Today, artificial intelligence is being actively introduced into the management of critical infrastructure: from overpasses and power grids to trunk pipelines and aviation. But is AI really so flawless when the safety of millions of people is at stake? In his report on infrastructure monitoring, Professor Pasquale Daponte argues that the main threat lies not in AI itself, but in what we “feed” it. 

The illusion of the digital twin

In modern engineering, it is customary to rely on the concept of “digital twins” — virtual 3D models of real objects that reflect their physical condition in real time. A classic example, according to the professor, is the monitoring system of the 25 de Abril Bridge in Lisbon, where hundreds of sensors continuously record load, vibrations, and tension in structural elements. The data flows into a single center, and artificial intelligence algorithms search for deviations that make it possible to detect metal fatigue long before visible damage appears.

However, Pasquale Daponte draws attention to the key link in this chain, where failures most often occur. The scientist reminds us of the basic process of error propagation: before data reaches a neural network, it passes through a physical transducer, primary signal processing, digitization, synchronization, and calibration. If distortion occurs at any of these initial stages — for example, if a sensor shifts, loses calibration, or its readings become desynchronized in time with neighboring instruments — all subsequent analytics become meaningless.

As the professor emphasizes, artificial intelligence is incapable of restoring information that was never correctly measured, synchronized, or calibrated in the first place. The result is a situation in which a mathematically flawless AI model processes unreliable input data and produces a seemingly convincing but fatally erroneous conclusion.

A return to the laws of physics

To overcome this vulnerability, Daponte proposes abandoning the use of AI as a “black box” that merely searches for statistical correlations in massive arrays of numbers. In critical systems, this approach is too dangerous, so the scientist insists on a transition to hybrid models — Physics-informed AI, so-called physics-based AI.

The logic is to integrate the fundamental laws of nature into the mathematical model of the neural network: equations of dynamic equilibrium of structures, laws of energy conservation, and formulas from the strength of materials. When an algorithm is trained with the physical model of a specific bridge or pipeline in mind, it gains the ability to verify the incoming data stream. If a sensor begins transmitting values that are physically impossible for a given structure under current conditions, the hybrid system instantly recognizes it as an instrument error rather than interpreting it as an emergency at the facility.

Why the “red button” remains in human hands

The logical continuation of this chain is the question of decision-making autonomy. Daponte divides the task of analytics into two fundamentally different stages: diagnostics and prognostics. The diagnostic AI module is capable of localizing an anomaly that has already arisen and determining the location of the fault. The prognostic module goes a step further and calculates the system’s remaining useful life, answering questions such as: “How many more hours can an airplane keep flying with this microcrack?” or “How many more days will the pipeline operate at the current level of corrosion?”

At the same time, Pasquale Daponte categorically objects to granting algorithms the authority to carry out automatic executive actions. Monitoring systems may automatically change the safety status from green to orange or red, warning of risks, but final verification and key decision-making must remain with humans. The professor explains this by noting that the cost of a false positive in critical infrastructure is incredibly high: the automatic shutdown of a main gas pipeline or the closure of a key transport artery due to an algorithmic failure entails enormous economic and social damage.

Shifting intelligence to the edge

To minimize such risks, Daponte describes the necessary evolution of the data collection architecture itself. Instead of sending colossal volumes of raw information from each sensor to a central cloud, a modern system should use the concept of edge computing.

In the professor’s view, primary signal denoising, interference compensation, data quality checks, and basic anomaly detection should take place directly on the sensor itself or in a local controller next to it. This not only reduces the burden on communication channels, but also ensures continuity of monitoring even if the external network is completely lost. Daponte proposes reserving cloud capacity exclusively for heavy tasks: global model training, historical analysis, and synchronization of the digital twin.

Summing up his report, Daponte outlines the main direction of the industry’s development. Artificial intelligence can become an exceptionally powerful tool in the hands of engineers, but its reliability is directly limited by the quality of the physical sensors’ performance. The future of infrastructure safety depends not on the abstract power of algorithms, but on our ability to ensure metrological accuracy at the very first meter of data transmission.