It was suggested in the past that if Henry Ford had asked his potential customers what they thought of the idea of a motor car, they would have probably told him to focus his efforts on getting them a faster horse.
The automotive manufacturing industry today is facing a quality control problem. Malfunctions in vehicle systems were responsible for hundreds of deaths and injuries in the last few years alone. Over 300 deaths were linked to a malfunction in the airbag deployment system, almost 100 people were killed as a result of unintended acceleration, and the horrible list goes on.
Some of these cases resulted in very high profile investigations and confidential settlement agreements. This is because testing procedures still rely on “horses”, and companies mistakenly seek to improve by looking for faster ones, when in fact they need to motorize.
Modern vehicles now rely heavily on electronic systems, giving rise to sophisticated problems accompanied by unprecedented negative consequences. In an attempt to tackle the problem, manufacturers strive to collect more and more data, create detailed test methods, and train engineers to identify and fix problems as they arise.
Any abnormality in end-of-line testing causes a halt in production, while engineers rush to identify (often, manually) the cause of the problem as losses begin to pile up. In some cases, the abnormality isn’t detected during testing, and damaged vehicles enter the market – leading to safety breaches, warranty claims, recalls, injuries, and sometimes death.
So is there a way to replace these inefficient “horses” with a “motor car” – an automated, smart, and efficient solution that would take the industry into a new era in quality control? Technology is now at a point where machines can be configured to learn from observations, just like humans, only better. What if we could replace one horse-power with hundreds of horse-power, but keep the driver?
Machine Learning to the Rescue
Traditionally, software was explicitly programmed to look for certain things based on a pre-determined model. As vehicle systems become increasingly more complex, it becomes virtually impossible for humans to model how each and every component should behave at any given moment.
Greta and her at team Acerta spent four years on applying machine learning for the purpose of anomaly detection. Their technology is built on the fact that severe malfunctions are very often preceded by minor abnormalities in data behavior. The software is programmed to scan through large amounts of data, create a model based on this data, and then use it to make predictions for future data.
The newly developed algorithm is unique in its ability to accurately detect hidden defects and malfunctions in real time, with an extremely low false-alarm rate. It far surpasses any human-based model that can be created and provides invaluable information to engineers. This will ultimately save manufacturers billions in recall, warranty, and downtime costs, and in many cases it will also save lives.
Disrupting the Industry
Prediction-driven anomaly detection is considered to be the most intricate type of data analysis to undertake; an algorithm that can predict failures with real certainty is extremely challenging to produce.
Some generic off-the-shelf machine learning algorithms, such as DATARPM and UPTAKE, might be able to provide some insight when working with various kinds of data. However, these types of solutions are not dynamic enough to handle vehicle systems, and tend to have unacceptable false-positive rates when applied to vehicle data.
For example, when a driver changes gear between Parking, Reverse, and Drive, all 3 are changes in the car’s state. A generic machine learning algorithm will most likely detect all of these as malfunctions, even though they clearly aren’t.
In the future of autonomous driving and fully-automated manufacturing, vehicle health cannot be left solely in the hands of humans. Acerta seeks to be a leader in autonomous health monitoring via smart algorithms. The algorithm and models are unique in their ability to swiftly and independently adapt to fresh data, even when the data sets are small and imbalanced (i.e. many examples of normal behavior but very few instances of true defects or malfunctions).
How Things Work
Acerta’s platform includes smart extraction of features from raw data. By analyzing recorded data, Acerta’s engineers are able to tweak the algorithm to take into account not only raw data, but also combinations of parameters and even transformations of the data. They examine different combinations to find specific characteristics of the data that are more indicative of problems. This makes their platform dynamic, adaptable to any vehicle subsystem and vehicle model, and is scalable to system complexity.
Since signal data is analyzed in real-time, engineers can quickly diagnose the severity and origin of anomalies, even within large batches of data. This leads to improved product quality, reduced downtime, increased efficiency, and better brand perception.
Timing is Everything
The perpetually increasing complexity of vehicle systems and the sheer volume of available data constitutes a compelling and pressing need for smart automation of product quality testing. Future manufacturers will not be able to rely on manual data analysis or generic multi-purpose AI solutions.
With platforms like Acerta, which integrate with manufacturers’ existing data collection systems using a dedicated API and cloud computing, they can now finally let go of their horses and jump on the motor-wagon.
If you liked what you read, please share it with friends.