The sexy model with a brain takes to the road
Paul Rubens Aston Martin's 2005 version of its DB9 has a mind of its own. The luxury sports car is equipped with an artificial neural network a microchip modelled on the human brain that has been “trained” to monitor its engine and detect misfires in its cylinders.
A misfire occurs when the fuel in one of the cylinders fails to combust completely, increasing the amount of harmful emissions. TThe usual cause is a faulty spark plug or ignition coil. . When an To comply with US and European emissions control rules, the DB9 must be able to detect misfires and alert the driver to check the engine when a certain level of misfires is reached. The regular tightening of emission controls means car manufacturers such as Aston Martin are trying to find increasingly fast and accurate ways of detecting misfires. Aston Martin has so far used a conventional microprocessor running an algorithm to process data picked up by sensors around the engine to detect misfiring but this method is becoming impractical. “We are running out of computing power for our existing misfire-detection program,” says Alan Bennett, Aston Martin's power calibration manager. “At 7500rpm, the top speed of our [six-litre] V12 engine, we have just 1.3 milliseconds to do the misfire calculation before the next batch of data comes in from another cylinder and we can't always distinguish between a fire and a misfire in that interval.” An artificial neural network can overcome this problem because it detects misfires differently. It looks at patterns in the data it receives and “spots” misfires by recognising misfire patterns it has already learned.
Constructing a neural network is actually quite simple. Neurons in the human brain make electrical connections with neighbouring neurons by “firing” during thought. Aston Martin's neural network starts with a row of artificial neurons, called the input layer. Data from various sources are fed into each neuron, and assigned a weight or co-efficient as an indication of how important that bit of data is. If the net value of the neuron exceeds a certain threshold, the neuron fires. All the neurons in the input layer are connected to another set of neurons, known as the hidden layer, and these repeat the process, calculating their own net values and firing if their threshold values are reached. There may be one or more hidden layers, but every neural network has one final layer the output layer, which consists of one or more neurons that fire when a given set of inputs is fed in. Data is fed in at one “end” of the network and provides an “answer” (either a firing neuron or a silent neuron) at the other end.
Aston Martin's engineers wanted to feed engine data into the neurons in the input layer, and observe the single neuron in the output layer fire up when the data indicate a misfire, or see it stay inactive when there is no misfire.
The appeal of a neural network is that to achieve this does not involve looking at or understanding the data it is simply a matter of training the neural network. Engineers at the Powertrain Research and Advanced Engineering facility in Dearborn, Michigan part of Aston Martin parent Ford developed a neural network for the DB9 and fed it with data from a misfiring engine. They trained it by altering the 469 co-efficients and the threshold values of its 23 input neurons and the neurons in the hidden layers until the output neuron always fired. Then they fed it with data from an engine that was not misfiring, and trained it so that the output neuron did not fire. The result is a highly specialised neural network program specifically trained to look at data received from a DB9's V12 engine, and it has proved far more effective than analysing the data using conventional engineering equations.
“We have found the results from the neural network to be much more robust than the results we got previously, so we can comply with emissions legislation while reducing the number of false detections of misfires,” Mr Bennett says.
The network has been hard-coded on to chips that cost about ￡3 to manufacture, and these are fitted in every new DB9. This is the first time a neural network has been used in a vehicle's control module, according to Ford.
For now, neural networks remain the preserve of Aston Martin. However, “emissions legislation is getting tighter every year, so the need for this technology is bound to filter down to vehicles with lower-specification engines,” Mr Bennett says.
装了“大脑”的跑车
阿斯顿?马丁公司(Aston Martin)2005年DB9车型有自己的“大脑”。这款豪华跑车配备了一个人工智能神经网络，做在一个模仿人脑的微型芯片上，“训练有素”可以监控引擎的运行，探测到气缸熄火。
当一个气缸中的燃料没有完全燃烧时就会发生熄火，有害的废气排放量随之升高。通常的起因是火花塞或点火线圈有缺陷。为了符合美国和欧洲的排放量控制规定，当达到某种程度的熄火时，DB9必须能够发现熄火，提醒司机检查引擎。对废气排放量不断加强的控制，意味着阿斯顿?马丁之类的汽车厂商正在努力寻找能更快更准发现熄火的方法。阿斯顿?马丁曾采用一种常规微处理器，借助算法程序处理由引擎周围传感器搜集的数据，以监测熄火，但是这种方法已变得不实用了。“电脑运算能力跟不上我们现有的熄火监测程序。” 阿斯顿?马丁动力校准经理阿伦?贝内特(Alan Bennett)说，“在V12(六升)引擎以每分钟7500转的最快速度运行时，我们只有1.3毫秒的时间在下一批另一个气缸的数据发来之前计算熄火，在这个间隔之间，我们不能始终将点火与熄火分得一清二楚。”人工智能神经网络可以克服这个问题，因为它发现熄火的方式不同。它收到数据后对数据的模式进行扫描，通过辨认已经掌握的模式“看出”熄火。
建一个神经网络其实相当简单。人思考时，大脑中的神经细胞通过“点火”与邻近的神经细胞产生电路。阿斯顿?马丁的神经网络从一排称为“输入层”的人工智能神经开始工作。不同来源的数据被填入每一根神经，然后被赋予一个重量值或系数，表明这个数据的重要度。如果这根神经的净值超过某个限度，神经就点火。在这个输入层中的所有神经都与另一组被称为隐藏层的神经相连接，这一组的神经重复这一过程：测算自己的净值，达到限度就点火。隐藏层可能有一两个，但每一个神经网络都有最后一层：输出层。它由一根或更多神经组成，只要有一组给定数据输入就会点火。数据从网络的一个“端点”输入，从另一个“端点”提供一个“答案”(要么是一根点火的神经，要么是一根没有反应的神经)。
阿斯顿?马丁的工程师想把引擎数据输入输入层的神经，然后观察输出层里单根神经在数据显示熄火时点火，或者在数据显示没有熄火时保持没有反应。
神经网络的吸引力，在于它不需要分析或理解数据就能发挥作用。唯一要做的是训练这个神经网络。阿斯顿?马丁的母公司福特汽车公司在密歇根州迪尔本动力装置研究及高级工程部门的工程师为DB9开发了一个神经网络，然后输入一个熄火引擎的数据。他们的训练方法是：不断调节469个系数、以及23根输入神经和隐藏层神经的门槛值，直到输出神经总是点火。然后他们再输入一个未熄火引擎的数据，同样调节各个环节，使输出神经保持不点火。如此训练的成果就是一个量身定制的神经网络程序，专门分析来自DB9车型V12引擎的数据，这比使用常规工程系数分析数据有效得多。
“我们发现神经网络的检测结果比以前我们得到的结果可靠得多。这样，我们就能在达到排放法规要求的同时减少误判熄火次数。”贝内特先生说。
这个网络已通过硬编码做到生产成本只需3英镑的芯片上，并装到每一部新的DB9型车上。福特公司表示，这是汽车的控制模块中首次使用神经网络。
目前，神经网络还是阿斯顿?马丁公司的专有技术。但是，“随着针对尾气排放量的立法一年比一年严格，对这项技术的需求一定会扩展到较低规格的汽车。”贝内特先生说。
