Multivariate time series (many to one relationship)

Question

I am working on a machine learning project for my summer internship and was hoping someone could help. I am new to the field of ML and am still learning so please bear with my attempt for an explanation.

I am given input (x value/features) continous numerical data for multiple bioreactors (20), which include PH, agitation rate, glucose level, etc. Data from each bioreactor is in intervals of 30 minutes and collected for 11 days. This yields about 500-600 rows of data per bioreactor. Each bioreactor has only one corresponding output (y value) that represents the product quality (final titer). So only one row as an out put per bioreactor, which is a total of 20 rows.

How do I train a ML model with an input and output relationship being many to one? In other words, how can I train my model to predict the product quality for each bioreactor with the 11 day continuos data? Any resources relating to this would be really helpful!

Answer 1

If I understand correctly you got 20 timeseries datasets (1 for every reactor) and after a timespan t a quality label is asigned for the whole timeseries.

You can take lots of different approaches from here. Doing some exploratory analysis first helps you to understand your underlying data better.

In other words: is your quality label determined by univariate characteristics or by the development of values in your series over time.

Assign the quality label to every row of your timespan for each bio reactor and do some plotting e.g. using seaborn scatterplots and plot every feature y over time x with the hue = your target variable, or use boxplots / pairplot to inspect your data without considering the time factor.

If the time factor is irrelevant to predict your label just remove the column and apply your y to every reactors observataion rows.