Nguyen Xuan Dung James, Liu Y A, McDowell Christopher C, Dooley Luke
Aspen Tech Center of Excellence in Process System Engineering, Department of Chemical Engineering, Virginia Polytechnic Institute and State University, Blacksburg, VA, 24061, USA.
Novonesis Biological, Inc., 5400 Corporate Circle, Salem, VA, 24153, USA.
Bioprocess Biosyst Eng. 2025 Jun 26. doi: 10.1007/s00449-025-03194-6.
This paper demonstrates an accurate and efficient methodology for fermentation contamination detection and reduction using two machine learning (ML) methods, including one-class support vector machine and autoencoders. We also optimize as many hyperparameters as possible prior to the training of the ML models to improve the model accuracy and efficiency, and choose a Python platform called Optuna, to enable the parallel execution of hyperparameter optimization (HPO). We recommend using Bayesian optimization with hyperband algorithm to carry out HPO. Results show that we can predict contaminated fermentation batches with recall up to 1.0 without sacrificing the precision and specificity of non-contaminated batches, which read up to 0.96 and 0.99, respectively. One-class support vector machine outperforms autoencoders in terms of precision and specificity even though they both achieve an outstanding recall of 1.0. These models demonstrate high accuracy in detecting contamination without requiring labeled contaminated data and are suitable for integration into real-time fermentation monitoring systems with minimal latency and retraining needs. In addition, we benchmark our ML methods against a traditional threshold-based contamination detection approach (mean 3 rule) to quantify the added value of using data-driven models. Finally, we identify important independent variables contributing to the contaminated batches and give recommendations on how to regulate them to reduce the likelihood of contamination.
本文展示了一种准确且高效的方法，用于使用两种机器学习（ML）方法（包括单类支持向量机和自动编码器）进行发酵污染检测与减少。我们还在ML模型训练之前尽可能多地优化超参数，以提高模型的准确性和效率，并选择了一个名为Optuna的Python平台，以实现超参数优化（HPO）的并行执行。我们建议使用带有超带算法的贝叶斯优化来进行HPO。结果表明，我们可以预测受污染的发酵批次，召回率高达1.0，同时不牺牲未受污染批次的精度和特异性，未受污染批次的精度和特异性分别高达0.96和0.99。单类支持向量机在精度和特异性方面优于自动编码器，尽管它们的召回率均达到了出色的1.0。这些模型在检测污染时具有很高的准确性，无需标记的受污染数据，并且适合以最小的延迟和再训练需求集成到实时发酵监测系统中。此外，我们将我们的ML方法与传统的基于阈值的污染检测方法（均值3规则）进行基准测试，以量化使用数据驱动模型的附加值。最后，我们确定了导致受污染批次的重要自变量，并就如何调节它们以降低污染可能性给出了建议。
