model generalization. Moreover, EIS measurements are sensitive to environmental and instrumental perturbations, which further constrains their practical applicability. To address these issues, this study proposes PCSC-Net, a Channel × Frequency (C × F) representation learning framework tailored to multidimensional EIS-based food quality assessment. The proposed Per-Channel Spectral Convolution module explicitly models the C × F feature matrix and disentangles frequency-response patterns across different electrical components. The Spectral-to-Geometry Pseudo-Image Generator further converts latent EIS representations into geometry-aware pseudo-images, enabling the extraction of higher-order structural correlations through convolutional learning. In addition, the Adaptive Uncertainty-Injection Module introduces sample-adaptive stochastic perturbations to improve robustness against measurement uncertainty. The effectiveness and generalizability of PCSC-Net were evaluated on three dedicated EIS datasets: hydrocolloid-adulterated beef detection, milk category classification, and apple juice freshness identification. PCSC-Net consistently outperformed representative machine-learning and deep-learning baselines, achieving classification accuracies of 0.9750 ± 0.0137, 0.9699 ± 0.0278, and 0.9525 ± 0.0146 on the three datasets, respectively. Under stringent grouped cross-validation that held out entire brand units, PCSC-Net retained the best performance (0.8333 ± 0.1179 accuracy for brand-wise milk evaluation), confirming generalization to previously unseen sources. Moreover, PCSC-Net demonstrated robustness under limited training samples, additive noise, and storage-induced distribution shifts, while maintaining a compact footprint of approximately 0.53 MB. A representation-level ablation under matched encoder capacity showed the C × F representation exceeding its flattened 1D counterpart by 13.94% (accuracy) and 14.93% (Macro-F1); channel/frequency permutations degraded only the former, confirming that the native channel-frequency topology cannot be recovered by flattening. These findings indicate that PCSC-Net provides a domain-specific representation learning strategy for multidimensional EIS data and offers a compact, robust, and deployable solution for rapid, non-destructive, and intelligent food quality assessment.