<p>The progression in integrated circuit technology has necessitated advanced solutions for the storage and rapid transmission of extensive data generated by electronic modules. Compared to traditional signal compression and transmission techniques, compressed sensing (CS) transcends the limitations of the Shannon–Nyquist sampling theorem by enabling low-frequency signal sampling, thereby becoming a extensively utilized approach in the signal processing. Recent advancements in Artificial Intelligence have further propelled the reconstruction efficacy of deep learning-based CS methods, mitigating certain constraints inherent in traditional CS approaches. Nonetheless, the existing deep learning-based CS methods are not optimally efficient for electronic data processing. In this regard, this study proposes a novel double-path multiscale adaptive compressed sensing network (DMA-CS). This network is structured around four key modules: signal compression, preprocessing, initial reconstruction, and secondary reconstruction. The signal compression module samples the signal for compression, the preprocessing module prepares the sampled signal for subsequent processing, the initial reconstruction module employs a double-path complementary network comprising a multiscale residual module fused with a multihead attention module and an inverse residual module for the initial reconstruction. Then, the secondary reconstruction module uses the adaptive dilated convolution residual module to adaptively adjust the size of the convolution kernel to ensure the high-quality reconstruction of different signals and combines it with the tree-like structure residual block for enhanced reconstruction. Our experimental evaluation on the P2020 module fault signal dataset and NASA Lithium Battery dataset demonstrates that our scheme attains the lowest percentage root-mean-square difference and the highest signal-to-noise ratio, demonstrating substantial enhancement in reconstruction performance and robustness. </p>