A study of excessive vibration of the drivetrain on roro ships operated by Wilhelmsen Ship Management* sheds important light on how digital twins can support condition monitoring of the drivetrain
Digital twins are dynamic digital models of machinery that can help predict faults and reduce downtime. The digital twin is a computer simulation typically accessed in the office, which gives an opportunity for real-time monitoring of onboard equipment and report generation. As more data is generated and sensed, the digital twin evolves. Communication between the twins is facilitated by web technology, notably Internet-of-Things (IoT).
The digital twin can help in making remaining life assessments, scheduling maintenance based on true operational profile, facilitating load management and early fault detection, and provide design feedback, support virtual inspection, and predict consequences of future operating profiles. On the digital twin, data on parts of the equipment not fitted with sensors can also be generated through advanced algorithms. Spare parts’ inventories can also be minimised due to the predictive capabilities and optimum scheduling.
According to the authors of a Wilhelmsen Ship Management study, titled Digital twin condition monitoring approach for drivetrains in marine applications, maintenance regimes are largely based on planned maintenance schemes (PMS), with a few exceptions on the offshore supplier vessel side of the industry. The frequency of maintenance depends on the maker’s specifications and the running hours of a system. The next step into condition monitoring (CM) is yet to be taken in a big way, despite the vast potential savings through decreased off-hire and downtime. Today, the control systems onboard vessels are largely used only to monitor and provide alerts for day-to-day operations.
Spare parts are a capital cost in the traditional PMS system. A vessel needs to have spare parts available at all times, in case of unexpected failures. A digital twin, or even the simpler CM system, would diminish the need for dead capital, the authors say. “By being able to predict failure, it would be possible to predict which port and what service personnel and spare parts are required. Accordingly, it would then be possible to reduce crew numbers and to sail the ship at minimum crew,” they add.
Wilhelmsen’s roro vessels used in the study had advanced to CM from PMS while adhering to CM standards set by classification societies. Data is collected from 84 sensors and a report with operational recommendations generated quarterly. This reduces off-hire time, but is a sub-optimal use of the system; the next level would be an online CM and digital twins.
Physics-based approach
Three different data analysis approaches exist for condition monitoring: data-driven; model-driven; and physics-based model-driven. The data-driven approach is sufficient in many cases, but for more complex and cost-heavy assets a model-based approach has a more holistic monitoring scheme, the authors say. The model-driven approach collects data and with enough data it would be possible to run machine-learning algorithms on the sensor data. Trends and correlations can be charted without having the domain knowledge usually required for such an analysis. The study on roro vessels used the physics-based model.
To model a digital twin of the drivetrain properly, contact analysis and power transfer through gear teeth contact should be considered. Researchers have proposed various methods to achieve this.
Vibration monitoring of the gearbox is performed by measuring vibration at the wheels and bearings of the gearbox and the generator. Depending on the component being analysed, different frequency ranges can be used. For low frequencies a position transducer is applied; medium frequencies employ velocity sensors, while high frequencies require accelerometers. When selecting a sensor it is important to evaluate the dynamic range and the sensitivity of the sensor. This is especially important for low frequencies, where the amplitude from acceleration can be small.
To evaluate the vibration severity, boundary layers defined by ISO 10816-1 can be used. RMS values of vibration in mm/sec are slotted in various zones depending on their levels: acceptable levels for unrestricted long-term operation; unsatisfactory for long-term operation and so on.
The study discusses various mathematical methods for fault detection. The authors say that using vibration data and transforming it from the time domain to the frequency domain opens up possibilities to see patterns in output that could be used not only to predict faults at the selected spot, but also elsewhere in the system. This is generally true in the model-based approach.
The data-driven approach has a sub-optimal performance benchmark and related alerts are then triggered. This approach involves collecting data points and comparing them to standards and recommendations laid down by the manufacturer and class.
The data model-based approach involves collecting data and considering correlations with faults. This approach requires large amounts of data and higher levels of competence in data science. It is a machine learning approach and could be used to detect issues and fault-driving parameters that have not yet been researched thoroughly.
A physics-based model, driven by data input, by utilising the equation of motion to predict behavior, is a good target for a digital twin, they authors say. Benefits can be obtained without generating large amounts of data sets, say the authors. “The physics-based model is more intense regarding domain knowledge and computational time. This approach has a model with universal validity and the benefits from collecting sensor data from anywhere on the twin is then achieved,” they say.
The steps towards building a digital twin involve collecting sensor data, going to a model that comprises the actual digital twin, and generating an output on useful life (RUL) and fault prediction. To achieve sufficient data collection one needs to have the correct sensor technology in place. Knowing where to place your sensors is essential.
To analyse the data the physical model uses equations of motion and fault-modelling algorithms. The output from the model is analysed for RUL and fault-prediction schemes.
To verify the method used in the study, a simple drivetrain test rig at the Marine Drivetrain Research Lab in NTNU was used. The rig employed optical sensors that detect displacement for vibration analyses.
* Digital twin condition monitoring approach for drivetrains in marine applications was first presented at the 2019 ASME 38th International Conference on Ocean, Offshore and Arctic Engineering by Sigrid S Johansen of Wilh. Wilhelmsen Holding ASA and Amir R Nejad of the Norwegian University of Science and Technology
© 2023 Riviera Maritime Media Ltd.