A total volatile basic nitrogen detection method for yellow croaker based on an electronic nose combined with an...

As a rapid detection method, the electronic nose exhibits enormous potential in food quality monitoring. However, the response data generated by electronic nose detection are highly complex time-series data. Traditional data analysis models struggle to fully resolve such long-sequence non-linear signals, leading to insufficient feature extraction and poor prediction accuracy.

We propose a novel total volatile basic nitrogen (TVB-N) prediction method based on the attention improved encoder long short-term memory (AIE-LSTM) hybrid network with a dual-stream feature fusion architecture. A modified informer encoder captures temporal dependencies via multi-head attention, and a temporal down-sampling layer compresses sequence dimensions at the same time as preserving key trend features. In addition, a bidirectional LSTM network processes raw voltage sequences and manually designed physical features separately. Finally, a multi-level feature fusion mechanism integrates these two types of features through fully connected layers to output predictions. Experimental results demonstrated that the AIE-LSTM model achieved the optimal TVB-N prediction performance across nine batches of electronic nose datasets, with an average coefficient of determination (R²) of 0.960, as well as the lowest relative standard deviation of R², root mean square error (RMSE) and mean absolute error (MAE). Notably, the model exhibited the best performance in the fourth batch, where the R², RMSE and MAE between the predicted and actual TVB-N values reached 0.979, 0.988 and 0.589, respectively.

This model significantly improves the prediction accuracy of fish TVB-N content, providing a reliable technical solution for the rapid, non-destructive and accurate detection of fish freshness. © 2026 Society of Chemical Industry.