The aeroengine fault-diagnosis system is critical to the safety of aeroengines. Traditional data-driven fault-diagnosis methods are characterized by high complexity and significant computational resource demands, which conflict with the real-time requirements and the limited computational and memory capacities of airborne equipment. Therefore, a dual-stream adaptive lightweight attention network (DALANET) for airborne deployment is proposed by combining the features of aeroengine gas-path fault data, which achieve collaborative optimization of diagnostic accuracy and computational efficiency by constructing a dynamic-fusion mechanism of global and local features. The monochannel global context attention module is developed to reduce the computational complexity of time series dependency modeling. The multiscale inverted residual module is improved to achieve efficient extraction of multiscale local features through the parallel structure of depth-separable convolution and grouped convolution, and a dynamic-fusion strategy is introduced to adaptively balance the global and local feature weights. Verification of aeroengine simulation data through dynamic fault injection demonstrates that DALANET achieves the maximum accuracy of 98.1% on the test set, with only 74,800 and 10.9 million floating-point operations per second of model parametric and floating-point operations, and a single inference time of 0.71 ms, reducing computational complexity while maintaining high accuracy.
Xiao et al. (2026) studied this question.