基于一维轻量化CNN的山地索道轴承故障诊断

为保证山地果园索道安全稳定运行，并在网络环境较差的山地果园实现对索道驱动系统轴承故障诊断，该研究提出了一种一维端对端轻量化CNN检测方法1D-MRL-CNN（one-dimensional mountain ropeways lightweight convolutional neural network ），直接对采集到的一维振动信号进行检测。基于残差结构（residual structure）和深度可分离卷积（deep separable convolution），引入BN（batch normalization）层，在保证检测精度的同时大幅度降低模型的参数量和复杂度，并提升鲁棒性和泛化能力，适用于索道的变负荷工作状态；采用改进stem block模块、h_swish激活函数并在主体模块最后一层添加通道注意力机制（squeeze and excitation, SE），提高网络模型的特征提取能力。为了验证模型的综合性能、变负荷工况下的稳定性以及抗噪声干扰性能，利用帕德博恩（paderborn university, PU）和凯斯西储（case western reserve university, CWRU）数据集进行试验验证。PU数据集试验结果表明，该方法故障分类准确率达99.43%，相比同类最优网络分类准确率提高0.97个百分点；参数量为83.44 kB，分别是Resnet18、VGG16、MobileNetV3-large和ShuffleNetV1模型的2.19%、0.81%、2.84%和3.31%。CWRU数据集试验结果表明，该方法在变负荷工况下的平均准确率达96.70%，比Resnet18、WDCNN和MobileNetV3-large网络分别高9.1、4.7和10.5个百分点；在4种噪声工况下的平均识别准确率达99.14%，比Resnet18、WDCNN和MobileNetV3-large网络分别高4.74、1.24和5.51个百分点。最后通过自建数据集对模型的实际工况故障分类效果进行验证，1 400个样本中仅有2个故障样本预测错误，准确率达99.86%。本研究的网络模型参数量小、准确率高，在变负荷和有噪声的工况下鲁棒性较高，适用于山地果园运输索道的轴承故障检测。

Fault diagnosis of mountain ropeway bearings based on one-dimensional lightweight CNN

China is the largest fruit producer and consumer in the world, while most orchards in the south of China are located in hilly areas. The ropeways can be expected for transporting orchards in mountainous areas. But the harsh working environment can often lead to machine failures in recent years. Therefore, it is of great significance to investigate and solve the problems at the initial stage of failures. Generally, the conventional bearing fault diagnosis system includes five links: signal acquisition, feature extraction, state identification, diagnosis analysis. and decision intervention. Deep learning is widely applied for bearing fault detection in recent years. Traditional machine learning also needs to manually extract the fault features at present, depending mainly on the deep professional knowledge. However, it is high demand for the high performance of fault diagnosis, particularly for simple structures and less calculation during feature extraction. This study aims to realize the fault diagnosis of ropeway drive system bearings in the mountain orchards with poor network environment, in order to ensure the safe and stable operation of ropeway in the mountain orchards. A one-dimensional end-to-end lightweight CNN detection, 1D-MRL-CNN was established to directly detect the one-dimensional vibration signals for the mountain ropeways. Specifically, the new model was established using residual structure and depth separable convolution. The depth separable convolution was applied to greatly reduce the parameter and calculation amount of the improved model. In addition, the residual structure was applied to make up for the accuracy loss caused by depth separable convolution. The parameter amount and complexity of the model were reduced significantly while ensuring the detection accuracy. The stem block and BN layer were then introduced to improve the robustness and generalization ability of the new model suitable for the variable load working state of the ropeway. Finally, the hard_swish activation function was also adopted in the model. The channel attention mechanism was added to the last layer of the main module, in order to improve the feature extraction ability of the network model. Two datasets (Paderborn University and Case Western Reserve University) were used to verify the comprehensive performance, stability under variable load and anti-noise interference performance. The Paderborn University dataset showed that the fault classification accuracy of the improved model was 99.43%, which was 0.56, 1, and 1.23 percentage points higher than those of the one-dimensional classical CNN, similar optimal network, and the lightweight CNN architecture optimal network. The parameters and floating-point calculations were 83.44kb and 0.20M, which were 2.19%, 1.17%, 2.84%, and 0.74% of the one-dimensional classical CNN classification network architectures (such as Resnet18, Resnet34, Resnet50 and VGG16), 6.19% and 30.40% of the same type of networks 1D-Lenet5 and 1D-Inception, while 4.22%, 2.07%, 2.84%, 3.31% and 5.14% of the one-dimensional lightweight CNN architecture MobileNetV1, MobileNetV2, MobileNetV3-Large, ShuffleNetV1 and EfficientNet-1. In addition, the Case Western Reserve University dataset showed that the average accuracy rate of the improved model was 96.70% in six load scenarios, which was 9.1, 4.7, and 10.5 percentage points higher than those of Resnet18, WDCNN, and MobileNetV3-Large, respectively. The average recognition accuracy was 99.14% under four noise conditions, which was 4.74%, 1.24%, and 5.51% higher than those of Resnet18, WDCNN and MobileNetV3-Large, respectively. Finally, the fault classification of the improved model under actual working conditions was verified by the established ropeway dataset, where only 2 fault samples out of 1 400 samples were predicted incorrectly. The new network model was suitable for the bearing fault detection of mountain orchard transport ropeway, due to the small parameters, high accuracy and robustness under variable load and noisy working conditions.