摘要
针对裂纹声发射信号的识别问题,基于大边缘自编码器(LMAE)和堆叠融合监督自编码器(SFSupAE)设计了自适应大边缘堆叠权重监督自编码(ALMAE-SWSupAE)算法.针对LMAE中的固定k值问题,引入自适应k值算法,修改h(s)运算方法解决数据溢出问题;在SFSupAE中引入子分类器的性能权重优化分配策略,并设计新的权重函数;使用铝合金试件进行拉伸裂纹实验,识别采集到的声发射信号.研究结果表明:所提出的ALMAE-SWSupAE算法法识别准确率达到98.89%,相较于SSAE、SDAE、CAE、StAE和SAE方法性能具有明显提升,并在消融实验中证明了其改进有效性.
Abstract
To address the recognition problem of crack acoustic emission signals,this paper designs an adaptive large margin stacked weight supervised autoencoder(ALMAE-SWSupAE)based on the large margin autoencoder(LMAE)and the Stacked Fusion Supervised Autoencoder(SFSupAE).First,to address the fixed k value problem in LMAE,an adaptive k value algorithm is introduced and the h(s)operation method is modified.This modification is crucial as it effectively tackles the data overflow problem that often plagues the systems.By implementing this adaptive algorithm,the model better adapts to different data characteristics and improves its performance and stability in handling crack acoustic emission signal recognition.Secondly,in the SFSupAE component,the performance weight of the sub-classifier is introduced.This helps to reduce the negative impact of poorly performing sub-classifiers on the overall model performance.Meanwhile,a new weight function is designed,which resolves the poor adaptability in different classification tasks when using the information entropy for weight allocation.This enhancement enables the model to handle a wider variety of classification scenarios and improve its generalization ability.Then,a tensile crack experiment is conducted with an aluminum alloy specimen.In the experiment,the collected acoustic emission signals are recognized by the proposed ALMAE-SWSupAE model.Our results are highly encouraging as the ALMAE-SWSupAE method proposed in this paper reaches a recognition accuracy of 98.89%,which is markedly higher than that achieved by SSAE,SDAE,and SAE methods.Finally,to verify the effectiveness of the method proposed in this paper,an ablation experiment is conducted on ALMAE-SWSupAE.In the experiment,the adaptive k value algorithm,the new weight allocation function and the performance weight of the sub-classifier are set as variables.Our results demonstrate the adaptive k value algorithm not only improves the model performance but also reduces the training time.The use of the new weight allocation function increases the accuracy of the overall model by approximately 3%.Moreover,the performance weight of the sub-classifier enables the model to maintain a good recognition rate even when there is a large performance difference among the sub-classifiers,effectively enhancing the model's robustness and reliability in applications.
国家科技期刊平台
NSTL
万方数据