industrial-ai-troubleshooting-agent: Command for Claude Code

ML Fault Detection

Run Isolation Forest anomaly detection and XGBoost feature importance analysis on the boiler historian data to identify fault periods and rank early-warning sensors.

Conventions

• Python: Always use the project's virtual-environment Python, never bare python (which on macOS may not exist or may point at a dependency-free system Python). Written as <venv-python> below — substitute .venv/bin/python on macOS/Linux or .venv/Scripts/python on Windows.
• Opening files: Use the command for the user's OS, written as <open> below — open on macOS, xdg-open on Linux, start on Windows.

Steps

1. Confirm data range

Call historian_get_data_range to confirm the dataset's time span and report it to the user before running.

2. Run the analysis

Run the ML fault detection script:

<venv-python> ml_fault_detection.py --downsample-minutes 1 --top-n-events 5

Capture the JSON output from stdout. The script prints progress to stderr and the JSON summary to stdout. If xgboost is unavailable for any reason — not installed, or installed but unable to load (e.g. missing libomp on macOS) — the script automatically falls back to GradientBoostingClassifier. Note which model was used from the feature_importance.model field.

If the script fails, diagnose the error from stderr and report it to the user.

3. Open key plots

All outputs are saved to plots/MLFault/MLFault_<date>/. Open the monitoring chart, feature importance chart, and fault report:

<open> <monitoring_plot_path>
<open> <feature_importance_plot_path>
<open> <fault_report_path>

4. Present model summary

Report to the user:

• Total observations and downsample resolution
• NOC training set size (rows where TE_8332A is in [530–545°C] vs total rows)
• Contamination parameter used and resulting anomaly rate
• Number of fault periods detected

5. Investigate worst fault events

For each of the top fault events in worst_events:

1. Call alarm_search_context(event.start, window_minutes=30) — what alarms were firing around this event?
2. Call alarm_get_active_at(event.start) — full list of active alarms at the fault peak
3. Note whether TE_8332A was outside 530–545°C at that time (cross-reference with monitoring plot)