lgdimaggio/predictive-maintenance-mcp/gear-diagnosis

Gear Fault Diagnosis

Detect gear faults through spectral analysis of vibration signals, focusing on gear mesh frequency (GMF) harmonics, sidebands, and modulation patterns. This workflow supports expert judgment — it never replaces it.

Prerequisite: The predictive-maintenance-mcp MCP server must be connected.

Background

Gear faults produce characteristic vibration patterns:

• Gear Mesh Frequency (GMF) = number_of_teeth x shaft_RPM / 60
• Healthy gears show GMF and low-level harmonics
• Faulty gears show increased GMF harmonics, sidebands spaced at the shaft

frequency, and broadband noise

Workflow

Step 1 — Signal Selection and Parameters

Load the signal with load_signal(filepath="gearbox_run1.csv", signal_unit="g") or pick a loaded one from list_signals(scope="memory"). Verify sampling rate and metadata with get_signal_info(signal_id="<id>"). For raw binary files (.bin/.raw/.dat), also declare sample_format and sampling_rate (or provide a companion <stem>_metadata.json) — see the signal-management skill.

Gather from the user (do NOT guess):

• Shaft RPM (or shaft frequency in Hz)
• Number of teeth on the gear of interest
• Gear ratio (if analyzing a gear pair)

Calculate:

• GMF = teeth x RPM / 60
• Shaft frequency = RPM / 60

Step 2 — Statistical Screening

Call analyze_statistics(signal_id="<id>").

Gear fault indicators (screening only, never diagnostic alone):

• Kurtosis > 3: impulsive content (possible tooth damage)
• Crest Factor > 5: localized damage
• RMS increase: overall vibration energy rise

Step 3 — FFT Analysis

Call analyze_fft(signal_id="<id>").

Inspect the spectrum for:

Systematic check of the expected GMF against the envelope spectrum:

Call check_bearing_faults(signal_id="<id>", rpm=1500, frequencies={"GMF": 800.0})

(The frequencies route accepts any labeled frequency — it is the out-of-catalog / gearbox path.)

Step 4 — Envelope Analysis (for Modulation)

Call analyze_envelope(signal_id="<id>", filter_low=500, filter_high=5000).

Center the demodulation band around GMF or its harmonics (keep it below Nyquist) to extract amplitude modulation. Sidebands in the envelope spectrum at shaft-frequency spacing corroborate gear fault modulation.

Step 5 — Advanced Spectral Analysis

Call compute_spectrogram_stft(signal_id="<id>") to check for time-varying patterns — localized tooth damage shows periodic energy bursts at the shaft rotation rate.

Call compute_power_spectral_density(signal_id="<id>") for a smoothed spectral view that highlights persistent gear-mesh patterns.

Step 6 — ISO Severity Context

Call assess_severity(signal_id="<id>", machine_group=2, support_type="rigid").

Report the zone (A/B/C/D) and urgency. The verdict requires a DECLARED signal unit — if refused, confirm the unit with the user and re-load with load_signal(filepath="<file>", signal_unit="mm/s", overwrite=True).

Step 7 — Diagnosis Summary

Classification rule (choose exactly one, cite the evidence):

• Gear tooth fault — strong evidence: GMF harmonics present AND at least

one clear sideband pair spaced at the shaft frequency AND a supporting statistic (Kurtosis > 3 or CF > 4)

• Possible localized tooth damage: partial sidebands or ambiguous

spacing — list the missing evidence needed for confirmation

• Uniform wear / increased load: elevated GMF amplitude WITHOUT sidebands

and normal impulsiveness

Each conclusion MUST cite: tools used, specific peaks (frequency and magnitude), sideband spacing vs the expected shaft frequency, and supporting statistics.

Step 8 — Report Generation

Call generate_fft_report(signal_id="<id>", max_freq=5000, num_peaks=15, rpm=1500) and optionally generate_diagnostic_report_docx(signal_id="<id>", sections={"summary": "..."}).

Important Notes

• GMF calculation requires an accurate tooth count and RPM — ask, never assume
• Multi-stage gearboxes: analyze each stage separately
• Gear and bearing faults can coexist; if bearing frequencies also appear,

recommend the bearing-diagnosis skill as well

• Use cautious language for all diagnoses; the final call rests with a

qualified engineer

• All processing is local — no data leaves the machine