Real-Time Performance Drift Detection

Introduction

Some production losses, such as unplanned stops and defective items, are easy to identify with existing monitoring systems. Slowdowns and other subtle losses, however, mostly go unnoticed.

Because there are many types of production lines, across different industry segments, it is very hard to build general rules that detect these losses. The goal was to develop an automated solution able to process high-speed data online, to identify these performance drifts in real time.

Methodology

In a fast-paced project, Mip Wise built an online loss detection model that raises alerts for the production lines. The proof of concept (POC) followed these phases:

• Discovery: understanding the problems and how the current system works, surveying the available data that could be used and gathering insights from the client’s team.
• Data cleaning and preparation: since the client’s datasets were not uniform, we built a preprocessing pipeline to handle irregularly spaced time series, outliers and variations between the different production lines.
• Exploratory data analysis (EDA): building on the previous step, we deepened our understanding of the time series and the other available data with charts and summaries, and formulated hypotheses that were discussed with the client’s team and tested on our side.
• Model development: still during the EDA, several statistical and machine learning models were considered. One of the statistical algorithms showed the most promising results. During development the model received several tailored improvements, and in the implementation phase the algorithm’s parameters were tuned by clusters of production lines.
• Deployment: the solution was delivered as a set of integrated modules and functions, designed to be built directly into the client’s production environment by its technical team.

Results

The solution identified latent throughput losses that did not show up in the dashboard views before, with detection within time windows that allow action.

• Accuracy: proven effectiveness in detecting subtle losses across heterogeneous lines.
• Strategic alignment: the client’s technical team is working on the internal integration and on validation with other datasets. Feedback on the results was positive, and the company’s CEO showed enthusiasm about the next steps.