SPE-192820-MS

Artificial Intelligence in HSE

Dayanch Hojageldiyev, Dragon Oil Turkmenistan Ltd.

Copyright 2018, Society of Petroleum Engineers

This paper was prepared for presentation at the Abu Dhabi International Petroleum Exhibition & Conference held in Abu Dhabi, UAE, 12-15 November 2018.

This paper was selected for presentation by an SPE program committee following review of information contained in an abstract submitted by the author(s). Contents
of the paper have not been reviewed by the Society of Petroleum Engineers and are subject to correction by the author(s). The material does not necessarily reflect
any position of the Society of Petroleum Engineers, its officers, or members. Electronic reproduction, distribution, or storage of any part of this paper without the written
consent of the Society of Petroleum Engineers is prohibited. Permission to reproduce in print is restricted to an abstract of not more than 300 words; illustrations may
not be copied. The abstract must contain conspicuous acknowledgment of SPE copyright.

Abstract
Artificial Intelligence (AI) is a relatively new concept for oil and gas industry. Yet, its agile nature and
massive potential give a promise for constantly evolving oil and gas business in tackling such issues
as environmental protection. AI systems utilize a variety of technologies making it possible to produce
comprehensive, nonetheless easy-to-use tools, which can be used by Health, Safety and Environment (HSE)
professionals in their daily activities. This paper examines a so-called HSE AI Assistant - another tool
aimed to provide support to HSE professionals in getting instant answers to questions on a variety of
regulatory topics in occupational health, safety, and environmental protection. This AI tool will utilize latest
NLP technologies to retrieve data from a bank of up-to-date HSE regulations and standards, and provide
answers to required questions. The tool is a much more efficient alternative to traditional search engines
and closed intranet data retrievals, as the AI system has an ability to understand human questions, hunt for
applicable data in attribute files, and produce the most relevant answer to the questions. The system utilizes
latest machine learning technologies, and has a capability of constant learning and evolvement. This will
be especially useful in terms of changing local legislation and updates to the HSE regulations. The HSE
AI Assistant can also be supplied with a capability to assist HSE specialists in classification of hazardous
materials and help to perform regular HSE audits in the areas where hazardous materials are used and
disposed. Developed locally in Turkmenistan by a team of HSE and IT professionals, the HSE AI Assistant
can become a very accessible and easy to maintain reference and solution tool for HSE specialists working
on COSHH audits and environmental impact assessment. The HSE AI can also assist those who control the
chemical stock in the warehouse, and who manage the hazardous waste at the disposal sites.

INTRODUCTION
"Some people call this artificial intelligence, but the reality is this technology will enhance us. So instead
of artificial intelligence, I think we'll augment our intelligence." - Ginni Rometty, IBM1
Studies show that human factors and errors contribute at least 92% to the root causes of workplace
accidents (R. Gordon, 1996)2. From near-miss slips and lapses caused by the personnel' inattentiveness
and knowledge-based mistakes, to major industrial accidents triggered by the organizational safety system
violations and decision errors - it is the human involvement in the process safety that HSE professionals are
striving to control, and sometimes, as we look in further, eliminate at all.
2 SPE-192820-MS

The hoary debate whether machines can replace humans at the workplace today settles with an inevitable
positive answer. There are numerous vivid examples where robots overtake the working environments,
increasingly in the manufacturing industry, leaving little for the humans to intervene. Robots are stronger,
faster, and more precise when it comes to a routine task. Moreover, these days robots are on the steady way of
replacing humans in the jobs that require creative and interpretive sensing within highly variable conditions.
Currently it takes around 83,000 processors operating for 40 minutes to run the equivalent of one second
of computations performed by just 1 percent of the human brain (A. Feinberg, 2013)3. However, machines
learn. They can sense their environment, think, and respond to what they sense, meeting programmed
objectives (PwC, 2018)4. Machine Learning (ML) systems learn how to combine input to produce useful
predictions on never-before-seen data (Google Developers, 2018)5.
Artificial intelligence today is a real thing. Thanks to the technology advancement, particularly growing
computer processing powers, availability of massive amounts of various structured and unstructured Big
data, and of course, ever-growing interest of major corporations and their investments, AI systems are
evolving dramatically transforming industries all around the World.
Artificial intelligent systems using machine learning are constantly improving in accuracy, presenting
an opportunity for constantly evolving oil and gas business in addressing such issues as environmental
protection.
The capabilities of AI systems have long outgrown the narrow decision making of online marketing and
translation services. Nowadays, automated intelligence tools complete repeated and labor-intensive tasks
in the workshops; help people in reviewing big batches of random data and identify important patterns
for further planning; develop viable models to assist examination of various event scenarios, predict and
mitigate emergencies. Quite successfully, AI systems are advancing in autonomous decision-making, which
eventually eliminates human supervision. The latter, however, raises a lot of questions in regards to the
safety of the AI systems themselves.
AI systems utilize a variety of technologies making it possible to produce comprehensive, nonetheless
easy-to-use tools, which can be used by Health, Safety and Environment professionals in their daily
activities.

BACKGROUND
AI IN ENVIRONMENTAL PROTECTION
In their 2018 report "Harnessing Artificial Intelligence for the Earth," PricewaterhouseCoopers (PwC)4
discussed a range of over 80 use case applications of AI for the environmental protection and
sustainable development. As shown, different types of AI agents could tackle the mankind's most pressing
environmental challenges such as climate change, biodiversity and conservation issues, healthy oceans,
water security, clean air, weather and disaster resilience. Machine learning algorithms and data-driven
decisions are already put for use in analysis of energy consumption, deriving information from millions
of smart meters, and assisting sustainability experts in increasing efficiency of the natural resources' use
and lowering environmental impacts. The data-driven tools are assisting farmers in increasing productivity
and improving crop yields by collecting data from various sensors, such as cameras and even drones;
allowing agriculture specialists to immediately react to anomalies and plan the field conditions. Artificial
intellect agents powered by the neural networks are gathering real-time feeds at the waste management
facilities and recycling stations to select and sort the right items from the belt; smart trash bins can
identify a wide range of waste composition to separate non-recyclable from recyclable items. Geographic
Information Systems (GIS) powered by the machine learning modelling gather information from the satellite
sensors to simulate the impact of different land-use activities, and map out planting patterns for natural
habitats protection. AI-enabled modelling is also used to predict dispersion of air pollutants under complex
environmental conditions. Finally, AI tools that combine data from open source sensors and social media
SPE-192820-MS 3

reports can use machine learning for real-time flood mapping providing accurate and up-to-date information
for governments and local residents in their planning and response.

METHODOLOGY AND LIMITATIONS

The variety of uses of AI powered agents in environmental protection has motivated a team of IT and HSE
professionals from Turkmenistan to think about applying the potential of AI capabilities in improving the
work of HSE professionals in oil and gas industry.
The Team started first with the analysis of pains and gains for the project. What kind of issues and
problems do HSE officers and supervisors face during their daily activities? What kind of tools would be
useful in improving and advancing their work? These questions were asked among many others in order to
identify the data environment, which would be used to breed the AI agent.
As with any data-driven AI tool, type and amount of data available is perhaps the decisive factor when
selecting the AI agent environment. Thus, the Team has also looked at the HSE fields where ample amounts
of workable data for machine learning could be gathered. These data sets were preferred to be structured in
order to save time, and provide for determined and stable machine learning. Finally, the Team was limited
to open source data only.

CONTROL OF HAZARDOUS MATERIALS AND ARTIFICIAL INTELLIGENCE

Industrial advancement and new technological approaches in oil and gas drilling and maintenance have
brought in thousands of new types of chemicals. These might offer an increased efficiency and convenience
to the processes. HSE specialists, however, have to think and offer ways for safe handling, use, transport,
and disposal of the chemicals, as most of the chemicals are not biodegradable, and pose direct harm to
health, safety and environment.
Control of substances hazardous to health (COSHH) is a vast field of attention in oil and gas industry.
The chemical risk assessment is the significant part of the control measures. General tasks such as weighing,
mixing, storage, transfer, and ventilation along with the dust extraction have to comply with international
standards and respond to local environmental and sanitary legislation.
In their COSHH Essentials guide, Britain's Health and Safety Executive Board (HSE)6 describe the
correlation between the factors in chemical risk assessment and appropriate control measures:

Basically, in order to determine the right and adequate control measure, the HSE professional needs to
execute the risk assessment combining health hazard with exposure potential factors to identify the exposure
ranges in different situations and determine the degree of control needed.
Further on, the COSHH Essentials guide7 is offering four steps for developing a chemical control scheme:
Step 1: Grouping the hazards that have "adequate control" at similar airborne concentrations.
Step 2: Grouping the "physical properties" with amounts used that show similar potentials for exposure.
Step 3: Assessment of the anticipated exposure from each control approach, for each combination of
physical property and amount.
Step 4: Linking Step 1 with Step 3 to form the generic risk assessment scheme.
How much risks are associated with any particular chemical depends on the characteristics of that
chemical, the availability of pathways for it to reach the receptor, and the characteristics of the receptors
(FEMA, 1993)8. The specialist working on the risk assessment needs to bear in mind that information can be
highly misleading. It is vital to understand the hazardous properties of the substances, and what effects can
it have on humans and the environment. Next, the concentration of the substance needs to be determined.
The higher quantity or concentration of a toxic substance is more dangerous. The risk posed by a toxic
4 SPE-192820-MS

substance entering the environment also depends on the pathways available to it, and to what extent the
concentration is reduced by the time it reaches receptors. The form of the substance is another factor to
consider - whether the substance is in large blocks or tiny particles, or whether it is a Iiquid or a vapor, the
form of the chemical will determine the pathways of how it could enter and affect the human body. The
chemical and physical characteristics of the substance determine the environmental pathways and speed of
spreading in the environment. The conditions and the area in which the substance is contained and stored
(e.g. rusty containers; proximity to natural water catchments and communities) determine the possibilities
for spill and leakage incidents, as well as the severity rate for disaster pollution and contamination. Finally,
the projected duration of exposure to the chemical also contributes to the health and environmental risks.
A lot of useful information that tackles most of the points raised above is duly described in the material
safety data sheets (MSDS). By standard, the chemicals leaving the manufacturer's production line has to
be supplied with this "safety passport".
The Code of Practice for the preparation of MSDS outlines the following structure of the document
("Chemwatch," 2014)9:
1. Product identifier and identity for the chemical;
2. Hazard identification;
3. Composition, information on ingredients;
4. First Aid measures;
5. Fire Fighting measures;
6. Accidental release measures;
7. Handling and storage;
8. Exposure controls, personal protection;
9. Physical and chemical properties and additional parameters;
10.Stability and reactivity;
11.Toxicological information;
12.Ecological information;
13.Disposal considerations;
14.Transport information; and
15.Regulatory information.
The data contained in the MSDS could serve as a pool of feature attributes and input variables for a data-
driven AI tools that operate conventional classification models. These models could assist HSE specialists,
as well as the material dispatchers in identifying characteristics of the chemicals and their exposure controls
during handling. There is also a possibility to include regression modelling into the AI tool, as the ML
engine would evolve throughout the time, and provide an opportunity for answering continuity questions
on hazardous substances' control. An example of the questions in this case would be drawing patterns on
the likelihood of environmental incidents with certain chemicals being sent to a certain facility for storage.
Based on the statistical relationship between historical data, the AI tool could indicate previous incidents
with the relevant types and amounts of chemicals coming from the same manufacturer in the same type
of containers, and being sent to the same facility for handling and storage, thus pointing on the need for
additional precautionary measures with that type of chemical batches.
SPE-192820-MS 5

THE HSE AI ASSISTANT

PROBLEM SCOPE
Perhaps, the major problem with executing the COSHH is understanding that COSHH is about control of
the risks associated with hazardous substances; and the risk assessment involved is really about evaluation
of risks, not the hazards.
The high-demand and time pressure during material-dispatching work in oil and gas industry might put
additional stress on to the people involved, and lead to work process errors. In this regard, an automated
COSHH tool could significantly ease the chemicals' dispatching and control work. Moreover, a bank
of up-to-date legislative data connected to the tool would ensure the compliance with the governmental
requirements and standards.
Work with MSDS requires time and close attention to the details, especially in cases when manufacturer
may modify certain parts of the data sheets in order to reflect production process changes, as well as
amendments to the current standards.
Errors may also appear in instances when delivered batches of chemicals are missing the MSDS. In that
case the material dispatchers and HSE officers need to track down and retrieve the MSDS for the said
chemicals, which requires time and administrative resources.
Another issue is that during the risk assessment for the chemicals, some of the conditions and situations
might be misinterpreted due to inattentiveness or wrong description. In this situation, the ML driven tools
could learn the patterns and offer determined decisions based on the input data during the risk assessment.

AI OPPORTUNITIES
Automation of the data retrieval from the MSDS and proper use of that data in the risk assessment process
could significantly ease the work of specialists working with COSHH. Ideally, an AI tool could support the
work of HSE and Logistics specialists serving as a knowledge base and instant situation-specific answering
system during the risk assessment and decision making for the use of control measures, maintenance, health
and environment protection, and further training.10
There is a number of automated solutions offered to the clients on COSHH analysis and risk assessment.
At a subscription fee, a so-called "COSHH-365™" tool11 claims to quickly create risk assessments based
on conditional input data, provide access to a large safety sheets database, store existing safety data sheets
online or request one through the system, and provide constant updates for the safety datasheets.
The COSHH Essentials e-tool6 of the Britain's Health and Safety Executive Board (HSE) provides direct
suggestions on the exposure control measures and actions in a form of factsheets. In order to produce such
an advice, the user needs to input some basic information about the substance by answering a queue of
prompted questions.
The HSE AI Assistant tool, which is being developed by the Team in Turkmenistan, will pilot with a
Chatbot designed to answer basically any question related to COSHH. The schematic illustration below
describes the fundamentals of the HSE AI Assistant's work:
6 SPE-192820-MS

Being a software that communicates with humans using a more or less natural language, chatbots provide
an opportunity beyond simple text or voice exchange. Chatbots also interact with users by the means of
graphical interfaces, exchanging image and video files.12
The advantages of chatbots include their availability 24/7; quick and stable access to large information
databases; speed and accuracy in narrow decision-making. Chatbots can certainly replace humans in some
routine tasks.13 Turkmenistan Team, however, is looking to produce an AI software that would support
human agents in their routine tasks by offering relevant information, which the human agent can then act
on finalizing a task.
The HSE AI Assistant would start as a scripted chatbot, i.e. conversations in this type of chatbot would
only follow predetermined paths. For each question from the user, the chatbot would return a few plain
options to choose from, so the user would determine the next step in the conversation. The returned options
might be presented to the user in different ways - as a text, voice or touch response.
The next step for the HSE AI Assistant would be use of artificial intelligence techniques with elements
of machine learning. The artificial intelligence modelling will allow the AI tool to be more flexible in terms
of the user input. They can accept input in the form of text or voice statements, and improve the decision
making as more labeled data would be exchanged by the users.

DATA-GENERATING PROCESS (DGP)

As mentioned earlier, the databank for the HSE AI Assistant would be comprised of the up-to-date MSDS
files available in the open sources and client databases. While the overall structure of the MSDS is following
one agreed outline, as described above, each manufacturer is offering their own design of the datasheets.
That poses a problem for the AI tool developers, as the data has to be structured up to one unified format.
Therefore, it was decided to develop a model datasheet along with an algorithm, which would re-format
and convert individual MSDS files, as well as MSDS batches, and provide a structured data for the AI
informational store.
The next issue to be resolved is the data labelling. The feature data has to be classified through the labels
in order to ease the machine learning process for the HSE AI Assistant and provide accurate output values,
predictions, and decisions.
Classification and Labelling determine criteria for chemicals, and answer questions on whether the
substance can cause harm to human health and the environment. It involves the identification and evaluation
of the physical properties of a chemical, along with its health and environmental effects, and then
communicating those hazards via a label.14
An extensive list of chemical labels is available in open access through the European Union's
"Classification, Labelling, and Packaging" (CLP) Regulation as part of European Commission's efforts
within the UN's Globally Harmonized System (GHS). The CLP Regulation provides several hundreds of
SPE-192820-MS 7

hazard labels (H-labels) and precautionary labels (P-labels) for substances and mixtures, including those
that are already on the shelf. The labels are provided in both graphic and alpha-numeric formats.
These labels would be used by the HSE AI Assistant in order to provide maximum clarity in identifying
chemicals' properties and impacts, as well as offer actions for handling, mixing, storing, and safety.
Another set of data proposed for inclusion is the current environmental, sanitary, and state standards
legislative procedures and rules. The legal database is still under compilation, and the developers are in
process of executing the structural design for this segment of the databank.

SAFETY OF THE ARTIFICIAL INTELLIGENCE

The ultimate goal of any AI and ML tool is to provide accurate and reliable predictions. In theory, the
algorithms used in the AI tools should be blind to prejudice, judge according to objective rules, and eliminate
biases. In reality, however, these algorithms actually might do the opposite (C. O'Neil, 2017)15. Since the
algorithmic models are based on historical data, they not only incorporate previous discriminations and
biases, but reinforce them.
In one of the interviews, Elon Musk, business magnate and advocate for regulatory control over the
AI technologies, said that: "The pace of progress in artificial intelligence is incredibly fast. Unless you
have direct exposure to groups like Deepmind, you have no idea how fast it is growing at a pace close to
exponential. The risk of something seriously dangerous happening is in the five-year timeframe. 10 years
at most."1
In 2018, the EU has enforced the General Data Protection Regulation (GDPR) that mandates all
companies and organizations offering goods or services to EU data subjects to develop and apply tight data
management systems and processes ensuring that no data is lost or compromised. The GDPR is considered
to be the first major international legal document that holds accountable the AI developers and requires the
AI to provide an explanation for decisions made (T. Revell, 2018)16.
One of the problems that AI faces these days is that many of the data-driven models confuse correlation
with causation, and generate decisions based on correlations alone, which is a big drawback in data analysis
(J. Pearl, 2010)17.
In his report on AI and Safety, S. Eldevik of DNV GL18, writes that the field of AI safety is broad, and
an active area of research, especially when implications of using AI and ML in low-probability and high-
consequence scenarios are evaluated, e.g. decisions in safety-critical systems.
S. Eldevik states that in order to have good prediction accuracy, it is important that the model is developed
based on sufficient relevant data from the data-generating process, it is also important that the DGP is
computationally feasible, and an algorithm can learn the DGP from a finite set of training data. For valid
predictions, the generated models and scenarios need to be stable with respect to the training data, which
in case of the high-consequence scenarios are scarce.
In the end, the researcher concludes that usage of ordinary data-driven methods only is inappropriate
for decision making. S. Eldevik describes several methodologies that need to be used along within the AI
decision modelling. The Tuning of all key hyperparameters and publication of the optimization results is
mentioned as one of them. Sliced Analysis that helps to break down performance measures by different
dimensions and categories of the data to unmask important accuracy and quality effects of testing the full
set of data. Next, the so-called ablation studies of all changes from prior baselines should be included to test
change in each of the components or a combination of those. Finally, the sanity checks and counterfactuals
are mentioned as a way to understand the model behavior. Analysis of counterfactual or counterusual data
outside of the test distribution may, for example, show how well an AI model performs on data from users
with different demographic distributions. The offered methodologies shall be used in AI and ML tools
incorporated into the safety-critical systems in order to ensure that the machines are trained on sufficient
and relevant data, and are able to produce predictions without potentially catastrophic scenarios.
8 SPE-192820-MS

Finally, DNV GL argues that ML and AI model predictions should also adhere to established physical
and causal knowledge about the test subjects. This implies the logical input to test the offered predictions
and analyze their modelling stability for further use.

CONCLUSION
Artificial intelligence technologies today offer a massive pool of decision-making models for oil and gas
industry. The tools developed based on the AI technologies and machine learning can help specialists
working in the fields of HSE and Logistics to advance their efforts in environmental protection. There are
over 80 areas in environmental protection initiatives where AI and ML models are actively used. These
tools prove to be more effective in executing routine tasks with higher accuracy and faster completion
rates. The AI is also progressing in unsupervised decision-making, which however, raises concerns among
the scientists and law-makers as algorithms using historical data may not only incorporate previous
discriminations and biases, but reinforce them. The HSE AI Assistant is another tool aimed to provide
support to HSE and other professionals in getting instant answers to questions on a variety of regulatory
topics in occupational health, safety, and environmental protection. Using the NLP and NLG technologies,
the AI tool will retrieve data from a bank of up-to-date MSDS and HSE regulations and standards, and
provide answers to required questions during the risk assessment, handling, mixing, storing, and disposing.
The tool will offer alternatives to traditional search engines and closed intranet data retrievals, as the AI
system will correspond with the user via chatbot technology, offering action prompts via user-friendly
graphic interface. The system will eventually evolve from the scripted chatbot technology with several
answer options and probability breakdown to a fully-fledged AI chatbot that would learn from each question
asked by users. The HSE AI Assistant will be supplied with a capability to assist HSE and Logistics
specialists in classification of hazardous materials and help them to perform regular HSE audits in the areas
where hazardous materials are used and disposed.

ACKNOWLEDGEMENTS
This manuscript has been developed for the ADIPEC Technical Conference, and describes a project in work.
This desktop research summarizes some of the latest achievements in the use of AI technologies in HSE
and COSHH. It also offers an overview of the issues and problems that the use of AI presents. Being a part
of the HSE AI Assistant Tool development team, the author describes the currently proposed design and
features of the tool. The author thanks his colleagues at the team and his employer company for encouraging
this research and providing support.

LIST OF ABBREVIATIONS
ADIPEC – Abu Dhabi International Exhibition and Conference.
AI – Artificial intelligence.
CLP – "Classification, Labelling, and Packaging" Regulation No.1272/2008.
COSHH – Control of Substances Hazardous to Health.
DGP – Data-generating Process.
DNV GL – A trademark name of the international accredited registrar and classification society
headquartered in Høvik, Norway
EU – European Union
FEMA – USA Federal Emergency Management Agency
GDPR – General Data Protection Regulation No.679/2016.
GHS – Globally Harmonized System.
HSE – Health Safety and Environment.
ML – Machine Learning.
SPE-192820-MS 9

MSDS – Material Safety Data Sheet.

NLG – Natural Language Generation.
NLP – Natural Language Processing.
PwC – PricewaterhouseCoopers Company.
UN – United Nations.

REFERENCES
1. B. Marr, 2017. "28 Best Quotes About Artificial Intelligence". Forbes Magazine online.
2. R. Gordon, 1996. "The contribution of human factors to accidents in the offshore oil industry".
Aberdeen University, Psychology Department. Kings College. Aberdeen AB24 2UB.
3. A. Feinberg, 2013. "An 83,000-processor supercomputer can only match 1% of your brain,"
Gizmodo online.
4. C. Herweijer, D. Waughray, 2018. "Fourth Industrial Revolution for the Earth. Harnessing
Artificial Intelligence for the Earth". A report of PricewaterhouseCoopers (PwC).
5. Google Developers, 2018. Machine Learning Crush Course online.
6. Britain's Health and Safety Executive Board (HSE), 2018. "COSHH Essentials" e-tool, available
at http://www.hse.gov.uk/coshh/essentials/index.htm
7. Britain's Health and Safety Executive Board (HSE), 2009. "The technical basis for COSHH
essentials: Easy steps to control chemicals". Online publication.
8. USA Federal Emergency Management Agency, 1993. "Hazardous Materials: A Citizen's
Orientation". Home Study Program.
9. Global Chemical Watch (Chemwatch) Network, 2014. "Structure of an MSDS" available at
https://jr.chemwatch.net
10. Diamond Environmental Ltd., 2011. "Common problems with COSHH assessments". Online
article available at https://diamondenv.wordpress.com
11. Sevron Ltd., 2018. "COSHH-365" e-tool available at https://coshh365.sevron.co.uk
12. Verloop.io. Conversational marketing and sales platform.
13. Botpress.io, 2018. "The Whats & Why of Chatbots".
14. Health and Safety Authority, 2010. "Classification, Labelling, and Packaging" (CLP) Regulation.
Online publication.
15. C. O'Neil, 2017. "Weapons of Math Destruction: How big data increases inequality and threatens
democracy." Broadway Books.
16. T. Revell, 2018. "After the Facebook scandal: The grand plan to hold AI to account". New
Scientist, 11. April 2018.
17. J. Pearl, 2010. "Causality: Models, Reasoning, and Inference". Cambridge University Press, New
York. 2nd Edition 2009.
18. S. Eldevik,PhD. "AI+Safety. Safety Implications for Artificial Intelligence," the HSE Now article
of DNV GL available at https://www.spe.org