Skip to main content

 

 

USING AI TO PREDICT MACHINERY FAILURES

Tuesday, October 16, 2018 

US company SparkCognition has recently carried out a failure prediction study for a large shipowner. Failure was accurately predicted in a blind test using sensor data from main propulsion units and generators.

Walter Mitchell, Maritime Director of SparkCognition explains below how the company applied its SparkPredict program to the task.

Whether ship operators are transporting valuable cargo or ensuring thousands of guests have a safe, enjoyable cruise, it is critical for them to invest in systems and methods that limit their exposure to asset failure. Taking cues from other industries, operators are looking at sensor data generated by shipboard machinery. The stakes are high: Failures at sea represent high opportunity costs, are expensive to remedy, and damage an operator’s reputation with their customer base.

Unanticipated failures can have serious consequences for the safety of the ship, its crew, and its cargo. Going off-hire due to an unplanned out-of-service period can result in an opportunity cost of up to several hundred thousand dollars per day. If undetected, the cost to repair a failure may also cost hundreds of thousands of dollars, as it can even involve expenses such as chartering a helicopter to fly technicians out to the ship, or towage from sea into port.

In addition, machinery failures generate negative publicity for the shipowner, especially if it is a ferry or cruise vessel or a ship transporting dangerous cargo (such as an oil tanker near shore).

Failures are an all-too-common occurrence. In one recent case, a coupling in a compressor became misaligned, leaving the vessel out of service for three weeks. The total cost of repairs (including lost hire) came in at several million dollars. So how can data technologies mitigate these costs?

One industry-leading ship operator decided to take a proactive approach to predicting failure in shipboard machinery, spearheaded by the technical manager of planning and operations. At the time, the operator had no predictive abilities for failures. This lack of prediction left marine superintendents and engineers in a reactionary mode for repairs or replacements rather than addressing these changes strategically. By improving the diagnostics and optimising the repair supply chain, the operator expected to not only reduce expenses and loss of hire but prevent those expenses from occurring.

The ship operator’s technical management team looked into applying machine learning techniques to the considerable volume of sensor data it had accrued over years. After searching for a firm deeply involved in artificial intelligence, it selected SparkCognition to undertake the project. Leading reasons for SparkCognition’s selection were its focus on maritime operations and proven successes in machine learning and natural language processing capabilities in utilities, oil and gas, and manufacturing.

The project scope evaluated how sensor data could be analysed with AI to predict when selected engine room assets would fail. The operator was interested in identifying failures that would cost over $100K or cause the vessels in the scope to lose time on the voyage.

The operator targeted two critical shipboard assets on which they were already collecting data:

  • Propulsion motors:  Diesel-electric drives with fixed pitch propellers
  • Generator alternators: Units that generate electricity for the vessel.

SparkCognition was provided with historical data for several ships in a class and assigned the task of predicting failures at least two weeks in advance. The data science team constructed a series of predictive models using neural network autoencoders to differentiate between normal and abnormal behaviour.

Over 600 data tags were provided by the ship operator. The SparkCognition data science team identified a few dozen tags for propulsion motors and alternators that were relevant and available for the models. Once the data streams were identified, the team leveraged SparkPredict to develop a model using data previously logged from the sensors. The model characterized the normal operating state of the assets with a high degree of confidence and differentiated failure patterns.

For propulsion motors, SparkPredict predicted failures up to 10 months in advance and identified eight main features that contributed to this prediction.

For alternators, SparkPredict predicted rotor pole failure 6 weeks in advance and identified three main features that contributed to this prediction. In addition, SparkPredict predicted an ancillary failure, even without that machine’s operating data.

This precise level of diagnostic accuracy was possible because SparkPredict leverages explainable AI algorithms that identify and communicate issues occurring within the asset. For example, instead of just knowing that the alternator would fail, the algorithms can pinpoint the rotor pole as the source of failure and flag it for remedial guidance. Technicians can then use this intelligence to order parts and plan maintenance before major degradation of the asset.

With a successful proof of concept—where failure lead times were predicted considerably earlier than the goal of two weeks—this operator is now strategising the operational rollout of SparkPredict.

Reader Comments (0)

There are currently no comments on this article. Why not be the first and leave your thoughts below.

Leave Your Comment

Please keep your comment on topic, any inappropriate comments may be removed.

Return to index

Web Analytics