By Steve Schwinke, VP of Customer Engagement at Sibros, winner of Cloud Awards 2021/2022 Disruptor of the Year
Imagine stepping into your car and based on your voice pattern it adjusts your seat, steering wheel and mirror position, temperature, and even starts playing your favorite playlist. It knows that you prefer the cabin temperature hotter in the morning and cooler after your evening gym session. It has learned your frequent hole-in-the-wall coffee shops, so it only recommends local brews when you visit a new city. Based on your driving patterns and scheduled calendar events, every day on your commute it preemptively offers route suggestions to avoid traffic. Imagine owning a vehicle that anticipates your needs, a vehicle equipped with machine learning.
Machine learning might have started as a simple training tool for artificial intelligence, but it has become so much more. Along with supplementing business models, search engine algorithms, and speech and facial recognition, machine learning has enormous implications for the future of the automotive industry.
Breaking Down Machine Learning
Machine learning is the process by which computers learn to program or adjust themselves based on previous experiences and parameters. Programmers begin by collecting, sorting, and labeling training data, which is either structured, such as tables, numbers, or equations, or unstructured, including text, photos, or videos. The data must be comprehensive and representative of the entire focus group to avoid human bias discrepancies. Once the training data is compiled, the programmer selects their desired training model: supervised, unsupervised, or reinforcement.
These models utilize labeled datasets provided by human programmers and are most often used for optimization purposes. In other words, a programmer knows what they are looking for and is trying to narrow down their results. Let’s say a manufacturer wants to teach a machine to monitor electric vehicle battery health. First, the programmer teaches the machine about normal parameters via labeled data. Eventually, the machine learns to identify normal parameters on its own. The problem with this technique is the machine won’t always be able to expand upon the subject. In other words, the machine would be capable of identifying a battery issue, but not the cause.
Unsupervised models are better when people are unsure of what they are looking for. Instead of labeled training data, the programmer exposes the machine to unlabeled data and allows it to find patterns on its own. Back to the battery example, a manufacturer could utilize unsupervised ML to help discover the source of the drain by allowing it to look for related trends. For example, the unsupervised algorithm might notice that the body control module (BCM) remains in the active position while the vehicle is off. Left unchecked, this type of fault will deteriorate the battery’s integrity.
The downside to unsupervised ML is the potential for irrelevant pattern discovery. Let’s say that the same vehicle is owned by a resident of the hilly city of San Francisco. Every night the driver turns the steering wheel to the right so the tires face the curb in case the parking brake fails. The machine might identify a trend between wheel position and battery power drain, even though the two are not connected. However there may be some additional correlation. If the driver uses the seat position motor to help them exit the vehicle when parked on a hilly street, it’s possible that the adjustment may keep the BCM active and that may have caused the unexplained battery power drain.
This model utilizes trial and error to encourage the machine to choose the best option. Machines make their own decisions and learn over time based on a combination of positive feedback and rewards. The pitfall to reinforcement ML is it requires a lot more input and attention on the part of the programmer.
Once a method is selected the programmer feeds most of the training data into the model and allows the machine to teach itself. It is vital to exclude some data during initial training as this will be used to evaluate the machine’s progress and test its ability to adapt and respond to novel information. All three models are malleable and provide programmers with the ability to adjust or add parameters to increase ML accuracy.
Where Is the Edge?
Bringing machine learning and data processing closer to the data source—the vehicle—has many benefits. It decreases bandwidth needs, reduces network latency, and overcomes cost limitations. This is also known as “the edge.” For the automotive industry, the ultimate goal is bringing the edge to the vehicle itself, yet there are too many unknowns to make this a practical solution.
One of the primary hurdles is enabling the vehicle with the power required to support ML. The limited memories and processing capacities of current options make this cost prohibitive. Another issue is adaptability. Static solutions within the vehicle are not equipped to keep up with the rate of automotive technological innovations.
The cloud, however, can grow and change with the increasing demands of the automotive industry. It has the storage and processing power required to make sense of the endless supply of vehicle data. Making this edge more efficient and effective is the next logical step in the evolution of automotive ML. This is where the vehicle comes in. Offloading some responsibilities onto the vehicle helps maximize cloud capabilities. Instead of waiting for data to transfer to the cloud, OEMs can utilize flexible data logging at the source to minimize transfer size and storage costs. By combining intelligent data collection with ML in the cloud, OEMs can rapidly realize new vehicle use cases, identify system faults, and innovate new and desirable features.
ML and ADAS
There is an almost limitless suite of advanced driver assistance systems (ADAS) features that can be enhanced by machine learning. Some of these include:
- Lane detection and correction
- Object detection, classification, and response
- Parking assistance
- Lane change assistance
- Backup assistance
- Parallel parking assistance
The way autonomous vehicles utilize machine learning mirrors that of the human brain. There are multiple levels of machine learning algorithms with designated functions. These interact and cooperate to perform advanced tasks. Let’s look at an object recognition example. The camera on the vehicle registers something in the road, one system determines what the object is, another the distance from the vehicle, and a third the steering correction necessary to avoid a collision. The third system also receives input from other systems, such as side and rear cameras, in case it needs to apply the brakes or move the vehicle into another lane.
Enhancing the Customer Experience
Most people don’t buy a new car because their old one doesn’t work. They buy it because they want the new features included with next-generation technology. Machine learning has the potential to enhance the customer experience by allowing the vehicle to grow and improve with current trends.
Each driver creates a user profile with their preferences, which are used as a baseline for the vehicle’s ML algorithms. Let’s say you buy a new car and you input your desired cabin temperature as 70 degrees Fahrenheit. This is fine on most days, but whenever the outside temperature is below 50 you pop the internal temperature up by a couple of degrees. You might not even notice you are doing it, but your car does. After a while, it makes the connection between the external temperature and your adjustments. Before you know it, your car is increasing cabin temperature before you’ve even noticed.
Personalization is only one aspect. Based on brake usage and driving behaviors, machine learning can anticipate fuel or EV battery needs. Although few people enjoy a backseat driver, ML will also be able to offer driving suggestions and alterations to improve fuel and battery efficiency.
Perhaps the highest level of machine learning is the ability to predict vehicle maintenance and service needs. Today’s maintenance alerts are based on rudimentary factors, such as miles driven or time between services. Once these numbers reach a pre-specified point they trigger a change oil light or a reminder message on the infotainment screen. The glaringly obvious issue with this setup is that time and distance are not the only factors that impact the vehicle’s health and maintenance needs.
Let’s look at an example. A 12V battery doesn’t fail because of the battery itself, but rather because the rest of the system is impacting it in a way that it shouldn’t. As a system’s complexity increases, so do the chances of unanticipated side effects. Unsupervised machine learning can identify the sources of the energy drain because it utilizes real-time data from every vehicle’s ECU to look at dependencies, discrepancies, and malfunctions. The ultimate goal of this type of approach is to identify and address issues proactively. Intelligent maintenance increases vehicle longevity and prevents unanticipated breakdowns.
There is, however, one missing piece to this well-oiled machine. Once the ML system identifies the vehicle’s needs, there is still the matter of the owner to follow-through. One survey showed that 64 percent of people ignore a check engine notification for at least a week, while almost 30 percent wait up to a year before taking their car in. Similar data exists in regards to vehicle recalls. The result is hundreds of thousands of unsafe vehicles on the roads.
This issue is largely solved with a vehicle-wide over-the-air software update and data management system. With this type of embedded firmware technology, manufacturers can correct software errors and address software-related recalls remotely. Instead of waiting for the owner to return to the dealership, the OEM can send an update package directly to the vehicle and correct the battery drain with the push of a button.
What if a vehicle doesn’t have cell reception to receive the update? Or the owner is on an extended vacation? Well, that’s another job for machine learning. Instead of wasting time and energy trying to connect two dead ends, ML can be trained to utilize vehicle data and information from cellular network providers to optimize update delivery by targeting active vehicles only. Meanwhile, another ML algorithm monitors inactive vehicles and follows up with update delivery and installation once they return to an active state. Implementing machine learning in this way maximizes recall and repair completion rates, saving OEMs and end-users money while increasing road safety and vehicle functionality.
A Single Source of Truth
Although the automotive industry still has a ways to go before machine learning becomes a vehicle standard, the ability to remotely update vehicle software and manage real-time data is already happening. Embedded firmware and SaaS solutions, like Sibros’ Deep Connected Platform (DCP)*, bring automotive OEMs a step closer to successfully integrating and utilizing machine learning.
This type of solution unlocks precision data from every vehicle ECU to enable OEMs with a comprehensive picture of vehicle functionality and eliminate limited narrative decision-making. It also allows manufacturers to perform deep-level diagnostics and implement system and software improvements as needed with OTA updates. These robust tools are the driving force behind today’s automotive innovation and will serve as the foundation for ML training, implementation, and success.