Gains:
- Ability to recognize the concept of predictive maintenance, vibration/temperature indicators and typical failure modes
- Ability to configure anomaly, trend and characteristic frequency analysis in sensor time series with AI
- Ability to cross-check AI interpretation with physical failure mechanism, calibration and field inspection
Predictive maintenance is one of the applications that provides the highest return in a facility by reducing planned downtime and preventing unexpected malfunctions by constantly monitoring the condition of the machine and predicting failure before it occurs. Instead of the classic "fix it when it breaks" or "replace it on schedule" approaches, we read the actual state of the machine from sensor data such as vibration, temperature, current and acoustics. Artificial intelligence (AI) is very powerful in this area: anomaly (deviation from normal) detection in time series data, trend analysis, interpretation of characteristic fault frequencies and report generation. But the “possible malfunction” given by the AI is a hypothesis, not a diagnosis. The sensor may be broken, its calibration may be off, the data may be noisy, and AI may assign meaning to a non-physical pattern. AI interpretation before a maintenance decision is made; The characteristic fault frequency should be cross-checked with sensor calibration, data quality and field inspection. In this unit, you will learn to safely interpret sensor data with AI.
Measured Quantities and Typical Failure Modes
The health of rotating machines (motor, pump, fan, bearing, gear) is read from a few basic quantities. Each quantity is sensitive to certain failure modes:
size
sensitive to what
Typical failure mode
Vibration (acceleration/speed)
Imbalance, misalignment, bearing/gear
Unbalance, misalignment, bearing damage
temperature
Friction, loss of lubrication, overload
Bearing heating, oil degradation
motor current
Electrical and mechanical load change
Rotor bar fracture, load anomaly
acoustic/ultrasound
Premature bearing damage, leakage
Oil film degradation, gas leak
The fundamental tool in vibration analysis is the frequency spectrum: we separate the time signal into its frequency components with the Fourier transform (FFT: the mathematical process that separates a signal into its component frequencies). Each failure mode produces a peak at a characteristic frequency. For example, imbalance appears at the rotation frequency (1X), axis misalignment mostly occurs at 2X, and bearing damages appear at special frequencies depending on the bearing geometry (such as BPFO, BPFI, BSF).
Tip: Before having the AI interpret the spectrum, give the machine's rotational speed (RPM) and bearing number, if applicable. Characteristic fault frequencies cannot be calculated without this information; without them, the AI's interpretation of "bearing failure" remains in the air.
Anomaly, Trend and False Alarm
AI does two things well in time series: anomaly detection (sudden deviations) and trend analysis (slow decay). But there are two traps. The first is a false positive: noise, a temporary load change or sensor vibration may appear to be the real malfunction. The second is a false negative: a slow, gradual deterioration may be missed because it remains within the "normal range". A good monitoring also tracks the change based on the machine's own history (baseline) rather than the absolute threshold.
Caution: A sensor itself may fail. While a temperature sensor that is out of calibration may give an "overheating" alarm, the machine is perfectly normal; or a loose accelerometer produces spurious vibration. What AI sees is just data; Field control tells you whether the source of the data is reliable. “Verify sensor first” is the first rule of predictive maintenance.
Step by Step: Interpreting Sensor Data with AI
- Give the context. Machine type, RPM, bearing/gear information, measuring point.
- Check data quality. Is the sampling frequency sufficient, are there any missing/outliers?
- Determine the baseline. What does a healthy state look like?
- Have the anomaly/trend extracted. Mark deviation and trend with AI.
- Request physical interpretation. Which failure mode, at what frequency, why?
- Verify in the field. Calibration, visual/tactile check, second measurement if necessary.
Spectrum interpretation prompt
Role: Vibration analysis (condition monitoring) specialist.Context: A centrifugal pump rotating at 1450 RPM, drive end bearing 6206.Data: peak frequencies in the vibration speed spectrum [list, Hz and amplitude].Task: Correlate each peak with the possible failure mode (1X unbalance,2X axis misalignment, bearing BPFO/BPFI etc.). Calculate the characteristic bearing frequencies with RPM and 6206 geometry, state the frequencies you show. Rule: This is a HYPOTHESIS; For each comment, write how I verify it in the field.
Anomaly/trend analysis prompt
Analyze the following temperature time series: [date, list of values].- What is the baseline (normal range)?- Are there anomalies (sudden deviation), when?- Is there a slowly rising trend, what is the slope?Rule: Assess the risk of false positives; What additional check would you recommend to distinguish between sensor error and real overheating?
Data quality and sensor verification prompt
The sampling frequency in this jitter measurement is 2 kHz. The highest fault frequency I'm interested in is ~800 Hz. Is this sampling sufficient (Nyquist), comment. Symptoms of sudden jumps and clipping in the data are as follows: [description]. Could these be real signals or sensor/recording errors, how can I distinguish?
Maintenance decision draft prompt
Write a draft maintenance recommendation based on the following findings: [findings]. Divide your recommendation into three levels: stop immediately, inspect at scheduled maintenance, continue monitoring. Write justification and field verification for each level. Rule: The competent maintenance engineer will make the final decision; state this.
Weak Prompt / Strong Prompt
Weak prompt:
Is this vibration high?
No machine type, RPM, measurement unit, baseline and bearing information; AI cannot say "high/low", even if it does, it is unsupported.
Powerful prompt:
1450 RPM centrifugal pump, bearing 6206, drive end metering. Vibration velocity spectrum peak frequencies: [list]. Associate each peak with a possible failure mode, calculate and display characteristic bearing frequencies. State that each comment is a hypothesis and write how to verify it in the field. Also consider the possibility of sensor error.
The second prompt contains the machine context, spectrum data, and verification request; It positions the interpretation as a hypothesis.
Three Mini Cases (By Numbers)
Case 1 - Characteristic frequency verification. On one pump the AI comments "inner ring failure (BPFI)". Engineer calculates BPFI with 1450 RPM (24.2 Hz) and 6206 bearing geometry: ~131 Hz. The peak in the spectrum is actually around ~130 Hz and has harmonics. The interpretation is consistent with the characteristic frequency; The engineer disassembles the bearing during planned maintenance and finds pitting on the inner ring. Confirmation turns the hypothesis into a diagnosis.
Case 2 - False alarm. On one fan, the temperature sensor gives a 95°C alarm, the AI says "bearing overheating". The field team does a manual check: the bed is warm, normal. The sensor connection is loose, the calibration has shifted. Actual temperature ~55°C. When the sensor is tightened and calibrated, the alarm disappears. If the AI interpretation were followed blindly, a solid bed would be dismantled unnecessarily. Lesson: verify the sensor first.
Case 3 - Catching the slow trend. The absolute vibration threshold is not exceeded in a gearbox, but AI shows amplitude increasing by ~6% per week on an 8-week trend. By extrapolation, it is predicted that the threshold will be exceeded in ~5 weeks. During the planned stop, the team analyzes the oil, finds the metal particle and replaces the gear in a controlled manner; Unexpected stopping is prevented. Lesson: the trend warns before the absolute threshold.
Common mistakes
- Asking for comments without context: Interpreting the spectrum without RPM and bearing information.
- Not verifying the sensor: Mistaking a calibration/connection error as a real malfunction.
- Mistaking a hypothesis for a diagnosis: Translating AI interpretation into a maintenance decision without field control.
- Looking only at the absolute threshold: Missing the trend of slow decay.
- Undersampling: Neglecting the Nyquist rule and overlooking the high frequency fault.
- Overreaction to false positive: Mistaking the noise for a real malfunction and stopping the healthy machine.
In summary
- Predictive maintenance reads the machine status from sensors such as vibration, temperature and current and predicts failure in advance.
- In the vibration spectrum, each failure mode produces a peak at a characteristic frequency; The comment requires RPM and bearing knowledge.
- The fault interpretation given by the AI is a hypothesis; The characteristic frequency must be verified by calibration and field inspection.
- The sensor itself may fail; "verify sensor first" is the first rule.
- The trend is as important as the absolute threshold; Slow degradation gives early warning. The maintenance decision belongs to the competent engineer.
Application task
Choose a rotating machine (pump, fan, motor). Write down the machine's RPM and bearing number, if applicable. Take an actual or sample vibration spectrum / temperature time series and have the AI interpret it with its context. For each failure hypothesis the AI shows: check the characteristic frequency yourself (verify BPFO/BPFI calculation if bearing), evaluate the possibility of sensor failure and write how you will verify it in the field. Finally, distill the findings into a three-level recommendation (stop immediately / inspect on schedule / continue monitoring) and state why the final maintenance decision remains with the qualified engineer.
checklist
- [ ] The machine context (type, RPM, bearing/gear, measuring point) is given.
- [ ] Data quality and sampling frequency (Nyquist) checked.
- [ ] The fault frequencies shown by AI were verified by characteristic frequency calculation.
- [ ] Possibility of sensor calibration/connection error evaluated.
- [ ] In addition to the absolute threshold, the trend was also examined.
- [ ] Comment confirmed by field inspection; The maintenance decision was left to the competent engineer.