Lag Effect Analysis VIP Exclusive
⏳ What is this?
Most correlation analyses assume one thing: the two variables are at the same moment in time.
But on a real production line, cause and effect are often delayed——you adjust the feed rate today, and it may take two hours before you see the response in product quality.
Lag effect analysis is used to answer this question:
"How many steps (or how long) does it take for a change in the source variable to show up in the target variable?"
🧐 How to read?
The chart title has the form Target ➔ Source (Lag: 20), showing the correlation strength between the target variable and each source variable at different lag steps.
- X-axis: lag step (from 0 to the configured maximum lag)
- Y-axis: correlation coefficient (or correlation strength)
- Each curve: the lag correlation between one source variable and the target variable
Key readings:
| Pattern | Meaning |
|---|---|
| The curve is highest at Lag = 0 | The two variables change in sync, with no obvious lag |
| The curve peaks at Lag = k | The source variable leads the target by k steps——there is a transmission delay |
| The curve is flat overall with no obvious peak | The relationship is unclear, or the lag window does not cover the real delay |
| The curve shows a negative correlation peak | The variable moves in the opposite direction to the target |
💡 The Lag at which the peak appears is the "transmission time" of the process. For example, if "feed rate" is found to correlate most strongly at Lag=12 and the sampling interval is 10 minutes, it means the feed adjustment takes about 2 hours to be reflected in the quality indicator.
🛠️ How to use?
Configuration items
| Configuration item | Description |
|---|---|
| Target variable (Target) | The "result" variable you want to observe, single select |
| Source variable (Source) | Variables that may lead the target, multiple selection supported |
| Maximum lag step (Max Lag) | How many time steps to look back, range 1 ~ 300, default 20 |
After selecting, click Draw to generate the chart.
💡 How to set the maximum lag step? First estimate an upper bound from business experience ("what is the longest delay in this process?"), convert it into sampling steps and then scale it up a bit. If it is set too small, you will miss the real peak; if it is set too large, there will be more noise and slower computation.
⚠️ Prerequisites (important)
This chart requires the model to have selected "non-random" dataset splitting when it was created.
The reason is that lag analysis is founded on time order. If the data is shuffled (shuffle = true), "k steps before" loses its physical meaning, and lag correlation becomes meaningless.
How to handle it:
- Go back to Step 1: Data Preparation and check the "random or not" selection
- If the current one is "random", you need to create a new "non-random" model before using this chart
- The chart will also give a prompt when the condition is not met
🎯 Typical uses
- Locate process transmission delays: provide lag parameters for feedforward control and soft-sensor modeling
- Troubleshoot "no response after adjustment": confirm whether the delay is too long, so that operators, unable to wait, repeated the adjustment
- Time series alignment: before building a dynamic model, first align each variable by its own optimal Lag
💡 The analysis results can be used directly for APC/RTO modeling in Industrial Control Systems—— dynamic MPC requires explicit lag times to build prediction models.