Process Safety PresenseTM in Eni

Idea, goal and innovation

Ensuring the safety of employees is an essential value of Eni and an absolute priority in all its activities.

The protection of people's physical integrity takes place through an articulated set of activities and tools that develops from the operating sites to the headquarters (HQ) with the aim of ensuring maximum effectiveness in the prevention, management, and control of risks, also through the dissemination of an increasingly mature HSE culture.

All HSE events, occurred worldwide in all Eni’ sites, that harm or have the potential to threaten the safety of people, assets, the environment, or the company’s reputation are carefully documented and recorded in the company databases managed and constantly monitored by HSE Functions.

Taking full advantage of this amount of information is strategic, but it is not easy to achieve without technological support capable of processing tens of thousands of texts, in different languages depending on the country of reference and characterized by technical language.

It is well known that behind the occurrence of an accident there is always the recurrence of many events of potential emerging problems (weak signals), which, if detected, analyzed and solved, could have prevented the accident from happening.

The objective of an effective accident prevention strategy is to intercept as many weak signals as possible and to be able to interpret them to recognize the hazard even before it becomes an accident.

The innovative idea behind “Safety Presense” is to assess in advance the occurrence of risk conditions based on the analysis of the description of HSE incidents and weak signals by exploiting the potential of Artificial Intelligence (AI) and predictive algorithms.

Safety Presense is one of the fundamental elements in the process of changing the HSE way of working and is part of Eni's zero-accident objective.

This project represents a paradigm shift, thanks to the conception and implementation of an innovative approach that aims to combine the potential of Natural Language Processing, Generative AI and predictive models with HSE experience, to move from reactive to preventive approach to safety.

The tool was firstly developed in 2022 to monitor past incident dynamics and anticipate future risks in Occupational Safety and then expanded to include Process Safety events such as Tier 1 and Tier 2 incidents, Tier 3.1, near misses, unsafe conditions, and maintenance data.

The launch of the Process Safety Module happened in July 2024 for all Eni Business Units and Companies having process safety risks.

This enhancement enables precise identification of risks, including anomalies and delays in safety-critical maintenance operations.

Main components and workflow of the Process Safety Module

It is possible to refer to the data recorded in a Eni's database containing all incidents with the term HSE events. These events are:

• Incidents: Tier 1, Tier 2;
• Weak signals: Tier 3.1, near misses and unsafe conditions, maintenance data.

These events can be represented through a pyramid, where at the base are weak signals, frequent events of limited severity, while at the apex are incidents, rare but severe events. Safety Presense is a digital product that, using the Safety pyramid concept, automatically extracts from weak signals recurrences and correlations relating to dangerous situations that have similarities with incidents that occurred in the past, to allow the implementation of targeted preventive actions.

The Safety Presense paradigm is based on five pillars:

1. Topic: risk categories used to classify the incident events (fire/explosion, well area accidents, gas or other release into the atmosphere, spill).
2. Keyword: for each topic, the most significant and representative words extracted from the (e.g. preventive maintenance, line, corrosion, pump, oil, valve). They were used to build the taxonomy at the base of Safety Presense.
3. Historical patterns: combinations of keywords that describe incident dynamics occurred in the past.
4. Latent patterns: combinations of recurring keywords, found in weak signals, which may retrace all or part of an historical pattern and therefore highlight a recurrent hazardous situation at Eni sites level. Each latent pattern is associated with statistical indicators that characterize its level of maturity.
5. Alerts and actions: reporting of a recurrent hazardous situation in a specific location followed by the determination of appropriate preventive actions and the verification of their effectiveness.

The Safety Presense workflow, based on the above pillars, goes through the following steps:

• Extract the Keywords from the description of incident events reported in HSE database.
• Define incident dynamic models starting from the keywords extracted from the previous step from the incident’s description. Those models are collected in the archive of incident dynamics shared at Eni level.
• Search the recurrence of incident dynamic models in the weak signals reported in HSE database combining them with maintenance data to detect recurrent hazardous situations at site level.
• All actions are monitored from input to implementation and evaluation of effectiveness and finally entered into a dedicated database of actions.

In addition, it is possible to extend alerts already present in Safety Presence to other potentially affected sites and to create alerts manually also to verify lesson-learned incidents.

AI engine

The Safety Presense AI engine processes data from Eni's database containing all incidents and maintenance data, generating outputs in terms of keywords, incident dynamics, potential hazardous situations, risk indicators and alerts.

Text Processing

This component represents the entry point of the Safety Presense AI engine and deals with the processing of the text of more than 70 thousand descriptions of safety events, standardizing the language and normalizing it to make it processable by subsequent components.

Keyword Extraction

This component is dedicated to keyword extraction and the construction of a business-specific taxonomy. To extract only the relevant information to describe the incident dynamics, algorithms have been developed with the support of HSE experts to extract representative keywords for given topic.

Pattern mining

Starting from the description of injuries, and based on the taxonomy of keywords just described, this component, using pattern mining algorithms, identifies groups of words that frequently co-occur within the description of incidents related to a specific topic. This chain of keywords extracted from the texts of incidents is called historical pattern and represents a description of a specific injury dynamics that occurred at an Eni site. The collection of this information has made it possible to build an archive of historical patterns shared at Eni level, that is the basis for searching for latent patterns. This step is done by searching for the keywords that make up the historical pattern in the description associated with each weak signal. This makes it possible to identify whether dangerous situations that have occurred in the past are occurring again at a specific Eni site.

Predictive Models

Finally, indicators were developed to describe the risk associated with each latent pattern. The maturity index is calculated by combining several factors that measure the frequency (recurrence), the degree of similarity of weak signals with dangerous events that occurred in the past (latency score), a measure of how recent the weak signals extracted are (lifetime) and the intrinsic criticality of the topic under examination (weight of the topic).

Application and results

Today, the Process Safety Module of Safety Presense allows HSE HQ to monitor all the Eni’s operating sites worldwide. The analysis of Eni’s database from July 2024 to date has led to collect more than 300 incident dynamics, leading to the identification of more than 2000 latent risk situations and sending 87 alerts to HSE Manager of operative sites, to implement appropriate corrective and preventive actions. The use of this tool has allowed HSE to improve its large-scale statistical analysis capabilities of the data recorded in HSE database and to strengthen the prevention mechanism that allow to intervene earlier in dangerous situations that could potentially lead to an incident.