Connected-Things Technology Chain
Sensors, big data, machine learning, artificial intelligence and robots are tightly interwoven. In the AI era, hardware and software coevolve and influence each other.
The trends of the Internet of Things, big data and robotics form a continuous loop. Each link in this chain affects the next, creating a reinforcing cycle. Sensors embedded in connected devices generate large volumes of data. Massive data enables machine learning, whose outputs lead to AI. AI then guides robots to perform tasks more precisely, and the robots' actions trigger sensors again. The result is a closed technical loop.
1. Sensors Generate Data
By 2014, the number of Internet-connected devices exceeded the world population. Estimates predict tens of billions of interconnected devices, most of which incorporate sensors, whether embedded or connected externally. These sensors produce unprecedented volumes of data.
2. Data Enables Machine Learning
As data volumes grow by orders of magnitude, both structured and, more commonly, unstructured data can be processed by machines to produce insights. Machine learning relies on that data to identify patterns and extract value.
3. Machine Learning Improves AI
Machine learning uses data processing and pattern recognition so computers can learn without explicit programming. Current large-scale data and compute capabilities are driving breakthroughs in machine learning, which in turn improves AI systems. For example, machine learning has been used to map detailed geographic information quickly at scale.
4. AI Directs Robot Actions
Computers now outperform humans in many structured domains, like board games and certain perception tasks. As sensor data increases, it improves machine learning algorithms, and logically, the integration of AI with robots will lead to exponential growth in robotic task performance.
5. Robots Take Action
Numerous companies develop robots for a wide range of tasks. Robots are becoming increasingly capable, and advances in AI enable robots to perform more complex tasks than before.
6. Actions Trigger Sensors
Robotic actions generate events that sensors capture, closing the loop and providing new data for learning and optimization. This continuous cycle forms the technical ecosystem of AI.
How AI Technologies Optimize Sensor Systems
AI technologies that can assist sensor systems include knowledge-based systems, fuzzy logic, automated knowledge acquisition, neural networks, genetic algorithms, case-based reasoning and ambient intelligence. These techniques are increasingly applied in sensor systems because of their effectiveness and the broader availability of computing resources.
Many AI techniques have low computational complexity and can be applied to small sensor nodes, single sensors or low-capacity microcontroller arrays. Proper application of AI methods can create more competitive sensor systems and applications. Advances in related areas such as data mining, multi-agent systems and distributed self-organizing systems also affect sensor systems. Ambient sensing technologies can integrate many microprocessors and sensors into everyday objects, enabling intelligent environments that communicate with other devices and interact with humans. Combining ambient intelligence with other AI technologies can enable powerful new capabilities, though the consequences of wide integration may be difficult to predict.
Designing Smarter Sensor Systems
AI can optimize sensor systems. As a branch of computer science, AI produced many powerful tools since the 1950s that help solve problems originally requiring human expertise. While industrial adoption has been gradual, AI promises improvements in flexibility, reconfigurability and reliability. New systems already exceed human performance in an increasing number of tasks. As humans and machines work closer together, combining human insight with computational capabilities supports analysis, inference, communication and problem solving.
AI combines technologies that enable learning, adaptation and decision making. These advances depend on neural networks, expert systems, self-organizing systems, fuzzy logic and genetic algorithms. AI techniques extend to many domains that require interpretation and processing of sensor information, including assembly, biosensing, building modeling, computer vision, tool diagnostics, environmental engineering, force sensing, health monitoring, human-machine interaction, networked applications, laser machining, maintenance and inspection, power-assisted systems, robotics, sensor networks and remote operations.
Recent research shows AI can support sensor systems in the following areas.
1. Knowledge-Based Systems
Knowledge-based systems, also called expert systems, integrate a large collection of problem-solving knowledge for a particular domain. They typically consist of a knowledge base and an inference mechanism. The knowledge base is often represented in "if-then" form, together with facts, frames, objects and cases. The inference mechanism operates on the stored knowledge to produce solutions, using methods such as inheritance, constraint handling, case retrieval, and rule application depending on the control strategy and search method.
Rule-based systems represent knowledge as "if-then-else" rules and are good at presenting knowledge and decisions in forms familiar to humans. Their strict rule structure makes them less suitable for highly uncertain or imprecise situations. A typical rule system includes a rule list or database, an inference engine or parser that derives information or actions from input and rules, a working memory, and user interfaces or other I/O mechanisms.
Case-based reasoning applies past solutions to current problems. Solutions are stored as historical cases representing expert experience. When a novel problem occurs, the system finds the most similar past case, adapts its solution, and updates the database based on success or failure. Case-based systems are often viewed as extensions of rule systems and can present knowledge in human-friendly formats while learning from past cases.
2. Case-Based Reasoning
Case-based reasoning typically follows four steps:
- Retrieve: retrieve relevant past cases from memory for the target problem. Cases include the problem, the solution and annotations explaining how the solution was obtained.
- Reuse: map the retrieved solution to the target problem, including adaptations for the new scenario.
- Revise: test the adapted solution in the real world or simulation and modify it if necessary.
- Retain: if successful, store the new solution as a new case in memory.
This approach relies on empirical evidence from past cases. Without statistical support, conclusions based on limited cases may be uncertain. Nevertheless, case-based systems are valued for applying past experience to current problems and for their ability to generate new cases from historical data.
Expert systems are among the more mature technologies, with many commercial shells and development tools available. Once domain knowledge is encoded, system construction becomes straightforward. Expert systems have been applied to sensor tasks such as input selection, signal interpretation, condition monitoring, fault diagnosis, machine and process control, machine design, process planning, production planning and system configuration. Additional applications include automated programming, intelligent vehicle control, inspection planning, hazard prediction, tooling selection, process planning and factory expansion control.
3. Fuzzy Logic
Traditional rule-based expert systems struggle with situations outside their knowledge base and may fail rather than degrade gracefully. Fuzzy logic introduces the imprecise, linguistic characteristics of human judgment, improving system adaptability. Fuzzy logic represents variables with linguistic values mapped to fuzzy sets, and decisions derive from these fuzzy descriptions.
Fuzzy expert systems handle incomplete or partially corrupted data by modeling human-like reasoning using the mathematics of fuzzy sets. Humans manage ambiguous semantic situations easily, but machines do not. Fuzzy controllers are widely used where knowledge is imprecise, structural or object matches are inexact, resolution is limited, or in image processing and reconstruction. Fuzzy systems are suitable for handling uncertainty and imprecision but typically lack learning ability because key parameters are preset.
Fuzzy logic has been successful in collaborative robots, automotive robotics, perceptual prediction, supply chain management and welding control.
4. Automated Knowledge Acquisition
Collecting domain knowledge to build a knowledge base is complex and time consuming, often becoming a bottleneck for expert system development. Automated knowledge acquisition techniques address this by using multiple training cases as input. Each case has attributes and class labels. A divide-and-conquer strategy can split case collections into subsets and induce decision trees that classify cases, yielding knowledge about specific case categories.
Another approach is the covering method, where induction seeks attribute sets that characterize a class of cases and formulates rules that cover matching cases until none remain. Inductive logic programming (ILP) uses predictive logic to describe training cases and background knowledge, enabling the induction process to produce first-order clauses with variables rather than propositional attribute-value pairs. ILP systems use top-down induction or inverse resolution.
Examples of learning programs include ID3 (divide-and-conquer), AQ (covering method), FOIL (ILP induction), and GOLEM (inverse resolution). Many algorithms produce explicit decision rules; some can produce fuzzy rules. Sensor systems and networks often provide training sets in the required formats, since attributes and classes are clearly defined, making automated learning technologies widely applicable. These methods suit cases with discrete or symbolic attributes more than those with continuous attribute values. Applications include laser cutting, ore detection and robotics.
5. Neural Networks
Neural networks can extract domain knowledge from cases, but the knowledge is implicit rather than represented as rules or trees. They handle continuous and discrete data and provide strong inductive capability similar to fuzzy systems. Neural networks model computation using interconnected units called neurons that operate in parallel.
The most common networks are multilayer perceptrons, which are feedforward networks mapping inputs to outputs at a single time. Recurrent networks include feedback loops that give dynamic memory, making outputs depend on current and prior inputs and outputs.
Implicit knowledge is embedded in the network through training. Supervised learning adjusts connection strengths by comparing actual outputs to target outputs. Backpropagation is commonly used in multilayer perceptrons to propagate errors back and update hidden-layer weights.
Artificial neural networks typically have input, hidden and output layers. They are flexible mathematical functions with configurable internal parameters trained to represent complex relationships. Once trained, a network can generalize to new inputs. Neural networks are suitable for function approximation, pattern classification and pattern completion.
Recent applications include feature recognition, heat exchanger inspection, weld spot inspection, spot-weld parameter optimization, power systems, haptic displays and vehicle sensor systems.
6. Genetic Algorithms
Genetic algorithms are stochastic optimization methods inspired by natural evolution. They can find global optima in complex multi-dimensional searches without domain-specific knowledge. Genetic algorithms have been applied in sensor systems for combinatorial or multi-parameter optimization, assembly line balancing, fault diagnosis, health monitoring and power steering optimization.
7. Ambient Intelligence
Ambient intelligence has advanced significantly over recent decades, enabling electronic systems that anticipate and respond to user behavior. The concept aims for seamless integration between humans and sensor systems to meet practical needs. Industrial adoption is still limited, but more interactive, intelligent systems are under research.
Extending Systems
AI can improve communication efficiency, reduce failures, minimize errors and extend sensor lifespan. Over the past 40 years, AI provided a set of powerful tools that are increasingly applied in sensor systems. Adopting appropriate AI methods will help build more competitive sensor systems. Due to existing technical barriers and limited familiarity among engineers, wider adoption may take more years, but research continues and many new sensor applications are emerging. Combined use of these techniques will enhance system capabilities.
Sensor applications continue expanding, from smart factory deployments to monitoring networks for power grids, air quality and transportation infrastructure. As AI-enabled devices enter homes and consumer products become smarter, sensors will become even more pervasive. For the sensor industry, this trend creates substantial opportunities but also challenges: developing sensors that meet market requirements, differentiating products and maintaining technological leadership. Success will depend on following technological trends, securing core capabilities and aligning development with market needs.
