AI/ML in Mechanical Engineering Projects

AI/ML in Mechanical Engineering Projects - A Motivational Booster

During my 50 years experience in system analysis, patented inventions, project management and business development, I realized that Mechanical Engineers (MEs) are excellent project managers of Mechatronics product development. They are highly capable of integrating machine controllers and sensors into their mechanical products and with a good multidisciplinary project team make it all work within the specs, on time and in the money. However, when it comes to Artificial Intelligence (AI) / Machine Learning (ML) most traditional ME's are quite helpless.

AI/ML may sound like a rocket science to some good old ME designers and business developers. The objective of this article is to discuss why it shouldn’t and boost the apatite for additional learning. 

We start with a simple definition of AI/ML in mechanical engineering - “The use of real-time data to improve machine performance through an optimal decision making process in uncertain environment” Boaz Eidelberg.

For example- A turbine in a power plant vibrates excessively. Should it be stopped for maintenance or reduce rpm and continue generate electricity?

The answer to this question depends on real-time information such as, for example, the following data:

Operational Variables

• Vibration amplitude
• Vibration frequency
• Bearing and shaft temperatures
• Input and output pressures
• Steam mass flow rate, etc’

Environmental Variables

• Floor vibration
• Ambient temperature, etc'

Design Parameters

• Shaft diameter, length
• Number of Blades
• Shaft Inertia, etc'

Data transmission needs sensors and communication, which MEs are much familiar with. Other mechatronics components, such as PLC, motion controllers and software, MEs know just as well. However, the uncertainty starts with optimal decision making process which require statistical modeling considerations. These are quite vague. The combination of all these aspects is what constitutes the subject of AI/ML mechatronics. Which in essence, is a machine, designed by MEs, which seeks an optimal performance within its design parameters, operational conditions and uncertain environment.

In our example, we assume that we have sensors, which transmit the controller, in real time, the signal values of the operational and environmental variables. The next question, in the decision process, is as follows: Based on all the real time data that we know of, what is the expected gain if the machine continues to operate versus the expected gain if the machine is stopped for maintenance? If we can calculate the total expected gain for each option the intelligent, common sense, decision would be to choose the option with the highest total expected gain.

To calculate the total expected gain for the decision option "continue", we need to know the probability that the machine will continue to operate fine, and multiply it by the value of the electricity generated. That is the expected gain of this option. We then need to subtract from it the expected loss, which is the product of the probability that the machine will fail (1- the probability that the machine will continue its operation fine) multiplied by the cost of downtime and repair. This is the total expected gain of the option to continue the machine operation. We now have to compare it with the total expected gain if we stop the machine for repair. That is equal, at a probability of 1, to the cost of repair plus the cost of downtime. 

The unknown parameter in these "Total Expected Gain" equations, each of which is also referred to as the Von Neumann–Morgenstern utility function, is the probability that the machine will fail given the set of operational, environmental and design values. Therefore, the next question in line would be - what is this probability?

To answer this question we need a model that relates the probability of turbine failure to its operational, environmental and design variables. And here is where machine learning comes into play. We write a neural network (NN) model, as shown in the above picture, where the input on the left side is a vector of some operational, environmental and design variables, and the output on the right side is the result, which in our case is the probability of turbine failure if it continues to run. In between the input output, inside the NN, there are neural network, inter correlated, weights of the process variables that add up, through successive matrix and vector multiplications, so called Forward Propagation, to yield the predicted output. These weights continuously improve as we update the model with more samples of known examples and their known results. We will try to clarify it in the following paragraph.

Let’s say, for example, that at a known power plant location, somewhere in the world, a turbine has failed (i.e stopped working with a probability of 1) and it happened at a certain set of input variables. We may now run our neural network model in a cloud (e.g. using AWS, IBM, Google or MS Azure, which are mostly free to start with). These clouds may input and process in real time millions, if not billions, of IoT signals, calculate the resulting predicted probability of failure which our NN model yields, and compare it with the real probability of failure, which was 1. Initially the NN weights are being guessed at and therefore the error in our prediction will be large. However, after each ML run, by using one of the many standard AI/ML languages in the cloud, such as Python and Tensor Flow, the gradient of the error is automatically calculated and the NN weights are being changed in the direction of minimizing the error, a so called Backpropagation process, where larger error gradients result in larger weight changes. In this way, after many runs, the model is expected to converge to the right representation of the probability of turbine failure as a function of its operational variables, environmental conditions and design parameters. Since many power plants in the world use steam turbines, a collaborative collection of datasets from many sources may assist in building reliable databases for developing such applications. In an analogy to this example, any high volume mechatronic device, as small as a rehabilitation exoskeleton, may use such AI/ML features and aim at an optimal performance, such as minimal rehabilitation period at minimal pain. Therefore, the opportunities in AI/ML innovations related to mechatronics product development are endless.

As mechanical engineers, who may be responsible for managing the development of innovative mechatronics solutions to real problems, we don’t necessarily need to become software or communication specialists, nor do we have to have an additional electrical engineering degree. What we must have is a good understanding of the machine we develop, including the relationship between our target machine performance and its design parameters, operating variables and environmental variables, through modeling, simulation, prototype testing and real machine testing. We need to be able to define the objective function for performance optimization, such as maximum productivity, minimum downtime, maximum precision, etc' and be able to present our requirements to our AI / ML team members to develop the proper AI/ML model. And there are many options which an expert would most likely know of. Then, with a good multidisciplinary team we will most likely run the AI/ML mechatronics project to a successful completion.