Adaptive end-side intelligent feature recognition system for acceleration measurement signals with strong adhesion interference under continuous multi-impact

The recognition of continuous multi-impact signals has long been an important and practical research topic in industrial and military applications. However, adjacent impact acceleration signals often become strongly coupled due to oscillations in electromechanical systems. Under the constraints of limited hardware resources on edge platforms, it is challenging to achieve both accuracy and timeliness in feature recognition algorithms. To address this issue, this paper proposes an end-side intelligent measurement system for strong adhesion acceleration signals under continuous multi-impact, based on a customized lightweight neural network and FPGA hardware. Firstly, in terms of algorithm design, key structural modifications and redundancy optimization were applied to SqueezeNet, transforming it into a lightweight CNN suitable for one-dimensional temporal shock acceleration signals. Second, for data filtering, a real-time processing mechanism using adaptive time windows was proposed. This method improves recognition accuracy by 19.15% to 32.06% compared to traditional approaches, while increasing network inference speed by nearly five times. For model optimization, a combined compression strategy of low-bit quantization and structural pruning was proposed, reducing the network size by 51.4% while keeping feature recognition accuracy nearly unchanged. For hardware deployment, a parallel computing strategy was implemented for the fully connected layers, enabling the compressed network to achieve an inference latency of 81.94 s on the hardware platform. Finally, a specialized dataset simulating multiple continuous impact scenarios was constructed using a continuous multi-impact experimental system. Field tests validated that the proposed end-side intelligent measurement system simultaneously achieves significantly improved accuracy, operational speed, and hardware efficiency compared to previous studies.