Engineers without signal-processing expertise can now employ reusable artificial intelligence (AI) models to quickly and accurately extract and select features for audio processing, predictive maintenance, and other applications in automotive, aerospace, defense, and industrial automation.
For instance, design engineers can use trained AI models to recognize and identify individual speech commands in an audio recording instead of relying on a time-consuming manual process. Johanna Pingel, product marketing manager at MathWorks, provided a walkthrough of the AI-driven system design workflow during a talk with EDN. She outlined the following four basic steps of the AI workflow.
In education and research, engineers spend a lot of time tweaking their models while playing with all the parameters. However, when they get into real-world designs, they spend a lot of time on data processing.
“Data cleansing and preparation represent most of the AI effort,” Pingel noted. “This is where tools like automatic labeling save engineers time while they are dealing with millions of images and can’t label every single image manually.”
These tools automatically generate synthetic data to improve data search and help automate the labeling process. Tools like Simulink enable engineers to generate synthetic data, improve datasets, and save weeks to months in signal-processing applications, Pingel added.
Once engineers have the hardware to accelerate, they can start with a complete set of algorithms and pre-built models. The AI modeling stage is more about being able to iterate on model design and understand the tradeoffs of the models.
“For non-experts, modeling helps in training, tuning, and visualizing the AI models,” Pingel said. With interactive point-and-click tools to iterate on the AI models, engineers can change the model’s input size by changing the parameters, which is a great time saver, especially for engineers who don’t want to code all the time for the models.
For example, the Classification Learner app tries different classifiers and finds the best fit for the dataset. Likewise, the Experimental Manager app runs deep learning experiments to train networks and compare results.
Simulation and test
The AI models need to exist within a complete system to integrate AI into a system-wide context. That also calls for simulation before engineers move to hardware and verify its effectiveness. “It’s all about making sure that the AI model works before engineers deploy it in the real world,” said Pingel.
Engineers must try all of their models before deployment to ensure that everything will work the way they expect. Take the case of a hydraulic pump for which engineers can simulate all different inputs and outputs to determine if and when things will fail.
Once AI models are ready for deployment, engineers should be able to deploy them to whichever device they have chosen, automatically generate the code, and run code efficiently on that platform.
Here, according to Pingel, being able to deploy to any processor without inheriting any coding errors is the key goal. “Once engineers are ready for deployment, they can deploy anywhere, anything between the cloud and FPGAs.”