Predictive Vehicle & Parts Maintenance

Predictive vehicle and parts maintenance is a machine learning use case with a wide range of applications. Predictive potential failures and the future health of vehicle and parts. It is used by rental, leasing, or fleet management companies to optimise when they should perform maintenance activities. It is as well used sometimes by Insurance companies to reduce their payout risks.

Predictive Maintenance programs normally try to balance the benefits of performing a maintenance activities, such as increase the asset lifecycle, decrease planned downtime, or reduced insurance premium with the cost of maintenance, be it the time and cost it takes to perform general maintenance activities or the cost of spare parts. This taking into consideration other factors such as safety, regulations, capacity etc…

An effective predictive maintenance program is usually more than just prediction, and goes into embedding an effective decisioning process around an asset lifecycle. By themselves a prediction that an asset has 5% chance of failure over the next year, doesn’t mean much. It needs to be embedded into a decisioning context, for instance does the chance of failure is close to pose a safety concern, does the vehicle rank in high priority in terms of maintenance vs. the rest of the fleet, do we have free available maintenance capacity that could be used to reduce the maintenance cost, …

Predictive maintenance allows companies to improve upon a baseline, typically a scheduled maintenance framework, but requires predictions to sit within an over-encompassing decisions framework to prove effective.

Predictive maintenance, typically relies on a mix of time series data and cross-sectional data in order to drive predictions and decisions.

One of the typical source of data for predictive maintenance tends to be IoT/Sensor data. Sensor data is information captured by sensors in IoT devices, converting physical attributes like temperature, motion, or light into electrical or digital signals. It provides a sequential recordings from sensors capturing changing attributes over time. This can be used to understand the health of vehicle, equipment or parts. In terms of data management, IoT/Sensor data is typically its own category — high volume, high throughput and low latency.

For vehicle maintenance, information such as speed, acceleration, milages, breaks, temperature sensors can be important to monitor and feed onto predictive maintenance applications.

In order to make the best use of IoT data, it is required to architect data processing applications to be fully enabled for real-time stream processing. Technologies such as Kafka, EventHub and Databricks delta live table provide a good technological background for achieving this type of challenge, be it by providing the ingestion capabilities or the processing framework convert noise onto signal.

To get started with predictive vehicle maintenance, however such a vast array of data might not be necessary, depending on the use case. Panel data measurement, which provides repeated measurement of a few variables across a period of time might prove sufficient. These type of measurement can for example be measurement typically obtained through scheduled maintenance activities.

Predictive maintenance, typically looks to build metrics and KPIs using machine learning models such as:

• Mean time to failure
• Mean Time To Repair
• Failure Propensity
• Remaining Useful life prediction (RUL)
• Sensor measurements

Predictive maintenance, typically tries to address failure of two types 1) Sudden failures 2) Failures due to Wear and tear.

Predictive maintenance relies on a wide variety of machine learning and statistical methods to be able to predict these metrics, taking on from regression methods such as Random Forest or Gradient boosted trees, classification methods, survival analysis, time series analysis or building digital twin.

Besides just focusing on predictions, some models will need to as well incorporate a variable for the treatment. For remaining useful life prediction for instance, what is often needed to know is how much additional useful life do we get by performing a maintenance activity now.

The decisioning stage is meant to take as input the different predictive metrics generated, merge them with other datasets and combine them in a decisioning process. The decisioning process stage might in turn make use of additional advanced analytics models such as ranking algorithms.

Within this stage, it is important to map out the decisioning actions, objectives and constraints, as these will dictate what data will be required and how the decisioning process will be implemented.

The above diagram maps out an example action, objective, constraints in the decisioning canvas. In the example above:

• 4 types of actions have been defined, continue driving — or essentially continue without maintenance activity, repair if a slot becomes available — essentially trying to capitalize of idle capacity to reduce cost, schedule a maintenance slot as soon as possible and stop driving as the vehicle has become unsafe to drive.
• The objective of the decisioning process is to maximize the net useful life of the vehicle — that is the value of the remaining useful life of the vehicle net of the maintenance cost.
• Two types of constraints and been laid out, one for safety reason (can’t drive if propensity to overheat >10%), the other due to business processes looking to slot in the available garage / maintenance slots over the next 2 weeks.

The decisioning will need to be able to take in a number of data points, and using these constraints and objective boil it down onto this set of 4 actions.

Wise Analytics is a data consultancy company comprised a team of seasoned data experts, including data engineers, data scientists, and data strategists. Wise Analytics is well-equipped to navigate the complexities of implementing end-to-end predictive maintenance applications. As a preferred Databricks partner, the company leverages cutting-edge technologies and tools, ensuring a robust foundation for processing vast amounts of IoT and sensor data in real-time.

WiseAnalytics’ team brings a wealth of experience in developing predictive models and in applying decisiong capabilities tailored to the specific needs of predictive maintenance. The expertise extends beyond mere predictions, as Wise Analytics specializes in integrating treatment variables into models, providing actionable insights into the impact of maintenance activities on asset longevity as well as mapping out the decisioning requirements, aligning them with the overarching goals of the maintenance program.

In collaboration with Wise Analytics, businesses can confidently embark on a journey to implement state-of-the-art predictive maintenance solutions, combining predictive modeling prowess with sophisticated decisioning capabilities to achieve operational excellence, minimize downtime, and optimize maintenance costs.

Extra Link: https://www.databricks.com/blog/2020/08/03/modern-industrial-iot-analytics-on-azure-part-1.html