Testing is often the most critical phase where vehicle systems matures! Most of the time , it is the testing maturity achieved that decides the product quality, reliability, and performance. But this is also the bottleneck for the entire vehicle development phase because of which the development time of the project is longer.
As the organizations try to keep up the competition and the demand of the customer, they are always on a lookout to reduce the development time! In the last decade, a lot of modularization, standardization has played a critical role in achieving this. But with the digital disruption literally sweeping across the functions forces the organizations to look for alternate ways to still achieve the same reliability, quality outcomes of the conventional testing methods however with the lower development time!
Data to the rescue:
Data driven test and simulation methods not just deals with the above constraint elegantly but thrives on! It is opening new value creation opportunities and influencing the way we work across multiple dimensions. Significant growth of AI/ML approaches, and organizational cloud adaptation is just acting like a catalyst to this situation.
The question remains is how to embrace this and ride this AI/ML wave to solve the very fundamental problems in the industry.
The above question basically translates into how to do it? What all the skills and approaches are required and most importantly how to drive the organizational acceptance and adaptation. I will highlight and discuss some of the opportunities in this space.
Testing of vehicle system- How to leverage AI/ML
Once the system specification is developed, the test cases are developed by the system engineers, and test teams at various levels like unit level, system level, vehicle level so forth. Though test automation is a well studied area and various tools provided by the tool vendors but still this automates the process of testing. But majority of the time and skill is consumed in test case generation and analyzing the test results to identify the root cause. To facilitate automatic test case generation, automatic analysis AI/ML can be used which are discussed below.
Automating test cases generation (AI/ML Approach)
Conventional approach :
All the system requirements in the automotive ECUs exhibit kind of entity relationships across the modules. Generally, test engineers prepare the test case against this also, they will re-use the test cases which are evolved through the lessons learned.
AI/ML approach :
With the advancement in the NLP ( Natural language Processing) , Text mining , and advanced machine learning algorithms from the specification document, it is possible to generate the graphical view of affected entities and simulate the effects of changes in one entity with respect to others. Using these relationships, test cases are mere rule generation exercise hence automating the test case generation.
Not only that, we have achieved the “executable specification” using which simulation is possible. This is no way replacing the human expert use case generation, rather it is easing or aiding the testers to achieve better performance.
Automating the test analysis & root cause analysis (AI/ML approach)
Conventional approach :
Lot of log data is getting generated during testing. We can declare the results of the testing as pass/fail automatically but finding the root cause and propagation of one failure mode and its manifestation of the same needs a deeper technical and domain expertise most of the time this is more time consuming.
Expert time is better utilized if the overheads on the data processing is reduced and prompting the expert to comment on the unexplained but important relationships. AI/ML based methods not just achieves that but also offers the automation of root cause analysis capability.
Using the AI/ML methods root cause analysis activity is solved. Also, for the deeper understanding and explainability reasons Probabilistic graphical modelling leveraging the Bayesian networks and Hidden Morkov model-based approaches are used.
The incubation of AI/ML practices can happen at different levels to drive the organizational acceptance starting from basic descriptive analytics to predictive, prescriptive analytics.
Advanced Driver Assistance systems and autonomy has been an ongoing theme in the industry, across geography. It is fair to say that without the help of AI/ML systems, it is impossible to deliver the ADAS systems. But geographic constraints make ADAS problem unique to the market. For example, complexity of the Indian market is completely different from the other markets like Europe or USA hence there is no one fit all solution.
ADAS Testing Bridging the impossible
Developing the perception layer in the ADAS system already using the advanced computer vision, Deep neural networks which are very data hungry applications. Testing of these ML models to be done using the unseen data samples which makes the overall system extremely data hungry. Practically data logging across diverse scenarios are very complex and costly and sometimes even impossible. Imagine a data collection of situations ranging from foggy, mist, rain, various sun positions, air quality and dust reasons. Synthetic data generation and scene generation seems to be the leading approach to solve this problem.
With the evolution of GAN (Generative Adversarial Networks) and advanced deep neural networks methods it is possible to simulate data which are required for the testing of the ADAS perception layer. More so, with the minimum data, you can scale up to meet the high data expectation. More so training ML models can be controlled or orchestrated to deliver the required based on the available data. As they say garbage in garbage out is the old saying which is true. With the advancement in the ML algorithms and cloud systems allows us to realize “Minimum in Maximum out”.
Complimenting the ADAS: Naturalistic driving study
Fundamentally ADAS implementation focuses mostly on offering the “assist ” feature to the driver with the intent to achieve improved safety and comfort to the driver. Basically, it involves various perception and control systems working in a coordinated way. The focus on the driver is still not fully leveraged. There are many use cases possible by including the drivers state in the loop. Discussing one such solution as below which we have developed in house.
Anticipating driver maneuver:
Driver’s the visual cortex is far more powerful than the various perception systems in the ADAS systems then why not leverage that! With that intent, we have developed the in-vehicle driver monitoring system which observed the driver facial cues and eyeball movement and predicted the possible actions (we considered 6 different actions basically capturing turns and overtaking combinations). This solution we can predict the diver’s actions on an average 4 seconds before the actual actions happen with the very good accuracy levels around 85%. The development of the system uses Recurrent Neural Network (RNN) based methods.
Combining the driver maneuver prediction information into the ADAS control strategy has a great potential for improved safety and reliability of the vehicle systems because the better control of the actuators is possible if the “action lookahead” is available. AI, ML methods make this possible.
Look ahead control of maneuver:
In the previous section, we discussed using the “Driver’s state” in ADAS application .With the connectivity revolution going on, we can leverage connectivity to even push the boundary for the ADAS there by experience of the driver.
ADAS perception layer is having the limited horizon which is limiting the overall performance boundaries. The ADAS perception horizon can be extended using the connected maps and 3D mapping information from the various map service providers. However, over all contextualization of the data from the map service providers, along with the vehicle trajectory, ADAS systems need intelligent services to be developed which we refer to as “Geo enabled ADAS services”.
This allows us to leverage the environment information’s beyond the perception layer horizon! Combining look ahead scene understanding using the geo maps, along with the levering driver’s intent makes an ADAS systems more safer and drivers experience is greatly enhanced as well.
This overall data orchestration, data stitching and consumable insights in the in -vehicle systems is possible by using the AI/ML approaches .
Primary objective of the simulations has been to understand the performance elasticity of the current design and set the targets. If we consider the engine modelling or any physical system modeling basically done using the CFD techniques which can be referred as “Physics based approach”. Straight way once sense that, the gap between the design and the real product because of the uncertainties in transients and environments, it is a difficult model. This gap can be bridged using the high complex models, but the cost and engineering efforts are too high. Data driven methods are an alternate to fill this gap with the lower cost and time.
Herewith highlighting some scenarios where AI/ML approaches can fill in.
Virtual sensing in stereo application (repurposing/multipurpose the available sensors):
As we know that visual stereo applications need more than one camera source. But more than one camera source in many contexts is costly. Can we map the camera capabilities to another sensor with lost cost then we can repurpose or multipurpose a lost cost sensor that still achieves the end application with the reliable performance?
We have solved the above problem where “camera information” is modelled using the spatial sensors like accelerometer, gyro as the application demeaned us only “depth” information. We have re-purposed the sensors to model the camera information. Basically, we have modelled the information gain achieved using the camera systems using the other sensors. The modelling in this case involves ensemble structures of various deep learning models.
An alternate approach is possible to simulate one visual source in the system using the GANs. Thereby reducing the overall BOM (bill of materials) in the system and cost and reliability of the system. Moreover, this is an improving solution with more and more observational data.
The similar re-purposing of the available sensors for the new applications are possible using AI/ML methods.
Accident data simulation:
Crash testing is done at the specified test conditions but road fatalities happen in multiple other forms. To investigate the accident information, generally there are companies which collect various information like the vehicle, impact locations, position, injury, error codes, other sensor data from the vehicle based on this the accident report is generated and shared with the authorities and development teams to gain insights and study the implementation. There are only a few rows of information available to describe the event.
Now with the help of powerful AI/ML approaches we can re-create the movement by simulating the actual events using the same data.
Moreover, this also allows us to simulate the various other boundary conditions involving the laden, passenger, trajectory, pedestrian information. To put it short, environment simulation with different shapes, trajectory pedestrians, vehicles can be created along with the vehicle simulation with different load and trajectory conditions. Solutions of this nature will act as perfect virtual proving ground which reduces the overall development time, cost and allows us to capture more scenarios.
Data driven methods involving AI/ML have already influenced all the sections of automotive business.
Testing and Simulation is a source of data for many development activities is no exception and has huge potential to use the AI/ML methods.
The fundamental objectives of the testing and simulation is not only met by AI/ML methods but also thrives on it and opens new avenues of value creation opportunities.
Analytics adaptation into testing and simulation needs focused vision and executional skills which are very diverse in nature.
Simulations are transformed by the data-driven methods.
This article only motivates the audience by explaining the various possibilities. Each use case discussed in this article is a topic of its own. I encourage the readers to reach out for a detailed discussion on specific topics.
V Senthil Murugan is a seasoned automotive professional with 14+ years of experience in diverse areas like product development, R&D, automotive body & security electronics, connectivity, and data analytics. Has worked in the various end to end vehicle development of various vehicle programs.
Currently he heads the automotive industry vertical at LatentView Analytics Pvt Ltd with primary responsibilities of driviing innovation, delivery and growth.
Published in Telematics Wire