基于离散Hopfield网络的自相关过程控制方法

对于自相关过程的统计控制，传统统计学方法虚发警报的概率较大，而BP人工神经网络方法权值训练困难，灵敏度不高。提出一种基于联想学习与离散Hopfield网络的自相关过程控制方法。不需任何训练样本，通过正交化编码将过程状态以吸引子的形式存储到Hopfield网络中，并利用网络的联想功能来检测自相关过程的阶跃型突变。算例研究表明，与Elman网络和EWMA方法相比，过程正常时，所提方法的平均链长（ARL）分别提高了27.9%和55.0%；过程异常时，所提方法的ARL分别降低了74.1%与81.8%以上。说明了方法的有效性与优越性。

Discrete hopfield networks-based approach of autocorrelated process control

Both traditional statistical methods and Back Propagation（BP） neural network methods have their limitations in monitoring autocorrelated industrial processes.Traditional statistical methods typically result in high false alarm rate.BP neural network methods usually lead to low sensitivity and cause difficulties in network weights training.Based on associative learning and discrete Hopfield network，a new statistical control method for autocorrelated processes is proposed.The method doesn’t need training samples.By using orthogonal coding，the process states are stored into Hopfield network in the form of attractors.Then the step shift of the autocorrelated process is detected out by using the associative phase of the Hopfield network.The case studies show that，the Average Run Levels（ARL） of the proposed method increase 27.9% and 55.0% when there is no step shift in the process and decrease at least 74.1% and 81.8% when step shift exists，respectively，compared with Elman networks and EWMA.