Data as of Sep 9, 2026 · Based on 287 AI responses from ChatGPT Search and Google AI Mode · See how Parse measures this
Named in 31% of answers
Named in 16% of answers
Named in 14% of answers
Named in 13% of answers
Source | Citation share |
|---|---|
Harnessing the Power of Predictive Maintenance in HVAC Systemshttps://ncd.io/blog/harnessing-the-power-of-predictive-maintenance-in-hvac-systems-vibration-analysis-of-cooling-towers/ | 4.2% |
Predictive Maintenance with Vibration Sensors: The Complete Starter Guide - NCD.iohttps://ncd.io/blog/predictive-maintenance-with-vibration-sensors-the-complete-starter-guide/ | 2.8% |
 Predictive Maintenance with Vibration Sensors White Paper | TE Connectivityhttps://www.te.com/en/whitepapers/sensors/predictive-maintenance-with-vibration-sensors.html | 2.5% |
Industrial IoT Vibration & Temperature | Predictive Maintenancehttps://www.iox-connect.com/wireless-industrial-vibration-sensor | 2.4% |
 Predictive Maintenance in HVAC Systems | TMA Solutionshttps://www.tmasolutions.com/insights/predictive-maintenance-in-hvac-systems | 2.4% |
 Predictive Maintenance - Reduce HVAC Downtime & Costshttps://www.exergenics.com/resources/predictive-maintenance-reduce-hvac-downtime-costs | 2.1% |
Optimizing HVAC Remote Monitoring Systems with Vibration Sensorshttps://ncd.io/blog/optimizing-hvac-remote-monitoring-systems-with-vibration-sensors/ | 2.0% |
 Future-Proof Your HVAC Service Business with Predictive Maintenancehttps://www.lessen.com/resources/future-proof-your-hvac-service-business-with-predictive-maintenance | 2.0% |
Staying Cool: Predictive Maintenance for HVAC Systems | Ambiqhttps://ambiq.ai/community/staying-cool-predictive-maintenance-for-hvac-systems/ | 1.9% |
 Maintenance sensors for condition monitoring & predictive maintenance solutions - ifmhttps://www.ifm.com/us/en/us/asset-health/maintenance-sensors | 1.9% |
Implementing a predictive maintenance (PdM) solution for industrial HVAC systems (such as large Air Handling Units, centrifugal/screw chillers, cooling tower fans, and large-scale exhaust blowers) using vibration sensors `requires a robust, end-to-end architecture`. A reliable industrial setup transitions your…
Implementing a predictive maintenance (PdM) solution for industrial HVAC systems (such as large Air Handling Units, centrifugal/screw chillers, cooling tower fans, and large-scale exhaust blowers) using vibration sensors requires a robust, end-to-end architecture.
A reliable industrial setup transitions your facility from reactive or calendar-based preventive maintenance to data-driven health monitoring.
If you'd like to narrow down or build out this solution, tell me:
I can help detail the bill of materials, network topology , or software integration specs.
Yes. For industrial HVAC, I’d design the solution as a **condition-monitoring platform centered on vibration**, rather than trying to predict every HVAC failure from vibration alone. ## Recommended architecture **Vibration sensor → Edge gateway → Signal processing → Asset health model → Alert/work order**
Yes. For industrial HVAC, I’d design the solution as a condition-monitoring platform centered on vibration, rather than trying to predict every HVAC failure from vibration alone.
Vibration sensor → Edge gateway → Signal processing → Asset health model → Alert/work order
Prioritize:
Vibration monitoring is particularly useful for detecting imbalance, misalignment, looseness, bearing degradation and mechanical wear in rotating equipment.
For each critical asset, use an industrial accelerometer mounted close to the bearing housing or motor bearing. A triaxial accelerometer is preferable where you want richer diagnostics.
I would collect:
Sensor bandwidth matters: accelerometers intended for bearing diagnostics need substantially higher frequency response than sensors used only for overall vibration severity.
For VFD-driven fans and compressors, speed-normalized analysis is especially important because the vibration signature changes with operating speed.
Don't simply use one universal vibration threshold.
During commissioning, capture several weeks of "healthy" operation across:
Then calculate a baseline for each asset.
The analytics should detect both:
Absolute severity
"Fan vibration is high." and
Deviation from its own normal behavior
"This fan's bearing vibration has increased 42% over its established baseline." ISO 20816 provides general guidance for measuring and evaluating machine vibration and explicitly considers both vibration magnitude and changes in vibration.
I'd avoid starting with a black-box "AI predicts failure" system.
Use three layers:
Examples:
Analyze the spectrum for signatures associated with:
Once you have enough historical data, train models to estimate:
0–1000–100The important distinction is that anomaly detection can work before you have a large library of failures, whereas supervised failure prediction requires labeled historical failures.
Vibration should be the core signal, but I'd strongly recommend adding:
| Signal | Main value |
|---|---|
| Vibration | Bearings, imbalance, misalignment, looseness |
| Temperature | Bearing/motor overheating |
| Motor current | Electrical and loading problems |
| RPM | Normalize vibration signatures |
| Pressure | Compressor/refrigeration problems |
| Supply/return temperature | HVAC performance |
| Power | Efficiency/degradation |
This makes the system much more reliable. For example, a vibration increase accompanied by rising bearing temperature is much more compelling than a single anomalous vibration measurement.
Existing BMS/PLC data can also be incorporated rather than installing duplicate sensors; industrial vibration systems can interface through mechanisms such as 4–20 mA, PLCs and other control infrastructure.
The dashboard shouldn't just say:
Fan 17: vibration high Instead:
AHU-17 Supply Fan — HIGH RISK Health: 38/100 Vibration: +47% vs. baseline Dominant change: 1× rotational frequency Bearing temperature: +8°C Probable fault: fan imbalance / developing bearing issue Recommended action: inspect fan wheel balance and bearings within 7 days. Then automatically create a CMMS work order if the confidence exceeds your threshold.
I'd start with a pilot rather than instrumenting the entire facility.
Pilot: 20–50 critical assets
For each:
Run the pilot for roughly 8–12 weeks, deliberately including different operating conditions. Establish healthy baselines first, then progressively introduce anomaly and fault classification.
A commercial vibration-monitoring deployment can use either wireless sensors or conventional industrial sensors; the latter can integrate directly into PLC/DCS infrastructure, which can be attractive in existing industrial plants.
HVAC ASSETS
│
├── Vibration
├── Temperature
├── RPM
├── Motor current
└── BMS data
│
▼
EDGE GATEWAY
│
▼
SIGNAL PROCESSING
├── RMS
├── FFT
├── Peak / crest factor
├── Bearing features
└── Speed normalization
│
▼
ASSET DIGITAL BASELINE
│
▼
ANALYTICS ENGINE
├── Rules
├── Anomaly detection
├── Fault classification
└── RUL / failure probability
│
▼
ASSET HEALTH SCORE
│
┌────┴────┐
▼ ▼
Dashboard CMMS
│
▼
Technician action
I would not promise "failure prediction X days in advance" initially. Establish detection accuracy first. Once you have enough labeled failures and maintenance outcomes, you can legitimately develop RUL/failure-probability models.
If you tell me your HVAC equipment types, approximate number of assets, whether you already have a BMS/SCADA system, and whether you want an on-premise or cloud solution, I can turn this into a concrete architecture with sensor specifications, sampling rates, communications, database/schema, ML approach, dashboard, and an estimated pilot BOM/cost.
AHU-17 Supply Fan — HIGH RISK Health: 38/100 Vibration: +47% vs. baseline Dominant change: 1× rotational frequency Bearing temperature: +8°C Probable fault: fan imbalance / developing bearing issue Recommended action: inspect fan wheel balance and bearings within 7 days. Then automatically create a CMMS work order if the confidence exceeds your threshold.
I'd start with a pilot rather than instrumenting the entire facility.
Pilot: 20–50 critical assets
For each:
Implementing a predictive maintenance (PdM) solution for industrial HVAC systems (such as large Air Handling Units (AHUs), centrifugal/screw chillers, cooling towers, and exhaust blowers) using vibration sensors requires a robust, end-to-end Industrial IoT (IIoT) architecture. 1. Hardware & Sensor Layer (The Edge)…
Implementing a predictive maintenance (PdM) solution for industrial HVAC systems (such as large Air Handling Units (AHUs), centrifugal/screw chillers, cooling towers, and exhaust blowers) using vibration sensors requires a robust, end-to-end Industrial IoT (IIoT) architecture.
Target critical rotating assets where mechanical failure typically originates: motors, fan bearings, gearboxes, and compressors.
To help tailor this architecture, let me know:
Absolutely. For industrial HVAC, I’d design this as a **condition-monitoring + predictive-maintenance platform centered on vibration**, but I would *not* rely on vibration alone. Fans, pumps, compressors, and motors have different failure signatures, and combining vibration with temperature, motor/load, and BMS data…
Absolutely. For industrial HVAC, I’d design this as a condition-monitoring + predictive-maintenance platform centered on vibration, but I would not rely on vibration alone. Fans, pumps, compressors, and motors have different failure signatures, and combining vibration with temperature, motor/load, and BMS data makes the predictions much more useful. Current HVAC condition-monitoring systems are taking this edge-sensor + analytics approach as well.
HVAC EQUIPMENT
┌─────────────────────────────────────────┐
│ Chillers │ AHUs │ Pumps │ Fans │ Motors │
└─────┬────────────┬─────────────┬────────┘
│ │ │
Vibration Temp Electrical
3-axis bearing current/power
│ │ │
└────────────┴─────────────┘
│
┌──────▼──────┐
│ Edge Gateway │
│ FFT / feature│
│ extraction │
└──────┬──────┘
│ MQTT / OPC-UA
▼
┌───────────────────┐
│ Analytics Platform│
│ │
│ • Baseline model │
│ • Anomaly score │
│ • Fault diagnosis │
│ • RUL prediction │
└─────────┬─────────┘
│
┌────────────┼─────────────┐
▼ ▼ ▼
Dashboard Alerts CMMS/EAM
Asset health Email/SMS Work order
Trends Teams/etc. generation
For each critical rotating asset, install a 3-axis accelerometer on the bearing/motor housing.
I'd monitor:
A 3-axis sensor is particularly useful because it captures axial, horizontal and vertical vibration from one mounting position. Industrial vibration sensors already support HVAC applications and can calculate indicators such as imbalance and bearing condition.
I'd add:
That gives you mechanical + electrical + process context rather than asking vibration to explain everything.
Don't send only a single vibration number to the cloud.
The edge device should calculate features such as:
Time-domain
Frequency-domain
Overall vibration is useful for detecting deterioration, while spectral information helps distinguish why the vibration changed. ISO 20816 provides the general framework for measuring and evaluating machine vibration, with specific parts covering different machine classes.
The system should classify likely failure modes rather than simply saying:
"Vibration HIGH." For example:
| Pattern | Likely problem | Recommended action |
|---|---|---|
| High 1× RPM | Imbalance | Inspect/balance fan |
| Strong 2× RPM + axial vibration | Misalignment | Check coupling/alignment |
| Multiple harmonics | Mechanical looseness | Inspect mounts/fasteners |
| Increasing bearing-frequency energy | Bearing degradation | Plan bearing replacement |
| High-frequency vibration + temperature | Lubrication/bearing problem | Inspect bearing/lubrication |
| Gear-mesh changes | Gearbox wear | Inspect gearbox |
| Vibration + rising motor current | Mechanical overload | Inspect driven equipment |
| Vibration increases only at certain VFD speeds | Resonance | Investigate operating speed |
These shouldn't be hard-coded as absolute rules only. The system should learn what healthy looks like for each individual asset.
I'd use a hybrid approach, rather than immediately trying to train a giant deep-learning model.
For each asset:
Asset #AHU-023
Normal operating conditions:
RPM: 1,480
Load: 72%
Vibration: 1.7 mm/s RMS
Temperature: 48°C
Healthy signature:
FFT spectrum
Bearing features
Operating-state information
The platform establishes a baseline over normal operation.
Calculate something like:
Health Score = 100 ─ anomaly_penalty
AHU-023
Health: 82/100
Trend: ↓
Risk: Medium
The anomaly model should account for operating state. A vibration level at 1,800 RPM shouldn't necessarily be compared directly with the same number at 900 RPM. Correlating vibration with motor speed is particularly important for variable-speed HVAC equipment.
Once an anomaly is detected:
Anomaly
↓
FFT / envelope analysis
↓
Operating conditions
↓
Historical behavior
↓
Fault classifier
↓
"Probable fan bearing degradation — 87%"
Once you have sufficient historical failure/repair data, add RUL:
FAN-017
Bearing degradation: HIGH
Probability of failure:
7 days: 3%
30 days: 18%
60 days: 54%
Estimated maintenance window:
3–6 weeks
Recommendation:
Inspect bearing during next scheduled shutdown
I would not promise a precise "failure in 23 days" prediction unless you have a substantial historical dataset. Initially, anomaly detection + degradation trending will generally be more credible.
The most important part is what happens after the AI detects something.
For example:
09:12
Fan vibration +32% above baseline
↓
09:13
AI identifies bearing-frequency increase
↓
09:15
Health score falls 91 → 64
↓
09:16
Maintenance alert
"AHU-14 supply fan bearing degradation suspected.
Inspect within 7 days."
↓
Technician inspection
↓
Bearing confirmed degraded
↓
CMMS work order
↓
Bearing replaced during planned shutdown
↓
New vibration baseline established
This closes the loop between sensor data and actual maintenance. That's crucial: predictive maintenance isn't particularly valuable if it just creates another dashboard full of alerts.
A practical implementation could look like:
Sensors
Edge
Data platform
ML
Integration
This architecture also avoids forcing you to replace the existing BMS. Commercial HVAC predictive-maintenance implementations increasingly combine vibration with BMS/process information rather than treating vibration as an isolated data source.
I'd start with 20–50 critical assets, rather than instrumenting an entire facility.
For example:
For each asset:
2 vibration sensors + temperature would be a reasonable starting point, with additional sensors where the mechanical configuration warrants them.
Then build:
After 3–6 months of operation, you can evaluate which faults you're actually detecting and determine whether you have enough labeled data to build more sophisticated failure/RUL models.
Don't make the product "an AI vibration sensor."
Make it:
"An HVAC asset-health system that uses vibration as its primary early-warning signal and combines it with operating context to tell maintenance teams what is failing, how urgent it is, and what they should do next." That's a much more defensible industrial solution.
If you're building this as an actual product, I can next lay out the complete technical design—sensor specifications, sampling rates, edge hardware, MQTT/OPC-UA data model, database schema, ML algorithms, dashboard, alert logic, and an estimated BOM/cost per HVAC asset.
For each asset:
2 vibration sensors + temperature would be a reasonable starting point, with additional sensors where the mechanical configuration warrants them.
Then build:
After 3–6 months of operation, you can evaluate which faults you're actually detecting and determine whether you have enough labeled data to build more sophisticated failure/RUL models.
Don't make the product "an AI vibration sensor."
Building a **predictive maintenance (PdM)** system for industrial HVAC using vibration sensors requires a solid data pipeline—from edge hardware to cloud analytics. Here is a blueprint to set up a comprehensive solution. 1. Hardware & Edge Layer (Data Collection) - Triaxial Accelerometers: Mount low-noise,…
Building a predictive maintenance (PdM) system for industrial HVAC using vibration sensors requires a solid data pipeline—from edge hardware to cloud analytics. Here is a blueprint to set up a comprehensive solution.
To tailor this architecture specifically for your project, let me know:
No sources cited
Yes. For industrial HVAC, I’d build this as a **condition-monitoring + predictive-maintenance platform centered on vibration**, but I would not rely on vibration alone. ASHRAE specifically recommends combining measurements such as vibration, temperature, load, pressure, and other operating data for complex HVAC…
Yes. For industrial HVAC, I’d build this as a condition-monitoring + predictive-maintenance platform centered on vibration, but I would not rely on vibration alone. ASHRAE specifically recommends combining measurements such as vibration, temperature, load, pressure, and other operating data for complex HVAC equipment.
Prioritize equipment where mechanical failure causes significant downtime:
For each critical motor/fan/pump, install 3-axis vibration sensors on or near the bearing housings, plus temperature sensing where practical.
The vibration system should measure:
ISO 20816 provides a useful framework for vibration measurement and evaluation of rotating machinery, including both vibration magnitude and changes in vibration over time.
A practical architecture is:
HVAC EQUIPMENT
┌─────────────────────┐
│ Motor / Fan / Pump │
│ Compressor / Gearbox│
└──────────┬──────────┘
│
3-axis vibration
+ temperature
+ RPM
│
┌──────▼──────┐
│ Edge Sensor │
└──────┬──────┘
│
LoRaWAN / Wi-Fi
Ethernet / Cellular
│
┌──────▼──────┐
│ Edge Gateway│
└──────┬──────┘
│
MQTT / HTTPS
│
┌─────────▼──────────┐
│ Analytics Platform │
│ │
│ Baseline + ML │
│ Fault detection │
│ RUL prediction │
└─────────┬──────────┘
│
┌──────▼───────┐
│ Maintenance │
│ Dashboard │
└──────┬───────┘
│
CMMS / Work Order
This is an important design decision.
For each asset, establish a healthy operating baseline. Then detect deviations from that asset's normal behavior.
For example:
Normal
│
│ vibration
│ ╱
│ ╱
│ ╱
│_______╱________________ Time
↑
degradation
Rather than saying:
"The fan will fail in 23 days." initially say:
"Fan AHU-17 has an 87% probability of abnormal mechanical behavior. Primary indicators: 1× rotational frequency increased 42%, bearing-envelope energy increased 3.1×, and temperature increased 8°C." That is much easier for a maintenance technician to trust and act upon.
ASHRAE describes predictive maintenance in essentially this way: compare current measurements with benchmarks/historical performance, trend degradation, and schedule corrective work before unacceptable performance or failure occurs.
| Fault | Typical vibration signature |
|---|---|
| Bearing degradation | Increasing high-frequency/envelope energy, kurtosis |
| Imbalance | Strong 1× rotational frequency |
| Misalignment | 1×/2× components, axial vibration |
| Mechanical looseness | Harmonics, broadband vibration |
| Belt problems | Rotational/belt-frequency components |
| Fan blade problems | Fan/rotational harmonics |
| Gear wear | Gear-mesh frequencies + sidebands |
| Cavitation | Broadband/high-frequency vibration |
| Resonance | Large amplitude at particular operating speeds |
| Structural problems | Changes in multiple sensor locations |
The really useful part is frequency-domain analysis. A raw RMS vibration value can tell you that something is getting worse; the spectrum can help tell you what is getting worse.
This is where the solution becomes substantially better than a generic vibration-monitoring product.
Pull operating information from the building/industrial control system:
Then the model can distinguish:
"High vibration because the machine is running at 100% load"
from
"High vibration despite normal load and speed."
ASHRAE notes that complex HVAC systems may require several conditions—including vibration and load—to properly assess equipment condition.
I'd use a layered approach rather than jumping directly to a deep-learning model.
Immediate detection of obvious conditions:
IF vibration > equipment limit
→ Critical alarm
IF vibration increases > X% over baseline
→ Warning
IF temperature + vibration both increase
→ Bearing investigation
Extract features from the raw accelerometer signal:
RMS
Peak
Peak-to-peak
Crest factor
Kurtosis
Skewness
FFT amplitudes
1× amplitude
2× amplitude
Bearing-envelope features
Spectral entropy
Band energy
Train an individual model—or asset-class model—on healthy operation.
Good starting algorithms include:
The output could be:
Asset Health Score: 72/100
Mechanical anomaly: HIGH
Bearing anomaly: MEDIUM
Temperature anomaly: LOW
Trend: deteriorating
Confidence: 91%
Recommended inspection:
Bearing / coupling / alignment
Estimated maintenance window:
7–21 days
Once you accumulate enough historical failure and maintenance data, introduce RUL prediction.
For example:
Failure
X
/
/
predicted/
/
/
-------------/----------------
↑
current state
← maintenance window →
Don't make RUL the centerpiece on day one. You need enough representative failure/degradation data to make those predictions credible.
The final product should not stop at an alarm.
A useful workflow is:
Sensor → anomaly → diagnosis → recommended action → work order → repair → verification
For example:
AHU-04 supply fan
Health: 🔴 38/100 Anomaly: Bearing degradation Evidence: High-frequency vibration + 1.8× increase in envelope RMS Trend: Increasing for 18 days Recommended action: Inspect drive-end bearing and lubrication Priority: High Suggested maintenance window: Next scheduled shutdown After the technician repairs it, the system should automatically check whether vibration returned to baseline.
ASHRAE specifically notes that condition monitoring can also be used after repair to verify that the underlying fault has actually been removed.
I'd make the primary screen asset-centric, not sensor-centric:
INDUSTRIAL HVAC HEALTH
─────────────────────────────────────────────
🔴 3 Critical
🟠 11 Warning
🟢 184 Healthy
Asset Health Problem Trend
─────────────────────────────────────────────
Chiller-01 91 None →
AHU-17 43 Bearing ↗
Pump-04 68 Misalignment ↗
Tower-Fan-03 31 Imbalance ↗
AHU-21 96 None →
Clicking an asset should expose:
I would not instrument an entire industrial facility initially.
Start with perhaps 20–50 critical rotating assets representing several equipment types.
Run the pilot for approximately 3–6 months, with the objective of establishing:
Then expand to the rest of the facility.
The key metric isn't "AI accuracy." It's:
How many actionable failures did we identify early enough to change the maintenance outcome?
A robust implementation could look like:
Sensors
↓
LoRaWAN / Wi-Fi / Ethernet
↓
Industrial Edge Gateway
↓
MQTT
↓
Time-series database
↓
Feature extraction / FFT
↓
ML anomaly detection
↓
Asset health engine
↓
REST API
├── Web dashboard
├── Email/SMS/Teams alerts
└── CMMS integration
For an industrial deployment, I'd also make cybersecurity, sensor calibration, timestamp synchronization, offline buffering, and role-based access first-class requirements rather than afterthoughts.
Use ISO 20816 as part of the vibration-evaluation framework rather than inventing arbitrary vibration limits. ISO For the HVAC operational side, ASHRAE's guidance supports continuous monitoring, anomaly detection, predictive maintenance, and integration with commissioning/operational baselines.
If you're building this as a commercial product, I can take this one step further and design the actual solution: sensor specifications → data schema → MQTT topics → ML architecture → database schema → dashboard screens → alert logic → CMMS integration → estimated BOM/cost per HVAC asset.
Then the model can distinguish:
"High vibration because the machine is running at 100% load"
from
"High vibration despite normal load and speed."
ASHRAE notes that complex HVAC systems may require several conditions—including vibration and load—to properly assess equipment condition.
Implementing a **predictive maintenance (PdM)** architecture for industrial HVAC systems using vibration sensors involves a structured pipeline: sensor selection, data acquisition, edge/cloud analytics, and integration with your Computerized Maintenance Management System (CMMS). 1. Sensor Selection & Edge Hardware -…
Implementing a predictive maintenance (PdM) architecture for industrial HVAC systems using vibration sensors involves a structured pipeline: sensor selection, data acquisition, edge/cloud analytics, and integration with your Computerized Maintenance Management System (CMMS).
To tailor this architecture specifically for your site, tell me:
Yes. A strong **industrial HVAC predictive-maintenance solution based on vibration sensing** should combine continuous vibration monitoring, signal processing, machine-learning anomaly detection, and maintenance workflow integration. ## Recommended solution architecture ```
Yes. A strong industrial HVAC predictive-maintenance solution based on vibration sensing should combine continuous vibration monitoring, signal processing, machine-learning anomaly detection, and maintenance workflow integration.
┌──────────────────── HVAC ASSETS ────────────────────┐
│ │
│ AHU fans Chillers Pumps Compressors │
│ │ │ │ │ │
│ Accelerometer Accelerometer Accelerometer ... │
└─────┬────────────┬────────────┬─────────────┬──────┘
│ │ │ │
└────────────┴──────┬─────┴─────────────┘
▼
Edge Gateway / PLC
│
FFT + filtering + features
│
MQTT / OPC UA
▼
┌──────────────────┐
│ PdM Data Platform│
│ │
│ • Time series DB │
│ • Asset history │
│ • ML models │
│ • CMMS interface │
└────────┬─────────┘
▼
Health Score / Alerts
│
┌──────────┴──────────┐
▼ ▼
Maintenance UI CMMS / Work Order
Start with the rotating equipment, rather than trying to monitor every HVAC component.
| Asset | Sensor locations | Faults to target |
|---|---|---|
| Supply/exhaust fan | Motor DE/NDE, fan bearing | Imbalance, misalignment, bearing wear, looseness |
| Chilled-water pump | Motor + pump bearings | Bearing wear, cavitation, imbalance, misalignment |
| Condenser-water pump | Motor + pump bearings | Bearing degradation, imbalance |
| Compressor | Motor/compressor housing | Bearing/mechanical faults, imbalance |
| Cooling-tower fan | Motor + gearbox/bearing | Gear wear, imbalance, bearing faults |
| AHU fan | Motor + fan bearings | Belt problems, imbalance, misalignment |
Vibration is particularly valuable because mechanical faults such as imbalance, misalignment and bearing degradation produce characteristic changes in the vibration signal. Recent HVAC research has specifically demonstrated the use of vibration accelerometers on pump and fan bearings for fault diagnosis and RUL prediction.
For industrial equipment, I'd use industrial IEPE/ICP accelerometers for wired installations or rugged wireless triaxial accelerometers where cabling is difficult.
A practical starting specification:
You don't necessarily need continuous high-frequency streaming. A cost-effective architecture can periodically capture high-resolution vibration "snapshots" while continuously tracking lower-bandwidth health indicators.
Don't send only raw acceleration to the cloud.
For each measurement window, calculate:
Time-domain
Frequency-domain
Envelope analysis
Also correlate vibration with:
This context is important: a fan operating at a different speed can legitimately have a different vibration signature.
ISO 20816 provides the general framework for measuring and evaluating machine vibration, including both vibration magnitude and changes in vibration over time.
I would not start with a complicated deep-learning model.
Establish a healthy baseline for every asset.
For example:
Asset 1042
Fan RPM: 1,780
Vibration RMS: 2.1 mm/s
Bearing temperature: 52°C
Health score: 94/100
The system learns that asset's normal behavior and flags deviations.
Good initial algorithms:
This is useful because industrial HVAC installations often don't have enough historical failure examples to train a supervised model.
Once you accumulate labeled maintenance events, train a classifier to distinguish:
Normal
↓
Abnormal
↓
Possible imbalance
Possible misalignment
Possible bearing degradation
Possible looseness
Possible belt problem
Random Forest/XGBoost is an excellent initial choice because it works well with engineered vibration features and is comparatively explainable. Research on industrial ventilation PdM has also evaluated FFT, wavelet and cepstrum features alongside ML models, with Random Forest performing strongly for fault detection.
Later, you can move to 1-D CNN/LSTM/transformer models if you have enough labeled vibration data. Real production HVAC research has demonstrated deep-learning fault classification using vibration and temperature sensors.
Don't give the maintenance technician:
"Anomaly score = 0.87." Give them:
AHU-07 Supply Fan — WARNING Bearing vibration increased 38% over baseline during the last 14 days. Dominant frequency indicates probable bearing degradation. Estimated intervention window: 2–4 weeks. Recommended action: Inspect fan bearings and lubrication at next planned shutdown. That's the difference between an analytics project and a useful PdM system.
I'd use a health score such as:
HEALTH
92
┌──────────┐
│ GOOD │
└──────────┘
Vibration █████████░ 91
Bearing temp ██████████ 96
Motor current █████████░ 89
Trend ████████░░ 82
Then have thresholds such as:
Importantly, thresholds should combine absolute vibration severity and rate of change, rather than relying on one universal vibration number.
For a new deployment, I'd consider:
Sensors
Edge
Data
Analytics
Dashboard
Integration
Integrating IoT sensor data with facility-management systems and maintenance records is particularly important because the maintenance history provides the labels needed to improve the predictive models.
I'd build the first version around 10–20 critical assets, rather than instrumenting the entire plant.
For example:
Phase 1 — 4–6 weeks
Phase 2 — 4–8 weeks
Phase 3 — 2–6 months
Phase 4
A recent 2026 study of HVAC predictive maintenance similarly combines IoT sensing, maintenance records, ML, and a digital-twin/decision-support layer rather than treating vibration prediction as an isolated sensor problem.
Don't build "a vibration sensor system." Build an asset-health system.
The sensor is only the first layer:
Vibration → features → baseline → anomaly → fault diagnosis → degradation trend → maintenance recommendation → work order → technician feedback → improved model
If you tell me the HVAC equipment you want to monitor (e.g. AHU fans, chillers, pumps, compressors), number of assets, and whether you want a wired or wireless deployment, I can turn this into a concrete hardware + cloud/edge architecture, sensor specification, data schema, ML pipeline, and estimated BOM/cost.
Yes. A good industrial HVAC predictive-maintenance system can be built around **vibration sensing + operating context + anomaly/fault analytics**, rather than relying on vibration thresholds alone. ## Recommended solution architecture ```
Yes. A good industrial HVAC predictive-maintenance system can be built around vibration sensing + operating context + anomaly/fault analytics, rather than relying on vibration thresholds alone.
Vibration Sensors
│
├── Motor bearings
├── Fan bearings
├── Pump bearings
└── Compressor bearings
│
▼
Edge Gateway / DAQ
│
├── Filtering
├── FFT / spectral analysis
├── RMS / peak / crest factor
└── Feature extraction
│
▼
Predictive Analytics
│
├── Asset-specific baseline
├── Anomaly detection
├── Fault classification
└── Remaining-useful-life estimate
│
▼
Dashboard / CMMS
│
├── Health score
├── Fault type
├── Severity
├── Trend
└── Recommended maintenance
Prioritize rotating equipment:
Put sensors on or near the drive-end and non-drive-end bearing housings where practical. ISO 20816 explicitly covers vibration measurements on rotating machinery and bearing housings and provides a framework for evaluating vibration magnitude and changes over time.
For industrial deployments, I'd generally use industrial accelerometers rather than inexpensive consumer-grade MEMS devices. Sensor selection should consider frequency response, dynamic range, temperature range, mounting, and noise. TE's guidance specifically identifies these as important parameters for vibration-based condition monitoring and discusses HVAC applications.
The valuable information is in the frequency spectrum and how it changes.
For each measurement, calculate features such as:
This lets the system distinguish, for example, fan imbalance from bearing deterioration rather than simply saying "vibration is high."
This is particularly important for HVAC.
A fan operating at 100% speed will naturally behave differently from one operating at 40% speed. The analytics should therefore ingest:
This allows you to build an asset-specific baseline conditioned on operating state.
I would avoid jumping immediately to a deep-learning model.
Stage 1 — anomaly detection
Learn what "healthy" looks like for each individual asset.
For example:
Fan AHU-17 normally has 2.1–2.8 mm/s RMS velocity at 45 Hz. If it gradually moves:
2.4 → 2.7 → 3.1 → 3.8 → 4.5 mm/s the system should recognize the trend before a fixed alarm threshold is reached.
NIST similarly emphasizes monitoring, diagnostics and prognostics as separate capabilities, with the goal of identifying when equipment performance will cross a threshold rather than merely detecting a failure after it occurs.
Stage 2 — fault classification
Classify the anomaly into likely causes:
| Signature | Likely problem |
|---|---|
| Strong 1× RPM | Fan/wheel imbalance |
| Strong 2× RPM + axial vibration | Misalignment |
| Multiple harmonics | Mechanical looseness |
| High-frequency/envelope energy | Bearing degradation |
| Blade-pass frequency | Fan/blade problem |
| Broadband/high-frequency vibration | Possible bearing/lubrication issue |
| Sub-synchronous vibration | Potential hydraulic/cavitation issue |
These should be treated as diagnostic indicators, not absolute rules; machine configuration and operating conditions matter.
The technician shouldn't receive:
"Anomaly score = 0.87." They should receive something like:
AHU-17 — HIGH
Supply fan bearing vibration has increased 42% over 21 days. High-frequency envelope energy is increasing while operating speed remains stable. Pattern is consistent with developing bearing degradation.
Recommended action: Inspect supply-fan drive-end bearing during next planned shutdown. Check lubrication, bearing condition and shaft alignment.
Confidence: 91% Estimated intervention window: 2–4 weeks That is what turns condition monitoring into predictive maintenance.
For a new deployment:
Sensors
Edge
Data platform
Analytics
Recent research also supports an edge-AI approach: a 2026 study using IEPE vibration sensors found that lightweight classical ML models such as Random Forest and SVM could perform vibration fault classification with very low inference latency, making edge processing practical.
I would not make this a "one AI model for all HVAC equipment" system.
Instead:
Asset → operating state → healthy baseline → anomaly → diagnosis → maintenance recommendation
A 20-year-old belt-driven AHU fan, a centrifugal chiller compressor, and a chilled-water pump have very different vibration characteristics. The system should maintain separate models/baselines by equipment class and preferably by individual asset.
Also, use standards such as ISO 20816 as part of the engineering framework, rather than allowing an ML model to invent vibration limits. ISO 20816 provides guidance around vibration measurement, evaluation criteria and operational limits.
I'd start with 10–20 critical HVAC assets rather than instrumenting an entire facility.
For each asset:
The resulting system can eventually provide fleet-level health ranking, e.g. "12 assets require attention, 3 are deteriorating rapidly, and 1 should be inspected immediately."
If you're building this as an actual product/system, I can also lay out the complete solution design—including specific sensor specifications, edge hardware, MQTT/OPC-UA architecture, database schema, ML pipeline, dashboard, and an example Python implementation for the vibration analytics.
A 20-year-old belt-driven AHU fan, a centrifugal chiller compressor, and a chilled-water pump have very different vibration characteristics. The system should maintain separate models/baselines by equipment class and preferably by individual asset.
Also, use standards such as ISO 20816 as part of the engineering framework, rather than allowing an ML model to invent vibration limits. ISO 20816 provides guidance around vibration measurement, evaluation criteria and operational limits.
I'd start with 10–20 critical HVAC assets rather than instrumenting an entire facility.
For each asset:
"Anomaly score = 0.87." They should receive something like:
AHU-17 — HIGH
Supply fan bearing vibration has increased 42% over 21 days. High-frequency envelope energy is increasing while operating speed remains stable. Pattern is consistent with developing bearing degradation.
Recommended action: Inspect supply-fan drive-end bearing during next planned shutdown. Check lubrication, bearing condition and shaft alignment.
Confidence: 91% Estimated intervention window: 2–4 weeks That is what turns condition monitoring into predictive maintenance.
For a new deployment:
Sensors
Edge
Data platform
Analytics
Recent research also supports an edge-AI approach: a 2026 study using IEPE vibration sensors found that lightweight classical ML models such as Random Forest and SVM could perform vibration fault classification with very low inference latency, making edge processing practical.
I would not make this a "one AI model for all HVAC equipment" system.
Instead:
Asset → operating state → healthy baseline → anomaly → diagnosis → maintenance recommendation
Building a predictive maintenance (PdM) solution for industrial HVAC systems using vibration sensors involves a mix of hardware edge data collection, IoT communication protocols, and machine learning analytics to catch faults like unbalance, misalignment, bearing wear, and looseness before catastrophic failure…
Building a predictive maintenance (PdM) solution for industrial HVAC systems using vibration sensors involves a mix of hardware edge data collection, IoT communication protocols, and machine learning analytics to catch faults like unbalance, misalignment, bearing wear, and looseness before catastrophic failure occurs.
To help tailor this architecture, let me know:
