基于轻量化1D-DenseNet的电磁超声导波缺陷识别

Electromagnetic Ultrasonic Guided Wave Defect Recognition Based on Lightweight 1D-DenseNet

摘要:

针对电磁超声导波检测中一维时序信号信噪比低、高精度深度网络难以在资源受限的便携式终端部署的问题,对基于轻量化1D-DenseNet的嵌入式缺陷识别算法进行了研究；通过采用BN参数折叠与INT8混合静态量化的关键技术,有效降低了网络推理阶段参数冗余与算力开销大的负担,在保留一维局部波包特征的同时实现了模型的大幅压缩；经ZYNQ-7020异构处理器硬件平台实验测试,实现了从信号高频采集到端侧模型推理的全链条部署验证；结果表明该方案在仅0.8 MB的存储开销下实现了96.08%的分类准确率，系统吞吐量达65.8 samples·s?¹；该方案较好地平衡了推理实时性与检测精度,为在役管道便携式智能无损检测的工程落地提供了一种可行的技术途径。

Abstract:

Aiming at the problems of low signal-to-noise ratio of one-dimensional time-series signals in electromagnetic acoustic transducer (EMAT) guided wave testing and the difficulty of deploying high-precision deep networks on resource-constrained portable terminals, an embedded defect recognition algorithm based on lightweight one-dimensional densely connected convolutional network (1D-DenseNet) was studied. The key technologies of batch normalization (BN) parameter folding and INT8 mixed static quantization were adopted. These technologies effectively reduced the burden of parameter redundancy and large computational overhead during the network inference stage. They achieved significant model compression while retaining one-dimensional local wave packet features. Through experimental testing on the ZYNQ-7020 heterogeneous processor hardware platform, the full-chain deployment verification from bottom-layer high-frequency acquisition to terminal model inference was realized. The results show that this scheme achieved a classification accuracy of 96.08% with a storage overhead of only 0.8 MB, and a system throughput of 65.8 samples·s?¹. This scheme effectively balanced inference real-time performance and detection accuracy. It provides a feasible technical approach for the engineering application of portable intelligent nondestructive testing for in-service pipelines.