Overhead Crane Digital Twin Virtual Sensing & Predictive Maintenance
Digital Twin virtual sensing technology leverages a multi-body dynamics model and finite element simulation of the overhead crane structure. It uses virtual sensing algorithms to estimate critical parameters in locations where physical sensors cannot be installed, and combines this with fault-injection simulation to deliver comprehensive Predictive Maintenance coverage.
Following the structural load-bearing capacity assessment methodology outlined in the ISO 4301 Crane Design Standard, many critical areas of an overhead crane are inaccessible for physical sensor installation. These include the top flange plate at the mid-span of the main girder (obstructed by the trolley and rails), the full length of the wire rope (where long-term strain gauge application is impractical), and the internal gear meshing surfaces within the reducer. Digital Twin virtual sensing technology addresses this by creating a high-fidelity multi-body dynamics model of the crane. This model is driven by real data from a limited set of physical measurement points to estimate key parameters at un-instrumented locations. Kelude's KL-DT-VS Digital Twin platform integrates a fault-injection simulation engine, enabling Predictive Maintenance throughout the entire crane lifecycle, from the design phase to the operational phase.
Digital Twin Construction Methodology
The overhead crane Digital Twin is built using a multi-layered modeling strategy:
Layer 1 – Rigid body multi-body dynamics model (using Simpack or ADAMS), capturing the rigid body dynamics of the Trolley Travel Mechanism, Crane Travel Mechanism, Hoisting mechanism, and the load handling system.
Layer 2 – Flexible body finite element model (using Ansys or Abaqus). The steel structure components, including the Main Girder, End Carriage, and Trolley Frame, are reduced using the CMS method (retaining the first 30 modes) and integrated into the multi-body dynamics model for rigid-flexible coupling simulation.
Layer 3 – Control logic model (using Simulink) that simulates the Variable Frequency Drive (VFD) operation, Brake response, and PLC Control logic.
Real-time synchronization between the physical and digital spaces is achieved via encoders at the reducer input, a PLC data acquisition card in the Operator Cab (collecting four-dimensional signals: Lifting Capacity, hook height, Trolley position, and Crane Bridge position), and vibration and strain sensors at 8 key measurement points. The Digital Twin steps forward every 100 ms, maintaining real-time synchronization with the physical crane. After initial model calibration through parameter identification, the thrust/torque prediction error under typical operating conditions is <5%.
Virtual Sensing Algorithm Fundamentals
The core of virtual sensing is a state observer based on the Kalman filter. The system inputs noisy physical sensor measurements yk into the state-space equations of the Digital Twin model. An Extended Kalman Filter (EKF) estimates the complete current system state vector x̂k|k. This state is then used in the output equation to calculate the physical quantity ŷkvir at the virtual sensor location.
For virtual strain estimation on the top flange plate at the main girder's mid-span, the algorithm uses strain sensor measurements from the End Carriages at both ends (physical measurement points). It fuses these with the four-dimensional load parameters (Lifting Capacity, Trolley position, Crane Bridge position) and applies the Influence Surface Method to estimate the maximum mid-span stress. In validation tests on a 50t/25m span crane, the deviation between the virtually sensed maximum mid-span stress and the values from physical strain gauges ranged from +4.7% to −6.2%, meeting engineering accuracy requirements. This algorithm is deployed on the edge side of the KL-DT-VS platform and can complete a single virtual sensing calculation within 1 ms.
The accuracy of the virtual sensing algorithm depends on the calibration of the Digital Twin model and the rationality of the physical sensor layout. If strain sensors are installed with positional deviation from the design coordinates or their sensitivity changes due to welding heat effects, the virtual sensing results can exhibit systematic bias. The KL-DT-VS platform includes a built-in sensor health self-check function. During low-load periods in the early morning, it injects a pseudo-excitation signal of known amplitude. By comparing the residual between the virtually predicted value and the theoretical expected value, it assesses the health of each physical sensor. If the residual for a specific sensor exceeds twice the baseline RMSE for three consecutive checks, the system automatically flags the sensor as "needs calibration" and sends a notification to maintenance personnel. This self-check mechanism ensures the reliable operation of virtual sensing throughout the equipment's entire service life.
Fault Injection Simulation Approach
Fault-injection simulation is a differentiating capability of the KL-DT-VS platform, allowing maintenance personnel to proactively inject various degradation faults in the digital space and observe how fault symptoms manifest in virtual sensor responses. The platform supports the injection of 12 common overhead crane faults: main girder fatigue cracks (stiffness reduction of 5%–30%, with selectable locations), reducer gear pitting (meshing stiffness degradation, with a time-varying meshing stiffness function superimposed with pulse functions), bearing raceway spalling (impact sequences superimposed along the transmission path, with a period equal to the reciprocal of the cage rotational frequency), wire rope broken wires (equivalent cross-sectional area reduction with linear stiffness degradation), motor rotor broken bars (torque pulsation injection at 2sf harmonic frequencies), and more.
Operators simply select the fault type, severity level (mild/moderate/severe), and injection time in the interface, and the system automatically modifies the corresponding physical parameters of the Digital Twin model to generate realistic virtual sensor signals containing fault signatures. These signals are used to train AI diagnostic models, optimize threshold settings, and develop maintenance and repair plans. In an application involving three Ship-to-Shore (STS) cranes at a port, a CNN diagnostic model trained on fault-injection simulation data achieved a classification accuracy of 93.8% in real fault scenarios—a 21-percentage-point improvement over models trained with limited field data alone.
Predictive Maintenance Strategy and Implementation Roadmap
The predictive maintenance approach based on Digital Twin virtual sensing is implemented in three phases:
Phase 1 (Baseline Period, Months 1–3) — Deploy physical sensors and calibrate the Digital Twin model to establish baseline vibration, strain, and temperature distribution profiles for each overhead crane.
Phase 2 (Operation Period, Months 4–12) — The Digital Twin model runs continuously, with virtual sensing calculating damage accumulation at unmeasured points. The system automatically generates a Health Index (HI) on a scale of 0–100 (new equipment = 100, recommended maintenance threshold HI ≤ 60).
Phase 3 (Optimization Period, Month 13 Onward) — Historical data is used to train degradation trend prediction models (based on an LSTM time-series network with a 30-day input window, forecasting Health Index trends for the next 14 days).
After deploying the KL-DT-VS platform across 28 overhead cranes at an aluminum company, unplanned downtime dropped from 17 to 3 incidents per year (an 82.4% reduction), annual maintenance costs fell from ¥2.86 million to ¥1.12 million (a 60.8% reduction), and the wire rope replacement interval extended from an average of 9 months to 15 months (a 66.7% improvement). The company adopted a phased deployment approach: the first year focused on Digital Twin construction and baseline calibration, the second year expanded to all 28 overhead cranes and integrated with the ERP system, and the third year established a plant-wide equipment health management dashboard.
| Comparison Parameter | Traditional Scheduled Maintenance | Digital Twin Predictive Maintenance |
|---|---|---|
| Maintenance Strategy | Fixing Periodic Replacement/Inspection | Health Index-Based HIDynamic Scheduling |
| unplanned downtime | Annual Average15~20Cycles/Per 100 Units | Annual Average2~4Cycles/Per 100 Units(80%) |
| Spare parts Inventory | Safety Stock30%redundancy | On-Demand Spare parts, Reduced to8%redundancy |
| Diagnosis Coverage Rate | Physical Sensor Coverage<30%Component | Virtual Sensor Coverage>85%Component |
| Fault Knowledge Base | Engineer Experience-Dependent | Automated Fault Injection Accumulation>12Fault Spectrum Classification |
| ROICycle | — | 12~18Months to Payback |
Frequently Asked Questions
Q: How is Digital Twin model accuracy ensured, and how often does the model require parameter recalibration?
A: Initial model accuracy is established through parameter identification and calibration. A Bayesian optimization algorithm adjusts key physical parameters (damping ratio, friction coefficient, stiffness coefficient, etc.) to minimize the root mean square error (RMSE) between model response and physical sensor measurements. Baseline calibration requires only five full-load test runs completed within the first two weeks after deployment. A model recalibration is then executed automatically every quarter: measured values from physical sensors are compared against the model's predicted values for the same period. If the RMSE exceeds 1.5 times the initial baseline threshold, the system automatically triggers a re-identification of model parameters. Field data shows that after one year of operation across all duty conditions, model drift stabilizes within ±12% of the initial RMSE.
Q: Can virtual sensing technology fully replace physical sensors? replace physical sensors?
A: Not entirely, but it can significantly reduce the number of physical sensors required. The Digital Twin model needs a certain number of physical measurement points to serve as driving inputs and calibration anchors. A typical configuration uses 8–16 key measurement points (strain, vibration, and temperature sensors distributed across critical crane locations). On this basis, virtual sensing can expand the equivalent monitoring coverage to more than 200 points. Without any physical sensors, the Digital Twin model would drift due to a lack of real-world input—a purely open-loop Digital Twin accumulates unacceptable prediction errors within 24 hours. The optimal strategy is therefore a hybrid architecture combining physical sensors with virtual sensing, striking the best balance between cost and coverage.
Q: How close is fault-injection simulation data to real fault data?
A: The gap stems from three main sources: the degree of simplification in fault models (e.g., fatigue cracks are simplified as stiffness reduction factors without fully simulating the two-dimensional crack propagation profile), unmodeled nonlinear system behavior (such as backlash in reducer gear drives and nonlinear friction characteristics), and manufacturing and assembly variability (dynamic characteristic differences between cranes of the same model due to manufacturing tolerances). In laboratory validation, diagnostic models trained on fault-injection data achieve approximately 80–90% accuracy on real fault data compared to models trained exclusively on real fault data. An incremental learning strategy is recommended: first train a baseline model on fault-injection data, then continuously fine-tune it with actual operational maintenance data. After 3–6 months of iteration, accuracy can reach levels comparable to models trained entirely on real data.
Q: What are the deployment timeline and ROI period for the Kelude KL-DT-VS platform?
A: Deployment for a single overhead crane takes approximately 5–7 working days: Days 1–2 involve installing physical sensors and the edge computing gateway (using scheduled maintenance windows); Days 3–4 cover dynamic testing and model calibration (including full-load test runs under three typical duty conditions); Day 5 is dedicated to deploying the fault-injection simulation module and AI diagnostic model; Days 6–7 complete system integration testing and operator training. For batch deployments (10 or more cranes), the average deployment time can be compressed to 3–4 working days per crane. Return on investment: for metallurgical/port overhead cranes classified as Work Duty A7/A8, the deployment cost for a single KL-DT-VS system is approximately ¥80,000–120,000 (about $11,900–$17,800), with expected annual maintenance savings of ¥150,000–250,000 (about $22,200–$37,100) from reduced unplanned downtime, extended component replacement intervals, and lower spare parts inventory. This puts the ROI period at roughly 12–18 months. The system includes a 3-year warranty and remote model maintenance services.
