Implementation of a Digital Twin of a Process Plant

In this chapter, we describe how to adopt different software components based on the previously presented conceptual solution and generate an integrating digital twin for a business workflow as a service extension. The focus of this software system lies in the implementation of methods from computer vision that can reliably recognize the existing objects in a process plant with their structure and interconnections. This intention benefits from numerous methods which were developed in the past years to tackle the challenge of the recognition of 3D objects under various conditions. A brief review of such methods is presented here, in particular concerning difficult environmental impact (vapor, dust, smoke, darkness, and dirt). Methods are distinguished according to data input: image, point cloud, or video. For several reasons, the implementation of an automatic object recognition procedure is realized by using existing convolutional neural networks, also known as deep learning. Literature review shows that effective and versatile automation capabilities of deep learning combined with large-scale processing may be an adequate means for the challenges of the extent and complexity of a process plant. Further is the recognition procedure described in more detail, in particular how the piping system is built up in its full complexity coming from singular components. This recognition runs iteratively in four steps with a mutual interdependence. The entire point cloud is processed by segmentation, where the pipe system is extracted. In two subsequent steps (clustering and classification) the point cloud is subdivided further to recognize the singular parts. While piping systems consist of forked structures with complex configurations and partial occlusion, the exact recognition of piping centerlines is conducted based on the position of singular parts. A robust clustering and graph-based aggregation yield a coherent pipe model before it is linked with piping and instrumentation diagram to the digital twin. Due to the high complexity and variance of the tasks, some individual tasks must run fully manually or with partial human assistance. Finally, we present how the singular steps are automated and orchestrated by using a workflow automation platform. Our workflow promises good results on pipe models with varying complexity and density both in synthetic and real cases.