To achieve artificial intelligence, two components are essential: firstly, the ability to measure a parameter and understand what the measurement means, and secondly, the ability to learn. The first part is about metrology, the scientific study of measurements. The second part is called machine learning (ML), which gives systems the ability to recognize when a measurement deviates from expectations and to change a process without being explicitly programmed to do so.

The ability to collect data

Metrology is about the in-depth understanding of a particular measurement. This measurement can be as simple and unambiguous as voltage, mass or temperature, or as multimodal as the function of aircraft control surfaces or complex production lines.

Depth of measurement: Regardless of whether a single parameter or multiple parameters are measured, the depth of measurement accuracy determines the level of programmability that can be achieved. For example, measuring a 3 V system to 1/10 V is not as informative as measuring to 1/1000 V.

Data feed: Measurement data is only useful to an algorithm if it is provided in a data feed. If, as in the example above, a sensor is able to measure to 1/1000 but its data feed output is limited to one decimal place due to data bus limitations, the additional precision is not available to the algorithm.

Multiple data feeds: Whenever possible, measuring multiple parameters leads to better decision making. For example, if the voltage can be measured with an accuracy of 1/1000 V and the temperature at the same time, it is now possible to detect voltage changes due to temperature fluctuations.

Machine learning

The ultimate machine learning feeds data from multiple sources into algorithms that mimic the way humans learn and gradually improve their accuracy. Once the data feeds are in place, there are three essential building blocks for ML: an algorithm to interpret the data, a table of expected results as well as reactive results, and a feedback loop.

The algorithm: The true 'intelligence' of any machine learning system is its ability to process data inputs, perform a series of calculations/instructions and interpret the results. Interpreting means the ability to recognize whether a result is within or outside the expected range and issue new instructions according to this result. In the previous example, the algorithm could activate an internal fan if a voltage measurement is far outside the expected range and the temperature is above the nominal value.

Expected results and reactive consequences: In their simplest form, expected results can be a 'lookup table' with combinations of data inputs and a series of reactive command statements. The more comprehensive the table, the more sophisticated and valuable the ML becomes. More interactive MLs can make incremental changes, such as changing the course of a drone based on real-time sensing, which requires both continuous sensing and constant adaptation.

Feedback loop: The final element is the feedback loop, which allows the system to check whether what it has done was sufficient or whether it needs further refinement. It also allows its parameters to be adjusted.

Adding multiple ML functions that focus on different aspects of larger systems, as well as adding more sensor data, enables ML at a more complex system level. Very advanced ML can add to its 'lookup tables' as it encounters new combinations of sensor inputs, execute variants of its reactive instructions and measure feedback on how sufficient the response was. This creates self-regulating algorithms that derive knowledge from data to predict outcomes. The more the algorithms are trained, the more accurate the results become.

Artificial intelligence

With trainable algorithms, we are now most of the way to deploying AI. This requires combining the results from the collection of ML engines with sufficient policies and iterations to allow the algorithm to make decisions in real time. Each time an AI algorithm processes data, iterates, weighs the iterative response with new data, and uses the combination to make its output decisions, it has reached the state of decision making. Through this continuous cycle, the AI is constantly learning and improving the quality of its decisions. This entire process can be very simple, like the example of the voltage and temperature sensor loop; or it can be as complex as the flight control system of a combat drone.

The DNA markers of AI

So how can we predict how well an AI algorithm will work?
Looking at the DNA markers helps. In its most basic form, implementing AI allows a machine to replace a human in the decision loop by simulating how we humans would perceive and process information and change a workflow for a given set of conditions. In essence, three common DNA characteristics should be considered:

Measurement and simulation: how well is the manufacturer able to measure, do they have sufficient knowledge and experience to create a digital twin of the environment?

Algorithms, analytics and insights: The depth of the developer's knowledge of the core properties of the signal and the relationship to the expected responses will determine the depth of the 'lookup table' of expected results.

Knowledge of workflow automation: the system-level understanding of how multiple iterative ML outputs can work together to optimize a desired outcome.

Jeff Harris VP, Portfolio and Global Corporate Marketing, Keysight Technologies, San Diego, USA.

© Keysight Technologies

The quality of an AI algorithm therefore depends on these characteristics: the depth of understanding of metrology in a particular area, the measurement technology and the number of technologies and standards for which it has this comprehensive knowledge.

So if AI is done well, it is not an overrated new technology. Rather, it is the only way for developers to cope with the exponential complexity of new designs. Or, in the words of futurologist Gray Scott: "There is no reason and no way that a human mind will be able to keep up with an AI-powered machine by 2035." Engineers have recognized this and have begun to incorporate ML and AI into their systems. So AI starts with smart, motivated engineers who understand measurement science and who know the expectations of system behavior enough to create digital twins for developers.