Modelling and Inference Systems in a Connected Car

Use of Analytic and Modeling in Automobiles

Automobiles are now fitted with OEM devices that provide telematics and analytical data.

Data for Analytic are emitted by different parts of automobiles and feed comes at periodic intervals. The number of vechiles with sensors are going to increase at 40% rate yoy. Think through a solution architecture and modeling technique?



Few insights that can be of use for a customer can be in area of preventive maintenance, security etc.

Engine oil change, tyre change, optimal speed to drive, suggestive drive modes like cruise, care etc, When electric charge needs to be recharged.

Say when to recharge is the dependent variable, independent variables can be capacity of battery, power available, charging rate, discharge rate etc.

Which modeling techniques to use?


The diagnosis of technical systems is a research field that is largely governed by modeling questions:

• Which modeling strategy is adequate, a deep behavior model or a shallow rule model?
• Is a distinction between a correct behavior model and a fault model necessary?
• Is the diagnosis process controlled by heuristic expert knowledge, and, if so, how shall it be represented?

Env of a vechile: The environment comprises the rest of the vehicle and includes the driver, the road, the weather, and the following vehicle domains:

• Powertrain Examples: engine, gear, drive shaft
• Chassis Examples: brakes, damping, steering
• Body Examples: light, wipers, climate control, key-less entry
• Safety Examples: airbag, active safety belts
• Telematics/Infotainment Examples: radio, navigation, telephone

Examples of these systems include driver assistant functions such as active cruise control (ACC), active front steering (AFS), or the lane keeping support. These systems are distributed over several ECUs and often rely on other, already existing software modules. Faults occurring at one module of such a distributed software system may trigger further faults within other connected software modules. 

In the field of machine learning, several algorithms exist. Two different learning algorithms are applied here:

• Linear Regression. This common method uses a least square estimation to get parameters a₁, . . . , a_nεR for the diagnosis function template
• Decision Trees. A decision tree uses a series of binary decisions to reach a classification. Learning such a tree is done by binary recursive partitioning of the input space using a given optimization criterion. 

•  some basic faults occurring in sensors and actuators were implemented. This is a realistic fault type in modern cars. In this case study, four different fault types were modeled:
• [0099]
  1. The signal is reduced to 90% of its correct value.
• [0100]
  2. The signal value used for the simulation is 110% of the correct value.
• [0101]
  3. Noise is added to the signal.
• [0102]
  4. The signal is dropped out.

After simulating the system in different scenarios with different faults (again, including the no-fault case), a machine learning approach has been applied to the simulation results as presented in Equation 3. Generally, the algorithm shall use the data of the no-fault simulation and the data of one fault simulation run (e.g., the noisy signal) to detect patterns in the data. Based on this detection, the algorithm develops a classification function which decides whether the accelerator pedal sensor was noisy or not. Ideally, this classifier can also be applied to new measured data of a real vehicle, which means that the classifier has to generalize the data. 

• The present invention introduces a new technology to the diagnosis of automotive systems as both a formal framework and a concrete implementation, comprising an experimental setup, simulation tasks, and classification results.
• [0052]
  The present invention is directed to intricate diagnosis situations in modern cars. The invention employs a mixture of model-based and associative diagnosis and an associated model compilation. In the present invention, a simulation database is built from module simulations of the system in various fault modes over a typical input range. From the simulation database, a simplified rule-based behavior model can be constructed where long cause-effect chains are replaced with much simpler associations. The behavior model is also optimized for a heuristic classification of the faults. The present invention applies the novel model behavior to the complex discrete event/continuous system models of modern cars.
• [0053]
  Model compilation is a diagnosis approach that combines the model-based paradigm and the associative (heuristic) paradigm within the following four steps:
• [0054]
  1. Simulation. A database, C, is compiled by simulating the interesting system in various fault modes and over its typical input range.
• [0055]
  2. Symptom Computation. By comparing the faultless simulation to simulation runs in fault modes a symptom database C_Δ is built up.
• [0056]
  3. Generalization. Using cluster analysis or vector quantization, the numerical values in C_Δ are abstracted towards intervals.
• [0057]
  4. Learning. Data mining and machine learning is applied to learn a mapping from symptoms onto the set of fault modes; the resulting classifier can be considered as a “diagnosis compiled model.”
• [0058]
  Since this process can be completely automated, the approach has the potential to combine the advantages of the model-based philosophy, such as behavior fidelity and generality, with the efficiency and robustness of a heuristic diagnosis system.

https://www.google.ch/patents/US20070283188

http://electronics.howstuffworks.com/gadgets/automotive/hughes-telematics-device2.htm

http://publications.lib.chalmers.se/records/fulltext/219421/219421.pdf

The telematic gateway unit is used in vehicles for communication with a back-office. The unit is an Electronic Control Unit (ECU) and the communication could for example be used to send the current position of the vehicle, the status of the vehicle components, or the current up-time of the vehicle to the back-office. The hardware resources on the telematics gateway unit are limited and with increasing functionality of the unit increasing amounts of hardware resources are needed. This thesis work was carried out on the onboard telematics department of Volvo Group Trucks and the main focus was to find a method to predict the required hardware resources needed when adding or changing functions on the telematics gateway unit. After studying around 60 different articles, 9 interesting approaches were found. Regression analysis was found to be the most suited for predicting the hardware resources. The method that is proposed in the report firstly, gather data regarding the resources used on the telematic gateway unit. Secondly, multiple regression is applied to the gathered data to make the predictions of the hardware resource chosen. Limited tests were carried out on the predicted values, the tests indicated that predictions on the hardware resources were possible. The predictions made with multiple regression can be used to roughly estimate the amount of resources a new function will require. However, implementing the prediction method requires measuring more resource data than is currently done by the onboard telematics department

------------------------
Snippet of few profiles found which cite in brief related work done...

Developed a vehicle model for the Mahindra Reva E2O in python taking care of the driver model , drive train, controller and energy storage.

Developed algorithms(in python) for State of Charge and Distance to End calculations for a vehicle at the end of the trip based on route information from google apis. Also using weather information to automate HVAC systems.

Automated the process of data acquisition from Telematics server to maintain a database of customer vehicle drive / charge information and developed prognostics algorithms to analyze them.