Project Summary
Worksite safety in the utilities sector has plateaued for 8 years.
Utility companies are facing the challenge of reducing costs, while improving standards of service to customers and employee safety. At least 10,000 working days were lost to injury in the sector in 2021. The network cannot afford the continued disruption.
Currently, the process for reducing lost-time injuries involves a large manual data-capture effort and experimental process changes. By the nature of this process, a worksite is already unsafe before anything is done to prevent it.
Instead of waiting for a site to become unsafe, FYLD and SGN want to analyse which conditions contribute the most to worksite safety, then multiply them throughout the network.
In 2021 SGN field teams recorded more than 31,000 video risk assessments using FYLD's AI-assisted technology, leading to a 20% decrease in safety events and £2.9m realised in related benefits. Of the safety events recorded, ~20% were failures to make the site safely accessible for teams and members of the public.
FYLD's vision is to assist every fieldworker to take corrective actions and put unsafe conditions right in real time, before they develop into something more serious.
During the Alpha phase, FYLD will build a machine-learning model to assess how effectively site controls have been deployed and determine which strategies lead to the safest outcomes. This model will be used to power an augmented reality proof-of-concept that will demonstrate how interventions can be made in real time -- with significant benefits to workers and members of the public.
In the SIF discovery phase, FYLD set out to determine whether the risk assessment data could be used to forecast the potential risk of a safety incident. We found a statistically significant inverse correlation between the number of risk assessments recorded and safety incidents logged. We tested 15 predictive machine-learning models and two showed potential - in both cases, recall surpassed the 50% threshold on multiple occasions.
However, we discovered that risk assessment data, alone, doesn't give the full picture. Fieldworkers at different sites can record nearly identical risk assessments, but only some of those sites will result in a safety event. This pattern presents even where the same control measures are, theoretically, applied to the same degree.
That's where the opportunity lies.
The earliest point that an intervention can be made to improve site safety is the moment after a risk assessment is completed.
Making worksites safer will improve the efficiency and resilience of the network, reducing time lost to injury and the disruption caused when incidents occur. SGN saved £240,000 in fines, in 2021, by simply recording evidence from the worksite.
SGN have been working with FYLD since the company's inception. In 2021, SGN realised £2.9m in benefits through using FYLD and both companies have recently entered a three-year innovation partnership targeting a further £16m in savings. At the last count, just over 2800 people at SGN are already using FYLD -- this has resulted in more than 143,000 point of work site assessments.
FYLD are best placed to assist SGN and bring this solution to market:
- A high-performing team with experience launching and maintaining AI/ML products
A demonstrable history of realising significant cost savings for utilities companies by deploying innovative solutions
Existing technology and datasets that can be built upon
In-house experience and expertise in change-management required for digital transformation, specifically within safety and productivity of utilities companies, at scale
Innovation Justification
The current strategies for reducing safety incidents aren't working.
One of the oldest safety theories is the Accident Triangle, which Is sometimes compared to an iceberg, where the visible part consists of reported injuries and fatalities, and the invisible part under water are all the unreported incidents and near misses. If you can capture the indicators, you can run experiments to reduce them -- reducing the more severe incidents that may occur as a result.
However, indicator events are significantly underreported. A number of modern, data led studies also question the validity of the accident Triangle (Marshall et al, 2018 and Moore et al, 2020).
In 2021: SGN had a shortfall of ~85,000 reports, capturing just 2% of what we estimate they should be; Thames Water (a company about 5 years ahead on the safety journey) captured 60% of the estimated events, but still had a shortfall of ~80,000 reports; The utilities industry (as a whole) should have recorded 33.6m indicator events
The reporting process is manual, difficult to train or incentivise and leads to inconsistent and low-quality data. Fieldworkers are busy and their workplace is distracting -- in SGN's case, the average description is just 260 characters long.
It will take years to address the delta, before any meaningful insight can be gleaned from it.
SGN are currently trialling methods to address the data delta; basing individual performance on the number of hazards and near misses reported by each individual.
However, Chris Trodd, the Head of Safety at SGN, is seeking a technological solution to capture this data in the long term, "When someone does a task repeatedly, they can't see the near miss they've missed. That's where machine-learning and [object-recognition] could come in. [It] presents the opportunity to say, 'just before you go and do that - is it such a good idea?'"
Rather than relying on fieldworkers to manually report safety events in the hope of preventing them next time, we believe technology is better placed to:
- Capture high-fidelity data; object recognition models can be trained on footage from body-worn cameras
- Recognise patterns in the data;
- Test strategies to lower the risk, at scale;
- Make Interventions in real time automatically triggered when thresholds are met
The combination of object-recognition and natural language processing used in FYLD's VRA technology has changed point-of-work risk assessments from a box-ticking exercise to a critical method for sharing up-to-date information about the worksite and driving behavioural change from the first time it is used by a field worker. Applying these same methods to the controls and outcomes will provide a real-time view of risk. By overlaying contextual data from other sources -- location, weather, traffic -- FYLD will also determine how these factors contribute.
By capturing the risks, controls and outcomes of all fieldwork in FYLD, our model will be able to:
- Assess the real-time risk of every worksite
- Determine the efficacy of controls in place to mitigate the hazards
- Suggest interventions that have provably lowered risk elsewhere
- Record outcomes and learn which strategies are most effective
- Share the findings between all parties involved
Benefits
There are three key areas where financial benefits will be accrued:
- Reduction of injuries and incidents and flow on impact to better management of network operations
- Lower cost to capture data about indicator events
- Reduction in fines (associated to permit condition breaches)
It would take
1) Reduction of lost-time injuries
The cost of lost-time injuries in the utility sector 2021 amounted to \>£125m. In 2021, SGN field teams recorded more than 31,000 video risk assessments using FYLD's AI-assisted technology, leading to a 20% decrease in safety events.
We hypothesise that FYLD's Predictive Safety Interventions will be able to make at least an additional 20% reduction in lost-time injuries. Discounting the additional savings related to reducing fatalities and other injuries, the total projected savings made possible amount to £59,201,640.00.
2) Lower cost to capture data about indicator events
Based on the number of lost-time injuries (LTIs), SGN should have reported 85,072 indicators.
It will cost SGN an absolute minimum of £280k (just considering field worker time, and not including program management) per year to address the data delta manually:
- 85,072 indicators to report
- 6 minutes per report (conservative estimate, likely takes longer)
- £0.55 per minute weighted average salary
We estimate it will cost the utilities industry ~£58m to record enough indicator events to have a meaningful impact on more severe incidents. This doesn't include the cost of training the workforce, analysing the data or the significant changes required to transform the safety culture.
FYLD helped SGN save 10,400 field work hours (across ~375 FTEs) in a year by transforming a paper-based process into an AI-analysed video recording.
In the alpha phase, we will determine how effectively the combination of body-worn cameras and an OR/NLP model could reduce the time taken to report an indicator event.
Impacts and benefits
The target benefits from the PSI model remain consistent with the forecast benefits in the Alpha submission, but with minor modifications from our learning during the Alpha project delivery, which will be taken forward into our Beta project forecast benefits.
The project will deliver a cost reduction in operating energy networks and wider energy systems. This will be delivered by:
1. Reduction of injuries and incidents
2. Lower cost to capture data and share learnings about indicator events
In addition, the project will deliver a new to market product and process. Use of predictive safety models to manage and mitigate risk is new to the Utility sector in the UK, with no established predictive models present in the market. This will deliver a core benefit in line with SIF target benefits of:
3. Improved risk visibility across network operations
1. Reduction of lost-time injuries
Based on HSE data, the cost of lost-time injuries in the wider utility sector in 2021-22 amounted to more than £160m (HSE data). We hypothesise that through accurately predicting safety events and intervening before a fatality or non-fatal injury occurs, we can substantially reduce the occurrences of fatalities and non-fatal injuries and reduce this cost.
In the Alpha phase, we demonstrated 57% accuracy of the PSI model, and we demonstrated a 35% improvement in response rate from workers to improve safety on site or review the site conditions. If the PSI model could predict 65% of incidents before they occur, with a 75% success rate, the potential annual saving could be more than £60m per annum when widely adopted across the industry.
2. Lower cost to capture data about indicator events
Based on the number of lost-time injuries, in the 21-22 financial year, SGN should have reported over 66,950 indicator events including near-misses, based on the Heinrich and Bird safety triangle, widely accepted in the industry. Manual reporting of near misses and safety indicator events takes a conservative estimate of 6 minutes. PSI has the capability to deliver substantial time and resource saving from replacing manual processes with an automated model with human validated data collection i.e., the PSI model will be able to detect when a likely safety incident has or will occur and request the worker to validate this. The assumptions in this saving are modest and do not include the extensive time require to analyse data and share learnings.
3. Improved risk visibility across network operations
In line with the target SIF benefit for new to market processes, PSI will deliver the qualitative benefit of improved risk visualisation through performance reporting against predictive safety analytics. Operational and safety leaders across the sector will have the newly developed ability to report on safety trends with live quantification of risk enabling a snapshot view into an AI quantified risk level across real-time sites and workstreams. The reporting and visibility of response to high-risk safety events has the potential to become the first of its kind in enabling cross-company benchmarking of risk through enhanced prediction-based risk visibility.
