Answers are grounded in this paper and cite their sources.
Biologically inspired neural computation offers a promising approach to processing complex temporal signals in autonomous engineering systems. However, acoustic fault diagnosis remains challenging under non-stationary operating conditions, particularly when multiple motors operate simultaneously and produce overlapping acoustic patterns that complicate fault identification. This study investigates a biologically inspired Liquid Neural Network (LNN) framework integrating heterogeneous Izhikevich spiking neurons, Spike-Timing-Dependent Plasticity (STDP), and Genetic Algorithm (GA)-based neuronal parameter optimization. The proposed architecture combines locally adapted recurrent synaptic connections with evolutionarily selected neuronal parameters to support fault-related temporal representation. Acoustic recordings representing healthy operation and four propulsion fault conditions were divided into 1500 non-overlapping one-second segments. Log-Mel acoustic representations and complementary descriptors were processed using a 64-neuron spiking reservoir and a regularized Ridge readout. Performance was evaluated using temporally blocked data partitions and a retrospective test subset of 290 segments across five random seeds. The proposed GA–STDP–LNN achieved an accuracy of 93.59 ± 0.87%, a macro F1 score of 93.60 ± 0.87%, and a Matthews correlation coefficient of 0.9214 ± 0.0107. Under synthetic noise, macro F1 decreased to 83.87% at 20 dB and 6.67% at 10 dB. These findings demonstrate the potential of combining biologically inspired neuronal dynamics, synaptic plasticity, and evolutionary optimization for acoustic fault classification, while highlighting the need for independent recording-based validation and improved robustness under severe acoustic disturbances.
No takes yet. Share an insight, caveat, or question.
Altinörs et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: