Predictive Fermentation Control of Lactiplantibacillus plantarum Using Deep Learning Convolutional Neural Networks

Abstract

The fermentation of Lactiplantibacillus plantarum is a complex bioprocess due to the nonlinear and dynamic nature of microbial growth. Traditional monitoring methods often fail to provide early and actionable insights into fermentation outcomes. This study proposes a deep learning-based predictive system using convolutional neural networks (CNNs) to classify fermentation trajectories and anticipate final cell counts based on the first 24 h of process data. A total of 52 fermentation runs were conducted, during which real-time parameters, including pH, temperature, and dissolved oxygen, were continuously recorded and transformed into time-series feature vectors. After rigorous preprocessing and feature selection, the CNN was trained to classify fermentation outcomes into three categories: successful, semi-successful, and failed batches. The model achieved a classification accuracy of 97.87%, outperforming benchmark models such as LSTM and XGBoost. Validation experiments demonstrated the model's practical utility: early predictions enabled timely manual interventions that effectively prevented batch failures or improved suboptimal fermentations. These findings suggest that deep learning provides a robust and scalable framework for real-time fermentation control, with significant implications for enhancing efficiency and reducing costs in industrial probiotic production.

Keywords: Lactiplantibacillus plantarum; bioprocess control; convolutional neural network; deep learning; fermentation prediction; probiotics.

Figure 1

Workflow of the probiotic fermentation monitoring and deep learning prediction system. The process begins with fermentation experiments and data collection, followed by preprocessing and feature vector generation. A CNN-based framework is then trained to provide predictive outputs, enabling real-time monitoring and optimization.

Figure 2

Preprocessing pipeline of fermentation time-series data. (a) Raw signal with noise and artifacts. (b) Outlier detection, with spikes flagged (red ×). (c) Linear interpolation of missing values, with arrows marking corrected points. (d) Truncation to the first 24 h of data. (e) Normalization of truncated signals using min–max scaling to the [0–1] range.

Figure 3

Representative early-stage pH trajectories across fermentation runs: (a) pH0, (b) pH1–pH2, (c) pH3–pH4.

Figure 4

CNN architecture for early prediction of fermentation outcomes, consisting of two convolutional layers (kernel size = 3; filters = 32 and 64; ReLU), each followed by MaxPool1D (pool size = 2), a flatten operation, two dense layers (each with 64 units, ReLU), and a final softmax classifier for three classes. The model ingests early-stage features of dimension 3 × 288.

Figure 5

Fermentation dynamics of L. plantarum. (a) Changes in viable cell counts over time (24, 48, and 72 h). (b) Effect of incubation temperature (30, 32, and 37 °C) on final cell counts. Error bars indicate the mean ± standard deviation.

Figure 6

Classification accuracy comparison among CNN, LSTM, XGBoost, and LightGBM on the internal test set using the first 24 h inputs. The CNN achieved 97.87%, outperforming all baselines.

Figure 7

ROC curves for the three outcome classes, failure (class 0, orange solid line), semi-success (class 1, blue solid line), and success (class 2, green solid line), based on CNN predictions on the test set. The macro-average ROC (yellow dashed line) and micro-average ROC (blue dotted line) yielded AUC values of 0.98 and 0.99, respectively, while the grey dashed line indicates random-chance performance (AUC = 0.50).

Figure 8

Confusion matrix of CNN classification results on the internal evaluation set (n = 47). Cell values are counts, while the color scale indicates row-normalized proportions (0–1). Classes correspond to fermentation outcomes: class 0 = failure, class 1 = semi-success (semi-successful), and class 2 = success.

Figure 9

Representative temporal motifs extracted by the CNN from the first 24 h of fermentation (288 time points). The four curves (Motifs A–D) illustrate normalized activation patterns (scaled to 0–3 a.u.) that the network used for early discrimination of fermentation outcomes. These temporal features highlight recurring oscillatory and trending signals within the early phase, supporting the model’s ability to achieve accurate predictions prior to the completion of the fermentation process.

Figure 10

Impact of intervention based on model prediction across operator groups. Final L. plantarum CFU/mL in control versus intervention runs for novice and expert operators. Axis uses scientific notation (×10⁹). Values correspond to group means reported in 3.8. (novice: 0.56 → 1.75; expert: 1.31 → 3.39 × 10⁹ CFU/mL).