InsightFab
Knowledge Base/Logistic Regression in Manufacturing: Defective Product Prediction Model
Statistical Analysis6 min read

Logistic Regression in Manufacturing: Defective Product Prediction Model

This article details a manufacturing scenario where a new machine's yield plummeted, with a Cpk of 1.08, leading to significant defective product output before issues were identified. It explains how a Logistic Regression model was implemented to proactively predict batches at risk of failure, enabling timely intervention and demonstrating the application of data science to solve real-world production challenges.

The Day the CPK Report Came Out, Silence Fell for Three Seconds

I still remember years ago, a new machine was just installed on our production line, and everyone was eagerly anticipating its high production capacity. However, shortly after trial production began, the yield started to plummet like a slide. The worst part was, after we ran a batch of products, my supervisor asked me to look at the Cpk report, which showed a number of 1.08. In the meeting room, besides the sound of the air conditioning, no one spoke. Everyone knew deep down that this number was a disaster. The boss's face turned green, and he directly asked, "Who can tell me where the problem is?"

Where Was the Problem?

Honestly, low yield can certainly have many causes. It could be machine parameters drifting, raw material batch issues, or even temperature and humidity fluctuations. But the most troublesome part is that usually, by the time we realize there's a problem, we've already produced a lot of defective products. This is like going to the doctor, who tells you you're sick, and *then* starts drawing blood and taking X-rays; by the time your results come back, you might be critically ill.

So, what's the point? What we need isn't hindsight, but a model that can "proactively" determine which batches might have problems. This is where "Logistic Regression" comes into play. Frankly, it sounds very academic, but at its core, it's a classifier. Based on various conditions you provide, it determines the probability of a batch being a "good product" or a "defective product."

Imagine feeding it data such as machine temperature, pressure, time, as well as raw material purity and supplier information. It learns to identify patterns from these data points. When a new production batch comes in, it can then give you a "probability that this batch will produce defective products."

How Is It Actually Done?

To build this kind of predictive model, you must have historical data at hand. This data should at least include:

  1. Various process parameters: Such as machine temperature, pressure, flow rate, rotational speed, and time. These numbers must be quantifiable.
  2. Raw material information: Supplier, batch number, and values from raw material inspection reports.
  3. Final outcome: Whether the batch is a "good product" or a "defective product." This is your target variable.

Feed these data points to the Logistic Regression model for training. Once trained, you can use it for prediction. For example, if your model predicts that a certain batch's yield is only 70%, then your DPMO (Defects Per Million Opportunities) would be 300,000! This number is alarmingly high. At this point, you can quickly stop the line for inspection before the entire production run is completed, instead of discovering a pile of scrap products only after everything is done.

So the key is that this tool can help you identify the critical factors among a multitude of parameters that truly affect yield. It will tell you which parameter has the greatest impact on the probability of defective products occurring.

Common Pitfalls

When implementing this type of model, I've most frequently encountered two pitfalls:

  1. Terrible data quality: Often, machine data is not properly collected, or it's incomplete. Some parameters are even recorded despite no one managing them. This leads to very poor model training results. Remember, Garbage In, Garbage Out. If the data is bad, even the most powerful model is useless.
  2. Assuming the model solves all problems: A model is just a tool; it gives you a probability, but it won't directly tell you *why* defects occur. If it predicts low yield, you still need to go back and examine the process to find the root cause. It merely helps you "narrow down the problem scope," it's not a magic wand.

Once, the model consistently predicted issues with a certain batch, but our parameter checks showed everything was normal. Later, we discovered that a particular operator habitually opened a valve to its maximum, leading to an instantaneous overflow. This operation wasn't recorded in any parameters, so naturally, the model couldn't catch it. This also reminded me that on-site "nuances" are still very important.

One Thing You Can Do Today

Go back and look at your product line with the worst yield, then gather the process parameters and yield data from the past three months.

Want to try it yourself?

Every tool mentioned in this article is available on InsightFab — just upload a CSV to analyze.

Go to Tools

More in this category

Correlation Coefficient: Pearson vs Spearman – Linear vs Monotonic Relationships

This article sheds light on a common pitfall in engineering data analysis: the misapplication of correlation coefficients. It explains why relying solely on Pearson correlation, which primarily detects linear relationships, can lead to incorrect conclusions when data exhibits non-linear, monotonic trends, especially when addressing issues like unexpected product yield drops.

Post-hoc Test: Choosing Between Tukey vs Bonferroni

This article addresses a common production line challenge: identifying the specific source of a problem when overall metrics like CPK fall short. It introduces Post-hoc Tests as a crucial tool to pinpoint significant differences between individual machines or groups, enabling precise problem-solving and immediate action.

Statistical Power: How to Ensure Your Conclusions Are Meaningful

This article addresses a common challenge for engineers: drawing conclusions solely based on observed data changes without accounting for statistical significance. It emphasizes the importance of using 'Statistical Power' to validate findings, ensuring that conclusions are robust and well-supported, particularly when presenting to management.