Frequently Asked Questions (FAQ)
📥 Data Import
❓ Which file formats are supported?
Currently Excel format is supported: .xlsx and .xls, with a single file no larger than 50MB.
⚠️ Importing CSV directly is not supported. If your data is in CSV, open it in Excel first and save it as
.xlsx.
❓ The header has two rows — how do I set that up?
The platform supports a two-row header:
- Title Row — holds the variable English names / tag names (such as
x1,Temperature), required - Name Row — holds the Chinese description (such as
温度,压力), optional
How to set it: simply click the row number. It doesn't matter if the header is not in the first two rows — just click the row number where it actually is.
❓ A column contains empty values and it says "please handle them first" — what should I do?
Click the Imputation button for that column and choose one of four methods:
| Method | When it applies |
|---|---|
| Mean | The distribution is fairly symmetric (most commonly used) |
| Median | More robust when there are extreme values and the distribution is skewed |
| Fill with 0 | The missingness itself has business meaning |
| Custom | Fill with a fixed value you specify manually |
❓ It says "empty values or non-numeric values exist"?
This means text has been mixed into that column. Common causes:
- Some "numbers" in Excel are actually in text format (with a green triangle in the top-left corner)
- Units have been mixed into the column (such as
25℃instead of25) - The column is descriptive text to begin with and was mistakenly selected as an X or Y column
🤖 AI Features
❓ Clicking an AI feature does nothing / it asks me to configure an API Key?
AI features require you to fill in an API Key first in Settings → LLM Configuration.
The prerequisites differ by feature:
| Feature | What it requires |
|---|---|
| AI Auto Modeling | LLM configuration + a non-free license |
| AI Cleaning / Chart Analysis / Analysis Chart Creation | LLM configuration + an activated and fitted model |
| AI Comparative Analysis | The above conditions + red and blue groups already selected |
| Scenario Analysis / Configuration Diagram Analysis | Only LLM configuration |
❓ Why does configuration diagram analysis fail?
A multimodal model must be used (one with image reading capability). A text-only model cannot recognize image content.
Recommendation: upload clear configuration diagram / DCS screenshots, at most 3, containing as complete variable names and units as possible.
❓ The AI says to remove a certain data point — can I trust it?
Not without verification. The platform will not automatically delete any data — the AI only gives suggestions, and whether to remove something is ultimately your decision.
It is recommended to verify in this order:
- Search for that sample ID in the Data Table and inspect the raw values column by column
- Confirm whether it is an obvious problem such as a unit error or data entry error
- Use process knowledge to judge whether it really is an abnormal operating condition
- After removal, run Fit again and compare whether Q²Y improves
⚠️ Large language models have hallucinations, and the analysis is limited by context length. Always combine it with business judgment.
❓ Does AI analysis consume a lot?
It consumes tokens, and the more charts and the larger the data, the higher the consumption. Suggestions:
- Close irrelevant charts first and keep only a few key ones before analyzing
- For long reports, prefer to run them during periods when your "cloud agent" quota is plentiful
📊 Modeling and Analysis
❓ Which models does the platform support?
Five: PCA, PLS, PLS-DA, OPLS, OPLS-DA.
- X columns only → PCA is used automatically
- X + Y (continuous) → PLS is used automatically
- X + Y (categorical) → PLS-DA is used automatically
- OPLS / OPLS-DA must be selected manually
❓ What is the difference between OPLS and PLS? Which should I choose?
OPLS splits the variation of X into two parts — "related to Y" and "unrelated to Y (orthogonal)" — peels off the noise, and then builds the model.
- Predictive ability: the two are usually close
- Interpretability: OPLS is clearly better (there is only 1 predictive component, so the score plot is cleaner)
- Recommendation: choose OPLS when the data has obvious systematic noise (baseline drift, environmental drift), or when you need to explain it to others
⚠️ Note: On OPLS, C+1 increases the "number of orthogonal components", not the predictive ability!
See Core Model Algorithms in Detail.
❓ Why won't the charts load / why are they so laggy?
It may be because the data volume is too large. Try taking the first 50,000 rows in Excel first — the platform mainly targets refined experimental data, not million-row big data scenarios.
Also note that the number of charts on the canvas has an upper limit, and too many charts will also slow down the response. You can use "Clear All" first and then drag in the charts you need.
❓ Why is R² very low?
- The data is too dirty: try AI Cleaning, or delete outliers in the T² plot
- The wrong variables were selected: check whether the X columns and the Y column really are related
- Too few components: try clicking C+1 a couple of times
- The relationship is simply weak: if Q² is also low and changing the model doesn't help, this batch of data may genuinely contain no strong pattern
❓ R² is very high but Q² is very low — what should I do?
This is a classic sign of overfitting. Handle it in this order:
- Reduce components first: click C-1 and see whether Q² recovers
- Then clean outliers: remove abnormal samples and run Fit again
- Check the variables: remove variables with VIP < 0.5 and rebuild the model
- Verify the gap: a healthy model should satisfy
R² − Q² < 0.2
❓ How exactly is the model's "component count" determined?
The platform selects it automatically through cross-validation (7-fold by default), and you can also fine-tune it manually with C+1 / C-1.
The behavior of the automatic search can be adjusted in Settings → Model Training Settings (seven parameters including the number of cross-validation folds, the maximum component limit, and the early-stopping tolerance). See Model Training Settings.
❓ How do I find abnormal samples?
Four routes, which can be combined:
| Method | Entry point |
|---|---|
| T² / SPE out of limits | T² plot, SPE plot, T²&SPE joint distribution plot |
| Manual lasso selection | Toolbar → Lasso |
| Isolation Forest / LOF | The detection button on the scatter plot / prediction plot |
| AI automatic cleaning | AI Analysis → AI Cleaning |
❓ The model didn't change after I deleted data?
The delete operation creates a new filtered model; the original model is still kept.
You need to: rename the new model → click to activate it (Active) → click Fit to retrain. See Step 3: Cleaning Data.
🔮 Prediction and Export
❓ How do I use a trained model to predict new data?
Use copy and paste: open Model Prediction, click inside the dialog, paste the data copied from Excel with Ctrl+V / Cmd+V, and click "Predict".
⚠️ The column order must match the X (+ Y) variable list — the platform reads by column order, not by matching column names. This is the most common usage accident.
You can click "Copy Results" and paste them back into Excel. See Model Prediction.
❓ Can the prediction panel import a file?
No. The prediction panel only supports pasting, which makes the "test each batch as it arrives" workflow faster.
If you need to run large amounts of data in batches, it is recommended to use Custom Download to export the model and predict in your own Python environment.
❓ Downloading a model says "not supported by the current license"?
Model download requires the download permission to be enabled in your license. You can check "Download permission: Allowed / Forbidden" in bottom status bar → License Info.
❓ What is the difference between .dimod and .joblib?
| Format | Purpose |
|---|---|
.dimod | Full-package migration, completely restoring the analysis environment on another computer with StarWayDI installed |
.joblib / .pkl | Can be loaded directly by Python, for secondary development outside the platform |
❓ Which items should I check when downloading a model package?
The minimal working combination: model + x_scaler (for regression models add y_scaler; for discriminant models add y_binarizer).
⚠️ Never forget to check
x_scaler— new data must be transformed with the same scaler, otherwise the prediction results will be completely wrong.
See Step 4: Exporting Results.
❓ Can I export images?
Yes, in two ways:
- Single chart: the camera icon 📷 in the top-right corner of the chart, saving a high-definition large image
- Whole screen: Save Screenshot in the canvas toolbar, capturing all currently arranged charts as a whole
❓ What formats can reports be exported to?
| Report type | Supported formats |
|---|---|
| AI Cleaning / Comparison / Chart Analysis | Markdown + Word |
| Scenario Analysis / Configuration Diagram Analysis | Markdown + Excel (multiple worksheets) |
🔒 Licensing and Deployment
❓ Does the software need an internet connection?
When you don't use AI features, it is completely offline. All data and algorithms run locally.
Only when you actively click an AI feature will data be sent to the API address that you configured yourself. Configure a custom Base URL to connect to a model deployed privately inside your enterprise.
❓ What license editions are there? How do the features differ?
The platform offers editions such as Free, Trial, Education, Enterprise, Personal, and Professional.
The main differences are:
- Chart permissions: some advanced charts (marked with a
Vbadge) require a specific edition - Model download permission: exporting model packages requires it to be enabled
- Maximum number of instances: the Free edition has a cap on the number of instances
- AI construction: not supported in the Free edition; requires a license upgrade
For the specific differences, you can check the actual permissions of your current edition in bottom status bar → License Info.
❓ Does it support domestic / Xinchuang environments?
Yes. The platform has been adapted to Xinchuang platforms such as Kylin ARM, and offers multi-platform versions for Windows / macOS / Linux.
🆘 Still Not Solved?
- Click Roaming Guide in the bottom status bar — an interactive step-by-step guide covering seven stages: "Instance Preparation → Modeling → Layout → Basic Charts → Advanced Exploration → Comparative Analysis → AI Features"
- Click Help Documentation — return to this site to consult the complete documentation
- Check whether the AI-related configuration is correct in Settings → LLM Configuration