During a strategic inflection, there comes a point when you must decide whether to embrace or reject new technology. This decision may seem easy at first but can be much more complex in reality. Revolutions don't always happen overnight; they can unfold over years, with most of the progress being hidden until a breakthrough brings it into the spotlight for everyone to see.
In 1986, the School of Computer Science at Carnegie Mellon University initiated the Navlab project, which demonstrated the potential of artificial intelligence. They developed a system that enabled a Chevrolet panel van to drive autonomously. This occurred during the early stages of development of computer vision and machine learning algorithms, which are now widely used across many applications.
During its development, it was considered groundbreaking but not yet ready to be a part of everyday life. However, in today's world, we are accustomed to hearing about, and on occasion even seeing, self-driving cars, which various companies are developing and selling to the corporate sector, with goals of marketing this technology to the general public in the not-too-distant future.
Autonomous driving, like most AI applications, requires algorithms and processors that can operate on very large data sets in real time. For example, large language models (LLMs) like the ones used by OpenAI in ChatGPT, Microsoft in Copilot, or other applications using Meta LLM Llama 3, need large data sets to recognize the patterns that create human-written sentences. Thus, at the core of machine learning, you will find that data, and the ability to rapidly process data are both necessary to work with these systems.
Let's consider a scenario where your factory needs to produce 1000 units per hour. Running at full capacity, if one of your machines goes down, you will not meet the target of 1000 units per hour. In such a situation, one of the key priorities is to expedite the maintenance cycle of the impacted machine. To prevent this line-down scenario, you may have a maintenance plan in place, and if following best practices, would be deploying a condition-based maintenance methodology. This involves measuring different variables on a machine, such as temperature and vibration, using sensors. When a specific limit is exceeded, alerts are sent out and you know that a particular part needs to be replaced prior to equipment failure.
For preventive maintenance, you will need a data acquisition system to collect mission critical variables and store them on a computer. Once collected, you can compare these data points against designated thresholds to determine when maintenance is required. Up to this point, it seems relatively straightforward. However, as replacement parts are integrated, the machine may exhibit different behaviors as it ages, potentially leading to false positives and unpredicted operational issues. This is when you realize that you have a substantial amount of data, including quality control data for the final product. It is at this juncture that you can implement a machine learning algorithm to effectively analyze and cross-reference all this data, enabling your system to autonomously make better maintenance decisions.
The best part is that learning models are scalable because they will acquire more data with each iteration and over time the overall machine learning algorithms improve. But that is not the whole story. Now that you have an evolving algorithm, you have the ability to implement it in systems that autonomously make decisions and improve over time.
In considering different architecture options, one approach involves using dataloggers like the CONPROSYS family from Contec that collect data and then send that data to a server for processing. The advantage of this approach is that the heavy data processing is offloaded to a centralized server. However, a significant drawback is the potential for generating a large amount of network traffic when sending the data. As the scale of operations increases, this could lead to issues and higher operational costs due to bandwidth limitations. In addition, the ability to react in real time to critical tasks is hampered by network delays to and from the server.
On the other hand, we can also deploy a PC-based DAQ system from Contec. This system allows you to locally run heavy algorithms, such as inferencing based on machine learning, while simultaneously obtaining the data directly on the same system. This capability brings processing to the edge and allows you to scale as needed without having to build infrastructure around servers. An interesting aspect of this PC-based DAQ is that you can incorporate remote data loggers, such as Conprosys products, to create a hybrid platform architecture.
Any artificial intelligence or machine learning algorithm will always need data to be useful. This is where the motivation to understand data acquisition is important, and finding the right partner like Contec to thrive in this revolution is crucial. Knowing your options and applying the best practices of data acquisition sets the cornerstone for getting the most out of artificial intelligence.