Sensor-fusion Strategies for Process and Performance Monitoring

Most of the major plant, factory, process, equipment and tool disruptions are preventable, and yet preventative fault detection and diagnosis section are not the norm in most industries. It is not uncommon to see simple and preventable faults disrupt the operation of an entire integrated manufacturing facility. For examples faults such as malfunctioning sensors or actuators, inoperative alarm systems, poor controller tuning or configuration can render the most sophisticated control systems useless. Such disruptions can cost in the excess of $1 million per day. Over the last decade the fields of multivariate statistics, and controller performance monitoring techniques have merged to develop powerful sensing and condition-based monitoring systems for predictive fault detection and diagnosis. These methods rely on the notion of sensor fusion whereby data from many sensors and units are combined to give a holistic picture of health of an integrated plant. Such methods combined with embedded digital intelligence are at a stage where such strategies are being implemented for off-line and on-line deployment. Introduction The purpose of this paper is to explore the area of information or sensor fusion in the context of the process industry. It is common knowledge that all process data is logged and archived digitally. The chemical process industry has been archiving data for more than two decades. We are therefore at a point where the process industry is rich in data to the point of deluge!; in short we have been experiencing an “information big bang”, an extraordinary explosion of data. It is fair to say that we are living through the most measured phase in the history of the world. In the context of the process industry, the questions to pause are: 1. Are we making the most effective use of this data? 2. Are we a data rich but information poor society or to put it in a more positive context: What information can be extracted from these massive volumes of data? 3. How do we assimilate/fuse all of this information? Over the last decade or so there has been a paradigm shift underway in process operation. The mantra for process engineers has been: ‘Listen’ to your data and use it to monitor the performance of your process. Data or sensor or information fusion means different things to different people. Most valuable sensor fusion technologies do the following: 1. Gather data automatically from multiple sources; 2. Filter data for errors and relevance; reconcile the data to make sure that the sensors are calibrated correctly; carry out time and/or event synchronization when data is assembled from different sources; and fill in missing data; 3. Look at or visualize and analyze all the data in a cohesive way; i.e. look at the data in a multivariate framework, i.e. data should not be examined as singular time series trends but rather in the multivariate temporal as well as spectral domains. 4. In analyzing process data there should be no disconnect between the data and the process, i.e. explore the process in a holistic way by fusing/combining information from many different sources, including process knowledge and flowsheet information, and presenting all of this information in a coherent way; 5. Integrate this information with business systems that utilize the same information and ensure that the data/information/sensor fusion technologies are well integrated with the natural work flow of a process engineer. For specific focus with respect to process data analysis, the technology of data or sensor fusion requires one to: 1. Use the data to build models of the process under normal operating conditions and then use such models to carry out predictive maintenance. The motivation in doing this is the realization that unplanned process upsets can constitute anywhere from 3 to 50 percent of production costs (IEE, 2005). In this respect the new operating paradigm is to do predictive or preventative fault detection and the realization that ‘fail and fix’ is not the right way of operating a process. 2. With this mode of operation, condition based monitoring (CBM) frees people’s time to do things that really matter in managing assets. The challenges of sensor fusion are illustrated via 3 separate industrial case studies. Each case study demonstrates several different aspects of data or sensor fusion methodology. The case studies serve to illustrate the challenges each application pauses. The main challenge is how to incorporate apriori process information with sensor data and the use of process flow sheet or connectivity information with process data for fault detection and diagnosis. Each case study is described in more detail below and illustrates the applications of the 5-step methodology for sensor fusion outlined above. 1. Imaging sensor fusion for monitoring interface level in an oil-sands operation This application analyzes image data to illustrate the concepts of sensor fusion. It specifically makes the case that each pixel in an image is a source of information and that image processing combines or fuses information from thousands of pixels to provide coherent information about the process. Bitumen extraction from oil sands is the core unit operation in the water based extraction process used to produce crude oil. This floatation process is carried out in large vessels called separation cells. In these cells, aerated Bitumen floats to the top as froth and sand, Clay settle to the bottom in these gravity separation vessels. Bitumen is then skimmed off and transported to later stages of the process for upgrading (Masliyah et al., 2004). In the extraction unit operation, three layers form inside the separation cell as shown in Fig 1. Of particular interest is the interface level between the Bitumen-froth and the Middlings layers, which is known to affect the froth quality and Bitumen recovery and thus heavily influence process economics. Figure 1: Schematic of the separation cell in the oils-sands extraction process For example, when this level is too high, fines (fine sand particles) escape into Bitumen-froth degrading its quality and when it is too low, Bitumen is lost to the Tailings ponds causing financial losses and environmental problems. A good regulation of the interface level also reduces its variability and hence allows for process operation closer to the optimal level set point, which results in economic gains (Jain, 2006). For these reasons, there has been much interest in the oil sands industry to control this interface at an optimum level for Bitumen recovery. Optimal control of the interface between Bitumen froth and Middlings in these cells can result in a significant improvement in Bitumen recovery resulting in large economic benefits. The major impediment in the implementation of such a control system is the lack of safe and reliable sensors for interface level detection. Traditional instruments such as nuclear gauges, capacity probes etc. are either unsafe or do not give reliable estimates. This work describes a novel sensor for interface level detection, developed using computer vision techniques on video frames captured from a sight glass camera. A simple edge detection method combined with modelbased particle filtering is used to provide estimates of the interface level and its quality. It is shown that the algorithm is robust to lighting changes and process abnormalities. The cropped image of the sight glass area of interest is 56 pixels wide and 222 pixels high; therefore in total there were 12432 bits of information. Each pixel is a source of information and it is the fusion of information from these pixels and extraction of the interface level that exemplifies the idea of data or sensor fusion. A particle filter algorithm with a process model was used to extract the interface level from the cropped image of 12432 bits of information. An example of the actual sight glass (middle portion), the processed and transformed image for edge detection (right hand portion) and the posterior distribution of the level with an indication of the variance is shown on the left portion of figure 2. Technical details of the image segmentation algorithm used for interface level are given in the paper by Jampana et al. (2008). Figure 2: Middle portion shows the cropped portion of the sight glass; right portion of the figure shows the processed and transformed image; left portion of the figure shows the distribution of the pixels (from the right portion of the image) after filtering. The final output of this image-based sensor has been used to regulate the interface level in the separations cell on line 6 at Suncor Energy Inc. with great success. Using approximately three weeks of laboratory data (one week of camera control, two weeks of capacitance probe control), it has been calculated that the Bitumen losses in the tailings stream dropped by 53.6%. A similar reduction of 29.12% has also been noticed in Bitumen losses to Middlings (Jampana et al. 2008). Process monitoring in a multivariate framework An important aspect of sensor fusion is that process information should not be ignored or that data should not be analyzed haphazardly or in isolation. Simply said, the multivariate nature of the process should be taken into account. Several important practical issues should be consdiered when information is to be extracted from many different sensors: 1. Due attention should be paid to process from which the data originates from. 2. The configuration of the data matrices for developing models for process monitoring should take into account the basics of unit operations. These ideas are illustrated below through a very simple example. Consider a simple ‘T-junction’ process where two streams of water, in liquid phase, with flow rates F1, F2, and temperatures Tcw and Thw respectively are mixed. The product stream has a flow rate F3 and temperature Tmix. An energy balance for the mixing node may be written as