Deploying ML into Mechanical Engineering - Three Applications

Introduction

In a Cloudforest article late last year, I took a look at how you can build data pipelines that capture, process and visualise data, focusing on how to design robust and scalable systems, but not talking so much about what else you can do with that data.

The right data can also be used to make predictions and decisions, using Machine Learning (ML) and Artificial Intelligence (AI), and this article looks at three ways these technologies can be used in mechanical engineering, R&D and product development projects.

ML and AI technologies are quickly finding application across a range of market segments, having entered the public conscious with high-profile applications in autonomous vehicles, drones and robotics systems. Open-source projects (Scikit-learn, TensorFlow) have made it possible for organisations of all sizes to embed ML and data analytics in day-to-day operations.

The Three Applications

All the applications here have a common theme: Models are developed to make predictions (by relating input data to an outcome), which would otherwise be impractical (i.e. difficult or expensive) to make through conventional means (e.g. from first principles).

I look at three cases - Generative Design, Predictive Maintenance and Control Optimisation - where ML/AI can be used to do something useful when designing, building or deploying a mechanical system.

Generative Design

Generative Design is an iterative design exploration process - A set of design goals and boundaries are defined, and an algorithm develops and evaluates a large number of design solutions, in the process learning what works and what doesn't. ML powers the design algorithm, creating unusual designs not easily conceived by a human designer, eventually arriving at a well-optimised solution.

A class of machine learning frameworks called General Adversarial Networks (or GANs) are often used for this task, as they train themselves to generate better and better solutions by comparing iterations in a goal-seeking exercise.

With generative design it is possible to (rapidly) arrive at unusual and novel solutions that precisely balance competing design requirements. Computers are able to evaluate and compare thousands and thousands of different designs in very short space of time, and the development of additive manufacturing has liberalised production constraints, allowing algorithms yet more degrees of freedom in design.

Generative design is now finding adoption in the design of components that have really well-defined design objectives, and need to balance strict cost, weight and loading requirements - Segments such as automotive, aerospace, industrial machinery and construction have all started to deploy this technique in component design.

Predictive Maintenance

In predictive maintenance, a model is developed to estimate how much lifetime a machine or system has before an intervention may be needed (i.e. before the engine breaks down!). This lifetime is expressed in operational units, which can be hours of operation/duty cycles/some other repetition, and they are used as a guide on when to next schedule maintenance.

Predictive algorithms are often deployed in the cloud, receiving data from the operating machine, which they then process and forward to operators as a remaining life estimate. These cloud-based systems can also alert operators to unusual behaviour, or even remotely shutdown systems before damage occurs.

Predictive maintenance can be applied to a wide variety of systems and devices, from factory machines to jet engines, cars to wind turbines, and seeks to reduce the cost of owning and operating machine assets by simultaneously reducing unnecessary maintenance and minimising unscheduled down-periods.

I wrote an in-depth article about developing an onset-to-failure model for an aviation turbofan in an article last year -https://www.cloudforesttechnologies.com/post/predictive-analytics-implementation-strategies.

Control Optimisation

Control optimisation with ML involves developing models to support or enhance control operation, reducing operating costs and improving operational performance and efficiency.

Examples include using ML models to replace temperature sensors in a motor (I've written about this too), assisting with flight stability and turbulence response in a drone control system, or tuning engine firing order in a cylinder deactivation strategy (for emissions control and efficiency).

In each case, the idea is to train a model to make quick and accurate predictions, which are then used as inputs to the main control strategy. Such models are deployed on local hardware, embedded in the control system itself, and must be small and fast.

ML for control optimisation will often improve the performance and efficiency of the parent machine (eg. the engine, motor, drone) by generating better inputs for controller decision-making. Additionally, using predictions to replace physical sensors reduces BoM costs and can boost reliability. However, issues can arise when real-world conditions move outside the model's parameters (i.e. An input temperature goes way off the scale).

Implementation comes from identifying an AI-augmentable control function, and then training a model with a suitable target and dataset. Once an accurate model has been developed, evaluation can be undertaken by Software In the Loop (SIL) and Hardware In the Loop (HIL) techniques - A good model needs to be sufficiently small and efficient to work on a microprocessor.

Summary

A few of months ago, I looked at what a data pipeline is, but I didn't talk about usage opportunities in machine learning. This month's article takes that on, and we've looked at how ML can be applied to engineering problems through three use cases: Generative Design, Predictive Maintenance, and Control Optimisation.

Although each application differs, the common thread is that a multi-input model is trained to make predictions, and these predictions are used for design optimisation, reducing maintenance costs and improve reliability, or enhancing machine efficiency or capability.

Thanks for reading, and as always please get in touch via LinkedIn, or our website, if you'd like to know more.

Paul