AI for Industry 4.0 – Cloud vs Edge

By Philippe Lambinet, Co-founder and CEO, Cogito Instruments

Philippe Lambinet, Co-founder and CEO, Cogito Instruments
Philippe Lambinet, Co-founder and CEO, Cogito Instruments

Industry 4.0 is about data driven manufacturing automation and is usually associated with the Internet of things and cloud computing in order to create the smart factory.

AI technologies, in particular Deep Learning, are used to build machine models or for image recognition, enabling applications such as predictive maintenance, visual inspection and trainable robots.

There are many documented use-cases of Deep Learning in the industry, but most have been published at proof-of-concept level and very few have actually been deployed in real life. In this article, we will look at the reasons why the use of Deep-Learning AI has not been as successful as advertised.

Limitations of Deep Learning in Manufacturing
Let’s see the main limitations of the mainstream Cloud-based AI technology and how these limitations impact some real life applications:

Availability of data
Deep Learning is a fantastic technology to analyze Big Data. The issue, when applied to Industry 4.0 applications is that in a lot of cases, there is not enough data available to train models properly.

Cloud-based learning
Because a lot of training data is needed and because the back-propagation process involves a lot of iterations, the creation of the deep neural network models requires tremendous computing power. It therefore happens in the Cloud, takes time and consumes a lot of energy. It is estimated that one training of a typical deep neural network has the same CO2 footprint as 5 cars during their entire lifetime. In addition to the cost and energy problem, there is often the issue of data privacy.

Cloud-based inference
Once the model has been created, in the Cloud, it can be used to perform prediction/recognition. Real-time data, captured in manufacturing machines by multiple sensors, is sent to wherever the execution of the model happens. Very often, it happens in the Cloud, so data needs to be sent to servers which takes more energy and bandwidth and therefore costs money and introduces latency. Sometimes, this data is analyzed locally, at the Edge of the network. Latency often becomes an issue because networks have limited performance and availability. When an edge device is used, the model, created in the Cloud, must be mapped onto this edge hardware execution unit which introduces performance limitations.

Explainability
Because no training is perfect, false positives or false negatives happen. Deep Learning models are so complex, with so many layers, that it is impossible to understand why a particular response was obtained. This may not be an issue when Facebook performs face recognition on billions of pictures uploaded by its billions of users. But, in manufacturing, it has serious consequences. First, if you don’t know why mistakes happened, it is impossible to fix them. This breaks the continuous improvement cycle which is so important in manufacturing. Second, there can be major safety issues if robots start to behave unpredictably and you cannot analyze why.

Skills
Ai technologies are more mainstream than ever. There are many platforms, such as PyTorch, Keras, TensorFlow, to “play” with these technologies. However, people should not underestimate the difficulty to properly use these tools. The world does not have enough data scientists and computer science specialists and the top talents are more likely to be recruited by the GAFAs and BATXs of this world rather than by more traditional industrial companies. This shortage of skills is a major limitation to the use of AI in Industry 4.0 applications.

Some real-life examples

Machine monitoring and predictive maintenance.

Company A has several manufacturing sites all around the world. Company A has engaged two years ago in a predictive maintenance project in order to improve the up time of their machines and reduce maintenance costs. Massive data sets have been gathered for the purpose of building predictive models able to identify risks of future failures in different sub-systems inside these machines. The pilot project was successful, and, in a lab environment, the models have been proven to provide very good accuracy in their prediction. However, when the solution was deployed to the multiple manufacturing locations, the number of false predictions was such that the downtime actually increased together with the cost of maintenance.

What happened?

This is what people call the “all machines are different” syndrome. In fact, the machines are not really different, but their working environments are different and therefore their behaviors are different and the “one-size-fits-all” predictive model approach does not work.

The solution is obviously to have an individual model for each machine, taking into account its individual environment, but then there is not enough data to feed the Deep Learning model generator so this AI methodology cannot be used.

Visual Quality Inspection

Company B is performing final visual inspection by human operators of their products at the end of the line. This inspection is performed by sampling because 100% inspection would be too expensive and slow. Company B would like to automate this process in order to move to 100% visual inspection without increasing costs and reducing throughput.

Company B trains its operators with a reference set of good and bad products and has seen that a dozen products for each defect category was sufficient to properly train humans to recognize defects.

Company B has tried a Deep Learning based inspection solution and has concluded that it was not usable due to the high number of false recognitions.

What happened?

A dozen pictures for each defect category is not enough to train a Deep Learning network. Deep Learning specialists estimate that they need tens of thousands or more pictures. There are techniques to reduce the number of pictures needed from millions to thousands, but nobody has ever demonstrated that a dozen pictures are enough to train a deep neural network.

Conclusion
With Industry 4.0, machines became capable of capturing data and transmitting data. This does not make them smart. Truly smart machines will exist when they will be able to “understand” i.e. analyze the data they produce and make decisions. As we have seen, outsourcing the decision to the Cloud, using Deep Learning, is not always the right approach. There is a need for machines that are able to:
– learn locally, from limited training data
– learn incrementally and explain their mistakes to enable continuous improvement
– keep their CO2 footprint and cost of ownership minimal.

This is why Cogito Instruments was created. We welcome you on this journey to make machines truly intelligent.