Around a year and a half ago, we visited a couple of manufacturing plants and what we saw completely CONFOUNDED us.
Being techies, we believed that with digitization revolution of the past 20 years along with recent adaptation of AI and IoT revolution – we would find shop floors either fully automated or on the verge of being so. At the very least, we believed that most of the automation problems were essentially solved already and it is just a matter of adoption.
To our surprise, 70% to 80% of the shop-floor processes are still manual. Most of these processes will continue to remain manual for the coming decades. This has nothing to do with the state of the technology – rather it is just the nature of the beast.
One of the confusion stems from the way the manual processes get defined.
The processes such as the one depicted on the left – where workers are either assembling or working closely with their hands are always considered as manual processes. But, the processes on the right where – workers are moving materials or equipment using a crane or forklift are less often described as manual processes. At the other end of the spectrum, the processes for which operators sit in a control room and make decisions to move stuff on the shop-floor are never really considered as manual processes.
But, as long as the decision to start a process is manual – it still is in a way "manually driven" process and suffers from many of the inefficiencies that come with limitations of human decision making.
The traditional way to monitor manual work is to add another layer of Supervisors on top of workers and it leaves serious gaps and adaptation of AI Supervision is yet to be seen.
The primary problem with the former is the fact that supervisors only have local context and are unable to make global optimizations across processes that are happening in different parts of the shop-floor. The other practical problem is that Supervisors are often overloaded with so much work that actual supervision becomes less than 10% of their job profile.
But, the un-intuitive fact is that a plethora of sensors is also not adequate to monitor manual processes.
By design, sensors are better suited in conjunction with machines but inappropriate for manual processes. Even the simplest manual process may have more than 100 steps and to cover them you may actually need 100 different sensors which are not feasible. As manual work keeps on evolving and there may be variation in actions of two workers – sensors just don’t fit the bill.
The practical problems make them untrustworthy – Workers often figure out a way to hack a sensor. E.g. If a sensor is just measuring weight, workers just press the sensor by hand and sensor being a dumb device captures that as well. Once, the sensor reading loses sanctity – it becomes useless. Another problem is that there is no proof to sensor reading – so even if the sensor gives out a number, those numbers often remain disputed.
Sensors are unviable if they require laying power and network cables right on expensive equipment. High temperatures and other hazardous conditions make them inappropriate at some factories. Many OEMs also do not provide sensors separately as add-ons – they, in fact, insist on buying new generations of equipment which come pre-fitted with sensors.
This is not to say that sensors and supervisors are useless – but that there remains a wide spectrum of manual processes which are NOT effectively monitored by current supervision mechanisms.
As manual processes remain poorly monitored it leads to four problems that directly and indirectly impact productivity
There are multiple processes going on in different parts of the shop-floor. As there is no real-time status available for any of these manual processes – there is no good way to do global optimizations.
E.g. Process Z may require outputs of both Process X and Process Y to start.
But, if Process X is running slow, it means that there will be a delay in starting process Z. Also, the team involved in process Y might have nothing to do but wait for process X to finish.
If real-time information, on the rates of work of manual processes, was available – resources of Y could have been diverted to process X to ensure smoother coordination across the floor.
There is always a huge gap between the planning and the actual sequence of events that unfold. The management always feels that supervisors and workers do not adhere to meticulous planning leading to delays. The truth is more complex.
In a dynamic environment of shop-floor, even small delay in any one small part of the process can cascade through the next sequence of steps making the entire days planning inapplicable. And as soon as that happens, the shop-floor starts with ad hoc decision making & firefighting which inevitably results in a poorer outcome.
As no real-time status of manual processes is available, there is no way to automatically shift the variables and change the day’s plan dynamically to ensure the optimal results.
So, what ends up happening is a series of 'Jugaads' on the shop-floor which is a valiant effort on the part of workers, but way inadequate and inefficient for driving productivity. This can be easily rectified by utilizing the new and advanced AI Technology that is available at our fingertips.
Most delays on the shop-floor are actually caused by a small list of easily solvable causes. All that is needed to avoid delays is that someone pays attention and resolves them at the right time.
But, as the status of these manual processes is not available in real-time - the delay doesn’t surface in time and it doesn’t get resolved until it becomes a crisis.
Also, in the absence of real-time monitoring often workers get a chance to relax. A simple data-point that we found across shop floors is that the productivity dips by as much as 20% when the Supervisors are absent from the floor.
All it needs is a mechanism to monitor the manual process in real-time to alert supervisors immediately to solve the delay as soon as it occurs.
If you don’t have real-time data it also means that you don’t have data for last week and last month. And in the absence of real data, the diagnoses of bottlenecks and causes of delays happen solely based on biases and conjectures of individuals.
For example, in one of the foundries, we found that management was convinced that the productivity was down because of poor planning while the workers strongly believed that lack of cranes and a smaller number of ladles and boxes led to delays.
And in the absence of data, there is no objective way to prove or disprove either. Likewise, it is impossible to scientifically diagnose and identify bottlenecks and solve them using the 'theory of constraints.
Time-motion studies done by third-party consultants, that happen once in a blue moon for the duration of one or two days may be sufficient to point towards a problem. But, in the absence of tools to monitor if the recommendations are being been adhered to, they fail to deliver expected productivity impact.
It was perplexing for us to observe that every shop-floor we visited – actual daily production was just 60-70% of the capacity leading to delays in client deliveries and piling up of undelivered orders.
And often the way management was trying to solve the issue, was to either create more capacity by purchasing more equipment or hiring more workers. Neither of these steps actually solves the underlying problem of a lack of a proper system to collect the real-time data to monitor manual processes.
One of the common ways to 'manage' in the absence of systems is to get time-sheets and log-books filled at workstations. These do record the data – but the data is not interpretable in real-time. Another common way is to use whiteboards to record data from various parts of the floor and use it to dynamically plan the day.
We came across two other hacks that got us thinking towards a solution:
We brainstormed and realized that perhaps recent advancements in AI Video Analytics can automate the job of watching CCTV feed and extracting time stamps of manual processes in real-time – just like that engineer we found at the steel plant.
The hypothesis is that – just as supervisors watch over the multiple areas of shop-floor, outputs of several CCTVs can be analyzed by AI algorithms and the data thus extracted can be leveraged for global optimizations and for driving productivity.
Another advantage with Video Analytics is that every alert comes with visual proof in the form of a video clip that can be tested. Also, as CCTV cameras can monitor from a distance – they are easy to install in hazardous environments. And regular existing CCTV camera infrastructure which is often already in place for security purposes can be re-purposed saving both cost and effort. One camera can also monitor multiple different processes simultaneously.
This information can then be made available across the shop-floor on dashboards right on their mobile phones of workers to enable real-time coordination.
Also, the solution can compare time-stamps against SOP times and daily planning - to identify deviations and help in dynamic planning.
Moreover, it can alert for any delay and send real-time alerts via WhatsApp – a communication tool which they are already comfortable with. An escalation system can also be built to automatically notify senior management if a delay persists beyond a threshold.
And, once this data is available for a month or so, it can always be used for systemic diagnoses of delays to identify bottlenecks.
This is exactly what - Sparrosense AI Supervisor offers and we are already seeing 10-20 % productivity impact with current deployments at steel plants, casting shops, auto-ancillaries, FMCG & textile plants - reaffirming our hypothesis and encouraging us to dig deeper.
Click here for a quick explainer video on how Sparrosense works.
Please do reach out to us at email@example.com if you wish to explore if Sparrosense can solve bottlenecks at your factory floor.