Overhead Crane Travel Data: Sensor Collection to Smart Analysis
The full data pipeline for overhead crane operations spans five layers: sensor acquisition (vibration/temperature/current/encoder, 8–16 sensors per crane), edge cleaning (3σ outlier removal + resampling + FFT feature extraction), time-series storage (InfluxDB/TimescaleDB, 10–50 GB/day per station), intelligent analytics (LSTM-based predictive maintenance / anomaly detection / optimized scheduling), and visualization dashboards (Grafana + mobile app push). Each overhead crane generates roughly 3–18 TB of raw data per year; effective analysis can reduce unplanned downtime by 40–60%. Kelude's big data platform is already connected to 200+ overhead cranes, with over 5 PB of operational data accumulated.
Overhead cranes (bridge and gantry cranes) are among the most data-dense assets in a modern factory—each unit is equipped with 8–16 sensors sampling at 100 Hz, continuously generating multi-dimensional signals covering vibration, temperature, current, and position. A single crane produces 30–50 GB of raw data per day, and a mid-sized plant with 20 cranes accumulates more than 200 TB per year. Yet most facilities use less than 5% of that data—sensor readings linger briefly inside the PLC and VFD and never make it into an analytics pipeline. This article breaks down the complete technical solution for overhead crane operational data, from acquisition and cleaning to storage, analysis, and visualization.
1. Sensor Data Acquisition: Multi-Dimensional Crane Monitoring
Data acquisition is where the analytics chain begins. Different sensors capture distinct physical quantities, covering the full set of operating parameters across the crane's mechanical and electrical systems. Kelude's standard data acquisition package includes the following sensors:
| Sensor Type | Measurement Parameter | Model/Specification | Sampling Rate | Quantity per Unit | Mounting Location |
|---|---|---|---|---|---|
| Vibration Sensor | Acceleration (mm/s²) | IEPE type, 0.5Hz~10kHz | 25.6kHz | 4~6pcs | Motor/Gearbox / Reducer/Wheel Axle Bearing/Drum |
| Temperature Sensor | Temperature(°C) | PT100, ±0.3°C | 1Hz | 4~8pcs | Motor Winding/Gearbox / Reducer Oil Sump/Brake/Crane Rail |
| Current Sensor | Current(A) | Hall Effect, ±1%F.S. | 1kHz | 3pcs | Hoisting/Trolley/Bridge VFD Output |
| Encoder | Encoder | Incremental/Absolute, 1024Pulses per revolution (PPR) | 100Hz | 2~4pcs | Hoisting Motor/Bridge/Trolley Drive |
| Weight Sensor | Load(t) | Column Type/Pin Type, 0.5%F.S. | 10Hz | 1pcs | Hook Block Sheave |
2. Edge Data Cleaning & Preprocessing — From Raw Signal to Usable Features
Raw sensor signals cannot be fed directly into analytical models. Edge computing nodes (NVIDIA Jetson Orin NX/AGX) perform real-time cleaning, resampling, and feature extraction close to the overhead crane, compressing raw 25.6kHz vibration data into second-level feature values. The cleaning pipeline consists of four steps:
Step 1 — Outlier Rejection: Using the 3σ rule (Pauta criterion), the mean μ and standard deviation σ are calculated for a sliding window (window length N=1,000 points) on each sensor channel. Data points falling outside μ±3σ are flagged as outliers and removed. Sporadic spike interference in vibration data (electromagnetic noise, mechanical shocks) accounts for approximately 0.1–0.5% of samples; removed points are reconstructed via linear interpolation.
Step 2 — Missing Value Imputation: Temperature sensors and current sensors sample at lower rates (1Hz–1kHz) and occasionally experience gaps due to communication packet loss. Gaps shorter than 100ms are filled using linear interpolation; gaps of 100ms or longer use KNN imputation (k=5, weighted average of the 5 most similar neighboring points within the time window). Kelude's edge gateway achieves a measured imputation accuracy of ≥98%.
Step 3 — Resampling & Alignment: Sensors with different sampling rates are uniformly resampled to a 1Hz base frequency. For vibration data downsampled from 25.6kHz to 1Hz: four statistical metrics — RMS, peak value, kurtosis, and peak frequency — are extracted per second, compressing data volume by a factor of 25,600. For current data downsampled from 1kHz to 1Hz: RMS and peak values are computed per second. All sensors are synchronized via GPS clock, with timestamp alignment accuracy of ±1ms.
Step 4 — Feature Engineering: Higher-order features are computed on the 1Hz base-frequency data: vibration kurtosis (reflecting impact characteristics), crest factor (reflecting the dynamic range of signal amplitude), and sideband energy ratio (used for gear/bearing fault diagnosis). Kelude's standard feature set comprises 128-dimensional time-domain and frequency-domain features, with each overhead crane outputting one 128-dimensional feature vector per second.
3. Data Storage Architecture — Time-Series Databases & Hot/Cold Tiering
A single overhead crane generates 30–50GB of raw data per day (continuous 25.6kHz vibration sampling), which is reduced to approximately 1.2GB/day of feature data after edge cleaning and compression. For a facility with 20 overhead cranes, this translates to roughly 24GB of daily incremental data, 720GB per month, and 8.6TB annually. Different data types require different storage strategies:
| Data Type | Storage Scheme | Daily Increment | Retention Period | Compression Ratio | Query Performance(1hRange) |
|---|---|---|---|---|---|
| Raw Vibration(25.6kHz) | Min IOObject Storage | 30~50GB | 30days | 2:1 (gzip) | — (File Read) |
| Feature Data(1Hz, 128Dimension) | Influx DB | 1.2GB | Permanent | 5:1 (Columnar Storage+delta) | ≤200ms |
| Event Log(Alarm/Operation) | Postgre SQL | 50~200MB | Permanent | — | ≤50ms |
| Model Training Data(Labeled Samples) | Postgre SQL+Min IO | 100~500MB | Permanent | — | — |
Tiered Storage Strategy: Hot data (last 7 days) resides in the InfluxDB memory and SSD tier, with query responses under 50 ms. Warm data (days 8–90) is kept on the InfluxDB HDD tier, delivering query responses below 500 ms. Cold data (older than 90 days) is automatically downsampled to 10-minute intervals and archived to MinIO object storage, cutting storage costs by 70%. Kelude Heavy Industry's big data platform ships with these three storage tiers configured by default, and users can customize retention periods to suit their needs.
4. Smart Analytics Models — Turning Data into Decisions
Once cleaned and stored, the data becomes the foundation for advanced analytics. The Kelude Heavy Industry big data platform comes with four built-in analytics models: fault prediction, remaining life forecasting, energy consumption optimization, and scheduling optimization. The performance of these models was validated using six months of field data from a 32t overhead crane at a steel mill:
| Model | Algorithm | Accuracy | Lead Time for Warning | False Alarm Rate | Deployment Location |
|---|---|---|---|---|---|
| Bearing Fault Prediction | LSTM-Attention | 96.3% | 7~14days | 4.2% | Edge |
| Gear Wear Prediction | Transformer | 94.1% | 14~30days | 5.8% | Cloud |
| Energy Consumption Anomaly Detection | Isolation Forest | 91.7% | Real-time | 3.1% | Edge+Cloud |
| Scheduling Path Optimization | PPOReinforcement Learning | 87.6% | Path Planning | — | Cloud |
Kelude Heavy Industry's big data platform integrates an AutoML pipeline that automatically retrains models every two weeks. Data drift detection relies on the Population Stability Index (PSI), with a PSI > 0.1 threshold triggering automatic retraining. Model inference results are pushed via MQTT to Grafana dashboards, with alerts sent through WeChat and SMS.
8~16 Sensors per Crane Raw Vibration/Temperature/Current/Encoder/Weight Five Categories | 25,600× Edge Data Compression Ratio 25.6kHz1Hz, RMS/Peak Value/Kurtosis Feature | 5PB+ Cumulative Operating Data Volume 200+Per Unitoverhead crane, Continuous Acquisition Exceeding3years |
96.3% Bearing Fault Prediction Accuracy LSTM-Attention, Advance7~14days | 40~60% unplanned downtime Reduce Predictive Maintenance+Comprehensive Benefit of Scheduling Optimization | 128Dimension Feature Vector Dimension per Second Time Domain+Frequency Domain, Full Coverage |
5. Visual Dashboards & Reporting System
The final output of data analytics is delivered through visual dashboards and reports. Kelude's big data platform is built on a Grafana + Prometheus monitoring stack, supporting real-time data refresh with latency under 2 seconds. The system can display plant-wide overhead crane operation overviews, individual crane health dashboards, historical trend comparisons, and fault alert lists on a single screen. Multi-tenant permission isolation is supported, so different roles—plant managers, equipment supervisors, maintenance technicians, and crane operators—each see dashboards tailored to their responsibilities.
Dashboard Tiers: The operations overview dashboard shows plant-wide crane availability, OEE, energy consumption trends, and alert counts. The single-equipment dashboard displays real-time vibration spectra, temperature trends, current waveforms, cumulative operating hours, and remaining life estimates. The historical analysis dashboard enables trend comparisons over any time window, event replay, and dataBacktracking for the 15 minutes preceding a fault. Alert notifications are delivered through WeChat Work, SMS, and email—three channels in total. Critical alerts (bearing temperature above 85°C, vibration above 10 mm/s) are automatically escalated to the equipment supervisor. Kelude's big data platform has already delivered 28 customized dashboards across 6 categories for a major steel plant, with monthly energy efficiency reports (PDF format) generated automatically, including per-crane energy rankings, maintenance recommendations, and cost estimates.
For more crane data application scenarios, see: Three Approaches to Crane Braking Energy Recovery (energy optimization) and Cloud-Edge Collaborative Control System Architecture for Overhead Cranes (edge computing support). Kelude offers the big data platform in two editions: Standard and Enterprise. The Standard edition includes core dashboards plus alert push notifications; the Enterprise edition adds AutoML, custom dashboard development, and bespoke AI model training.
Frequently Asked Questions
Q: Do I need to retrofit my existing overhead cranes for data collection?
A: Kelude offers two options. For new cranes, sensors and an edge computing box are pre-installed at the factory—plug and play. For existing cranes, a retrofit kit is available: sensors are mounted using magnetic bases or fixtures without altering the existing electrical system. The edge box reads current, speed, and other existing signals through the VFD communication interface (RS485/Profinet), while new vibration and temperature sensors transmit wirelessly. A typical retrofit takes 3–5 days per crane and does not interrupt normal production.
Q: We generate 30–50 GB of data per day—is our network bandwidth sufficient?
A: Raw data is never uploaded directly to the cloud. The edge computing box performs compression and feature extraction on-site, uploading only the extracted feature data—approximately 1.2 GB per crane per day. For a site with 20 cranes, that's roughly 24 GB of daily uploads. The plant needs an internet connection of at least 100 Mbps. The actual uplink bandwidth required for feature data is about 2 Mbps (24 GB ÷ 86,400 s × 8 = 2.2 Mbps), which a standard business broadband connection can easily handle. Raw data is retained locally on-site for 30 days and then automatically purged; optional local NAS storage is available for longer retention periods.
Q: Your model claims 96% accuracy, but false alarms still disrupt production. What then?
A: Kelude uses a three-tier alert strategy. Tier 1 alerts (model confidence ≥ 95%) automatically generate a maintenance work order and push it to the supervisor's phone. Tier 2 alerts (80% ≤ confidence < 95%) are flagged as "needs attention" on the dashboard with a recommendation to inspect during the next scheduled downtime. Tier 3 alerts (confidence < 80%) are logged in the historical report only and are not proactively pushed. In addition, false alarms are fed back into the training loop—samples flagged as "false positive" by users enter a hard negative mining pool, and every two weeks during retraining, the model automatically lowers anomaly scores for similar patterns. After six months of this feedback cycle, the false alarm rate dropped from 4.2% to 2.1% in field deployments.
Q: How is the Kelude big data platform priced?
A: The Standard edition is billed per crane: a one-time data integration fee of ¥3,000–8,000 per crane (depending on the number of sensors) plus an annual platform service fee of ¥2,000–5,000 per crane per year, covering data storage, dashboards, alert push notifications, and model maintenance. The Enterprise edition is quoted on a project basis and includes AutoML, custom dashboards, and private deployment. Kelude offers a free 14-day trial of the Standard edition—connecting up to 2 cranes during the trial period incurs no charges.
The end-to-end technical solution for crane operational data—from collection to decision-making—is the foundational infrastructure for factory intelligence. Kelude's big data platform already serves customers operating more than 200 overhead cranes across steel, chemical, building materials, and machinery manufacturing industries. To learn more about the platform or to schedule a free demo, contact the Kelude technical team for a customized proposal.
