Combined with wavelet time frequency analysis and lightweight deep convolutional networkintelligent recognition of coal rock acoustic emission signal

During deep coal mining operations, coal-rock instability disasters frequently occur. However, real-time monitoring is constrained by challenges such as significant noise interference, low recognition accuracy of traditional methods, and insufficient computational efficiency of models. To address these issues, this paper proposes a lightweight acoustic emission signal recognition method based on the improved MobileNet V2 framework. First, Morlet wavelet transform is employed to construct acoustic emission time-frequency diagrams. Through energy entropy minimization criteria, scale parameters are optimized to compress 200 linear scales into 80 logarithmic distributed feature scales, enhancing high-frequency resolution to 0.01 seconds while effectively suppressing power frequency interference-induced spectral aliasing. Second, a channel mean fusion strategy compatible with single-channel inputs is designed, incorporating a dynamic expansion factor mechanism. This approach reduces feature expression capacity while decreasing shallow-layer module parameters by 33%, compressing overall parameters to 2.1 MB (a 38.2% reduction). In 30,000 coal-rock sample tests, the enhanced model achieved 94.0% classification accuracy with a single inference requiring only 14.3 ms. The accuracy decreased by merely 5.2% under high-noise conditions, demonstrating excellent real-time performance and robustness. Research findings indicate that this method significantly improves model lightweighting and adaptability to complex working conditions while maintaining precision, providing a viable technical solution for intelligent monitoring and early warning of coal-rock instability disasters in mines.