Category: Uncategorized

Axle bearing failure on rolling stock is one of the most consequential failure modes in rail operations. A seized axle bearing at speed generates enough heat to cause axle fracture, and a fractured axle can lead to derailment. Even short of fracture, severe bearing degradation forces emergency stops, disrupts service schedules, and triggers regulatory investigations. The challenge is that axle bearings — tapered roller or cylindrical roller designs in most applications — operate under high radial loads, significant speed variation, and exposure to contamination that accelerates raceway wear.

The Historical Approach: Wayside Hot Box Detectors

The traditional tool for catching failing axle bearings is the hot box detector (HBD), also called a journal temperature detector. Wayside HBDs use infrared sensors aimed at passing axle journals as trains move through detection zones at line speed. A bearing that is generating excessive friction heat will show an elevated journal temperature, triggering an alarm.

Hot box detection has real value — it has prevented numerous derailments since its widespread deployment in the mid-20th century — but it has fundamental limitations that continuous onboard monitoring addresses:

• Latency: A bearing must be actively overheating as the train passes the detector. Early-stage defects that have not yet reached the thermal threshold pass undetected.
• Spatial gaps: HBD installations are spaced many miles apart. A bearing can fail catastrophically in the gap between detectors.
• No trending: Wayside detectors provide a pass/fail reading at a point in time, not a progressive picture of bearing health. There is no early warning before the alarm threshold is crossed.
• Temperature lag: Heat generation in the bearing must conduct through the axle and journal housing to reach the external surface. There is significant thermal lag between internal bearing temperature and the external temperature the HBD reads.

Continuous Onboard Vibration Monitoring

Accelerometers mounted on axle bearing housings provide a fundamentally different type of information than thermal detectors. Vibration-based monitoring detects the mechanical signatures of bearing damage — the repeating impulses at ball pass frequencies (BPFO, BPFI) — before thermal effects become significant. A surface fatigue spall on an outer race will generate detectable BPFO impulses from its earliest formation, often weeks or months before the bearing generates measurable excess heat.

The specific defect frequencies for railway axle bearings depend on bearing geometry and axle rotational speed. At typical freight speeds (60–90 km/h) with a 920mm wheel diameter (approximately 0.92m circumference), the axle rotates at roughly 5–8 Hz. BPFO for a typical axle bearing with 15 rollers and a geometry factor of approximately 0.4 works out to:

BPFO ≈ 0.5 × 15 × (1 − 0.4) × 7 Hz ≈ 31.5 Hz

This is well within the detection range of standard accelerometers, but it requires that the sensor can distinguish this frequency from the broadband vibration generated by wheel-rail contact, track joints, and vehicle suspension dynamics — a significant signal processing challenge at full line speed.

Health Scoring and Windowed Sampling

Continuous raw vibration storage at high sampling rates is impractical for battery-operated onboard sensors — the data volumes are too large and power consumption too high. Practical railway bearing monitoring systems use windowed sampling: short bursts of high-frequency data captured at regular intervals or triggered by specific events (speed thresholds, shock exceedance).

From each sample window, a set of condition indicators is computed on the sensor’s embedded processor: RMS acceleration, peak-to-RMS ratio (crest factor), kurtosis (sensitive to impulsive events), and where processing power allows, envelope spectrum peak levels at expected defect frequencies. These computed indicators are far smaller than raw waveforms and can be transmitted wirelessly without significant power penalty.

A health score — a normalized composite of these indicators over a rolling time window — provides the trending information that HBDs cannot. A bearing whose health score has declined consistently over 30 days is a very different situation from a bearing that shows a one-time anomalous reading, and a condition monitoring system can distinguish between them.

BLE Mesh and LTE Gateways for Remote Monitoring

Rolling stock presents a unique connectivity challenge: the asset moves continuously, operates in areas with variable cellular coverage, and may spend extended periods in yards or depots where direct maintenance access is available. A tiered wireless architecture addresses this:

• BLE (Bluetooth Low Energy) connects individual axle bearing sensors to an onboard concentrator. BLE operates at 2.4 GHz with low power draw and provides adequate range for intra-vehicle sensor networks. BLE Mesh extends single-hop BLE to multi-hop networks, allowing sensors throughout a long consist to communicate with a single gateway.
• LTE cellular connects the onboard gateway to cloud infrastructure. LTE coverage on main lines is generally adequate for periodic data uplink, and modern LTE modems with antenna diversity handle the handover between cell towers at rail speeds reliably.
• Yard WiFi provides high-bandwidth data transfer when the vehicle is stationary at a depot or inspection facility, enabling full waveform upload for detailed analysis.

The gateway aggregates data from all axle sensors, manages connectivity switching between LTE and WiFi, and provides local alarming if connectivity is lost. Critical alerts (bearing temperature or vibration exceeding thresholds) can be transmitted via SMS or satellite backup channel if the primary LTE link is unavailable.

When Bearing Failure Triggers Disputes

Railway bearing failures do not occur in a vacuum. A seized axle bearing may involve claims between the rolling stock operator, the bearing manufacturer, the maintenance provider, and the insurer. Each party has an interest in the failure data and a potential motivation to interpret it in their favor. Standard monitoring data — health scores, trend charts, alarm logs — is useful for operational decision-making but is not designed to serve as forensic evidence in a dispute.

For high-value liability situations, what is needed is a tamper-evident, cryptographically sealed record of the raw vibration signals at and around the failure moment. Systems designed for this purpose, like Fault Ledger, capture a fixed window of pre-event and post-event raw data, seal it on-device using multi-party cryptographic keys, and maintain an immutable chain-of-custody log. This transforms the bearing failure data from an operational record into forensic evidence that can withstand legal scrutiny.

The combination of continuous health trending for operational decisions and event-triggered forensic capture for dispute resolution provides a complete picture of axle bearing condition — from early fault detection through failure event documentation. As railway operators face increasing regulatory scrutiny and higher insurance claims after bearing-related incidents, this two-layer approach represents the state of the practice for serious axle bearing monitoring programs.

For operators evaluating Fault Ledger’s railway bearing monitoring solution, the key differentiator is not the health scoring capability (which many systems offer) but the forensic capture architecture that preserves the failure event data in a form that is defensible, neutral, and independently verifiable — without requiring the vendor to be trusted as the arbiter of what the data shows.

The shift from wired to wireless vibration sensors for bearing condition monitoring has been underway for over a decade, but the proliferation of wireless protocols has made sensor selection more complex, not less. BLE, LoRa, and LTE each offer distinct tradeoffs in range, power consumption, data bandwidth, latency, and infrastructure requirements. Choosing the wrong protocol for a deployment context can result in a system that is either chronically offline, battery-depleted in months rather than years, or incapable of transmitting enough data to be diagnostically useful.

The Core Tradeoffs

All wireless protocols for IoT sensors navigate the same fundamental tension: radio frequency communication consumes power proportional to transmit power and time on air, and higher data rates require either more bandwidth (spectrum) or shorter range. No protocol dominates all three dimensions simultaneously.

• BLE: Short range, low power, moderate data rate
• LoRa: Long range, very low power, very low data rate
• LTE: Medium range, high power, high data rate

BLE: Bluetooth Low Energy

BLE operates in the 2.4 GHz ISM band with 40 channels of 2 MHz bandwidth each. Typical transmit power ranges from 0 dBm to +8 dBm, with peak current draw of 5–15 mA during transmission. At maximum transmit power and with a clear line of sight, BLE 5.0 achieves ranges of 100+ meters; in steel machinery spaces with walls and obstructions, practical range drops to 10–30 meters per hop.

BLE’s primary advantage for bearing monitoring is its ecosystem and power profile. The nRF52 and nRF53 family of BLE SoCs (Nordic Semiconductor) include hardware accelerators for signal processing, enabling computation of condition indicators (RMS, kurtosis, FFT peaks) on-sensor before transmission. This dramatically reduces the data volume that must be transmitted — instead of sending raw waveforms, the sensor transmits only computed indicators.

BLE’s data rate of up to 2 Mbps (BLE 5.0 coded: 125 kbps to 2 Mbps depending on PHY selection) is adequate for streaming short waveform bursts when full waveform upload is required. The coded PHY option trades data rate for extended range at the same power level, reaching 4× the range of 1M PHY at the cost of 1/8th the throughput.

Best use cases: Close-range installations (marine machinery spaces, small industrial facilities), applications requiring multi-sensor networks within a single room or hull compartment, battery-operated sensors requiring multi-year life.

BLE Mesh: Extending Range for Multi-Hop Networks

BLE Mesh (Bluetooth Mesh Profile) extends BLE from point-to-point to multi-hop broadcast networks. Messages are flooded through the network using a managed flood model — each node retransmits received messages with a configurable time-to-live, effectively relaying data across distances that a single BLE hop cannot cover. A 50-node BLE Mesh network can span several hundred meters through multiple walls.

The tradeoff is increased power consumption (each node participates in relaying for others) and latency. BLE Mesh is not suitable for low-latency applications or for high-bandwidth waveform streaming across multiple hops. It is well suited for distributed condition indicator reporting — periodic transmission of computed health metrics from sensors throughout a large asset.

LoRa: Long Range

LoRa (Long Range) modulation, commercialized by Semtech, uses chirp spread spectrum (CSS) in sub-GHz frequency bands. In North America, LoRa devices operate in the 902–928 MHz ISM band (LoRaWAN AU915/US915 plan), which provides better building penetration than 2.4 GHz and longer free-space range at the same power level.

LoRa’s spreading factor (SF7 through SF12) controls the tradeoff between range and data rate. At SF12 and 125 kHz bandwidth, LoRa achieves a link budget of approximately 157 dB (versus ~100 dB for WiFi) with a data rate of only 250 bps. At SF7, data rate rises to ~5.5 kbps but link budget decreases to ~137 dB. In practice, most industrial LoRa deployments operate at SF9–SF10 for a balance of range and data rate.

LoRaWAN networks operate over public infrastructure (The Things Network, Helium, carrier LoRaWAN) or private gateways. A single LoRa gateway with an omnidirectional antenna can cover an entire industrial campus or a 2–5 km radius outdoors, making it extremely cost-effective for widely distributed asset monitoring.

The critical limitation for vibration monitoring is bandwidth: LoRa cannot transport raw vibration waveforms in any practical sense. A 2-second vibration sample at 25 kHz sampling = 50,000 samples × 2 bytes = 100 KB. At SF9 (1.7 kbps effective throughput), this would take nearly 10 minutes to transmit — and LoRaWAN duty cycle restrictions (1% in EU, varies in US) would prevent it entirely. LoRa for bearing monitoring is strictly a computed-indicator protocol: RMS, kurtosis, crest factor, temperature, and envelope spectrum peaks can be encoded in tens of bytes and transmitted in a single LoRa packet.

Best use cases: Industrial facilities with widely distributed assets (pumps, fans, conveyors across a large plant), outdoor installations, situations where gateway infrastructure cost must be minimized, and where computed indicators rather than raw waveforms are acceptable.

LTE: 4G/5G Cellular

LTE (and its successors LTE-M and NB-IoT for IoT applications) provides wide-area connectivity using licensed spectrum through cellular carriers. LTE-M (Cat-M1) targets IoT applications with peak data rates of 1 Mbps uplink and moderate power consumption; NB-IoT achieves even lower power but at reduced data rate (250 kbps uplink max).

For bearing monitoring, LTE-M is the relevant variant. Its 1 Mbps uplink rate means a 100 KB waveform can be transmitted in under one second. This enables raw waveform upload from remote or mobile assets without the data rate constraints of LoRa. Peak current draw for LTE-M transmission is 200–500 mA — one to two orders of magnitude higher than BLE or LoRa. Battery life for LTE-M sensors is measured in weeks to months at frequent sampling intervals, rather than years.

LTE’s geographic coverage makes it uniquely suitable for mobile or remote assets: vehicles, vessels, assets in locations outside WiFi or LoRa gateway coverage. A bearing sensor on a mobile generator, a rail car, or a remote pump station can report directly to cloud infrastructure without local gateway infrastructure.

Best use cases: Mobile assets (vehicles, rail, vessels), remote assets without local infrastructure, applications requiring raw waveform upload, and high-criticality assets where latency in receiving bearing data must be minimized.

Protocol Comparison Summary

Characteristic	BLE 5.0	BLE Mesh	LoRa (US915)	LTE-M
Frequency	2.4 GHz	2.4 GHz	902–928 MHz	Licensed band
Typical range	10–100 m	100–500 m (multi-hop)	1–10 km	Carrier coverage
Peak data rate	2 Mbps	<1 Mbps	0.25–5.5 kbps	1 Mbps
Tx current	5–15 mA	5–15 mA	25–125 mA	200–500 mA
Infrastructure	Gateway	Mesh + gateway	LoRaWAN gateway	SIM + carrier
Waveform upload?	Yes	Limited	No (impractical)	Yes
Best for	Close-range, marine	Large buildings	Distributed industrial	Mobile/remote

Multi-Protocol Architectures

The most capable bearing monitoring deployments use multiple protocols in a tiered architecture. BLE collects data from sensors within a machinery space; a BLE-to-LTE gateway aggregates and uploads to cloud infrastructure. This combines the power efficiency and ecosystem of BLE at the sensor with the wide-area reach of LTE at the gateway — neither protocol alone achieves both properties.

Systems like Fault Ledger take a multi-protocol approach specifically because no single wireless standard addresses all deployment environments. Marine installations benefit from BLE’s low power and metal-space performance; railway rolling stock needs LTE for connectivity in motion; large industrial facilities may use LoRa for campus-wide sensor aggregation. A platform that supports multiple protocols avoids forcing the deployment environment to adapt to the sensor’s connectivity constraints, rather than the reverse.

For engineers evaluating wireless vibration sensors, the protocol selection should be driven by the specific combination of range, data rate requirements, power budget, and available infrastructure — with the recognition that raw waveform access (required for forensic analysis and deep diagnostics) effectively requires BLE or LTE, while LoRa works well for operational trending applications where computed indicators are sufficient. Visit Fault Ledger to explore how multi-protocol sensor design applies to your specific bearing monitoring environment.

Among the variables that determine whether a vibration-based bearing monitoring system can actually detect early-stage defects, sensor coupling — the mechanical interface between sensor and structure — is consistently underweighted in system design. Engineers spend considerable effort selecting accelerometers with suitable frequency response and configuring sampling rates, then mount the sensor with foam tape or a flexible bracket that nullifies those choices. The physics of vibration transmission from bearing to sensor determines what information is available for analysis; no amount of signal processing can recover information that was attenuated before it reached the sensor.

The Physics of Vibration Signal Transmission

A rolling-element bearing defect generates a mechanical impulse — a brief, high-amplitude force pulse — each time a rolling element passes over a defect. This impulse propagates as a stress wave through the bearing housing and surrounding structure. The stress wave contains energy across a broad frequency range: from the low-frequency bearing defect frequency (which may be below 100 Hz) up through the high-frequency structural resonances excited by the impulse (which may extend to 30 kHz or beyond).

The high-frequency components of this stress wave are where early-stage defect signatures are clearest. A fresh surface fatigue spall produces sharp, high-amplitude impulses; as the defect grows and the edges of the spall round off under rolling contact, the impulse becomes broader and less energetic at high frequencies. The early-stage signal is predominantly at high frequencies, which means that any attenuation of high-frequency content in the signal path directly reduces sensitivity to early fault detection.

How Compliant Mounts Attenuate High-Frequency Signals

A mechanical mounting interface behaves as a low-pass filter. The cutoff frequency of this filter is determined by the stiffness of the interface and the mass of the sensor:

f_cutoff ≈ (1 / 2π) × √(k / m)

Where k is the interface stiffness (N/m) and m is the sensor mass (kg). For a 50-gram sensor mounted with a foam adhesive pad (stiffness roughly 10⁵ N/m), the cutoff frequency works out to approximately 225 Hz. Any bearing defect frequency content above 225 Hz is attenuated by this interface — and the attenuation increases at 40 dB per decade above cutoff (for a simple single-mass system).

This means that a sensor mounted with foam tape is effectively blind to the high-frequency structural resonances that carry the most diagnostic information in early-stage bearing defects. By the time the bearing damage has progressed far enough that low-frequency indicators (RMS acceleration, crest factor at the raw defect frequency) show significant change, the defect is already well-developed — and the window for low-cost corrective action may have passed.

Typical Interface Stiffness Values

• Steel stud mount (threaded into housing): ~10⁹ N/m → cutoff: ~2.25 MHz (effectively transparent to all bearing frequencies)
• Epoxy adhesive mount: ~10⁸ N/m → cutoff: ~700 kHz (adequate for all practical bearing monitoring)
• Magnetic mount (direct metal contact): ~10⁸ N/m → cutoff: ~700 kHz (comparable to epoxy for rigid magnets)
• Beeswax mount: ~10⁷ N/m → cutoff: ~225 kHz (adequate for most applications)
• Foam adhesive pad: ~10⁵ N/m → cutoff: ~225 Hz (attenuates nearly all diagnostic content)
• Handheld probe: ~10⁴ N/m → cutoff: ~71 Hz (essentially useless for high-frequency analysis)

Direct Metal-to-Sensor Transmission

The highest-fidelity coupling approach for non-permanent sensor installations is direct metal-to-metal contact between the sensor housing and the bearing housing surface. When a rigid metal sensor shell is placed in direct contact with a clean, flat metal surface, the contact stiffness is determined by the surface area, surface roughness, and contact pressure — all of which can be designed to achieve effective stiffness in the 10⁷–10⁸ N/m range.

Magnetic mounting with a rigid stainless steel housing achieves this. The magnet provides the contact force (pull force typically 50–200 N for sensor-scale magnets), and the rigid shell ensures the interface is metal-to-metal rather than polymer-to-metal. The stainless steel shell also provides the environmental protection required for industrial and marine environments, making direct coupling and environmental durability complementary rather than conflicting requirements.

The key design requirement is that the sensor’s sensing element (the MEMS or piezoelectric accelerometer chip) must be mechanically connected to the exterior shell with minimal compliant material in the signal path. Potted sensors — where the chip is embedded in epoxy directly connected to the metal shell — perform better than sensors where the chip is mounted on a PCB isolated from the shell by rubber shock mounts.

Impact on BPFO and BPFI Detection Sensitivity

The practical consequence of coupling quality on defect frequency detection can be quantified. Consider a bearing with BPFO at 150 Hz and a structural resonance excited at 8,000 Hz. In the envelope spectrum (bandpass filtered around the resonance, then FFT of the rectified signal), the BPFO peak amplitude reflects the strength of the impulse as detected at the sensor.

With a rigid stud mount, the 8,000 Hz resonance is transmitted with minimal attenuation; the envelope spectrum shows a clear BPFO peak at a defect severity level of 10% of failure severity (approximately). With a magnetic mount at 10⁸ N/m, the 8,000 Hz content is attenuated by a factor of roughly 1.3 dB — barely perceptible. With foam tape at 10⁵ N/m, the 8,000 Hz content is attenuated by approximately 61 dB — the resonance is completely eliminated from the spectrum, and envelope analysis cannot function at all.

The minimum detectable defect severity scales inversely with the signal attenuation. A sensor on foam tape may require 30× more defect severity to produce a detectable signal than a rigidly mounted sensor on the same bearing. The clinical implication: conditions that would trigger an early-warning threshold on a well-coupled system are simply invisible to a poorly coupled sensor until much later in the fault progression.

Coupling and Forensic Evidence Quality

Beyond operational condition monitoring, coupling quality has direct implications for forensic analysis of bearing failures. When a bearing fails and the vibration record is examined to reconstruct the fault progression, the quality of the coupling determines whether the pre-failure defect signatures are recoverable from the data. A well-coupled sensor provides a record where the initiation and growth of the defect can be traced through bearing defect frequency amplitude over time. A poorly coupled sensor provides an ambiguous record that may be useless for determining when the defect first became detectable — a question that is central to warranty and liability disputes.

Systems designed for forensic capture, like Fault Ledger, specifically engineer direct coupling into the hardware design because the forensic value of the captured data depends entirely on signal fidelity. A magnetically mounted 316L stainless steel housing with internal accelerometer directly bonded to the shell provides the rigid, direct signal path required to preserve high-frequency defect signatures from the failure event. The goal is a sensor whose mechanical design does not attenuate the evidence before it can be preserved.

For engineers specifying bearing monitoring systems, coupling should be treated as a first-order design parameter alongside accelerometer frequency response and sampling rate. The cheapest sensor on the best coupling typically outperforms the best sensor on a poor coupling. Mounting method selection deserves the same rigor as sensor selection. Learn more about how Fault Ledger’s direct coupling design preserves bearing defect frequency content for both operational monitoring and forensic capture.

When a critical bearing fails — in a marine propulsion system, a railway axle, an industrial gearbox, or a wind turbine drivetrain — the mechanical failure is rarely the most costly part of what follows. The bearing itself is replaceable. The downtime is quantifiable. What is often far more expensive, and far more protracted, is the dispute about why it failed.

In high-value failure scenarios, multiple parties share the loss and the liability: the equipment operator, the bearing manufacturer, the OEM who specified the bearing, the maintenance provider who most recently serviced it, and the insurer who underwrites the risk. Each party has incentives to reconstruct the failure narrative in ways that minimize their exposure. Without an authoritative record of what actually happened at the bearing in the moments before and during failure, these disputes are settled by leverage rather than physics — by which party can sustain the longest legal engagement, not by what the evidence shows.

Why Standard Monitoring Data Is Insufficient

Most installed bearing monitoring systems generate operational data: trend logs, health scores, alarm histories, periodic vibration spectra. This data is genuinely useful for maintenance planning. It is not, in general, useful as forensic evidence in a contested failure investigation, for several reasons.

The Data Resolution Problem

Trend data is averaged and compressed. A daily or hourly health score tells you that something changed, but the raw high-frequency vibration record — the actual physical signal from which defect frequencies are computed — is typically not stored. By the time a failure occurs, the detailed signal that would allow a forensic analyst to determine the fault type, progression rate, and sequence of events has been discarded as part of normal data management.

The Survivability Problem

Industrial data historians and monitoring gateways are not designed to survive catastrophic failures. A bearing failure that takes out a gearbox may also take out the sensor, the local data concentrator, and the power supply for the monitoring system. The standard failure mode is that the most critical data — from the final moments before failure — is precisely what is unavailable because the monitoring infrastructure did not survive the event it was monitoring.

The Integrity Problem

Even when data exists and is accessible, standard monitoring data typically lacks the integrity properties required for adversarial legal use. Anyone with administrative access to a monitoring platform can alter or delete records. The data has no cryptographic signature binding it to the specific sensor at the specific time. An opposing expert can reasonably argue that the data has been selectively presented, manipulated, or is not actually from the sensor claimed. Without a chain of custody that can be independently verified, the data’s evidentiary weight is severely limited.

The Timing Problem

Determining what the bearing was doing before the failure requires data from before the failure. Standard monitoring systems record data in real time and transmit or store it on a continuous basis — but they do not specifically preserve a bounded window of high-resolution data centered on the failure event. Pre-event data may exist in logs, but its relationship to the failure moment is imprecise, and the log entries nearest the failure are often corrupted or missing because the failure disrupted the logging system.

What Forensic-Grade Bearing Failure Evidence Requires

A bearing failure record suitable for legal or insurance purposes needs to satisfy requirements that standard monitoring systems do not address.

Pre-Event and Post-Event Capture Windows

A forensic recorder must continuously buffer high-resolution vibration data and, upon detecting a terminal event (a physical discontinuity consistent with catastrophic bearing failure), preserve a fixed window of data from before and after the event. The pre-event window captures the defect progression leading to failure; the post-event window captures the acoustic and vibration signature of the failure itself. Together, they provide the complete failure chronology.

The pre-event buffer must be stored locally on the sensor — not in a remote data system — because the failure event may disrupt connectivity and power before transmission can occur. Local battery-backed storage with no dependency on external infrastructure is the minimum survivability requirement.

Cryptographic Sealing

Evidence integrity must be provable without requiring trust in the party who preserved it. Cryptographic hashing of the evidence package — SHA-256 or better — produces a fingerprint that changes if any single bit of the evidence is altered. When the evidence is sealed on-device at the moment of capture, before any transmission or human access, the hash can be verified at any subsequent time by any party with a copy of the data. If the hash matches, the data is unaltered.

Sealing must occur on-device, not in a cloud system, because cloud systems are accessible to the vendor and therefore cannot demonstrate independence. A tamper-evident record must be sealed by a system that no single party controls.

Multi-Party Key Control

If the evidence is encrypted, the encryption key structure must prevent any single party from unilaterally accessing or suppressing the evidence. A two-key architecture — where decryption requires both a device key (generated on the sensor hardware and stored in tamper-resistant memory) and a custodian key (held by the asset owner or designated evidence custodian) — means that neither the vendor nor the operator can access the evidence without the other’s participation. This structural neutrality eliminates the possibility of one party suppressing evidence that is adverse to their interests.

Chain-of-Custody Metadata

Every access event — extraction, verification, inspection — must be logged immutably within the evidence package itself. The chain-of-custody log records that access occurred, when it occurred, and which key was presented. This log is sealed with the evidence and cannot be altered. In a dispute, any party can inspect the custody log to verify that the evidence was not accessed, altered, or selectively disclosed prior to the point where they received it.

Use Cases: Where Forensic Evidence Matters

Warranty Disputes

An OEM claims the bearing failed because of improper installation or operation outside rated parameters. The operator claims the bearing was defective from manufacture. Without a high-fidelity record of what the bearing was doing in the hours before failure — whether defect frequencies consistent with a manufacturing defect (inner race or ball defects) were present prior to any overload event — the dispute is resolved by assertion. A forensic record of the pre-failure vibration signature can often distinguish between a defect that originated in the bearing (present from early operation) and a defect that resulted from an operational event (appearing suddenly with overload signatures).

Insurance Claims

Marine hull and machinery insurers typically require evidence that the insured asset was operated and maintained within prescribed conditions. A tamper-evident failure record provides this evidence in a form that the insurer cannot reasonably dispute: it was captured by an independent device, sealed before anyone could access it, and the sealing can be cryptographically verified. This accelerates claims settlement and reduces the need for extended expert investigation.

Regulatory Investigations

In rail transport and aviation, catastrophic bearing failures trigger regulatory investigations that may involve subpoena of all available data. A forensic bearing failure recorder that produces a tamper-evident, chain-of-custody evidence package is specifically designed for this scenario: the data is preserved in a form that meets evidentiary standards without requiring after-the-fact reconstruction.

The Architecture That Delivers This

The requirements above — local pre-event buffering, on-device sealing, multi-party key control, immutable custody logging — are architectural requirements, not feature additions. They cannot be retrofit onto a standard monitoring platform. They require a system designed from the outset around forensic evidence as the primary deliverable, not operational trending.

Fault Ledger is built around exactly this architecture. It is not a predictive maintenance system, a health scoring platform, or a condition monitoring dashboard. It is a forensic recording system: it buffers high-frequency vibration data continuously, detects terminal events, seals a bounded evidence package on-device using multi-key cryptography, and maintains an immutable chain-of-custody log — all without requiring network connectivity or external power at the moment of failure.

For operators of high-value rotating machinery where bearing failure triggers multi-party disputes — marine propulsion, rail axles, wind turbines, industrial drivetrains — the question is not whether bearing failures will generate contested claims, but whether the evidence needed to resolve those claims efficiently will exist. Standard monitoring systems answer that question with a gap. Fault Ledger is designed to close it.

Vibration-based bearing condition monitoring depends on one foundational insight: every defect in a rolling-element bearing generates a repeating mechanical impulse at a predictable frequency. These frequencies — collectively called bearing defect frequencies — are deterministic functions of bearing geometry and shaft speed. Understanding them is the prerequisite for interpreting any FFT spectrum from a bearing sensor.

The Four Fundamental Defect Frequencies

Four characteristic frequencies describe the four major fault locations in a rolling-element bearing. Each is expressed as a multiple of shaft rotational frequency (in Hz), so they scale linearly with speed.

BPFO — Ball Pass Frequency, Outer Race

BPFO is the rate at which rolling elements strike a defect on the fixed outer raceway. Because the outer race does not rotate, each ball contacts the defect once per ball pass.

BPFO = (N / 2) × (1 - (Bd / Pd) × cos α) × RPM / 60

Where N is the number of rolling elements, Bd is ball diameter, Pd is the pitch diameter (bearing center-to-center), and α is the contact angle. BPFO is typically the most easily detected defect frequency because the outer race is stationary and the load zone is fixed, producing a consistent, repeatable impulse each ball pass.

BPFI — Ball Pass Frequency, Inner Race

BPFI describes impacts on the rotating inner race. Because the inner race rotates, the defect passes through the load zone once per cage revolution, which modulates the amplitude of BPFI harmonics. This amplitude modulation — at shaft frequency — is a key diagnostic signature.

BPFI = (N / 2) × (1 + (Bd / Pd) × cos α) × RPM / 60

BPFI defects are generally harder to detect than BPFO because the load modulation smears energy across sidebands.

BSF — Ball Spin Frequency

BSF represents the spin rate of an individual rolling element (ball or roller). A spall on a ball surface generates two impulses per spin — once on the inner race, once on the outer.

BSF = (Pd / (2 × Bd)) × (1 - (Bd / Pd)² × cos² α) × RPM / 60

In practice, BSF defects are the most difficult to detect. Slip between rolling elements and raceways introduces frequency smearing, and the double-impact pattern makes amplitude modulation analysis essential.

FTF — Fundamental Train Frequency (Cage Frequency)

FTF is the rotational frequency of the cage (retainer), which holds the rolling elements in angular spacing. Cage defects, cage-raceway rubs, or lubricant starvation produce energy at FTF and its harmonics.

FTF = (1 / 2) × (1 - (Bd / Pd) × cos α) × RPM / 60

FTF is always sub-synchronous (below shaft frequency) and typically falls between 0.35× and 0.48× RPM. Energy at FTF without corresponding BPFO or BPFI often points to lubrication issues rather than raceway damage.

How Bearing Geometry Shapes the Frequencies

All four frequencies share the same geometric factors: ball diameter (Bd), pitch diameter (Pd), contact angle (α), and ball count (N). The ratio Bd/Pd — sometimes called the bearing geometry factor — directly controls how far BPFI departs from BPFO. Larger balls relative to pitch diameter push BPFO and BPFI further apart. Contact angle matters most in angular-contact bearings (common in high-axial-load applications), where α may range from 15° to 40°, shifting all four frequencies measurably.

Bearing manufacturers publish these parameters in their datasheets, and most modern bearing catalogs include the computed defect frequency multipliers (in orders of shaft frequency) so engineers do not need to compute from scratch.

FFT Analysis and Spectral Signatures

Fast Fourier Transform (FFT) analysis converts a time-domain vibration waveform into its frequency components. A healthy bearing produces a broadband noise floor with no tonal content at the defect frequencies. As damage initiates, narrow peaks appear at BPFO, BPFI, BSF, or FTF and then grow as harmonics emerge (2×, 3×, 4× of each fundamental).

Early-stage defects are best detected using envelope analysis (also called demodulation). The raw acceleration signal is bandpass-filtered around a resonance excited by the impacts, rectified, and then re-transformed via FFT. The result — the envelope spectrum — reveals the modulating impulse rate even when the raw FFT shows no obvious peaks.

For outer-race defects (BPFO), the raw FFT is usually sufficient once damage has progressed beyond initiation. For inner-race defects (BPFI), envelope analysis is almost always required because of the load-zone amplitude modulation. For ball defects (BSF), envelope analysis combined with cepstral analysis is recommended.

Sensor Placement and Frequency Resolution

Detecting bearing defect frequencies requires sensors with adequate frequency range and resolution. High-speed shafts (3,000–30,000 RPM) generate defect frequencies in the hundreds to low thousands of Hz range, while low-speed machinery (10–100 RPM) may produce defect frequencies below 5 Hz — a demanding requirement for sensor sensitivity and FFT resolution.

Accelerometers with a flat response to at least 10 kHz are standard for most industrial applications. The sensor must be rigidly coupled to the bearing housing with the shortest possible mechanical path; compliant mounts (rubber pads, foam tape) attenuate the very high-frequency impulses that carry defect information.

Platforms like Fault Ledger are designed with direct metal-to-sensor coupling specifically to preserve these high-frequency signatures intact. When investigating a bearing failure, the quality of the vibration signal captured at the moment of the event determines whether the defect frequencies are recoverable from the data at all.

Frequency Resolution and Aliasing Considerations

FFT resolution (Δf) is determined by the sample length: Δf = fs / N, where fs is the sampling rate and N is the number of samples in the FFT block. To resolve BPFO on a 60-RPM machine (~0.5 Hz) with confidence, you need Δf well below 0.5 Hz, which requires at minimum a 2-second sample window at any sampling rate. At 10,000 samples/second, a 2-second capture yields 20,000 points and Δf = 0.5 Hz — just adequate.

The Nyquist theorem requires sampling at least twice the maximum frequency of interest. For bearing diagnostics targeting frequencies up to 10 kHz, a minimum sampling rate of 20 kHz is required. Most industrial-grade vibration sensors sample at 25–51.2 kHz to provide margin.

Putting It Together: A Diagnostic Workflow

• Obtain bearing datasheet parameters (Bd, Pd, α, N)
• Compute theoretical BPFO, BPFI, BSF, and FTF at operating speed
• Acquire a high-resolution vibration sample (minimum 2 seconds, ≥20 kHz sampling)
• Compute the raw FFT and check for peaks at defect frequencies and harmonics
• Apply bandpass filtering and envelope analysis for early-stage or inner-race detection
• Track trend over time — rising peak amplitude at a defect frequency indicates progressing damage

The physics of bearing defect frequencies is well established. What determines whether a monitoring system can actually detect a developing fault is the quality of the signal chain from bearing to analysis: rigid coupling, adequate sampling rate, sufficient frequency resolution, and consistent measurement timing. Systems designed around these requirements — such as Fault Ledger, which captures high-frequency raw vibration data at the moment of a terminal event — preserve exactly the spectral content needed to reconstruct which defect frequency drove the failure.

For rotating machinery engineers, building familiarity with these four frequencies transforms a raw FFT from an opaque spectrum into a diagnostic map with clearly labeled landmarks.