This paper aims to evaluate the performance of multiple non-linear classification techniques. Time-series classification has become central to many sensor-driven applications, where capturing temporal patterns efficiently and accurately is critical. In this paper, we evaluate and extend the proposed MiniRocket framework for classifying multivariate sensor data, focusing on a practical “Fish Freshness” monitoring use case. Precise quantification of post-harvest fish freshness underpins food-safety compliance and public-health protection while curbing spoilage-induced economic losses across the cold-chain. In operational practice, high-fidelity freshness classification enables dynamic inventory routing, just-in-time processing, and trustworthy quality labelling, thereby reducing food waste and reinforcing consumer confidence throughout seafood supply networks. MiniRocket is a highly efficient convolution-based feature extractor that transforms each raw time series into a fixed-length vector of pattern-frequency features (Proportion of Positive Values). We demonstrate that, when paired with a simple linear classifier such as Ridge Classifier, MiniRocket achieves state-of-the-art accuracy while requiring orders of magnitude less computation than deep-learning alternatives. We have demonstrated that combining MiniRocket with Random Forest Classifier yields greater than or equal to 90% classification accuracy, particularly improving performance on ambiguous classes like “Semi-Fresh”. The model’s simplicity, scalability, and efficiency make it suitable for real-time applications. Our findings validate MiniRocket as a practical solution for robust, low-latency time-series classification in resource-constrained environments.