Vibration Fault Diagnosis

Diagnostic Methodology

Always follow this structured diagnostic workflow:

Step 1 — Gather Context

Ask or determine:

• Machine type (pump, motor, fan, compressor, etc.)
• Shaft speed in RPM — ask the operator; do not guess values
• Bearing type if known (designation like 6205, 6306, etc.)
• Symptom description (noise, temperature, vibration increase)
• Machine group for ISO 10816 (group1–group4)

Step 2 — Acquire or Load Data

Option A — Live acquisition from STWIN.box (preferred):

1. datalog2_connect to the board
2. datalog2_start_acquisition(duration_s=5) — timed mode prevents overshoot
3. load_signal(file_path="<acquisition_folder>", sensor_name="iis3dwb_acc")
   • The acquisition folder path is returned by the start tool
   • The SDK decodes .dat binary format automatically
   • Sample rate is auto-detected from device config

Option B — Load from file:

• load_signal(file_path="path/to/data.csv", sample_rate=26667)
• For DATALOG2 folders: load_signal(file_path="path/to/acquisition_folder")

Step 3 — Spectral Analysis

1. Compute FFT with compute_fft_spectrum
2. Find peaks with find_spectral_peaks
3. Identify shaft harmonics: look for peaks at 1×, 2×, 3× shaft frequency
4. Compute PSD if noise assessment is needed

Step 4 — Shaft-Frequency Interpretation

Use the fault-signature reference table below:

Step 5 — Bearing Analysis (if applicable)

1. Calculate bearing frequencies: use calculate_bearing_frequencies or lookup_bearing_and_compute
2. Envelope analysis: use compute_envelope_spectrum
   • Band-pass around a resonance (typically 2–5 kHz for small bearings)
3. Check for BPFO/BPFI/BSF/FTF: use check_bearing_fault_peak for each frequency
4. Interpret:
   • BPFO (outer race) — most common, modulation at cage frequency
   • BPFI (inner race) — modulated at shaft frequency
   • BSF (ball defect) — less common, often with 2× BSF
   • FTF (cage) — low frequency, rare and severe

See fault-signatures.md for detailed patterns.

Step 6 — Severity Assessment

• Use assess_vibration_severity for ISO 10816 classification
• Compare against baseline if available

Step 7 — Report

Present findings with:

• Overall severity (ISO zone)
• Identified faults with confidence level
• Evidence (which peaks, which harmonics)
• Recommendations (actions to take)
• Suggested follow-up measurement interval

Quick Diagnosis Tool

For a fully automated diagnosis, use diagnose_vibration:

• With RPM: full shaft-frequency features + bearing analysis + ISO classification
• Without RPM: basic statistics only (RMS, kurtosis, crest factor) + ISO severity — do not invent an RPM value; ask the user if it matters

Use the multi-step approach above when more control or explanation is needed.

Important Caveats

• Always state the confidence level of each diagnosis
• Multiple faults can coexist — don't stop at the first finding
• Bearing faults progress: defect → spalling → catastrophic. Early-stage defects may only show in the envelope spectrum, not the raw FFT
• Electrical faults (2× line frequency, rotor bar pass) can mimic mechanical faults
• Always recommend verification through complementary methods (temperature, acoustic, visual inspection) before major maintenance decisions
• If RPM is not provided, ask for it or explicitly omit shaft-synchronous conclusions

⛔ Mandatory Rules — Honesty About What the Data Shows

RPM and shaft-frequency analysis

If diagnose_vibration was called without RPM:

• The tool returns only basic statistics (RMS, kurtosis, crest factor) + ISO severity + spectral peaks
• You MUST NOT add shaft-frequency fault diagnoses (unbalance, misalignment, looseness) that the tool did not return
• You MUST NOT guess RPM from spectral peaks. A dominant peak at, say, 24.5 Hz does NOT mean "the shaft speed is 1470 RPM" — it could be a fan blade pass frequency, an electrical frequency, a structural resonance, or many other things
• If you observe prominent peaks and want to mention them, say: "Dominant spectral peak at XX Hz — origin unknown without shaft speed information"

Bearing fault hypotheses

If no bearing data was provided (no designation, no geometry, no fault frequencies):

• The tool cannot compute BPFO/BPFI/BSF/FTF
• You MUST NOT hypothesize bearing faults based on spectral peaks alone
• If kurtosis or crest factor is elevated, you may note impulsive content but must say: "Elevated impulsiveness may indicate bearing degradation, but confirmation requires bearing fault frequency analysis with known bearing geometry"

Spectral peak interpretation without context

A spectral peak is just a frequency with energy. Without knowing:

• The shaft speed (RPM)
• The number of fan/impeller blades
• The gear tooth count
• The bearing geometry
• The electrical supply frequency

…you CANNOT attribute it to a specific fault. State what was measured, not what you think it might mean, unless the MCP tool explicitly classified it.

Distinguish MCP tool output from general knowledge

When providing recommendations or context based on your general engineering knowledge (e.g., ISO 1940 balancing grades, typical speed ranges for machine types, maintenance intervals) rather than MCP tool output, you MUST label it:

⚠️ General engineering knowledge — not from sensor data analysis.

Never present textbook knowledge as if it were a finding from the MCP analysis pipeline. The user must know what came from measured data vs. your training data.

Fault Classification Script

See classify_fault.py for the rule-based classification logic.

Evidence & Assumptions Protocol

When presenting results, always separate:

1. Measured evidence (tool outputs, frequencies, amplitudes, ISO zone, statistics)
2. Inference (diagnostic interpretation with confidence)
3. Assumptions / prior knowledge (catalog values, typical fault heuristics, missing machine metadata)

If assumptions are used because required inputs are missing, declare this explicitly and ask for the missing data when practical.