Mobile-Menu

Definition What is Machine Learning?

By konstruktionspraxis | Translated by AI 9 min Reading Time

Learning from data and always making the right decisions, and doing so much faster than humans? That's what machine learning can do. But what exactly is machine learning, how does it work, and what types of machine learning are there?

Machine learning is a subset of artificial intelligence that enables computer systems to learn from data and experiences and improve independently.(Image: © WrightStudio - stock.adobe.com)
Machine learning is a subset of artificial intelligence that enables computer systems to learn from data and experiences and improve independently.
(Image: © WrightStudio - stock.adobe.com)

The term "Machine Learning" has essentially existed for more than 70 years. Its origins lie in the 1950s and are closely tied to the beginnings of artificial intelligence (AI). Arthur Lee Samuel, an American electrical engineer and computer scientist, described "Machine Learning" in 1959 as the ability of computers to learn from experience without explicit programming. The evidence for this was a program developed by Samuel that could play the board game checkers at championship level by learning from its own gameplay experiences.

The actual roots go back even further, as early as 1943, when Warren Sturgis McCulloch, an American neurophysiologist and cybernetician, and Walter Pitts, an American logician working in cognitive psychology, laid a crucial foundation for later developments in machine learning with their model of artificial neurons.

The Art of Self-Optimization

Machine Learning is a subfield of artificial intelligence that enables computer systems to learn from both data and experience and to improve independently without being explicitly programmed for each task. The focus is on the use of algorithms that identify patterns and relationships in large datasets to derive predictions and decisions from them.

The central task of machine learning is to solve complex tasks in a data-driven and automated manner—usually much faster and more precisely than humans.

Not least due to the AI hype of recent years, the topic of machine learning has gained enormous relevance and has become an immensely broad field that cannot be comprehensively and exhaustively covered for every technological area in a single article. To somewhat narrow down the subject, the following explanations focus on the field of electronics.

How Does Machine Learning Work?

The typical functionality of machine learning (ML) in electronics takes place in four steps:

  • Data collection,

  • Data analysis,

  • Model training and

  • Real-time deployment.

During data collection, sensors and controllers in production and energy systems, for example, continuously gather a large amount of data such as temperatures, current flow, production parameters, etc.

Special ML algorithms analyze this data and identify patterns, correlations, as well as deviations to detect problems in production processes at an early stage. To make predictions or automate decisions, these algorithms are additionally trained with historical data.

After training, the ML models are capable of reacting to new data in real-time to detect errors before failures occur and to automatically adjust production systems to current conditions.

What Are the Specific Application Areas for ML?

Examples of the use of ML can be found in electronics: In quality assurance, manufacturing quality can be optimized, for instance, through sensor data and image analysis. In the development of complex electronic components, algorithms can also assist by learning from a multitude of design iterations and suggesting optimal solutions.

In energy systems, ML can be used to analyze vast amounts of data, ultimately leading to more efficient energy production and distribution.

In the field of maintenance, specifically predictive, condition-based maintenance, ML predicts the maintenance needs of machines and detects emerging issues early, thereby reducing downtime and costs. However, this is just a small glimpse of the multifaceted possibilities for the application of ML in electronics alone.

What Types of Machine Learning Are Distinguished?

There are various types of ML that differ in terms of their functionality and areas of application. The types of ML described below can be applied in many different technological fields. The fundamental distinctions are as follows:

  • Supervised learning

  • Unsupervised learning

  • Reinforcement learning

  • Deep learning

  • TinyML

—Supervised Learning

A system or algorithm is trained with predefined categories or target values, such as sensor data with known errors, to automate specific tasks like classification or prediction.

This form of learning is often used for solutions in quality control or to predict failures as well as necessary maintenance in predictive maintenance.

Supervised learning takes place in multiple steps:

  • First, electronic systems like sensors or cameras collect data such as images of components, measurements, or specific production parameters. Each data sample is tagged with a label describing the desired output, for example, "OK" (in order) or "NOK" (not in order) for components. This process is also referred to as data collection and labeling.

    Subscribe to the newsletter now

    Don't Miss out on Our Best Content

    By clicking on „Subscribe to Newsletter“ I agree to the processing and use of my data according to the consent form (please expand for details) and accept the Terms of Use. For more information, please see our Privacy Policy. The consent declaration relates, among other things, to the sending of editorial newsletters by email and to data matching for marketing purposes with selected advertising partners (e.g., LinkedIn, Google, Meta)

    Unfold for details of your consent

  • In the next step, data preparation takes place, where the collected data is prepared for training and sometimes supplemented with additional information to highlight relevant features for the algorithm, such as marking defect areas on images.

  • Subsequently, the algorithm can analyze the labeled data and learn to recognize patterns and relationships between the input data and the corresponding labels. The goal of the algorithm's training is to identify rules that enable the classification or prediction of new and still unknown data in the future.

  • After training the algorithm, the created model is tested with new, previously unknown data to evaluate how accurate and reliable the system is in its predictions (testing and validation).

  • The trained model is then integrated into ongoing operations, such as for automated quality control, using new data generated during daily production to further improve the model and adapt it to changing situations or conditions.

In summary, the individual steps of supervised learning:

  • Data collection and labeling

  • Data preparation

  • Training the algorithm

  • Testing and validation

  • Deployment in the specific process

  • Continuous optimization

—Unsupervised Learning

In unsupervised learning, algorithms analyze unlabeled data and independently detect patterns, structures, and/or relationships. There are no external specifications about which features or groups might be relevant. The algorithm is thus capable of independently identifying order within the data.

The typical process of unsupervised learning is similar to that of supervised learning:

  • First, data collection takes place again, but as previously mentioned, without labeling:

  • Special "unsupervised learning" algorithms then search the data for similarities, recurring patterns, and differences. Key terms in this context include clustering, dimensionality reduction, and association analysis.

  • Frequently, for example, data points are grouped (clustered) based on similarity, without any prior determination of how many clusters there should be or what they should specifically look like.

  • In structure and pattern recognition, the algorithms can eventually detect relationships that are not immediately apparent to humans, such as unusual patterns indicating defects or failures. An example would be a system grouping sensor data from a production line and automatically recognizing different operating states or anomalies.

Unsupervised learning is used, among other things, for the automated monitoring of machines, systems, and even power grids (condition monitoring). In electronics production, this form of ML can help segment large volumes of measurement data to identify new quality standards or suggest process optimizations.

In summary, the individual steps of unsupervised learning:

  • Data collection

  • Analysis by algorithms

  • Grouping/Segmentation/Clustering

  • Structure and pattern recognition

  • Deployment in the specific application

A particular advantage of this type of ML is that, for example, no predefined target values or categories are required for training, which saves time and costs. Additionally, unknown patterns and relationships can be identified that might have been difficult or even impossible to detect with traditional methods.

Unsupervised learning is particularly suitable when there is little prior knowledge about the data or when data structures frequently change.

—Reinforcement Learning

If something is done correctly, there's a reward. This is how reinforcement learning can be summarized in a nutshell. This form of ML works in electronics through an interactive learning process, where an algorithm, also called an agent, independently makes the best decisions in a dynamic environment to maximize rewards in the long term.

In reinforcement learning, an agent, such as the control system of a production plant, interacts with its environment, including actuators or sensors. The agent continuously observes the current state of the plant, including parameters such as power consumption or temperature, and selects an action accordingly. Such an action could involve, for example, adjusting the production speed.

Each action taken by the agent leads to a change in state and, as a result, a reward value. The reward value evaluates the action performed: for example, if the action saves energy through optimized cycle time, it results in a positive reward; however, if it leads to overheating of a component, it is met with a negative reward. Based on this trial-and-error process, the agent develops a strategy that maximizes the sum of positive rewards over time.

The key components for reinforcement learning in electronics are:

  • the state space—it describes the possible environmental conditions, such as sensor data or operating states

  • the action space—it defines the available control options, such as adjusting cycle time or voltage or activating safety mechanisms

  • the reward function—it determines which actions of the agent are advantageous in which states

For reinforcement learning in electronics, various algorithms and methods are available:

  • Q-Learning—An agent learns a table (Q-table) that stores the value of each action in each state. Over time, it adjusts these values to make optimal decisions.

  • Deep Q-Networks—Combines reinforcement learning with neural networks to handle complex state spaces.

  • Policy Gradient—Directly optimizes the strategy by adjusting probabilities for actions.

Machine learning based on the reinforcement learning approach enables systems to learn autonomous and adaptive behavior, ranging from optimizing energy-efficient circuits to precise control of industrial robots. However, incorrect decisions by the agent can always lead to damage in critical systems.

What is Deep Learning?

Deep learning is, in a sense, the "royal class" of ML, as this method uses artificial neural networks with multiple layers, known as layers, to recognize complex patterns in data and make or automate decisions.

These layers consist of:

  • the input layer—The input layer receives raw data, such as sensor data, measurements, or images of components

  • hidden layers—In the hidden layers, this data is analyzed through weighting and transformation to extract abstract features

  • the output layer—The output layer ultimately delivers the result, such as a classification like "defective" or "intact"

The neural network is trained with examples where the input data and the desired output are known (labeled data). Through a process called backpropagation, also known as error backpropagation or backward propagation, the weights of the neurons are optimized to minimize errors. This method is used to adjust the weights of a network so that the model's predictions align as closely as possible with the actual target values.

After training, the model processes new, previously unknown data in real-time, such as defect detection on a conveyor belt (inference).

The basic principle of deep learning thus consists of the following steps:

  • Layer construction of neural networks

  • Training with labeled data

  • Inference

The key technologies for deep learning in electronics are:

  • Convolutional Neural Networks (CNNs)—They are ideal for image processing, such as defect detection in electronics manufacturing.

  • Recurrent Neural Networks (RNNs)—They are used for processing temporal sequences, such as analyzing data streams in production.

  • Frameworks—TensorFlow, PyTorch, or Keras enable implementation on embedded systems

Deep learning has virtually revolutionized electronics through automated pattern recognition, precise defect detection, and adaptive control systems. However, deep learning requires, among other things, powerful hardware and a large amount of labeled data for training.

What is Tiny Machine Learning (TinyML)?

Resource-constrained electronic devices such as microcontrollers, sensors, or IoT edge devices can also benefit from ML through TinyML, which combines optimized algorithms, energy-efficient hardware, and lightweight software to analyze data directly on the device without cloud connectivity.

To do this, sensors first capture raw data within the milliwatt power range, such as vibration, structure-borne noise, or temperature data for condition monitoring via a microcontroller. The models are then initially trained on powerful systems to recognize specific patterns, such as anomalies in sensor data.

Subsequently, the model is optimized and compressed, for example, through quantization, which reduces numerical precision, and pruning, which removes irrelevant neuron connections. This creates models just a few kilobytes in size that can run on microcontrollers.

The optimized model is then deployed to the microcontroller, where inference occurs directly on the device. The model continuously analyzes local sensor data and makes decisions, such as triggering an alarm or shutting down a device when a fault is detected.

Summarized, the individual steps of TinyML:

  • Data collection

  • Model development and training

  • Model optimization

  • Deployment to edge devices

  • Real-time inference

The key advantage of TinyML is that this approach enables real-time decisions through local data processing, with systems functioning reliably on battery power alone for years. Moreover, no sensitive information is transmitted over the network due to local data processing. However, TinyML has only limited memory in the KB to MB range and low computational power (MHz clock frequency).

Related content