Large, supply-chain heavy industries face increasing safety challenges along with higher expectations and stricter regulations for safe work spaces. There are slippery floors and heavy machinery, difficulties on the road and communication breakdowns. AI offers practical solutions by analyzing and predicting risks. A typical example of the challenges of a dairy company and six-step approach for using AI to monitor and improve safety conditions as well as empower workers with smarter tools.

AI Professional – Fiona Passantino, early December, 2024

Slippery Business

How can we use AI to improve the working lives of people in a typical, large industrial company?

Imagine a typical, large dairy supplier with a combination of deskless and headquartered individuals. What are the three biggest safety risks for these types of workers out in the field, at the factory, or with the customer? They are far from headquarters, often on the road or working in shifts.

Consider the dairy industry. These factories are often wet, involve heavy, complex machinery and time pressure due to the short shelf life of the product. So, things have to follow a wide array of protocols and happen fast. According to the European Commission, after the construction industry, the tasks involved in transport and storage, distribution, manufacturing and agriculture make up nearly 30% of all workplace injuries and 43,9% of fatalities[i].

There are injuries related to slips, trips and falls. There are often wet and slippery floors due to spills and frequent cleaning. There might be uneven surfaces or obstacles in high-traffic areas. There might be icy or difficult conditions in cold storage areas or outdoor loading docks that people loading and unloading might need to be aware of.

There are physical ergonomic injuries. The work demands repetitive movement, getting through packaging and processing, lifting heavy crates, drums and maneuvering machinery. The machinery itself also presents risk; workers can become entangled, crushed, rammed with a forklift or burned during pasteurization. The chemicals used to clean, break down or sterilize can also cause burns or scalding.

What AI Does Well

AI does three things very well; analyze and predict, analyze and advise and analyze and create. These are the three basic pillars we use it for.

Algorithmic self-learning tools are great at crunching through vast amounts of data and offering advice, predictions or new output based on its library of information. After all, the main mandate described in all text generative AI is to “simply” predict the next word in a sentence. This extrapolates to the next pixel in an image, the next line of code in a program and the next number in a database.

How can AI help on the ground, in the factories? The first thing it needs is data. Who are the people in the factories, and what are their specific challenges? What are they doing, and how are they behaving? What types of accidents happen most frequently, and why?

What basic safety rules are being followed, and which are being ignored?

Easy steps to a safer workspace

Step one: Start a data library

Collect data you already have; gather and structure as much as possible. For data to have any value, it needs to be collected over a period of at least three years and come from a variety of sources that overlap. This includes a safety audit of current warehouse conditions, historical data on accidents, near-misses, and reported incidents, age of equipment and so on.

Factual data include the number of accidents, reasons, number of hours worked per shift, and number of shifts. Demographic data includes employee information, years of service, countries of origin. These can be augmented by anecdotal information which describe employee surveys, employee engagement and team sentiment.

Step two: Create new streams of data collection

Installing a few sensors and cameras to collect real-time workplace conditions gathers video data that can be analyzed by computer vision. What is actually going on in the factory, day and night? Sensors can detect changes in temperature or other hazardous conditions.

Your workers might be so well-trained that there are not enough images that show undesirable behavior. This means the staff can enjoy a fun day of creating fake video that show people “working” without protective gear, creating fake spills or other conditions that are plausible, just to teach the model what to look for. As long as it’s a fun day with doughnuts served afterwards.

Step three: Train an off-the-shelf model

Once there are profiles in place, new data streams and video, there’s enough material to begin to train AI vision. The data will have to be labeled (“desirable” and “undesirable”) and this takes patience and personnel. But it will allow you to see when and how often safety violations are occurring, categorize them, understand when equipment fails, or when spills and wet surfaces create situations of elevated risk.

Model training will require a small team of external AI specialists to set up and create the program, and to train your internal teams to maintain the system.

Step four: Define your training goals

imagine you are training the model for four main metrics:

• Safety violations (types and frequency)
• Equipment failures (types and frequency)
• Use of safety clothing and equipment (helmets, hairnets, earphones, gloves)
• Security failures (break-ins, lapses, types and frequency)

Your team will feed the (off-the shelf) base machine learning model specific streams of “good” and “dangerous” examples to build the training library.

Step five: link your model to the data collection

Now we’re ready to link the AI vision models straight to the incoming streams of visual data in real-time, and this is where the magic happens. The same high-resolution cameras and sensors that you installed for the baseline data is now the source of food for a fully trained model, which can signal up the chain when any of the “undesirable” conditions are seen.

Step six: onboard with AI-Driven Safety Training

Now that you have hours and hours of good and bad video, this is easily turned into an interactive, web-based training simulation that any new hire can follow before they ever reach the factory floor. Thanks to AI, this can be done in any language, with any local cultural setup. Essentially, you are training staff to “see something, say something”, and giving real-life images of what this looks like.

Need help with AI Integration?

Reach out to me for advice – I have a few nice tricks up my sleeve to help guide you on your way, as well as a few “insiders’ links” I can share to get you that free trial version you need to get started.

[i] Eurostat (2021) “Accidents at work statistics”. European Commission. Accessed August 19, 2024. https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Accidents_at_work_statistics#Analysis_by_activity