Get warmed up for the Cloudflight AI Coding Contest
Like in many other industries, AI is turning several fields of manufacturing upside down. While (mechanic) automation has been a topic of innovation for many years now, the focus is moving towards (intelligent) automation. This means that, in addition to following a limited set of deterministic rules, complex systems and processes must learn to adapt to various situations.
Let’s first look at computer vision. This is a common set of technologies that allows systems to recognize their environment in a human-like manner. It is therefore well suited for retrofitting into processes without having to completely redesign them. Computer vision is a descriptive method that acquires structured information that would otherwise be input manually by shop floor workers. On one hand, existing decision rules such as quality gates or the correctness of an assembly can be checked automatically. On the other hand, additional data can be made available to decision-makers in order to enhance quality and performance.
An important application of computer vision in manufacturing is quality assurance, where visual inspection is still one of the most important methods. Use cases include the detection of surface imperfections (e.g., scratches or paint chippings), assembly errors (e.g., missing parts), and foreign objects that have entered the process by mistake. In addition to mere detection, value is also often added by classifying different sub-types of errors. Another innovation is to rely on not only the visible color spectrum but also other image data such as heat maps (infrared spectrum) or advanced non-destructive testing instruments.
In addition to quality assurance, process monitoring can be supported by computer vision. This ranges from simply counting or measuring objects along the production process to generating extensive documentation.
Computer vision can also act as an enabling technology, especially in the field of robotics. As machines become increasingly more autonomous, they need to be aware of their environment and context. Important tasks here are the identification of objects (e.g., a pallet to be lifted) and obstacles (e.g., differentiating between areas that can be entered or driven on) and the recognition of potential dangers. Certain machines are safer if they are operated at a lower speed when humans are nearby but more efficient if they are operated a higher speed when there are no humans in the danger area.
From this list of applications, it becomes apparent that computer vision is a powerful technique. A major strength is that it helps replace various other sensors by a single one – the camera. Thus, with only a software update, additional features can be added without having to invest in procuring, installing, and maintaining new hardware.
Times Series Analytics
A second important field of data science in manufacturing is time series analytics. Many important signals such as temperature, speed, and pressure are vital to production. Some are exogenous (i.e., determined by an external force) yet still influence the target system. An obvious example is the outside temperature or weather influences in general. Although these cannot be influenced, some measures might be required to equalize their impact. In contrast, other signals are system-internal ones that can easily be influenced by control loops.
A typical machine learning task on such data is pattern recognition. Based upon the evolution and correlation of different signals, various states of a machine or process can be distinguished. Time series forecasting is another task that focuses even more on the process background of such data.
Anomaly detection is a technique that works for both the computer vision and time series domain. The goal is to automatically detect events in which a process deviates from the norm without specifying possible anomalies in advance. This latter aspect is what makes the technique so powerful compared with rule-based approaches. In the field, there are numerous corner cases or unforeseen conditions. When developing a control system, it is likely that not all of them will be considered.
The motivation behind predictive maintenance is to find the sweet spot between avoiding wear-related damages or risks and avoiding the costs of overly frequent maintenance. It is technically built upon time series analytics and forecasting. In a shallow implementation, the actual load on certain wear parts is inferred. A manually engineered model is then applied in order to infer the remaining time of life based on the load statistics.
In a deeper approach, the current wear of some parts is directly observed from near real-time data. Examples are the relation between certain pressures and speeds within a machine. If, at the same applied pressure (and a temperature), the speed of a compactor machine increases or decreases over time, this is a good indicator.
Shorter term indications that maintenance is due can also be extracted from structure-borne noise. Like human experts that can distinguish between states of a machine from only its sound, those noise patterns can be classified or used in anomaly detection.
Computer vision, time series analytics, and anomaly detection are descriptive approaches. They look at unstructured data and extract structured information from it. In contrast, time series forecasting and predictive maintenance are predictive approaches. With reinforcement learning, there is another type of approach – the prescriptive one. This means that it does not generate information but rather learns the actions necessary to achieve a certain state.
While this field of AI was previously known only from a game context, there are now many industrial applications. A prominent example is intelligent heating and cooling, where machine learning based control mechanisms have significantly outperformed traditional ones. This is just one example that can be generalized to autonomous machines designed to find optimal strategies to various tasks in the context of a changing environment and external influences.
The AI Coding Contest
The next Cloudflight Coding Context will take place on November 5, 2021. In the AI part of the contest, we will ask you to solve a machine learning challenge on several levels. Our goal when designing the challenge was to allow the best participants to finish it within four hours. We therefore considered approaches with training times that are feasible within the limited amount of time. We also want the contest to be as inclusive as possible and do not want to limit participation to experts in a particular machine learning specialty. Are you up to the challenge?
We hope to see you soon at the Cloudflight Coding Contest.