Abstract:

Data-driven fault diagnosis technologies can help operators find and detect process abnormalities in a timely and effective manner, having emerged as one of the hot topics in the current integration of industry and big data. The deep convolutional neural network (DCNN) approach is the most commonly used data-driven fault diagnosis model, but its activation process suffers from the mismatch of positive and negative values and the problem of parameter redundancy resulted by inefficient information flow. In this study, a novel activation mechanism based on the maximum smoothing unit (MSF) function was proposed to overcome the shortcomings of the previous activation functions, and the attention mechanism combined with the gated recurrent unit (GRU) was introduced to overcome the problem of parameter redundancy by improving the efficiency of information flow in DCNN. The as-established model of enhanced deep convolutional neural networks (EDCNN) exhibited significantly improved performance, which was verified by its applications in two industrial processes, the industrial actuator control system and the industrial acid gas absorption process. The average fault diagnosis rate in both processes exceeded 99.0%.

Key words: fault diagnosis, enhanced deep convolutional neural networks, process control, systems engineering, activation function

摘要：

基于数据驱动的故障诊断技术可以帮助操作人员及时有效发现和检测异常情况，是当前工业与大数据融合的热点领域之一。深度卷积神经网络（deep convolutional neural networks，DCNN）是最常用的基于数据驱动的故障诊断模型，但其激活过程存在正负值计算不匹配以及信息流通效率低导致的参数冗余问题。本文提出一种基于最大平滑单元（maximum smoothing unit，MSF）函数的新激活机制克服传统激活函数的缺点，并且引入注意力机制 （attention mechanism）结合门控循环单元 （gated recurrent unit，GRU）提升DCNN的信息流通效率克服参数冗余问题，以综合提升传统DCNN模型的故障诊断性能。强化深度卷积神经网络（enhanced deep convolutional neural networks，EDCNN）的现有模型表现出显著提高的故障诊断性能，这在工业致动器控制系统和工业酸性气体吸收过程中的应用得到了验证。两个过程的平均故障诊断率均超过99.0%。

关键词: 故障诊断, 强化深度卷积神经网络, 过程控制, 系统工程, 激活函数