Model Based System Monitor Using Artificial Neural Networks (ANN)

Abstract

Nowadays, with the integration of diverse mechatronic components, a mechatronic system is capable of performing more difficult and complex tasks. At the same time, it becomes more and more difficult to develop a system to ensure the safety of these complex mechatronic systems. This paper outlines a novel method for fault detection, which uses a neural network monitor based on a feed forward back propagation (FFBP) algorithm. The goal of this study is to develop a real-time capable method for detection of abnormal or faulty behavior of an automation system. An example is used to illuminate the capability of the neural network in modeling and real-time data in this work. Furthermore this method is validated on a test bed, which is developed at the institute of product development Karlsruhe (IPEK). This method will allow the monitor to detect critical state of the system due to unexpected influences in real-time, so that safety hazard to personnel and equipment caused by the fault can be prevented or mitigated.

Neural Networks in Process and Automation

“…However, many applications of NNs are intended to model continuous response variables, and in this case a sigmoidal (s-shaped) transfer function is more useful. The transfer function illustrated in Figure 13.5 and Equation 13.9 is a bipolar sigmoidal function, with the bipolar meaning that Step 2 in the excitation of a network of nerve cells. Individual neurons are assembled in a configuration such as that illustrated in Figure 13.6, in which there are two process inputs, x 1 and x 2 , and one process response, y.…”

“…In another example, the equation might represent the installed valve characteristic (flow rate as a function of valve stem position) and its inverse would be used in automation to determine the controller output given a desired flow rate. For this example, the derivative of the model, Equation 13.3, describes process gain, and might be used to gain schedule the flow rate controller:…”

“…The architecture of an NN refers to the number of neurons in each layer, the transfer function used, the location of the bias influence, and other structural aspects to be discussed. Architecture to an NN would be similar to the choice of powers and crossproducts chosen for a conventional regression model, such as Equation 13.4. Each neuron would use Equation Set (13.10) to convert its input signal values and weights to its output value.…”

“…If the process data have a different scale, then have the NN output model a scaled variable. Use Equation 13.11 to scale the output process variable from −0.8 to +0.8 so that (1) the "z" signal internal to the NN does not have to go to ±infinity, and (2) it can cope with new data that might extend the range: Input data should also be scaled. Consider the bipolar sigmoidal transfer function.…”

“…Large values for the weights on the bias, or small values for weights on the input signals could relocate the summed z-value to be within the −2 to +2 range to create a nonsaturated neuron response, but scaling input variables to a −1 to +1 basis accelerates the network optimization (training) by effectively limiting the weight values within a −2 to +2 range. Use Equation 13.12 to scale the input process variable:…”