基于多模型级联的VSV调节机构装配智能检错方法

An Intelligent Fault-Detection Method for Assembly of VSV Adjusting Mechanism Based on Multi-Model Cascade

在航空发动机可调静子叶片(Variable stator vane,VSV)调节机构的装配过程中,目前仍需要人工检测其连杆防松钢丝的装配正确性,效率低且易出错。为替代人工检错,提出了一种基于多模型级联的智能检错方法。该方法是多个卷积神经网络级联的模型集成,其中包含检测模块、分类模块以及后处理比对检错3个部分。首先在检测模块上提出混合不同尺寸卷积核的深度可分离卷积与轻量化解耦头来对YOLOv5s进行改进,改进的YOLOv5s在测试集上的平均精度达到97.9%,相较于YOLOv5s、YOLOv8s分别提升了3.4%、1.5%。其次在分类模块上使用7×7深度卷积替代全局平均池化以改进ConvNeXt分类头,改进后性能得到提升,在连杆数据集和螺纹数据集上的准确率分别达到97.5%和95.4%。最后在后处理模块对两个分类模型的结果进行匹配,得出装配检测结果。利用现场装配车间采集的图片数据集对该智能检错方法进行验证,结果显示该方法平均精度达到92.7%,进一步验证了智能装配检错方法的可靠性。

The assembly process of the variable stator vane(VSV)adjusting mechanism of aero-engine requires manual detection of the anti-loosening wire assembly correctness of the connecting rod,which is inefficient and errorprone.An intelligent fault-detection method based on multi-model cascade is proposed to replace the manual detection operation.The method is a model integration of multiple convolutional neural networks,which consists of three parts:detection module,classification module,and post-processing of comparison&fault detection.Firstly,the depthwise separable convolution with lightweight decoupling head mixing different sizes of convolutional kernels is proposed on the detection module to improve YOLOv5s,and the improved YOLOv5s achieves an average accuracy of 97.9%on the test set,which is improved by 3.4%and 1.5%compared to YOLOv5s and YOLOv8s,respectively.Secondly,the ConvNeXt classification head is improved by using 7×7 deep convolution instead of global average pooling on the classification module,and the performance is improved,reaching an accuracy of 97.5%and 95.4%on the connecting rod dataset and the thread dataset,respectively.Finally,the results of the two classifcation models are matched in the post-processing module to obtain the assembly detection result.The intelligent fault-detection method is verifed by the image dataset collected from the feld assembly workshop,and the results show that the average precision of the proposed method reaches 92.7%,which further verifes the reliability of the proposed method.