Safety consideration: Neural Network 01
[Author]
Renhong WENG,
1.Artificial Intelligence safety investigator
2.Automotive hobbiest
This article is for neural network safety consideration in automotive. Present, worldwide there are some neural network safety consideration from IEC 61508, but unfortunately, it is not directly applied into automotive fields.
We have to describe the Neural Network in the way of automotive safety technology SOTA.
First: neural network scope
originated from human brain structure, scientists had built the Neural Net or Network as following:
And after that, present, the neural network will be structured in following way more detail:
we normally use the neural network in following automotive areas:
perception system
cognition and human machine interface
potential support for future decision module, by using graphical neural network
For layers in neural network, there are four:
(1)input layer
1.inputs
(1)hidden layer
1.weights
2.artificial neuros
3.activation function
4.output
(3)output layer
1.weights
2.artificial neuron
3.activation function
4.output
Second: Safety criteria for Neural Network in automotive
based on the reference-01, we derive out the basic safety principles for Neural Network in automotive, red zone is automotive highly related:
For above we shall allocate the corresponding solutions for each Goals, but due to time limit, i just show one example of goal:
Synthesis speaking, the goals for the neural network listed as following:
Third: Software Development Process for Neural Network in automotive
Proposal as following:
Fourth: SOTA standards application
1.ISO 26262:
we get it from ISO 26262, Chapter 6, statistics as following:
(1)Table1
| TOPIC | If applicable to NN |
| 1a | x |
| 1b | |
| 1c | |
| 1d | x |
| 1e | x |
| 1f | x |
| 1g | x |
| 1h | x |
| 1i |
(2) Table2
| TOPIC | If applicable to NN |
| 1a | |
| 1b | |
| 1c | |
| 1d | x |
(3)Table3
| TOPIC | If applicable to NN |
| 1a | x |
| 1b | x |
| 1c | |
| 1d | |
| 1e | |
| 1f | x |
| 1g | x |
| 1h | x |
| 1i | x |
(4) Table4
| TOPIC | If applicable to NN |
| 1a | |
| 1b | x |
| 1c | |
| 1d | x |
| 1e | x |
| 1f | |
| 1g | |
| 1h | x |
(5) Table 5
| TOPIC | If applicable to NN |
| 1a | |
| 1b | |
| 1c | |
| 1d | x |
(6) Table 6
| TOPIC | If applicable to NN |
| 1a | x |
| 1b | |
| 1c | x |
| 1d | x |
| 1e | x |
| 1f | x |
| 1g | x |
| 1h | x |
| 1i | |
| 1j |
(7) Table7
| TOPIC | If applicable to NN |
| 1a | |
| 1b | x |
| 1c | x |
| 1d | |
| 1e | x |
| 1f | |
| 1g | x |
| 1h | |
| 1i | x |
| 1j | x |
| 1k | |
| 1l | x |
| 1m | x |
| 1n | x |
(8) Table 8
| TOPIC | If applicable to NN |
| 1a | x |
| 1b | x |
| 1c | x |
| 1d | x |
(9) Table 9
| TOPIC | If applicable to NN |
| 1a | |
| 1b | |
| 1c | x |
(10) SW integration and further tables
not applicable in NN, they are in more higher hierachy.
2.UL4600 Official release version
3.ASPICE PAM 3.1
only for the BP in each sub group, and rough listed the results:
Thanks for your reading, and any comments welcome.
[Ref]
1.Developing artificial neural networks for safety critical systems
2.ISO 26262-6
3.UL4600
4.ASPICE PAM3.1
- 用户评论