Paweł presents Cloud Your Car tests results
Paul did four attempts to record a video about result of the Cloud Your Car plug tests and each time someone interrupted him. Finally he made it, so now we all may see the results. Enjoy!
Please note that these tests were the first we have conducted and for sure not the last.
The questions that drove us to conduct tests were:
Does the particular kind of sensor we currently use, with specific parameters is possible to distinguish four different events characteristic for any car ?
1. Driving over road holes
Additionally Paul tested two data-processing algorithms, which were converting measured signals into normalized dimensions of acceleration, filtering and smoothing, transforming the coordinate system of the device position to the coordinate system of the vehicle and finally making frequency analysis.
Paul also prepared two different decision algorithms, which were supposed to help when conducting tests. One of algorithms turned out to be an absolute non-sense, because of improperly floating cut-off mechanism (cut-off threshold) that Paul implemented into the algorithm. Let’s put it simple, the output result of this algorithm was a uniform distribution of all the decisions in their extent - we had the same number of events of every type. The second algorithm results satisfied Paul and were used in developing cyber-neural network, which automatically classified incoming data to a specific event.
It cannot be dismissed that conducted tests were too short to bring much data about the car, so the transformation of coordinate system could not define the vehicle orientation. This means that it was impossible to explicitly assess the front, back or sides of the vehicle. Paul took that fact into consideration and analyzed characteristic factors of the gathered signals and data sets like duration, amplitude, frequency and spectrum diversity all this allows to recognize important events.
The analysis referred to system parameters validation which was responsible for processing of measured data to sharpen the information sent to the decision system about four considering states of vehicle which Cloud your car is supposed to detect. The selected parameters were: frequency range, processing accuracy, range of constants and limit values. In the next stage of analysis results, based on established parameters, were used to teach various types of cyber neural networks - mentioned above. These networks vary in number of layers and neurons in an individual layer, activation functions, learning method and abilities of implementation on microcontroller. Step by step elimination tests helped to decide what kind of neural network is the best in this case. The idea of using cyber neural networks is still being improved and developed by Paul as he is using the data gathered during the past tests. Now, Paul is preparing the implementation of whole neural network on the microcontroller. It will be tested and recorded as well. Stay tuned!
To summaries, with the data gathered offline, it is much simpler to make better decisions about actually working algorithm. But it should be remembered that the factor influencing the quality of decision is experimenter's personal observation form of tests and it’s results. Perhaps, the next test goal would be to enrich the learning database for mentioned cyber neural networks, Paul said.
with strong cooperation with Paul
