onboard_condition

01 — System Architecture

Six-layer onboard CMS stack

The complete pipeline from physical sensor to maintenance decision, running on the bogie in real-time.

Layer 01

📡

Physical Sensing

MEMS triaxial accelerometers and acoustic emission sensors mounted directly on each axlebox. Sample rate 25.6 kHz. Captures vertical, lateral, and longitudinal vibration from the direct mechanical path.

ICP Accelerometer AE Sensor RTD Temperature Speed Encoder

Layer 02

⚡

Signal Conditioning

Anti-aliasing low-pass filters, charge amplifiers, and ADC conversion. Gain-stage normalization relative to train speed. Synchronous sampling triggered by wheel encoder pulses to track angular position.

Anti-alias filter Charge amplifier 24-bit ADC Trigger logic

Layer 03

🖥️

Edge DAQ & Processing

Ruggedized embedded compute unit (ARM Cortex-A72 or NI CompactRIO class) performs real-time FFT, EMD, and feature extraction. Operates in harsh railway environments: −40°C to +85°C, IP67 rated.

Real-time OS Onboard FFT Feature vector Local storage

Layer 04

🧠

AI Fault Diagnosis

Lightweight deployed ML models (quantized CNN or ONNX Runtime) classify bearing fault type and severity in real-time. VMD + envelope spectrum analysis identifies BPFI, BPFO, BSF, and FTF characteristic frequencies.

Quantized CNN ONNX Runtime RUL estimation Anomaly detection

Layer 05

📶

Communication & Telemetry

GSM/LTE or Wi-Fi offload at stations. Compressed feature vectors and alarm states transmitted to central server. Full raw waveform stored locally and synced during maintenance windows via USB/Ethernet gateway.

LTE/4G MQTT protocol Data compression Local buffer

Layer 06

🔔

Maintenance Decision

Cloud-based dashboard aggregates fleet-wide health. Alarm thresholds trigger driver cab alerts and automatic work-order creation in the CMMS. Reports severity stage (1–4), fault location, and recommended intervention window.

Fleet dashboard CMMS integration Driver cab alert Maintenance API

02 — Development Roadmap

Build phases: concept to fleet deployment

A structured 5-phase approach from hardware prototyping to full-scale commercial operation.

Requirements & Hardware Selection

Months 1–2

Define system specifications based on bearing geometry, operating speed range (40–300 km/h), sampling requirements, and environmental constraints. Select and procure sensor hardware. Establish data format standards and communication protocols.

Define bearing geometry parameters: pitch diameter, ball diameter, contact angle, number of rolling elements
Calculate characteristic frequencies: BPFI, BPFO, BSF, FTF for target bearing types
Select accelerometer (IEPE/ICP) with bandwidth ≥ 10 kHz and dynamic range ≥ 80 dB
Design IP67-rated enclosure for axlebox mounting in railway environment (EN 50155)
Select embedded processing platform and establish software architecture
Define data acquisition rate (minimum 25.6 kHz), storage format, and compression strategy

SRS Document Bearing Frequencies Calc Sheet Hardware BOM v1 Sensor Test Report

Lab Prototype & Signal Pipeline

Months 3–5

Build a laboratory test rig with a target bearing under controlled speed and load. Collect baseline and seeded-fault vibration datasets. Develop the complete signal processing pipeline and validate characteristic frequency detection.

Assemble bench test rig: motor, gearbox, target bearing, load application, speed control
Seed outer race (BPFO), inner race (BPFI), and rolling element (BSF) faults using EDM or wire saw
Collect vibration dataset: 3 health states × 5 load levels × 5 speed levels × 3 trials
Implement time-domain features: RMS, kurtosis, crest factor, skewness, peak-to-peak
Implement FFT envelope analysis, spectral kurtosis, and Kurtogram for resonance band selection
Implement VMD and EMD for non-stationary signal decomposition
Validate BPFI/BPFO/BSF sidebands visible in spectra; document SNR vs speed

Lab Dataset (labeled) Signal Processing Library Validation Report Kurtogram Results

AI Model Training & Validation

Months 6–9

Use lab dataset to train, tune, and validate machine learning models for fault classification and severity estimation. Augment with public datasets (CWRU, PU Bearing). Optimize models for edge deployment via quantization.

Augment lab data with CWRU and Paderborn University bearing datasets for model generalization
Train CNN-1D on raw vibration segments (256–1024 sample windows) for fault classification
Train LSTM for temporal trend modeling and Remaining Useful Life (RUL) regression
Implement semi-supervised GAN to generate synthetic fault data for rare fault classes
Quantize models to INT8 using ONNX Runtime for edge deployment (≤5ms inference latency)
Cross-validate on hold-out test set; target accuracy ≥ 95% for 4-class fault classification
Build adaptive thresholding system using PCA-based statistical process control

Trained Model (ONNX) Validation Accuracy Report Edge Inference Benchmark RUL Model

Onboard Integration & Field Trial

Months 10–16

Install prototype system on a test vehicle (locomotive or passenger coach). Conduct field trials on operational track under real operating conditions. Collect in-service data, validate detections against ground truth inspections, and refine the system iteratively.

Fabricate final enclosure: vibration-isolated, EMC-shielded, EN 50155 compliant mounting
Instrument 4 axleboxes with triaxial accelerometers and temperature sensors
Install embedded DAQ/edge compute unit in equipment bay with UPS power conditioning
Configure GSM/LTE telemetry module with secure MQTT connection to cloud backend
Run 10,000 km field trial; log all detections and compare with quarterly workshop inspection
Measure false alarm rate; tune detection thresholds to achieve FAR < 2%
Collect real-service dataset for model retraining with operational noise floor

Field Trial Report 10,000 km Dataset Updated Model v2 FAR/FNR Analysis

Fleet Rollout & Continuous Learning

Months 17–24+

Scale to full fleet. Deploy federated learning pipeline to continuously improve models from in-service fleet data without sharing raw waveforms. Integrate with CMMS/ERP for automated work-order generation and maintenance scheduling optimization.

Deploy system to full fleet; standardize installation procedure and quality checklist
Build federated learning pipeline: model gradients shared, raw data stays on vehicle
Integrate with CMMS (SAP PM / Maximo) via REST API for automated work order creation
Implement digital twin model: physics-based bearing fatigue + data-driven correction
Deploy fleet-wide health dashboard with per-axle RUL, fault map, and trend analytics
Establish ISO 13373-compliant reporting and railway authority certification process

Fleet Dashboard CMMS Integration Federated ML Pipeline ISO 13373 Report

03 — Signal Processing Pipeline

From raw vibration to fault diagnosis

The complete signal chain from sensor to classification, with key parameters for each stage.

📡

Raw Acquisition

Triaxial acceleration time-series from ICP accelerometer

🔧

Pre-processing

Bandpass filter, detrend, speed normalization, resampling

📊

Decomposition

VMD / EMD into IMFs; Kurtogram resonance band selection

🔬

Feature Extraction

RMS, kurtosis, BPFI/BPFO/BSF amplitudes, cepstrum

🧠

Classification

CNN-1D / LSTM: healthy, outer race, inner race, rolling element

📋

Prognosis & Alert

RUL estimate, severity (1–4), CMMS work order trigger

Sample rate

25.6 kHz

Window size

1024 samples

Overlap

50%

Update rate

Every 3 min

Inference time

< 5 ms

IMF modes (VMD)

K = 6–8

Fault classes

4 + healthy

Bearing characteristic fault frequencies

BPFI

Inner Race Fault

BPFI = (n/2) × f_s × (1 + d/Dcosα)

Impacts generated as each rolling element passes the defect on the inner race. Modulated by shaft rotation frequency. Most common fault type (40–50% of bearing failures).

BPFO

Outer Race Fault

BPFO = (n/2) × f_s × (1 − d/Dcosα)

Fixed outer race produces a regular impacting pattern as rolling elements pass the defect. Not modulated by shaft speed. Detectable at lower SNR than BPFI.

BSF

Rolling Element Fault

BSF = (D/2d) × f_s × [1 − (d/Dcosα)²]

Ball/roller spall creates two impacts per revolution. Frequency is non-integer multiple of shaft speed. Often manifests with cage frequency (FTF) sidebands.

FTF

Cage / Train Fault

FTF = (f_s/2) × (1 − d/Dcosα)

Cage slip, deformation, or impact. Lowest frequency component — typically 0.3–0.5× shaft speed. Often indicates lubrication failure or advanced degradation.

n = number of rolling elements · f_s = shaft rotation frequency · d = rolling element diameter · D = pitch circle diameter · α = contact angle

Technology stack

⚙️

Signal Processing

Core DSP Libraries

SciPy / NumPyFFT, filtering, stats

PyWaveletsDWT, WPT, VMD

MNE / PyEMDEMD, EEMD, CEEMDAN

Kurtogram (C extension)Resonance band detection

🧠

Machine Learning

Training & Inference

PyTorch / TensorFlowModel training

ONNX RuntimeEdge inference (INT8)

Scikit-learnFeature selection, SVM, kNN

TensorRT (optional)GPU-accelerated edge

🔌

Embedded / Edge

Onboard Compute Platform

FreeRTOS / ZephyrReal-time OS for DAQ

Raspberry Pi CM4 / JetsonEdge inference host

NI CompactRIOIndustrial DAQ (optional)

C/C++ DSP coreCMSIS-DSP for ARM

☁️

Cloud & Backend

Fleet Management Layer

InfluxDB / TimescaleDBTime-series data store

Apache KafkaReal-time data streaming

GrafanaFleet health dashboard

FastAPI + PostgreSQLREST API & metadata

04 — Hardware Bill of Materials

Complete hardware specification per axle

Recommended components for a production-grade onboard CMS installation. Costs are indicative per-unit.

Component	Part / Model	Specification	Category	Unit Cost (USD)
ICP Triaxial Accelerometer	PCB 356A32 / Kistler 8763	±50g, 0.5–10,000 Hz, 100 mV/g, IP67, stainless steel	Sensor	$350–$600
Acoustic Emission Sensor	Physical Acoustics PICO / WD	100–400 kHz, 34 dB gain preamp, magnetic mount	Sensor	$180–$320
PT100 RTD Temperature Sensor	Omega PR-10-3-100-1/8-6-E	−50°C to +200°C, ±0.3°C accuracy, axlebox mount	Sensor	$45–$80
Wheel Speed Encoder	Hengstler RI58-O / Kübler 5823	1024 PPR, HTL output, IP67, −40°C to +85°C	Sensor	$120–$250
Charge Amplifier / Signal Conditioner	Kistler 5134 / PCB 480E09	4-channel, ICP power supply, anti-alias filter, ±10V output	Processing	$400–$900
Data Acquisition Module	NI USB-6366 / Measurement Computing	8-channel, 24-bit ADC, 500 kS/s aggregate, simultaneous sampling	Processing	$800–$2,000
Edge Computing Unit	NVIDIA Jetson Orin NX / CM4 + HAT	8-core ARM, 8GB RAM, eMMC storage, −40°C rated, DIN rail	Processing	$500–$1,100
Industrial SSD	Innodisk 3IE7 / Swissbit S-56	256GB, SLC NAND, −40°C to +85°C, power-loss protection	Processing	$180–$350
4G LTE Industrial Modem	Sierra Wireless RV55 / Teltonika RUT955	Cat-12 LTE, dual SIM, MQTT, GPS, EN 50155, −40°C	Communication	$350–$650
Isolation / UPS Power Module	Phoenix Contact MINI-PS / Bicker BEP-450	24VDC input, 100W, supercapacitor UPS, EN 50155 certified	Power	$200–$450
IP67 Ruggedized Enclosure	Spelsberg TK / Rittal KL	316 stainless, IP67, vibration-isolated DIN rail, −40°C to +80°C	Mechanical	$150–$350
Cabling & Connectors	Lemo 00 / M12 A-coded shielded	Shielded coax for AE/ICP, M12 data connectors, railway-rated	Mechanical	$100–$250

Estimated total per axle (prototype grade) $3,375 – $7,250

05 — AI / ML Model Stack

Diagnostic and prognostic model architecture

Layered AI models from classical feature-based classifiers to deep learning for RUL prediction.

Baseline

Statistical + SVM

Time-domain feature extraction (kurtosis, RMS, crest factor) combined with Support Vector Machine classifier. Fast, interpretable, works with small datasets. Ideal for initial deployment.

Accuracy (lab)91%

Inference<1 ms

Training data needed~500 samples/class

Primary

1D-CNN Fault Classifier

Convolutional neural network on raw vibration segments (1024 samples). Automatically learns discriminative frequency patterns. Achieves high accuracy across fault types without manual feature engineering.

Accuracy (CWRU)98.5%

Inference (INT8)3–5 ms

ClassesHealthy, BPFI, BPFO, BSF, FTF

Temporal

LSTM Trend Monitor

Long Short-Term Memory network tracks feature evolution over time. Detects gradual degradation trends that snapshot classifiers miss. Input: rolling 50-sample feature window over last 2.5 hours.

Fault detection lead6–48 hours advance

Inference8 ms

Sequence length50 feature vectors

Prognosis

RUL Regression (Transformer)

Transformer encoder with attention over degradation trajectory. Predicts Remaining Useful Life in kilometers or hours. Output is a probability distribution over RUL — providing confidence intervals for maintenance scheduling.

RMSE (validation)~12% RUL

OutputRUL ± CI (km)

Update frequencyEvery 3 min

Anomaly

Autoencoder (Unsupervised)

Trained only on healthy bearing data. Reconstruction error spikes when novel faults occur — providing a zero-shot anomaly detector that catches fault types not seen during training. Outputs anomaly score 0–100.

True positive rate93%

False alarm rate<2%

Training dataHealthy only

Data Aug.

Conditional GAN (cGAN)

Conditional generative adversarial network to synthesize rare fault class vibration data. Overcomes class imbalance (real bearing failure data is scarce). Generated samples validated by domain expert review before inclusion in training set.

Classifier accuracy boost+3–7%

Use caseOffline training only

OutputSynthetic waveforms

06 — Dashboard Interface

Operator monitoring interface

Real-time health overview for fleet operators and maintenance engineers.

Fleet health score

94.2

▲ +0.3 from yesterday

Active alarms

Axle 3R — Stage 2 BPFO

Vibration RMS (Axle 1L)

0.84 g

Normal range <1.5g

Kurtosis (Axle 3R)

6.8

Threshold >4.0 — elevated

Vibration envelope — Axle 3R (last 60s)

Axlebox health — all 8 axles

Axle 1L

● Healthy

Axle 1R

● Healthy

Axle 2L

● Healthy

Axle 2R

● Healthy

Axle 3L

● Healthy

Axle 3R

⚠ Stage 2

Axle 4L

● Healthy

Axle 4R

● Healthy

Fault alert & notification system

CRITICAL

Stage 4 bearing degradation — immediate intervention required

Outer race defect frequency (BPFO) confirmed at 6.8× background level. Thermal rise +18°C above ambient. Kurtosis 14.2. HABD corroboration recommended at next trackside detector.

Train 2241 · Axle 5R · BPFO @ 124.6 Hz · Temp: 72°C · Speed: 140 km/h

Create WO →

WARNING

Elevated kurtosis — Stage 2 inner race defect developing

BPFI sideband amplitude rising trend detected over past 240 km. Autoencoder anomaly score: 68/100. RUL estimate: 1,400 ± 320 km. Schedule inspection at next planned depot visit.

Train 2241 · Axle 3R · BPFI @ 178.4 Hz · Kurtosis: 6.8 · Anomaly: 68

Schedule →

INFO

Lubrication anomaly — cage frequency elevated

FTF component rising 22% above baseline. Consistent with early lubrication degradation. No structural fault detected. Recommend relubrication at next scheduled maintenance window (est. 800 km).

Train 2241 · Axle 2L · FTF @ 14.8 Hz · Anomaly score: 34

Log →

07 — Key Performance Indicators

System performance targets

Measurable KPIs to validate system performance across detection, operational, and business dimensions.

Fault detection accuracy

≥ 95%

4-class fault classification on held-out field validation dataset. Measured per fault type and overall weighted accuracy.

False alarm rate (FAR)

≤ 2%

Percentage of healthy bearings incorrectly flagged as faulty per 1,000 operating hours. Tunable via adaptive threshold.

Detection lead time

≥ 48 h

Minimum advance warning before functional failure for Stage 1–2 faults. Enables scheduled depot intervention before service disruption.

RUL prediction error

≤ 15%

Root Mean Square Error of Remaining Useful Life estimates as a percentage of actual remaining life. Validated on field run-to-failure data.

System uptime (on-vehicle)

≥ 99%

Availability of the onboard DAQ and edge compute system, excluding planned maintenance. Supported by hardware WDT and UPS.

Edge inference latency

≤ 5 ms

Time from vibration window completion to fault classification output. Enables real-time continuous monitoring at update rate of every 3 minutes.

Onboard Vibration Condition Monitoring System for Rail Axle Bearings

Six-layer onboard CMS stack

Physical Sensing

Signal Conditioning

Edge DAQ & Processing

AI Fault Diagnosis

Communication & Telemetry

Maintenance Decision

Build phases: concept to fleet deployment

From raw vibration to fault diagnosis

Bearing characteristic fault frequencies

Technology stack

Complete hardware specification per axle

Diagnostic and prognostic model architecture

Operator monitoring interface

Fault alert & notification system

System performance targets