Assembly Line Failure Detection

Use Case Overview

This article demonstrates how to build a classification model for early detection of faulty parts in an assembly line. Early detection increases productivity and reduces component and operational costs, by remove the affected part from the assembly line before it reaches later processing stages.

Data Requirements

We use simulated process and test data to illustrate the concepts in this Playbook, encompassing 12 months of operations, two workstations, and two million manufactured parts. Features include a unique part identifier, date and time, test measurements, and the outcome for the part (whether it passed the tests and whether it was scrapped). The raw data are contained in XML and TXT documents.

Machine Logs Extraction

First template workflow for text extraction.

Our first step will be to extract and standardize the data from the two workstations. Workstation 10's data are stored in XML documents, while 20's data are stored in a custom log format. This flow extracts both data formats into a tabular representation for further processing using our Text Extractor operator and Variable operators with regex patterns that isolate the individual features.

In some cases, the incoming data can be in a binary or unstructured format. In that case, developers can build custom connectors with the Extensions SDK  to convert the incoming data into a structured form.

Data Transformation and Modeling

Second template for data transformation and modeling. 

The second step is to join part data across both stations into a final combined feature set, then use a range of classification algorithms to predict whether a part will be scrapped. The main transformation we apply is a series of aggregations by part ID - instead of tracking each individual pressure reading, for example, we capture the minimum and maximum pressure each part was exposed to.

We evaluate a range of classification models: logistic regression, naive bayes, and random forest, using the first 11 months of operational data as a training set, and the final month as a hold out validation set. Logistic regression is the best-performing algorithm, with a scrap prediction accuracy of 98% resulting in a ~3% cost savings from pulling scrap parts off the assembly line earlier in the process.

Accuracy Monitoring

Third template for monitoring accuracy of trained models on new data.

 Using the Load Model operator, this flow applies the most recently trained classification model to new data samples. If the fresh model performs better than the model running in the production scoring engine, the fresh model is pushed out to take its place.

Self-Serve Reporting with Touchpoints

Touchpoint interface for parameterizing failure rate query.

The Touchpoint included with this Playbook demonstrates how to wrap a data visualization flow in a simple interface for business users. Users select which stations and time ranges they would like data from, and as output, they get a series of bar charts that summarize the observed failure rates.

Check It Out!

For access to this Playbook, including its workflows, sample data, a PowerPoint summary, and expert support from Spotfire® Data Science data scientists, contact your Spotfire® Data Science sales representative.