(191b) Enhancing Data Science in Agrichemical Manufacturing: Challenges and Solutions

In the modern agricultural chemical (Ag-Chem) manufacturing landscape, the integration of sensors and data logs into daily operations has ushered in an era of abundant data. This wealth of data presents a fertile ground for data science applications, with the potential to bring about substantial improvements in productivity, operational efficiency, and the resolution of unexplainable plant behaviors. Data science techniques, including machine learning predictions, anomaly detection, and root cause analysis, offer significant promise, provided one can effectively navigate the challenges posed by the unique nature of manufacturing data.

Manufacturing data in Ag-Chem poses multifaceted challenges that demand rigorous pre-processing and curation before it can be effectively leveraged for data science applications. The primary context behind these challenges lies in the fact that this data is not obtained from controlled experiments, but instead, is a by-product of plant operations which are primarily geared toward achieving a consistent routine. This results in limited availability and variability of data in features and target variables, which are essential components of well-structured data science problems. Consequently, data scientists confront challenges on two fronts. On one hand, there are specific data preparation challenges, including addressing data noise caused by inaccurate measurements, harmonizing data points distributed across multiple vessels and timestamps, and integrating features from multiple systems with unsynchronized time resolutions. On the other hand, broader challenges exist related to data availability and usefulness. These encompass the scarcity of data variability and the lack of contextual information about background changes within the plant, which can lead to a high quantity but low-quality dataset for deriving insights. Additionally, data can lack specific features due to the absence of corresponding sensors or routine sampling procedures, or the influential features are not controllable by the operators giving rise to the risk that the insights from data science studies may not yield actionable solutions.

This presentation aims to shed light on each of these challenges, highlighting the resulting disadvantages they impose on data science efforts within manufacturing. Moreover, it offers practical solutions and insights whenever possible, such as procedures for pre-processing data to circumvent the unintended misrepresentations that may lead to underperforming models or inaccurate insights. Real-world case studies from projects undertaken by the Advanced Analytics team in the Crop Protection Manufacturing division at Corteva Agriscience are presented to illustrate the impact of proper data collection and curation, highlighting the transformation of outcomes. The information highlighted in these studies helps in gaining a deeper understanding of the intricate challenges inherent in Ag-Chem manufacturing data and practical techniques to overcome these hurdles. The use of real-world case studies ensures that the insights and solutions are grounded in tangible experiences, making this presentation a valuable resource for data scientists, engineers, and professionals involved in manufacturing and data analysis.