According to Clarksons Research, spares and maintenance costs in the shipping industry have been steadily on the rise and can constitute approximately 18% of a vessel’s Operating Expenses (OPEX).
Costs due to unscheduled repairs for equipment failures and spare part unavailability can adversely impact total maintenance expenditure. Specifically, it is observed that breakdown repairs can be significantly higher than planned ones due to increased acquisition, service and forwarding needs. In addition, all maintenance cost elements can further deteriorate as the vessel ages.
Within a volatile tramp market, where rates tend to fluctuate wildly across all segments, i.e. tanker, dry and container fleets, the need to set maintenance costs under tight control is indisputable as it is an urgent matter for ship owners. The current legacy approach implemented by Technical departments in rightsizing such costs is a combination of a preventive and a corrective maintenance system. Crew on board performs inspections and actions, such as cleaning of main components, based on specific running hours or calendar intervals. In case an equipment fails, it is repaired or replaced provided sufficient stock exists onboard. Periodically repeated maintenance actions, such as cleaning of scavenging area and lubricating of main components, keep equipment in decent working condition but an adequate prevention of machinery failures cannot be guaranteed.
The proper processing of the data which results from monitoring systems and after factoring various parameters, e.g. if a spare was sourced via a maker or parallel market supplier, the vessel’s age and prior maintenance history, the onboard crew’s experience etc., can yield a much more accurate assessment of a component’s condition. Such a framework would contribute in the prediction of potential equipment failure, would prevent it through dynamic adjustments of the maintenance plan and the subsequent reevaluation of the P-F curves (Potential Failure-Functional Failure graph).
From an organizational perspective, adhering to a robust predictive maintenance framework is a cross functional effort necessitating active participation from critical front-line departments such as Technical, Purchasing and Safety & Quality. Collaboration would be enabled via a structured process that entails stakeholders across departments reaching consensus on a maintenance forecast for the coming 12 months, performing strategic sourcing initiatives to secure spare part supply and service engineer availability and booking forwarding capacity to ensure timely delivery in a cost efficient way. The set-up would be further completed through the monitoring of specific KPIs such as forecast accuracy, volume discounts achieved, breakdown ratio etc.
Properly identifying the equipment that will partake in such a bulk ordering process is always a challenge for shipping companies, in particular for those whose sister-vessel overlap is small. The Planned Maintenance System (PMS) houses important information on equipment history such as frequency of work orders, defects, downtimes. In order to be able to efficiently handle this very granular data, an aggregation of equipment into larger categories or “forecast groups” is needed. PMS contains the full function hierarchy of components per vessel with additional details such as the maker and various interdependencies which can further help in the segmentation effort. For example, an easy rule of thumb commonly is: all vessels carrying Diesel Generators built by the same manufacturer are sister vessels and belong to the same group.
Data analytics, in the form of unsupervised learning, has a pivotal role to play in properly identifying the forecast groups to focus on within a predictive maintenance framework. Classifying components in sizeable groups, e.g. main engine, diesel generators etc., enables shipping companies to untap strategic sourcing initiatives, increase bargaining power against suppliers and gauge significant volume discounts.
Further deploying clustering analysis within each forecast group, the underlying spares can be further classified into specific segments based on similar attributes e.g. maker, vessel category, price and number of unique vessels requiring the same part. For instance, a Diesel Generator forecast group can have numerous segments highlighting a varying degree of criticality which can then be used to define inventory policies per spare part in alignment with global policies, such as TMSA 3, or other regulations.
From that point onwards companies can switch gear into supervised learning, and deploy predictive models to predict future demand of each spare part cluster under each forecast group. Data availability and degree of automation could dictate the underlying method, i.e. Random Forest if data is scarce or Deep Learning with enough breadth and depth of time series data. In any case, any ensemble method approach could benefit from utilization of exogenous variables such as equipment running hours, onboard sensor data, vessel age, crew experience etc. The end result would be the forecasted quantity per group, disseminated to the underlying critical spares.
The analytics framework around predictive maintenance is enhanced by the inclusion of a prescriptive model that leverages all associated costs to “prescribe” the consensus forecast to be procured. These models pit the predictive model quantities against inventory, forwarding, stock-out and acquisition costs and use operations research algorithms to optimize the exact spare part quantities that ensure the commercial availability of the vessel whilst minimizing the associated upstream supply chain spend.
To close the loop of a differentiated predictive maintenance framework, high frequency telemetry data can be consumed in real-time by the predictive models to constantly improve forecast accuracy and lead to the recalibration of the P-F curves. Telemetric systems onboard can amass and transmit automatically data pertaining to high temperature detection, oil analysis, vibration recording etc. Feeding that data to any machine learning model can guide the future redesign of the P-F curves and can lead to different conclusions regarding the systems, equipment, spare part demand and time intervals that need to be included in the predictive maintenance plan.
For instance, if sensors installed on main engine flowmeters indicate high fuel consumption for a prolonged period of time, maintenance measures will be triggered and equipment will need to be reevaluated both in terms of quantity and quality of spares, running conditions and hours as well as service engineer capability.
The aforementioned predictive maintenance approach can yield significant benefits for the shipping companies that choose to deploy it. We expect that the first phase alone, which doesn’t necessitate any advanced analytics or telemetry technologies, can lead to an approximate 15% reduction in acquisition costs through volume discounts from vendors. Deploying phase 2 can further bring down the entire supply chain costs by an approximate 10%. Whereas phase 3 can additionally reduce maintenance costs by 7%. Furthermore, this predictive maintenance framework can be transformative not just in terms of costs, but also in the way maintenance is perceived and performed by the office and the crew onboard leading to less firefighting and stress, less rescheduling, better upkeep of working hours and enhanced resiliency across the entire organization.