Signal Analysis for Condition Monitoring: Decoding Asset Health Beyond the Noise

1. THE REAL PROBLEM: We Are Drowning in Data and Starving for Wisdom

The fundamental promise of the Industrial Internet of Things (IIoT) and Industry 4.0 was that sensors would solve our reliability problems. We were told that if we just strapped enough accelerometers, thermocouples, and current clamps to our critical assets, the "data" would tell us when the machine was going to fail.

That promise has largely been broken.

The real problem in modern maintenance isn't a lack of data; it is a lack of signal clarity. Most organizations today are collecting terabytes of vibration readings, acoustic samples, and amperage logs, yet they still suffer from unplanned downtime. Why? Because they are confusing data collection with signal analysis.

Collecting data is passive; it is merely recording the symptoms. Signal analysis is active; it is the translation of those symptoms into a diagnosis.

The industry is currently suffering from "threshold fatigue." Maintenance teams set generic ISO alarm limits on overall vibration levels (RMS), get bombarded with alerts every time a pump cavitates for three seconds, and eventually tune out the system entirely. This is not condition monitoring; it is noise monitoring.

True success in this field does not look like a dashboard with a thousand green lights. It looks like a reliability engineer looking at a specific frequency peak in a spectrum and saying, "The outer race on the non-drive end bearing of Conveyor 4 has a spall, but we can run it for three more weeks at 80% load before we need to intervene."

That level of precision requires moving beyond simple trendlines and understanding the physics of the signal. It requires us to stop treating sensors as magic black boxes and start treating them as stethoscopes that require a trained ear to interpret. This guide is for the practitioners who are tired of generic predictive maintenance sales pitches and want to understand the mechanics of translating voltage into value.

Dive Deeper: For more on selecting the right hardware for your strategy, see our guide to Industrial Sensor Monitoring Systems.

2. FOUNDATIONAL CONCEPTS: The Physics of Failure

To master signal analysis, we must first dismantle the jargon that vendors use to obscure the simplicity of the underlying physics. At its core, signal analysis is the study of energy release. When a machine degrades—whether it’s a bearing pitting, a shaft unbalancing, or a stator winding shorting—it releases energy in specific patterns. Our job is to capture that energy and decode the pattern.

The Time Domain vs. The Frequency Domain

The most critical mental model for any reliability engineer is the relationship between the Time Waveform and the Fast Fourier Transform (FFT).

The Time Waveform (The "What"): This is the raw signal as it happens. If you look at a vibration signal in the time domain, you are seeing the complex summation of every force acting on the machine—imbalance, misalignment, gear mesh, bearing defects, and background noise—all jumbled together. It looks like a chaotic squiggly line.

• Why it matters: Many software solutions hide the time waveform, but it is essential for detecting transient events like impacting or looseness. An FFT averages data, which can smooth out the very "clicks" and "pops" that indicate early-stage failure.

The Frequency Domain / FFT (The "Who"): The Fast Fourier Transform is a mathematical algorithm that breaks the complex time waveform apart into its constituent frequencies.

• The Analogy: Imagine you are listening to a symphony orchestra. The Time Waveform is the combined sound of the music hitting your ear. The FFT is the sheet music that tells you exactly which note the violin is playing, which note the cello is playing, and how loud each is.
• The Application: If we know the rotational speed of a motor and the geometry of its bearings, we can calculate exactly what frequency a defect will generate. If the FFT shows a spike at that exact frequency, we know who is playing the bad note.

Resolution and Sampling: The Trap of "Good Enough"

A common failure mode in predictive maintenance implementations is poor sampling strategy. You cannot analyze what you do not capture.

• Fmax (Maximum Frequency): This is the bandwidth of your analysis. If you are looking for gear mesh faults in a high-speed compressor, but your sensor only captures up to 1,000 Hz, you are legally blind to the problem. You will see a flat line while the gearbox destroys itself.
• Lines of Resolution: This determines how detailed your spectrum is. Low resolution blurs adjacent frequencies together. If your motor speed (1x) and a bearing cage frequency are close together, low resolution will merge them into one lump, making diagnosis impossible.
• Aliasing: This occurs when the sampling rate is too slow for the frequency being measured, causing the signal to appear at a false, lower frequency (the "wagon wheel effect" where wheels look like they are spinning backward). Anti-aliasing filters are mandatory, yet often overlooked in cheap wireless sensors.

The Concept of Stationarity

Most standard signal analysis assumes the machine is running at a constant speed and load (stationary). However, in the real world of variable frequency drives (VFDs) and cyclical loads, machines are rarely stationary.

• The nuance: If a machine speeds up during the data collection window, the spectral peaks will "smear" across the graph, ruining the analysis. Advanced practitioners use Order Tracking, which normalizes the signal against the shaft speed, ensuring that the "1x" peak stays at "1x" regardless of RPM changes.

The Missing Dimension: Phase Analysis

While the FFT tells you what frequency is vibrating and how much (amplitude), it does not tell you how the machine is moving relative to itself. This is where Phase Analysis becomes the tie-breaker in difficult diagnoses.

Phase measures the delay between a reference point (like a tachometer pulse) and the peak vibration signal. It is measured in degrees (0° to 360°).

• The "Tie-Breaker" Scenario: You see a high vibration peak at 1x running speed. Is it imbalance? Is it misalignment? Is it a bent shaft? In a standard spectrum, these all look identical—a big spike at 1x.
• How Phase Solves It: By placing sensors on both the vertical and horizontal axes (or across the coupling), you can compare the phase.
  • If the vertical and horizontal phase readings differ by roughly 90°, it is likely imbalance.
  • If the phase readings across the coupling are 180° out of phase (moving in opposite directions), it is likely misalignment.
  • If the phase is unstable and shifting, you may be dealing with resonance or a loose footing. Without phase data, you are often just guessing at the root cause of a 1x vibration, leading to "corrective" actions (like balancing a misaligned machine) that solve nothing.

Dive Deeper: For more on advanced signal processing techniques, see our guide to MSB Signal Processing for Fault Diagnosis.

3. HOW IT ACTUALLY WORKS: From Sensor to Decision

Let’s walk through the technical reality of how signal analysis functions across different modalities. We will focus on the three pillars: Vibration, Motor Current, and Ultrasound.

Vibration Analysis: The Heavy Lifter

Vibration is the most common application of signal analysis because mechanical force is the primary destructive agent in rotating equipment.

1. Displacement, Velocity, and Acceleration:

   • Displacement (mils/microns): Measures how much the part actually moves. It is critical for low-speed assets (below 600 RPM) and sleeve bearings where the concern is physical clearance (e.g., the shaft hitting the housing).
   • Velocity (in/sec or mm/s): Measures how fast the part is moving. This is the standard for general machine health (10 Hz to 1,000 Hz) because it correlates best with fatigue. It captures imbalance, misalignment, and looseness.
   • Acceleration (g): Measures the rate of change of velocity. This is high-frequency energy. It is the only way to see early-stage bearing defects and gear mesh issues. If you are monitoring a gearbox solely with Velocity RMS, you are waiting for the catastrophic failure phase.

2. Envelope Analysis (Demodulation):

   • When a bearing has a defect, it doesn't just vibrate; it "rings" the bearing housing like a bell with every impact. This ringing happens at a very high natural frequency, but it repeats at the slow rate of the shaft rotation.
   • Envelope analysis filters out the low-frequency noise (imbalance, etc.), focuses on that high-frequency ringing, and then "demodulates" it to reveal the repetitive pattern of the impacts. It is the most powerful tool for early bearing fault detection.

3. The Four Stages of Bearing Failure: To apply these concepts practically, you must understand the chronological progression of a bearing failure in the signal data. This framework allows you to estimate Remaining Useful Life (RUL).

   • Stage 1 (The Invisible Phase): The defect is microscopic. There is no visible change in the standard velocity spectrum. However, Ultrasound or high-frequency acceleration (Spike Energy/PeakVue) will show a rise in the 20 kHz to 60 kHz range. Action: Note it, grease it, and trend it.
   • Stage 2 (The Ringing Phase): The defect begins to "ring" the bearing's natural frequency. You still won't see much in standard velocity, but Envelope Analysis will clearly show peaks at the specific bearing defect frequencies (BPFO, BPFI). Action: Plan for replacement at the next convenient shutdown.
   • Stage 3 (The Harmonics Phase): The damage is now physical and visible. Defect frequencies appear in the standard velocity spectrum. More importantly, you see harmonics (multiples) of the defect frequency and sidebands appear. Action: Order parts immediately. Monitor daily.
   • Stage 4 (The Chaos Phase): The bearing geometry is gone. The distinct spectral peaks disappear and are replaced by a raised "noise floor" (often called a "haystack"). Paradoxically, the specific defect amplitude might drop, but the overall RMS energy skyrockets. Action: Shut down immediately to prevent shaft damage or fire.

Motor Current Signature Analysis (MCSA)

While vibration looks at the mechanical shaking, MCSA looks at the electrical "heartbeat" of the motor. The rotor and stator act as a transducer; mechanical variances in the air gap induce current variances in the power cable.

• The Mechanism: By clamping a current transformer (CT) around one phase of the motor lead, we can analyze the AC waveform.
• Rotor Bar Health: If a rotor bar cracks, it changes the magnetic flux, creating sidebands around the line frequency (e.g., 50Hz or 60Hz). MCSA is the undisputed king of detecting broken rotor bars and eccentricity problems that vibration sensors often miss until it’s too late.
• The Context: MCSA is particularly valuable for predictive maintenance on pumps that are submerged or inaccessible, where you cannot physically mount a vibration sensor. The motor control center (MCC) becomes your window into the asset.

Acoustic Emission and Ultrasound

Vibration measures the motion of the mass; ultrasound measures the friction and stress waves.

• The "Pre-Failure" Domain: Before a bearing generates enough physical movement to trigger an accelerometer, it generates microscopic friction. This emits high-frequency ultrasound (20kHz - 100kHz).
• Signal Heterodyning: Since humans (and standard audio equipment) cannot hear 40kHz, ultrasound devices "heterodyne" the signal—shifting it down to the audible range while preserving the signal characteristics. This allows an analyst to "hear" the friction of a dry bearing or the turbulence of a leaking valve.

Dive Deeper: For specific applications in difficult environments, read our guide on Submersible Vibration Sensors for Wastewater or solving the Vertical Turbine Pump Resonance Puzzle.

4. IMPLEMENTATION APPROACHES: Strategy Over Technology

Knowing the physics is useless if the deployment strategy is flawed. There are three distinct approaches to implementing signal analysis, and the "best" one depends entirely on asset criticality and failure modes.

1. Route-Based Data Collection

This is the traditional approach. A technician walks around with a handheld analyzer once a month.

• Pros: High-resolution data. The technician uses their five senses (smell, sight, sound) alongside the sensor. It forces a human inspection.
• Cons: It is a snapshot in time. If the machine fails the day after the route, you miss it. It is labor-intensive and dangerous for assets in hazardous areas.
• Best For: Balance-of-plant assets, non-critical pumps, and equipment where failure development time (P-F interval) is measured in months.

2. Wireless "Traffic Light" Sensors

The market is flooded with cheap, coin-sized vibration sensors that send overall RMS values to the cloud.

• Pros: Cheap, easy to install, provides 24/7 trending.
• Cons: The Data Trap. Most of these sensors have low Fmax (bandwidth) and poor resolution. They are excellent at telling you a machine is already broken (lagging indicator) but poor at diagnosing why. They often lack the battery power to transmit full waveforms, sending only calculated averages.
• Best For: Asset management of non-critical assets where you just need to know if the machine is running or shaking violently.

3. Continuous Online Monitoring (Protection Systems)

Hard-wired, high-speed systems (like Bently Nevada racks) that feed directly into the PLC or SCADA.

• Pros: Real-time protection. Can trip the machine in milliseconds to prevent catastrophic destruction. Full waveform buffering.
• Cons: Extremely expensive ($5k-$10k per channel). Complex integration.
• Best For: Turbines, large compressors, and assets where failure costs millions of dollars or risks human life.

The Hybrid Strategy: The "Watchlist" Approach

The most mature organizations use a hybrid model. They use cheap wireless sensors to monitor trends on Tier 2 assets. When the wireless sensor shows a deviation in the trend, it triggers a work order for a Category 1 analyst to go out with a high-end handheld unit to perform a deep-dive signal analysis. This optimizes labor while ensuring diagnostic precision.

Dive Deeper: Choosing the right tech stack is critical. Explore the Latest Software Tools for Fault Detection and understand why Moving All Maintenance Data to the Cloud might be a mistake.

5. MEASURING WHAT MATTERS: Metrics That Don't Lie

If you report "Average Plant Vibration" to your plant manager, you will be ignored. If you report "Assets with <30 Days Remaining Useful Life," you will get budget. You must measure the right signal characteristics.

The Fallacy of RMS (Root Mean Square)

RMS is a measure of the total energy in the signal. It is the industry standard for alarms (ISO 10816), and it is dangerously misleading for early warning.

• Why: A bearing defect starts as a short, sharp spike. Because RMS averages the signal energy, these short spikes get washed out by the dominant lower frequencies (like shaft speed). By the time RMS increases significantly, the bearing is often already in late-stage failure.

Crest Factor and Kurtosis

To detect early faults, you need metrics that are sensitive to "spikiness," not just total energy.

• Crest Factor: The ratio of the peak value to the RMS value. A healthy bearing has a smooth hum (low crest factor). A damaged bearing has impacts (high peaks, low RMS), resulting in a high crest factor.
• Kurtosis: A statistical measure of the "tailedness" of the probability distribution. In plain English, it measures how many outliers exist in the data. Pure random noise has a Kurtosis of 3. As bearing impacts begin, Kurtosis rises to 4, 5, or higher.
• The Trade-off: As the bearing damage gets worse, the "spikes" become so frequent that they start to look like continuous noise. Paradoxically, Crest Factor and Kurtosis can actually drop right before catastrophic failure. This is why relying on a single metric is fatal.

The True KPI: The "Save"

The ultimate metric for signal analysis is not technical; it is economic. You must track "Avoided Cost."

• When signal analysis identifies a gearbox defect 3 months early, allowing you to order parts via standard shipping (vs. air freight) and schedule the repair during a planned outage (vs. breaking the schedule), you must calculate that delta.
• Use an ROI calculator to quantify the difference between the reactive repair cost and the planned repair cost. This is the only signal the CFO cares about.

Illustrative Scenario: Anatomy of a Save

To illustrate this, consider a representative scenario involving a critical cooling tower fan.

• The Signal: A wireless vibration sensor flags a slight increase in overall vibration (from 0.15 to 0.22 in/sec). A generic alarm would ignore this as "acceptable." However, the reliability engineer performs a spectrum analysis and notices a low-amplitude peak at the Gear Mesh Frequency (GMF) of the gearbox.
• The Diagnosis: Crucially, the GMF peak has "sidebands" spaced at the speed of the input shaft. This specific pattern—sidebands around the mesh frequency—is the fingerprint of a misaligned or worn pinion gear.
• The Decision: The team estimates several weeks of remaining life. They order a replacement gearbox and schedule a crane for a planned outage.
• The ROI Principle:
  • Reactive Scenario: If the gearbox seizes, the fan could shatter, destroying the stack. Emergency costs including rental equipment, rush shipping, and unplanned production throttle can easily reach six figures.
  • Planned Scenario: Parts, crane rental, and labor at standard rates typically represent a fraction of emergency costs.
  • The Save: The cost difference between planned and reactive maintenance on critical rotating equipment often represents 5-10x savings. This type of analysis demonstrates the power of translating signal to value. Individual results vary based on equipment criticality and facility operations.

Dive Deeper: To better understand the financial impact of reliability, learn How to Reduce Downtime in Manufacturing or explore our specific guide on Predictive Maintenance for Spiral Conveyor Belts.

6. COMMON MISTAKES AND HARD TRUTHS

The gap between the theory of signal analysis and the reality of the shop floor is where most programs die.

Mistake 1: The "Black Box" AI Delusion

Vendors will tell you that their AI predictive maintenance tools can ingest raw data and output perfect work orders without human intervention.

• The Truth: AI is incredible at anomaly detection (spotting that a pattern has changed). It is currently mediocre at precise diagnostics (telling you it's a loose fit on the inner race). AI requires training data. If your plant hasn't experienced a specific failure mode on a specific machine type while recording data, the AI doesn't know what it looks like. You still need a human analyst to validate the AI's findings. For a realistic view, read about Data Science in Manufacturing.

Mistake 2: Sensor Mounting is Non-Negotiable

You can have a $50,000 analyzer, but if you mount the accelerometer using a magnet on a painted motor fan shroud, you are wasting your time.

• The Physics: Every interface absorbs high-frequency energy. A magnet mount acts as a low-pass filter, often cutting off data above 2,000 Hz. A hand-held probe is even worse.
• The Fix: For critical signal analysis, you must use stud-mounted sensors on bare metal spots prepared with a spot-facing tool. If you can't stud mount, use high-strength epoxy on a flat surface.

Mistake 3: Ignoring Process Data

Signal analysis does not exist in a vacuum. A pump vibrating at 0.5 in/sec might be in critical danger, or it might just be cavitating because operations throttled a valve.

• The Context: If you analyze vibration without looking at process parameters (pressure, flow, temperature, load), you will make false diagnoses. You cannot diagnose a machine if you don't know how it is being operated. This is why Root Cause Analysis must go beyond just the machine itself.

Mistake 4: Analysis Paralysis

Engineers love data. It is easy to fall into the trap of spending hours analyzing a spectrum to determine if a defect is on the inner race or outer race.

• The Hard Truth: Does it matter? If the bearing is failing, the work order is "Replace Bearing." Whether it's the inner or outer race rarely changes the corrective action. Don't let the pursuit of academic perfection delay the release of the PM procedure.

Mistake 5: Neglecting Signal Integrity and Cabling

A frequently overlooked source of "ghost" signals is the cabling itself. In industrial environments, electromagnetic interference (EMI) is rampant.

• Ground Loops: If your sensor is grounded at the machine and your analyzer is grounded at the panel, and there is a voltage potential difference between them, you will see a massive 50Hz or 60Hz "hump" in your spectrum. This is not the machine vibrating; it is electrical noise.
• Triboelectric Noise: In portable analysis, cable movement can generate internal static charges (triboelectric effect) that the analyzer interprets as low-frequency vibration. This often looks like a "ski slope" at the beginning of the spectrum (the "thermal transient" look).
• The Fix: Use twisted-pair, shielded cables. Ensure the shield is grounded at one end only (usually the instrumentation side) to prevent ground loops. Treat your cables as precision instruments, not extension cords.

7. GETTING STARTED WITHOUT GETTING OVERWHELMED

If you are starting from zero, or restarting a failed program, do not go out and buy software. Follow this sequence:

1. Criticality Analysis First: Rank your assets. You only have the resources to perform deep signal analysis on the top 10-20% of your assets. Everything else should be on a "run-to-failure" or generic PM schedule. Learn how to Match the Strategy to the Asset.
2. Establish Baselines: You cannot know if a signal is "bad" if you don't know what "good" looks like. Capture full spectrum baselines on your critical assets when they are running under optimal load.
3. Start with Pilot "Bad Actors": Pick 5 machines that give you the most headaches. Implement rigorous signal analysis (vibration and MCSA) on just these 5. Solve their chronic issues.
4. Build the Business Case: Document the savings from those 5 machines. Use those wins to justify the budget for scaling the program.
5. Integrate with Workflow: The best analysis is useless if it sits in a PDF report. The output of your signal analysis must be a work order in your CMMS. If the data doesn't trigger action, it is waste.

A Note on Certification and Training

Finally, do not underestimate the learning curve. Buying a violin does not make you a violinist, and buying a vibration analyzer does not make you a reliability engineer.

• ISO 18436 Standards: Look for training that aligns with ISO 18436-2 for vibration analysis.
  • Category I: Understands data collection and basic alarms.
  • Category II: Can diagnose common faults (imbalance, misalignment, bearings) and set up databases.
  • Category III/IV: Advanced diagnostics, resonance testing, and modal analysis.
• Mentorship: Signal analysis is as much art as science. If possible, pair junior analysts with external consultants or senior mentors to review complex spectra. The cost of training is always lower than the cost of a misdiagnosed critical asset.

Signal analysis is a language. It takes time to learn, but once you speak it, the machines will tell you exactly what they need. Stop guessing, stop reacting, and start listening.

Dive Deeper: Ready to advance your program? Review our Predictive Asset Management Maturity Model or use our Checklist Generator to standardize your inspections.

Related Guides

• Root Cause Analysis Strategy: Why Fixing the Machine is Not Enough
• Data Science in Manufacturing: Beyond the Hype to ROI
• Outcome Driven Maintenance: Why 99% Uptime is Meaningless
• Industrial Data Gravity: Why Your ERP Cannot Be the Single Source of Truth
• Arc Flash Assessment: From Engineering Theory to Compliance