Fig. 1.

Use cases of artificial intelligence to accelerate and improve the

synthesis of P&IDs.
Sequential node prediction using recurrent neural networks
into an one-hot vector, i.e., a vector consisting of zeros except a single entry
that is set to one (Gulli and Pal, 2017). The sum of the vectors allows for
determining exactly the used one-hot vectors, which were needed for its generation.
Digital Chemical Engineering
International Organization for Standardization 2012b
and applies it to the sum-MLP combination in Eq. (6) (Hamilton, 2020).
Table 1. Higher level node classes and their quantity in the training dataset.
Google Scholar
Google Scholar
Pumps 69
message of an aggregated neighborhood of a node in a graph
Oeing, 2022
Download : Download full-size image

Under a Creative Commons license

layer of a GNN / iteration step in message passing
Graph neural networks: a review of methods and applications

View PDFView articleView in ScopusGoogle Scholar

Artificial intelligenceProcess synthesisP&IDProcess engineeringDetail engineering

Tables (1)
D.P. Kingma, J. Ba
(6)

(9)
Yoshinari Hashimoto, Hiroto Kase
J. Zhou, G. Cui, S. Hu, Z. Zhang, C. Yang, Z. Liu, L. Wang, C. Li, M. Sun

K. Xu, W. Hu, J. Leskovec, S. Jegelka

X. Hu, P. Balasubramaniam
Proceedings of the NIPS (2017), p. 2017
Fig. 4. GUI of the DEXPI-2-graph converter.
output
Acknowledgment

TOWARDS AUTOMATIC GENERATION OF PIPING AND INSTRUMENTATION DIAGRAMS (P&IDS)

WITH ARTIFICIAL INTELLIGENCE

A. Gulli, S. Pal
Readers:
Download : Download full-size image
GraphML Project Group, 2017. GraphML specification [WWW Document]. URL
http://graphml.graphdrawing.org/specification/dtd.html (accessed 2.10.22).
Since the structure of the graphs in GNNs is learned by aggregating the neighbors,
it is important that the graphs have no missing links. As mentioned before, this is
not always the case in the data base from chapter 2.2 due to inconsistencies in
real P&ID drawings and their DEXPI exports. For this reason, a new dataset of
laboratory plants as well as industrial distillation plants is used. The dataset
contains 13 P&ID graphs and have a sufficiently high density of cross-links such
that they represent the original P&IDs well. These 13 P&ID graphs with 2020 nodes
and 2283 edges are used to train the GNNs in the following. Within the 13 P&IDs,
there are a total of 47 different equipment classes. In this context, a feasibility
analysis must be performed to investigate the potential of classifying P&ID
equipment by GNNs and to identify further challenges. Since different components in
a P&ID can fulfill the same process engineering function and since the use of
meaningful functionalities at the existing positions is to be examined for a
consistency check, it is recommended to divide the components within the P&ID into
meaningful classes. This has the additional advantage that even single components
can be used as training data influencing the classification. Hence, the 47
different components in the training data set are sorted into 9 superior classes,
which are shown in Table 1 below. The number of components per class is also shown
in relation to the training data.

Chen, 2021
arbitrary node u of a graph
Fig. 2. Structure of the Python DEXPI-2-graph converter
Volume 4, September 2022, 100038
Recurrent Neural Networks
Abstract

GraphML Project Group 2017

with trainable parameters
Deep sets

Fig 3
N
https://doi.org/10.1016/j.dche.2022.100038
Murphy, 2012
Figures (10)
PHASuite: an automated HAZOP Analysis tool for chemical processes
Manaswi, 2018
Digital Chemical Engineering, Volume 3, 2022, Article 100032

Elsevier
Packt Publishing, Birmingham (2017)
Turau and Weyer, 2015
Inductive representation learning on large graphs
Zhou et al., 2020
Elsevier logo
In the second step, the structure of the node sequences is learned with the aid of
recurrent neural networks (RNN). For this purpose, four different approaches are
used: a simple RNN (Chen, 2021), a Bidirectional RNN (BRNN) (Schuster and Paliwal,
1997), a Gated Recurrent Units (GRU) (Cho et al., 2014) as well as a Long-short-
term memory (LSTM) (Hochreiter and Schmidhuber, 1997). An RNN, as shown in Fig. 5,
has an input layer, a hidden layer, and an output layer. The RNN represents a
replication of a neural network with the same dimensions. For each replication, a
state is transferred to the next layer. In this way it is possible to learn local
relationships in the sequences. Basically, RNNs have a disadvantage that should not
be underestimated. The training of weights by more distant information is
difficult, since their errors explode or diminish during the backpropagation.
(Chen, 2021) To get around this there are further possibilities such as the use of
BRNNs. These take into account future information in addition to previous
information to increase the accuracy (Schuster and Paliwal, 1997). Another option
is the use of GRUs or LSTMs. Both consists of single cells and using cell states
and gates to decide which information will be processed and which forgotten. This
allows to get the behavior of a short-term memory. GRUs consist of a reset and
update gate (Chen, 2021; Cho et al., 2014). A LSTM uses one input, one output, and
one forget gate. The gates control the information flow and decide, which
information is necessary to make a prediction. A kind of short-term memory is
created, which also gives the network its name (Chen, 2021; Hochreiter and
Schmidhuber, 1997). Both the GRU and the LSTM are used in state-of-the-art deep
learning applications. GRUs have fewer tensor operations, which leads to faster
training. For this reason, they are used for modeling in this paper to investigate
which of them leads to better results for predicting P&ID equipment based on
sequential data.
Social Media

View in ScopusGoogle Scholar

N.K. Manaswi
Apress, Berkeley (2018), 10.1007/978-1-4842-3516-4
Download : Download full-size image
Process control equipment (PCE) 547
Fig. 2. Structure of the Python DEXPI-2-graph converter.
M. Zaheer, S. Kottur, S. Ravanbakhsh, B. Poczos, R. Salakhutdinov, A. Smola
Another option is to form the message via a set pooling approach (7), which is
similar to the aggregation of a Graph Isomorphism Network (GIN) (Xu et al., 2019).
According to Zaheer et al. (2017), this uses an
Further opportunities are safety assessment and HAZOP studies, where machine-
readable HAZOP scenarios could be mapped to graph-based plant topologies using
search algorithms. Furthermore, in future a possible detection of subgraphs in
P&IDs i.e. functional equipment assemblies could facilitate the mapping between PFD
simulations based on unit operations and P&IDs. This could provide the foundation
for automated generation of PFD simulations from a DEXPI plant topology.
M. Schuster, K.K. Paliwal
Journals & Books
To make the structure of the P&ID available for further processing, the respective
DEXPI file of the P&ID is converted into a graph using Python. The plant topology
of the P&ID, including parameters relevant for modeling, is stored in a directed
graph in the form of a GraphML file (GraphML Project Group, 2017). A directed
graph, per definition, consists of a set N of nodes and a set E of edges. The edges
are directed, meaning that each edge is defined by ordered pairs of nodes (start
and end node) (Turau and Weyer, 2015). The P&ID to be processed is stored in the
DEXPI format, which is based on a Proteus XML schema (Proteus XML, 2017) and
contains three levels of information relevant for topology extraction. The
Equipment, which contains a listing of the components present in the P&ID. The
PipingNetworkSystem, which describes the piping system and the interconnections
between the various equipment, and the PipingComponents, which contain components
embedded in the piping system, like valves. The classes of equipment and components
are uniquely defined in DEXPI via the EquipmentClass and ComponentClass.
Furthermore, the InstrumentationFunction provides information about process control
equipment (PCE) and their connections. Fig. 2 shows the procedure used to convert
the P&ID via Python.
M. Wiedau, G. Tolksdorf, J. Oeing, N. Kockmann
Copyright © 2023 Elsevier B.V. or its licensors or contributors. ScienceDirect® is
a registered trademark of Elsevier B.V.
(1)
Captures
Show more
M. Khosla, V. Setty, A. Anand
Search ScienceDirect

Zhao et al., 2005

Wiedau et al., 2019
hu

Grover and Leskovec, 2016

Datasets
Process. Saf. Environ. Prot., 83 (2005), pp. 509-532, 10.1205/psep.04055

Declaration of Competing Interest

Google Scholar
Keywords
arbitrary node in the neighborhood of node u
In this paper, AI algorithms are trained to learn patterns in P&IDs using the DEXPI
standard, which can be used as a basis for intelligent and efficient process
development. To achieve this, DEXPI P&IDs are converted to graphs, hence are made
available for graph theory methods. Based on these graphs two different AI-
supported use cases for assisted P&ID synthesis are developed and explained in the
following section. The first use case uses sequential processing by recurrent
neural networks for the prediction of P&ID equipment. The second use case uses
pattern recognition in P&IDs with the aid of graph neural networks (GNN) for
consistency checks.
AI-based suggestions can be used to speed up the process of drawing P&IDs. A
sequence of drawn and connected components is used to learn their course with the
help of an RNN. Recurrent neural networks are neural networks that can model time
series (or in this case linearly structured sequences) in training data based on
their structure (Hu and Balasubramaniam, 2008). Based on the learned correlations,
a node prediction is carried out, which returns the most probable subsequent P&ID
components based on an input sequence. The workflow of the modeling is explained in
more detail in the following.
Elsevier logo with wordmark
To quantify how well GNNs are suited to classify individual pieces of equipment
based on their location in the P&ID topology, the recursive GNN presented in
chapter 3 is trained with the P&ID graphs presented before. For this purpose, the
models were implemented and trained using the Deep Graph Library (version 0.7.2)
with PyTorch framework in Python (version 3.7). The calculations were performed on
an Intel® Xeon® W-2155 (3.31 GHz) CPU in combination with 128 GB RAM. The
aggregation functions listed previously are used in the GNN and compared against
each other. The prediction accuracy is used as an evaluation metric, which was
defined in chapter 3.2. For the examined models, the accuracy is shown below in
Fig. 9.
Fig. 10. Normalized confusion matrix of the recursive GNN with sum aggregation for
the test data set.
Google Scholar
Fig. 5. Workflow of an RNN-based model for predicting subsequent equipment in
P&IDs.
Google Scholar
Safety valves 93
Fig. 4. GUI of the DEXPI-2-graph converter
M. Grabisch, J.-L. Marichal, R. Mesiar, E. Pap
Keywords
The design and engineering of piping and instrumentation diagrams (P&ID) is a very
time-consuming and labor-intensive process. Although P&IDs show common patterns
that could be reused during development, the drawing is usually created manually
and built up from scratch for each process. The aim of this paper is to recognize
these patterns with the help of artificial intelligence (AI) and to make them
available for the development and the drawing process of P&IDs. In order to achieve
this, P&ID data is made accessible for AI applications through the DEXPI format,
which is a machine-readable, manufacturer-independent exchange standard for P&IDs.
It is demonstrated how deep learning models trained with DEXPI P&ID data can
support the engineering as well as drawing of P&IDs and therefore decrease labor
time and costs. This is achieved by assisted prediction of equipment in P&IDs based
on recurrent neural networks as well as consistency checks based on graph neural
networks.

v
Proteus XML, 2017. Proteus schema for P&ID exchange [WWW Document]. URL
https://github.com/ProteusXML/proteusxml (accessed 5.18.22).
(5)
When parsing the DEXPI files, it becomes apparent that the level of detail in the
P&ID description varies greatly depending on the user. Thus, different numbers of
attributes of the XML files are filled in. At the same time, the use of the
attributes leaves some room for interpretation, such that synonymous information
was mapped to different attributes. This requires a certain degree of robustness,
which has been considered in the DEXPI-2-graph implementation. Therefore, several
attributes (e.g. design temperature, pressure, material, …) are deliberately
searched until the desired information for the respective node is found.

As a last investigated variant, the graph attention approach (8) is applied, where
additionally for each
Download : Download high-res image (542KB)
Synth. Lect. Artif. Intell. Mach. Learn., 14 (2020), pp. 1-159,
10.2200/S01045ED1V01Y202009AIM046
As mentioned before, the information from the P&ID is interpreted in the form of a
graph. This makes it possible to store the relationships between components and the
topology in an unambiguous and machine-interpretable way. However, to learn the
graph structure as a whole and to solve tasks such as node classification, edge
classification or link predictions, machine learning methods of graph analysis are
required that can deal with non-Euclidean data structures such as graphs. The
modeling of graph structures is particularly interesting in the field of P&ID
engineering. By learning connections (e.g. piping, signal lines, …) or components
(e.g. valves, equipment, …) based on their neighborhood with the help of AI, it
will be possible in the future to perform consistency checks in P&IDs and detect
errors in P&IDs. This could reduce the amount of time for drawing P&IDs, which will
shorten the time for developing a plants documentation. To achieve this goal, Graph
Neural Networks can be used for modeling Graph Neural Networks (GNN) can be used
for modelling, which have become increasingly important in recent years (Zhou et
al., 2020). A GNN is based on a message passing algorithm that aggregates arbitrary
information from the neighborhood of a node, which will convolve the graph
(Hamilton, 2020). In general, the message passing of a GNN is analogous to the
Weisfeiler-Lehman algorithm to test the isomorphism of two graphs (Weisfeiler and
Leman, 1968), which was introduced in 1968 and in which information is aggregated
from the neighborhood of each node.
Results - GNN node classification
Hochreiter and Schmidhuber, 1997
Add to Mendeley
The workflow of the node classification is shown in Fig. 8. First, all nodes of all
P&ID graphs in the used dataset are divided into a training dataset (80%) and a
test dataset (20%) using a mask. The neural network is then provided with
information about the topology of the graph, as well as attributes of the nodes and
edges, e.g. equipment class, connection type, etc… From this information, the
network generates an embedding for each node and the predicted node class. This is
compared with the real node class and the error is reduced via backpropagation.
After the training is finished, the trained network can be used for node
classification of unseen data (nodes).
Piping equipment 41
Download : Download full-size image
References
Tandoh Henry, Yi Cao
StellarGraph 2020
Heat exchangers 86
CrossRefGoogle Scholar
(8)
Hamilton et al., 2017

Google Scholar
RELX group home page

Google Scholar
List of symbols

ISO 10628-2 - Diagrams for the Chemical and Petrochemical Industry – Part 2:
Graphical Symbols
Bahdanau et al., 2014
DEXPI is a machine-readable P&ID exchange format under development by the DEXPI
Initiative. The initiative consists of owner operators, engineering, procurement &
construction companies, software vendors and research institutions. The latest data
model and the associated DEXPI specification 1.3 (Theißen and Wiedau, 2021) were
published in 2021. Within the specification, different international standards for
the description of engineering relevant data for P&IDs are combined (e.g. ISO 15926
(International Organization for Standardization, 2013), ISO 10628 (International
Organization for Standardization, 2012a), IEC 62424 (International Electrotechnical
Commission, 2016), ISO 10209 (International Organization for Standardization,
2012b). In particular, these include plant breakout structures, instrumentation,
properties of equipment and components, and piping topology. The DEXPI information
model is already offered by some manufacturers and is exchangeable via a Proteus
XML schema (Proteus XML, 2017). At the same time, DEXPI provides the possibility to
be used as a platform for digital plant data in process industry (Wiedau et al.,
2019), which can significantly reduce the development time of chemical and
biotechnological production plants. Additionally, interoperability increases due to
the continuous integration of DEXPI into existing engineering software (Fillinger
et al., 2017). The uniform and machine-readable format as well as the increasing
acceptance of the DEXPI format in the process industry improve the potential for
the application in the field of data science and allow the application of
artificial intelligence (Wiedau et al., 2021).

This research work was supported within the KEEN project (grant number: 01MK20014S)
and has been funded by the Federal Ministry of Economic Affairs and Climate Action
(BMWK).
International Electrotechnical Commission 2016
Fig. 1. Use cases of artificial intelligence to accelerate and improve the
synthesis of…
Table 1
Fig. 3. P&ID topology representing GraphML structure used for further training
View in ScopusGoogle Scholar
m
Separation units 15
Fig 4
MIT Press, Cambridge (2012)
Results – node prediction

Fig. 3. P&ID topology representing GraphML structure used for further training.
In the following, several P&IDs in the standardized DEXPI format are used as
training data, which were exported using the program PlantEngineer from the
software vendor X-Visual Technologies GmbH and converted to graphs in GraphML
format (GraphML Project Group, 2017) according to chapter 2.1. In total, 35 P&ID
graphs from third parties (laboratory and industrial plants) with 1641 nodes and
1410 edges are used. The data set contains 92 different equipment classes (valves,
pumps, vessels, instrumentation, etc.) based on the DEXPI specifications (Theißen
and Wiedau, 2021) and has three different classes of edges (pipes, signal lines,
process connection lines). The ratio of nodes/edges shows that, as expected for
P&IDs, these are very linear graphs with rather low connectivity structures. At a
closer look there are usually many single nodes along a pipeline (e.g. valves,
vessels, pumps, heat exchangers, measuring points, etc.) which results in a kind of
dead ends. Additionally, some P&IDs show inconsistencies in their drawn structures,
which in some cases lead to isolated nodes or several, smaller graphs. However,
these inconsistencies were deliberately included in the data set, as the data is
intended to represent the current state of machine-readable P&IDs in the process
industry to obtain representative results. The influence of the inconsistencies on
the results is examined in more detail in chapter 4.
Improving interoperability of engineering tools - data exchange in plant design

Download : Download high-res image (216KB)

Fig 6
Download : Download full-size image
Advertise
The computational graph shown in Fig. 7 visualizes how it is possible to obtain
information of a component based on its neighborhood through the neural networks
AGGREGATE(k=1) and AGGREGATE(k=2). If the neighborhood components of the vessel are
known, it is possible to predict the class of the vessel or to analyze in a
consistency check how probable it is that a vessel is located at exactly this
position in the graph.
View in ScopusGoogle Scholar
R. Batres, M.L. Lu, Y. Naka
Oeing, J., 2022. DEXPI2graph converter application [WWW Document]. URL
https://github.com/TUDoAD/DEXPI2graphML (accessed 5.18.22).
weight by which the features of nodes influence each other
The main diagonal shows that most components of each class are correctly
classified. This way, all separation units are correctly classified. Process
control equipment (PCE, 96%) and piping components (93%) are also almost completely
correctly assigned. With over 87% prediction accuracy, valves and check valves are
also classified sufficiently well, although it is noticeable that there is
confusion between valves and piping equipment and valves and safety valves.
However, with less than 10%, this is still within tolerable limits. The
classification of the remaining components is much more difficult for the GNN.
Thus, 39% of safety valves are identified as piping equipment and 16% as standard
valves. This is not surprising since both classes are usually found in similar
positions of a P&ID. Classification of pumps (67%), vessels (55%) and heat
exchangers (75%) is only reasonably satisfactory. It is noticeable that all three
classes are mainly classified as valves or PCEs. Classes which are particularly
strongly represented in the data set according to Table 1. The GNN models should
therefore be further optimized in the future. At this point, it is conceivable to
integrate the comparatively underrepresented classes more strongly into the
training by introducing weighting factors. Furthermore, it would be conceivable to
use a larger k, which would aggregate more information. However, this results in a
larger computational effort.
View in ScopusGoogle Scholar
To verify the feasibility, a node classification based on a GraphSAGE model is
applied in the following. GraphSAGE (Hamilton, 2020; Hamilton et al., 2017; Khosla
et al., 2020) is a GNN, which is characterized by a good generalization for unseen
structures in graphs (Hamilton et al., 2017), which makes it particularly suitable
for the currently small amount of machine-readable P&ID data. A challenging aspect
is that GraphSAGE aggregates information of all nodes including the target node.
This becomes a problem if the equipment class is part of the aggregated features,
since the own equipment class is aggregated. This inevitably leads to an accuracy
of 100%, too, since the GNN can read the information in the node. In the case of a
consistency check, this does not make sense, since the information is available but
must not serve as input for the network, since this is what is to be checked.
Hence, it is essential that the features of the target node are not aggregated, but
only their neighborhood. The remedy is a recursive neural network, which was
modified from the GraphSAGE algorithm. For this purpose, a calculation rule (3) is
used, which considers in a case distinction whether information of a node should be
included for aggregation. Mathematically the variation can be described as follows,
where in contrast to the classical GraphSAGE (Hamilton et al., 2017) it has to be
recognized that the model determines the embedding from the starting parameter
The need for data-driven modeling and optimization is growing in the process
industry due to increasing digitalization of processes and tools. Previous work
already shows an acceleration of process development by agent-based environments,
which can be understood as the first steps of intelligent process development
(Batres et al., 1997). Further work shows the potential for using intelligent
process information models by describing chemical plants and processes in terms of
a machine-readable format (e.g. colored petri nets), which enables accessibility
for deterministic algorithms (Zhao et al., 2005). This paper aims to develop
applications that accelerate the development of P&IDs using artificial intelligence
(AI). This requires an accelerated development and application of standardized and
machine-readable file exchange formats to ensure a sufficiently large and highly
available database for the application of AI in P&ID development. In the field of
P&IDs, the DEXPI (Data Exchange in Process Industry) standard is becoming
increasingly popular., as it enables the uniform description of P&IDs and ensures
the vendor-independent exchange of information (Theißen and Wiedau, 2021). At the
same time, DEXPI provides the possibility to be used as a platform for digital
plant data in process industry (Wiedau et al., 2019), which can significantly
reduce the development time of chemical and biotechnological production plants.
Additionally, interoperability increases due to the continuous integration of DEXPI
into existing engineering software (Fillinger et al., 2017).
DE GRUYTER, Berlin (2015), 10.1515/9783110417326
View PDFView articleView in ScopusGoogle Scholar
Schuster and Paliwal, 1997
node2vec

Google Scholar
Google Scholar
Get rights and content

Fig 9
Conclusion & outlook
Download : Download full-size image
Preprocessing – DEXPI-2-graph
K.P. Murphy
Download : Download full-size image
Chem. Ing. Tech., 93 (2021), pp. 2105-2115, 10.1002/cite.202100203
Fig. 8. Workflow of a learning P&ID components using GNNs.
The accuracy is calculated by dividing the sum of true positive (TP) and true
negative (TN) by the sum of true positive (TP), true negative (TN), false positive
(FP) and false negative (FN). The computations were done on an Intel® Xeon® W-2155
(3.31 GHz) CPU in combination with 128 GB RAM. The results are shown below in Fig.
6. The training accuracy and validation accuracy are shown. In addition, the
accuracy5 indicates the correctness, with which the real output of the validation
dataset is predicted, when the five most probable outputs are returned. This score
is of particular interest to investigate whether the trained models are suitable
for a suggestion system that can be used, for example, in a drop-down menu to speed
up the drawing process of P&IDs. Furthermore, both the calculated loss and the
training and validation accuracy over 60 epochs as well as the training time needed
to calculate the 60 epochs are given.
L. Chen
Citations
a weight factor

Wiedau et al., 2021

1
Google Scholar
Google Scholar
Cho, K., van Merrienboer, B., Bahdanau, D., Bengio, Y., 2014. On the properties of
neural machine translation: encoder-decoder approaches. arXiv Prepr.
arXiv1409.1259.
Google Scholar
Adam: a method for stochastic optimization
Download : Download high-res image (235KB)
View PDF

The model is programmed as a recursive function, since the function CONCAT for
concatenating the embedding calls itself to compute the message m during iteration
via all neighboring nodes.
Graph representation learning
AI-based processing of P&IDs
How powerful are graph neural networks?
Theißen and Wiedau, 2021
Neural machine translation by jointly learning to align and translate
Chem. Ing. Tech., 89 (2017), pp. 1454-1463, 10.1002/cite.201700032

y
Outline

Proceedings of the NIPS (2017), p. 17

Mathematically, the message passing of an GNN (Eq. (2)) can be described as follows
(Hamilton, 2020):
(7)
Download : Download high-res image (199KB)
Conclusion & outlook
AI-based processing of P&IDs
Towards a systematic data harmonization to enable AI application in the process
industry
AI Open, 1 (2020), pp. 57-81, 10.1016/j.aiopen.2021.01.001
Previous article in issueNext article in issue

Digital Chemical Engineering

number of nodes
Graph-based P&ID formats are a promising way to improve machine readability of
important process information. The standardized DEXPI format, which is defined and
will be improved continuously by the DEXPI Initiative, is able to store the
topology of process plants as well as all apparatus specifications in a structured
way and make them available in a machine-interpretable manner. This standardization
enables the retrieval of information and a simple application of AI models. Thus,
in addition to the possible accessibility of P&ID data, two approaches have been
presented in this work, which make it possible to recognize patterns in P&IDs and
use AI to help support and increase efficiency in P&ID synthesis. First, P&ID
graphs are decomposed into sequences, and subsequent components can be predicted
based on the drawn partial structure by RNNs. In this regard, high accuracies were
achieved and, in particular the use of AI-based suggestion algorithms was
quantified. On the other hand, the use of GNNs allows for learning correlations in
plant topologies. In the present case, existing components were classified into
P&IDs for a consistency check. In this case, too, good results were achieved in a
first feasibility analysis with a recursive GNN and the data basis is expandable
for the application of AI. However, further fields of application for GNNs are
possible. For example, the prediction of connections is possible, e.g. pipelines,
signal lines, as well as the detection of subgraph structures which could help to
automatically detect functional equipment assemblies (as a part of the engineering
of modular plants).

Proteus XML, 2017

LEARNING FROM FLOWSHEETS: A GENERATIVE TRANSFORMER MODEL FOR AUTOCOMPLETION OF
FLOWSHEETS
Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery
and Data Mining, ACM, New York, NY, USA (2016), pp. 855-864,
10.1145/2939672.2939754
Google Scholar
NTI Ser., 2 (1968), pp. 12-16

Data augmentation for machine learning of chemical process flowsheets

Fig 1
Fig. 1 shows the two use cases identified in this paper for AI-assisted P&ID
synthesis and their respective modeling approaches. In the first use case, a node
prediction generates suggestions about subsequent components based on a recurrent
neural network (RNN). These suggestions can support the user and decrease the time
of the drawing process. The second approach uses graph neural networks (GNN), which
are neural networks especially developed for the modeling of graphs. They can learn
the topologies of process plants, which are stored in the form of a graph, and
enable a consistency check during drawing by comparing the models with drawn P&IDs.
The use of a neural network offers the advantage that patterns and rules for
generating P&IDs can be learned from existing plant topologies. Therefore, no
explicit heuristics need to be stored. Furthermore, the models can be adapted to
the user's requirements and preferences by re-training them with data from the
user. GNNs can be used to perform various classifications and predictions, which
then can be used for consistency checking. The node classification allows for
predicting information, such as the equipment classes of individual nodes to check
whether components within a P&ID are present at meaningful positions. In contrast,
edge classification focuses on the prediction of edge information. In relation to
the P&ID, these are for example the connection types (piping or signal lines) or
the information, whether a pipeline is insulated or heated. In parallel, a link
prediction can be used, which provides information about the probability of a
possible link between two components. This enables the validation of connections
within a P&ID as well as the suggestion of connections during the drawing process.
In the following, the preprocessing for converting the P&IDs into graphs as well as
the two modeling concepts is introduced in more detail. The results for a node
prediction via RNNs as well as a node classification based on a GNN is presented.

V. Turau, C. Weyer
M. Wiedau, L. von Wedel, H. Temmen, R. Welke, N. Papakonstantinou
Abbreviations
Learning from flowsheets: A generative transformer model for autocompletion of
flowsheets
2023, arXiv
Fig. 9. Results of the node classification in a P&ID graph via recursive GNN
grouped by the applied aggregation functions.
The results for all models show deviations among each other. The accuracy for the
training for all models is between 75.4% (sum-MLP) and 87.3% (sum) while the
accuracy for the test data varies between 74.5% (sum-MLP) and 81.5% (sum). It is
striking that the gap between the test and training accuracies for the sum
aggregation as well as the attention aggregation is larger than for the remaining
aggregation functions, at about 6 percentage points. Additionally, the results show
that simpler aggregation algorithms such as sum and arithmetic mean achieve higher
training accuracies than the more complex aggregations using attention, set pooling
or sum-MLP. It is hypothesized that this is due to the fact that all neighborhood
information is equally important in predicting the component class. For this
reason, learning the individual P&ID components works particularly well when the
neighborhood information is aggregated with the same weight, i.e., equally
important. This is especially true for the sum or mean.

Shopping cart
Remote access
View PDF
D. Bahdanau, K. Cho, Y. Bengio
A. Grover, J. Leskovec
Declaration of Competing Interest
Fig. 5. Workflow of an RNN-based model for predicting subsequent equipment in P&IDs
Author links open overlay panelJonas Oeing a, Wolfgang Welscher b, Niclas Krink b,
Lars Jansen a, Fabian Henke a, Norbert Kockmann a

Show abstract
Aggregation Functions (Encyclopedia of Mathematics and its Applications)
Google Scholar
Bidirectional recurrent neural networks

Google Scholar
Download : Download high-res image (499KB)
Download : Download high-res image (217KB)
Fig. 6. Results of the training of following P&ID equipment with different RNN
models.
Fig 5
E
Modified smith predictor for slug control with large valve stroke time in unstable
systems
Download : Download high-res image (872KB)
A comparative study for unsupervised network representation learning
(4)
Cited By (4)
Workflow – node prediction

International Organization for Standardization 2012a

Abbreviations
For a wide acceptance and usage of AI in process engineering, it is necessary that
digital, machine-readable formats are further implemented and are integral part of
software packages. The analysis of the current, still very small data base shows
that P&IDs are often drawn in such a way that they are available as a pdf document
or printout, while important information such as interconnections (e.g. piping,
signal lines) are often missing in the digital representation due to incorrect
drawn P&IDs by the user. If there is a continuous increase in machine-readable
P&IDs in the future, the approaches demonstrated in this paper show the potential
to accelerate the engineering process and save labor time and costs, when
integrated into P&ID software. Primarily, it is recommended to suggest components
during P&ID synthesis and to identify and reduce possible errors through
validation. In addition, further AI solutions can be developed that recognize
functional groups in P&IDs and consider the implementation of these modules when
developing processes.

Citation Indexes:
ArXiv ID 1412.6980
Fillinger et al., 2017
Node classification using graph neural networks
Batres et al., 1997
Fig. 6. Results of the training of following P&ID equipment with different RNN
models
The first step is sampling, during which the graph with its networked structure of
nodes and edges is sequentially transformed into linear input data. These input
data consist of a list of contiguous nodes, which contain the interconnected graph
in linear representations. In the sampling process, all possible turns based on the
number of output edges are made at branches to obtain a reliable representation of
all node interconnections via random walks (Grover and Leskovec, 2016). The
sampling is performed with the function randomBiasedWalk, which is part of the
Python library StellarGraph (package: stellargraph.data.BiasedRandomWalk / version:
v1.0.0rc1) (StellarGraph, 2020). The random biased walk requires four input
parameters. The number of walks defines how many walks are generated from each node
in the graph. The walk length specifies how many nodes are considered per walk.
Important special features of the biased random walk are the return hyperparameter
p and the in-out hyperparameter q, which guide the walk. Thus, 1/p defines the
probability of reversing the sampling direction during the random walk, while 1/q
describes the probability of discovering new nodes in the graph. In this way, the
depth of the search can specifically be controlled (Grover and Leskovec, 2016).
Since the generated samples should represent a clean and linear section of the
plant topology, the parameters must be chosen in a way that the random walk jumps
back as rarely as possible and continuously explores new paths. In this respect,
previous investigations have shown that convincing results can be achieved with
values of p = 1000 and q = 1. Smaller values of p, lead to an undesired probability
of sampling against the flow direction. The sequential samples represent the actual
training data for AI modeling and have a previously defined length l. They are
divided in such a way that the first l-1 entries represent the input sequence x,
while the entries at position l are the corresponding output y. The dataset used in
this work is composed of a total of 4923 sequences, each consisting of six nodes.
For validation, 20 % of the data set are randomly retained as a test set. The
remaining 80% are used to train the RNN.
Hamilton, 2020
References
Google Scholar

(Bahdanau et al., 2014).

Xiaochi Zhou, …, Markus Kraft
The reduction of a graph to canonical form and the algebra which appears therein
Chollet, F., 2020. Keras API - documentation vers. 2.4.0 [WWW Document]. URL
https://keras.io (accessed 2.20.22).

Gulli and Pal, 2017

Download : Download high-res image (186KB)
Digital Chemical Engineering, Volume 3, 2022, Article 100028
An agent-based environment for operational design
plumX logoView details
Strubell et al., 2019
k
2022, arXiv
Article Metrics

Google Scholar

E. Strubell, A. Ganesh, A. McCallum

Proceedings of the ACL (2019), p. 19
Sequential node prediction using recurrent neural networks
is defined (Bahdanau et al., 2014; Hamilton, 2020). In the context of this work, a
softmax approach (9) is used to calculate the weight factor
S. Fillinger, H. Bonart, W. Welscher, E. Esche, J.-U. Repke
Theißen, M., Wiedau, M., 2021. DEXPI - P&ID Specification [WWW Document]. Version
1.3. URL https://dexpi.org/specifications/ (accessed 3.1.22).
Google Scholar
Khosla et al., 2020

Using artificial intelligence to support the drawing of piping and instrumentation

diagrams using DEXPI standard
Google Scholar
The authors declare that they have no known competing financial interests or
personal relationships that could have appeared to influence the work reported in
this paper.
IEEE Trans. Knowl. Data Eng., 33 (2020), pp. 1807-1818, 10.1109/TKDE.2019.2951398
Proceedings of the ICLR (2014), p. 2015
Show 4 more figures
International Organization for Standardization
Vessels 75
Google Scholar
InTech (2008), 10.5772/68
input
AIartificial intelligenceBRNNbidirectional recurrent neural networkGNNgraph neural
networkGRUgated recurrent unitLSTMlong short-term memoryMLPmultilayer
perceptronP&IDpiping and instrumentation diagramPCEprocess control
equipmentRNNrecurrent neural network
Google Scholar
Cambridge University Press, Cambridge (2009), 10.1017/CBO9781139644150
Abstract
We use cookies to help provide and enhance our service and tailor content and ads.
By continuing you agree to the use of cookies.
Download : Download full-size image
Cite

The sequential node prediction can be divided into three parts and its workflow is
shown in Fig. 5.
Thermal control of coke furnace by data-driven approach
About ScienceDirect
2023, Computers and Chemical Engineering
Energy and policy considerations for deep learning in NLP
Introduction
Google Scholar
Machine Learning : A Probabilistic Perspective
Xu et al., 2019
Long short-term memory
Fig 2
l

Google Scholar
Algorithmische Graphentheorie, De Gruyter Studium
bias
Terms and conditions
Show 3 more articles
Google Scholar
length of the directed walks
Download : Download full-size image
Tweets:
Weisfeiler and Leman, 1968
(2)
View in ScopusGoogle Scholar
Privacy policy
Share
Deep Learning with Applications Using Python
Chem. Ing. Tech., 91 (2019), pp. 240-255, 10.1002/cite.201800112
B. Weisfeiler, A. Leman

W.L. Hamilton
Introduction

The recursive GNN is used based on the GraphSAGE algorithm presented in the
previous chapter 4. The number of layers is k = 3. For the activation function, the
ReLU function (Manaswi, 2018) and a subsequent normalization are applied. To
achieve the most efficient prediction accuracy, different state-of-the-art
aggregation functions are used and compared against each other. Basic variants in
this respect are the calculation of a sum (4) or an arithmetic mean (5) (Grabisch
et al., 2009).
Deep Learning with Keras: Implementing Deep Learning Models and Neural Networks
With the Power of Python
and CONCAT concatenates the individually calculated aggregations of each node.
Cited by (4)
Node class Quantity
4
IEEE Trans. Signal Process., 45 (1997), pp. 2673-2681, 10.1109/78.650093
To check how well the classification can be performed for the different classes,
the confusion matrix of the model with sum aggregation using the test data set is
also considered, see Fig. 10. The columns in the matrix describe the predicted
classes, while the rows represent the real classes. Consequently, the main diagonal
displays the number of correctly classified components (TP).
Question answering system for chemistry—A semantic agent extension
Workflow - GNN node classification
Proceedings of the ICLR (2014), p. 2015

Hu, 2008

Download : Download full-size image

International Organization for Standardization, 2012b. ISO 10209, technical product
documentation – vocabulary – terms relating to technical drawings, product
definition and related documentation. Geneva.
View PDF

It appears that the classes have an uneven distribution, which is unavoidable since
this data set is a representative cross-section of all components in a process
plant. In the context of this work, it is important to investigate to what extent
the unequal distribution of training data will affect the results of the
classification.

S. Hochreiter, J. Schmidhuber
© 2022 The Authors. Published by Elsevier Ltd on behalf of Institution of Chemical
Engineers (IChemE).
Kingma and Ba, 2014
(3)
Download : Download high-res image (310KB)
International Organization for Standardization, 2013. ISO 15926-2 - industrial
automation systems and integration – integration of life-cycle data for process
plants including oil and gas production facilities – part 2: data model. Beuth
Verlag, Geneva.
2022, arXiv
ENPRO data integration: extending DEXPI towards the asset lifecycle

Check valves 60
Comput. Chem. Eng., 21 (1997), pp. S71-S76, 10.1016/S0098-1354(97)87481-9

embedding/feature of a node u
The results show that RNNs are generally able to learn patterns in sequences from
P&ID graphs. It is noticeable that the SimpleRNN provides the best results with a
validation accuracy of 78.36%. In the case, where the equipment is part of the five
most likely predictions, even 95.2% accuracy is achieved. The BRNN reaches an
accuracy of 94.39%, while predicting the five most suitable equipment types. The
LSTM and GRU have slightly lower accuracy, suggesting that the effect of the
diminishing gradient for the short sequences involved does not have a significant
effect on the training. At the same time, it should be noted that training for the
GRU took less than one-third the time of a SimpleRNN model. Given the current small
amount of data, this is not a decisive factor with the current setting. However,
should the training of the models be done in the future on large data sets or
continuously, it is recommended to give more attention to this aspect, as the use
of GRUs or LSTMs can save time and resources (Strubell et al., 2019), which should
be considered with respect to a sustainable process development.
Fig 10
x

View PDFView articleView in ScopusGoogle Scholar

Chollet, 2020

Google Scholar
Google Scholar
In the following, the different RNN models are used and trained with the in chapter
2.2 generated P&ID graphs according to the presented workflow. The implementation
is done in Python using the keras library (Chollet, 2020). The "Adam" optimizer
(Kingma and Ba, 2014) is used for all trainings and the calculation of the loss is
performed by the "categorical cross entropy" (Murphy, 2012). The prediction
accuracy is used as an evaluation metric and is defined as follows.
4
International Organization for Standardization (2012)
Deep Learning and Practice with MindSpore, Cognitive Intelligence and Robotics
View in ScopusGoogle Scholar
W.L. Hamilton, R. Ying, J. Leskovec
Neural Comput., 9 (1997), pp. 1735-1780, 10.1162/neco.1997.9.8.1735
stands for the embedding of a node u at iteration step k. UPDATE and AGGREGATE are
arbitrary, differentiable functions, where the aggregation of the neighborhood N(u)
of node u represents the actual "message" m. The parameter k defines the number of
iterations, at which the message passing proceeds, thus represents the number of
hidden layers of the GNN. Since the aggregation of the neighborhood information
must be independent of the order, it is important that the AGGREGATION is a
permutation-invariant function. Based on the embedding for each iteration step k, a
final embedding for each node u can subsequently be determined using a final layer
(Hamilton, 2020).
u
International Electrotechnical Commission, 2016. IEC 62424, Representation of
process control engineering – requests in P&I diagrams and data exchange between
P&ID tools and PCE-CAE tools. International Electrotechnical Commission, Geneva.
To better understand the modeling of plant topology by message passing GNNs, an
example is given in Fig. 7 that relates the aggregation of neighborhood information
to a snippet of a P&ID. The example shows the aggregation by a two-layer neural
network. Since the plant topology is to be learned, we focus in the following on
the equipment information, such as the classes of each component in the P&ID. Thus,
in a first step (k = 1), inferences can be made about the vessel based on the
information from the valve and the heat exchanger. In a second step (k = 2), a
valve's and a temperature sensor's information can be aggregated for the embedding
of the heat exchanger, while the valves´ embedding is influenced by the connected
drive and flow control.
International Organization for Standardization 2013

b
Download : Download high-res image (630KB)
open access
Grabisch et al., 2009
Proceedings of the ICLR (2019), p. 2019

Skip to main contentSkip to article

In addition, there is the possibility of adding a multilayer perceptron (MLP) over
the states to be aggregated, as shown in Eq. (6) for the sum-MLP (Xu et al., 2019).
The idea is that an MLP can transform each state
StellarGraph, 2020. StellarGraph machine learning library - documentation [WWW
Document]. URL https://stellargraph.readthedocs.io/en/stable/README.html (accessed
3.1.22).
C. Zhao, M. Bhushan, V. Venkatasubramanian
Recommended articles
Original Article
Zaheer et al., 2017
Valves 1034
Acknowledgment

Cho et al., 2014

Fig. 7. Example of the neighborhood aggregation of a GNN using a P&ID.
Fig 8
For better application, the DEXPI-2-graph converter is equipped with a graphical
user interface (GUI), which is shown in Fig. 4. The figure also shows a
visualization of an extracted P&ID graph. The path folder containing the DEXPI
P&IDs to be converted is selected and the conversion is started via buttons. A
console window shows the progress, errors and the generated GraphML files. In
addition, a plot window is used to directly check the generated P&ID graphs. The
converter including GUI is published as an open-source application and is available
on Github as a Python application at https://github.com/TUDoAD/DEXPI2graphML
(Oeing, 2022).
Contact and support
number of edges

The basis is an XML-parser (python package: xml.etree.ElementTree / version 3.3) to

read out all relevant information from the DEXPI-file into Python. In the first
step, the parser searches for all Equipment and PipingComponents and creates a
separate node in a graph for each component. At the same time an equipment list is
created, which documents the respective EquipmentClass and ComponentClass of the
individual nodes. In a second step, the PipingNetworkSystem is searched for
connections. These are implemented as directed edges between the already created
nodes of the graph. Similarly, the InstrumentationFunction of the DEXPI P&ID is
scanned and PCE components are added as additional nodes and associated signal
lines are added as edges. This allows for considering control loops and signal
streams in the graph information model. Subsequently, the graph and the associated
equipment list are loaded into a data store and saved for further use. The
structure of the P&ID representing graph is shown in Fig. 3.

Show full outline

Fig 7
Springer, Singapore, Singapore (2021), 10.1007/978-981-16-2233-5

Digital Chemical Engineering, Volume 2, 2022, Article 100010