The new method, described in the journal Medical & Biological Engineering & Computing by Wanchao Yao, Tianshu Gu, and Fuwang Wang, combines three technologies into a single pipeline called MVMD-FCASNN: multivariate variational mode decomposition, a frequency and channel attention spiking neural network, and a carefully designed signal-processing front end. The system reads multichannel electroencephalogram, or EEG, signals collected noninvasively from the scalp and classifies whether the operator is mentally fatigued in real time. In benchmark tests, the model achieved a classification accuracy of 98.81 percent and, critically, maintained strong performance even when the EEG recordings were deliberately corrupted with noise at a signal-to-noise ratio of minus six decibels, a level of interference far harsher than most laboratory experiments ever attempt.

The problem the researchers set out to solve is fundamentally one of signal integrity. EEG is famously vulnerable to contamination. On a construction site, the electrical systems of large machinery, radio transmissions, vibration transmitted through the operator’s body, and the operator’s own muscle movements and eye blinks all inject artifacts into the faint electrical potentials measured at the scalp. Classical decompositions of EEG signals, such as empirical mode decomposition, have long been used to tease apart the mixed frequency content of brain recordings, but they are notoriously sensitive to noise. When noise corrupts the decomposition process itself, the resulting components lose their physical meaning, and downstream classifiers inherit the error. Previous work, including studies on driver fatigue detection, has struggled with exactly this weakness, which is why many published fatigue-detection systems perform well in quiet laboratories but falter in the field.

The team’s answer to this challenge is multivariate variational mode decomposition, or MVMD. Unlike methods that process each EEG channel independently, MVMD jointly decomposes all channels at once, searching for a set of intrinsic mode functions, or IMFs, that are frequency-aligned across every recording electrode. In practical terms, this means that the alpha-band oscillation detected over the frontal cortex and the alpha-band oscillation detected over the central electrodes are extracted as matching components of the same decomposition, preserving the spatial relationships between brain regions that conventional approaches tend to scramble. Because variational mode decomposition formulates the problem as an optimization with explicit bandwidth constraints, it is also far more robust against noise corruption than empirical mode decomposition, a property that earlier studies in motor-imagery brain-computer interfaces had already hinted at. Applying it to fatigue detection in a genuinely noisy setting is one of the paper’s key contributions.

Once the EEG signals have been decomposed, the question becomes which of the resulting components actually carry the signature of fatigue. Not all frequency bands are equally informative, and the answer can shift from person to person and from moment to moment. To handle this, the researchers designed an IMF attention mechanism that adaptively evaluates the contribution of each extracted frequency component, automatically weighting the bands that best discriminate fatigue from alertness. Neuroscientifically, this makes good sense: mental fatigue is typically accompanied by changes in relative spectral power, such as increases in theta and alpha activity over frontal regions and alterations in beta activity associated with sustained attention. By letting the network learn which IMFs matter, the system avoids the brittleness of hand-picked frequency features.

Alongside frequency attention, the architecture incorporates a channel attention module, inspired in part by squeeze-and-excitation networks developed in computer vision, that emphasizes critical brain regions. The authors report that the model places its strongest weight on frontal and central areas of the scalp, precisely where neurophysiological research locates much of the neural activity associated with cognitive fatigue, vigilance decrements, and the executive control of sustained attention. This spatial selectivity strengthens the model’s representation of the brain’s fatigue signature while suppressing electrodes that contribute mostly noise. The combination means the network effectively learns both where to look on the scalp and which rhythms to listen for, two questions that traditional pipelines answer with fixed, hand-engineered rules.

The final stage of processing is perhaps the most futuristic element: a spiking neural network, or SNN. Unlike conventional artificial neural networks that transmit continuous-valued activations, spiking networks communicate through discrete spikes, mirroring the way biological neurons actually fire. This choice serves two purposes. First, spiking neurons are inherently event-driven, meaning they respond to changes in their input rather than to every sample, which makes them energy efficient and well suited to portable, battery-powered monitoring hardware that an operator could realistically wear on a job site. Second, spiking dynamics preserve the temporal structure of the EEG signal, retaining information about how brain activity evolves over milliseconds, information that frame-based networks tend to discard. The researchers point to earlier work on low-power driving fatigue monitoring with spiking networks as a foundation for this design philosophy.

The experimental results are the reason this study is attracting attention. On mental fatigue classification tasks, MVMD-FCASNN reached 98.81 percent accuracy, a figure the authors describe as significantly outperforming traditional methods. More striking is the model’s resilience: at a noise level of minus six decibels, where the interference is actually stronger than the underlying EEG signal, the system continued to classify fatigue reliably, while conventional comparison methods degraded sharply. For a technology intended to run continuously in environments where electrical noise from cranes, welding equipment, and generators is unavoidable, this anti-interference capability may matter more than headline accuracy. The study also reports informed consent from all participants and ethics approval consistent with the Helsinki declaration, and the authors declare no conflicts of interest.

The broader context explains why this line of research is urgent. Tower crane accidents remain a persistent source of fatalities and injuries in the construction industry, and studies in Hong Kong, Indonesia, and elsewhere have tied a substantial share of unsafe hoisting behavior to operator fatigue and excessive mental workload. Earlier detection systems have relied on eye-tracking glasses, facial landmark analysis, smart cushions, heart rate variability, and other physiological proxies, each with tradeoffs in comfort, invasiveness, and robustness. EEG offers a direct window into cognitive state, but until now its practical deployment in high-noise industrial settings has been hampered by exactly the interference problem this new architecture is built to withstand. Related work by overlapping research groups, including flexible dry electrodes designed for real driving environments, suggests a growing ecosystem of wearable brain-monitoring hardware maturing alongside the algorithms.

The researchers, funded by the National Natural Science Foundation of China and supported by Northeast Electric Power University and the Jilin City Science and Technology Bureau, argue that their approach offers a reliable and efficient solution for real-time mental fatigue monitoring in high-noise construction scenarios. If such systems can be integrated into crane cabins as a routine safeguard, an operator whose brainwaves betray dangerous drowsiness could be alerted, or their work schedule could be flagged for review, before a lapse in attention translates into a dropped load. The work also points beyond construction: the same pipeline of joint multichannel decomposition, dual attention mechanisms, and spiking computation could plausibly be adapted to truck drivers, pilots, air traffic controllers, and any profession where sustained vigilance is a matter of life and death. For now, the result stands as a striking demonstration that brain-inspired computing can withstand the messiness of the real world, not just the laboratory, and that the human brain’s own warning signs of exhaustion can be read clearly enough to act on them in time.

Subject of Research: EEG-based detection of mental fatigue in tower crane operators using multivariate variational mode decomposition and a frequency and channel attention spiking neural network in real construction environments

Article Title: Study on mental fatigue detection of crane operators based on frequency and channel attention spiking neural network with multivariate variational mode decomposition in real construction environments

Article References: Yao, W., Gu, T., & Wang, F. (2026). Study on mental fatigue detection of crane operators based on frequency and channel attention spiking neural network with multivariate variational mode decomposition in real construction environments. Medical & Biological Engineering & Computing. https://doi.org/10.1007/s11517-026-03669-y

Image Credits: AI Generated

DOI: 10.1007/s11517-026-03669-y

Keywords: mental fatigue, crane operators, EEG, spiking neural network, multivariate variational mode decomposition, attention mechanism, construction safety, brain-computer interface, signal processing, neural decoding, occupational health, machine learning