A sensor-equipment diagram is a type of engineering drawing used in the industrial practice that depicts the interconnectivity between a group of sensors and a portion of an Oil & Gas facility. The interpretation of these documents is not a straightforward task even for human experts. Some of the most common limitations are the large size of the drawing, a lack of standard in de ning equipment symbols, and a complex and entangled representation of the connectors. This paper presents a system that, given a sensor-equipment diagram and a few impositions by the user, outputs a list with the reading of the content of the sensors and the equipment parts plus their interconnectivity. This work has been developed using open source Python modules and code, and its main purpose is to provide a tool which can help in the collection of labelled samples for a more robust arti cial intelligence based solution in the near future.
A sensor-equipment diagram (SED) is a type of engineering drawing which is commonly used in the Oil & Gas industrial practice to depict how a group of sensors are interconnected to a certain section of an oil rig or a plant. These drawings are composed of a central main grid with multiple pieces of equipment, plus a series of circular shapes which represent sensors. An example is shown in Figure 1. Notice that there are two types of sensors: Local, which are connected to the main grid through a solid line, and Panel Mounted, which are connected to the local sensors through dashed lines.
In recent years, the Oil & Gas industry has shown a particular interest in developing systems that can digitise and interpret a large amount of SEDs in order to migrate printed drawings towards a paperless environment. Nonetheless, experts have realised the di culty of the automation of this task due to several factors, most notably poor image quality, the large size of a SED, a lack of clarity on the delimitations of an equipment symbol and the complexity of understanding the connectivity. For instance, the SED shown in Figure 1 is an image of 4460 2544 pixels (approx 1.5 Mb) which contains 164 sensors (not all of them connected to the equipment parts) and four pieces of equipment which can only be identi ed empirically, since there is no particular standard for differentiating and delimiting equipment parts. Moreover, notice that connectors may overlap or gap each other.
Digitisation and interpretation of engineering drawings from the Oil & Gas
industry is not a new problem. In fact, the rst literature dates back to the
1980's [
In 2016 Banerjee et al. [
Most recently, Moreno-Garcia et al. presented a heuristics-based method to
detect and segregate a series of commonly found shapes in P&IDs. Afterwards, a
state of the art digitisation methodology called text/graphics separation [
The tool starts by importing the image as a grayscale bitmap. Then, this
image is binarised using a standard thresholding algorithm [
Once the area of interest is segregated, a circle detection algorithm based on
OpenCV2 Hough circles, which was succesfully implemented in previous work
for P&IDs [
To segregate the text inside the sensors, the largest and outermost contours of each sensor image are discarded, assuming these to be the sensor shape and any noisy pixels surrounding the sensor. In addition, for the case of Panel Mounted sensors, an additional detection step for large and elongated contours is implemented. This allows the identi cation and removal of the horizontal line dividing the text within the sensor. Then, a connected component analysis (CCA) is applied on the remaining image to obtain the height and width of each text character. This will be useful in a later stage to detect the text naming each equipment part. Finally, the sensor text is read using the Pytesseract OCR3 module.
To automatically detect the text naming each equipment part, a CCA is
run on each equipment part area to detect all shapes which are approximately
the height and width of an average text character found in sensors. Then, a
morphological brushing operation [
Finally, the image containing the connectors and the main grid is analysed
using a line detection method similar to the one porposed in [
This paper presents a system which, given an engineering drawing known as a SED, produces a list of the connectivity between the sensors and the equipment parts contained on the main grid. By requesting the user to select the area of interest and the equipment parts, the system automatically nds the sensors, reads the content of the sensors and equipment parts, and deduces the connectivity between these shapes. The method has been successfully showcased
to industrial partners of the Oil & Gas sector and has been mounted in one of their servers for testing in future projects4.
It is important to note that this work opens a niche area of engineering
drawing analysis in the light of deep learning advancements. Literature
regarding the application of machine learning for the digitisation and interpretation
of engineering drawings is still scarce, and most importantly, there is an insu
cient amount of labelled data which can aid on training automated systems to
perform these tasks. While some work based on neural networks has been
presented for the digitisation of other assets such as circuit diagrams [
c rs e o m l p
i g u n q
e