The health monitoring and fault diagnosis of heavy-duty engines are increasingly important for energy storage ecosystem. During operation, vibration characters corresponding to the specific fault need to be extracted ...The health monitoring and fault diagnosis of heavy-duty engines are increasingly important for energy storage ecosystem. During operation, vibration characters corresponding to the specific fault need to be extracted from the overall system vibration. Faulty characteristics emanating from one single cylinder are also mixed with those from other cylinders. Besides, the change of working condition brings strong nonlinearities in surface vibration. To solve these problems, an improved deep residual shrinkage network (IDRSN) is developed for detecting diverse engine faults at various degrees using single channel surface vibration signal. Within IDRSN, a wide convolution kernel is utilized in first convolution layer to capture the long-term fault-related impacts and eliminate the short-time random impact. The residual network module is adopted to enhance the focus the relevant components of vibration signals. Mini-batch training strategy is used to improve the model stability. Meanwhile, Gradient-weighted class activation map is adopted to assess the consistency between the learned knowledge and the fault-related information. The IDRSN is implemented to diagnosing a diesel engine under various faults, faulty degrees and operating speeds. Comparisons with existing models are analyzed in terms of hyper-parameters, training samples, noise resistance, and visualization. Results demonstrate the proposed IDRSN's superior performance on fault diagnosis accuracy, stability, anti-noise performance, and anti-interference performance. An average accuracy rate of 98.38 % was achieved by the proposed IDRSN, in comparison to 96.64 % and 93.56 % achieved by the DRSN and the wide-kernel deep convolutional neural network respectively. These results highlight the proposed IDRSN's superiority in diagnosing multiple faults under various working conditions, offering a low-cost, highly effective, and applicable approach for complex fault diagnosis tasks.展开更多
针对滚动轴承退化性能难以评估、寿命状态难以识别的问题,提出一种基于特征噪声能量比(Feature-to-noise energy ratio,FNER)指标及改进深度残差收缩网络(Improved deep residual shrinkage network,IDRSN)的滚动轴承寿命状态识别新方...针对滚动轴承退化性能难以评估、寿命状态难以识别的问题,提出一种基于特征噪声能量比(Feature-to-noise energy ratio,FNER)指标及改进深度残差收缩网络(Improved deep residual shrinkage network,IDRSN)的滚动轴承寿命状态识别新方法。首先,将全寿命轴承信号进行希尔伯特(Hilbert)变换和快速傅里叶变换(Fast fourier transform,FFT)得到包络谱,根据故障特征频率及其倍频计算包络谱幅值的特征能量比(Feature energy ratio,FER);然后,根据自相关函数(Autocorrelation function,AF)得到包络信号的总能量,将故障特征能量和噪声能量的比值作为轴承性能退化指标,之后按照FNER指标曲线划分轴承寿命状态和实现样本标签化;随后,使用标签化样本训练引入了密集连接网络的IDRSN得到轴承寿命状态识别模型。为了提高抗干扰能力,将DropBlock层引入第一个大型卷积内核,在全局平均池化之前引入Dropout技术。最后,运用两个滚动轴承全寿命周期数据集验证FNER指标和IDRSN模型的实用性和有效性,结果表明所提方法能更准确地实现滚动轴承寿命状态识别。展开更多
滚动轴承是风电机组关键部件,其运行工况复杂,故障类型难以准确识别。针对传统深度神经网络在强噪声环境下特征学习能力不足的问题,提出一种基于稠密连接模块的改进深度残差收缩网络(Deep residual shrinkage network based on dense bl...滚动轴承是风电机组关键部件,其运行工况复杂,故障类型难以准确识别。针对传统深度神经网络在强噪声环境下特征学习能力不足的问题,提出一种基于稠密连接模块的改进深度残差收缩网络(Deep residual shrinkage network based on dense block,DB-DRSN),实现强噪声、不同负载工况下滚动轴承故障的高效诊断。首先,将添加不同等级噪声的振动信号间隔采样并矩阵化,构建二维灰度图作为输入样本。然后,基于Dense block构造稠密连接残差收缩模块层(Residual shrinkage block unit based on dense block,DB-RSBU),利用Bottleneck层替代残差收缩模块中的卷积隐层,并加入Concat连接,达到对浅层和深层特征的充分利用。在每次稠密连接后通过1×1卷积进行降维,利用注意力模块和软阈值对逐通道特征赋不同阈值并降噪。最后,输入样本经过卷积池化层和DB-RSBU层堆叠的网络得到分类结果。试验表明,DB-DRSN模型在CWRU与PU滚动轴承数据集上不同噪声等级下的平均诊断准确率分别达到99.80%和96.44%,相比其他模型有更高的准确率、更快的收敛速度和更强的抗干扰能力。引入稠密连接核心思想对网络结构的改进可为基于数据驱动的风电机组滚动轴承故障诊断方法提供新思路。展开更多
基金funded by the National Key R&D Program of China(Grant No.2021YFD2000303)Tianjin Research Innovation Project for Postgraduate Students in China(Grant No.2021YJSB182)Weichai Power Co.,Ltd.in China(Grant No.WCDL-GH-2023-0147).
文摘The health monitoring and fault diagnosis of heavy-duty engines are increasingly important for energy storage ecosystem. During operation, vibration characters corresponding to the specific fault need to be extracted from the overall system vibration. Faulty characteristics emanating from one single cylinder are also mixed with those from other cylinders. Besides, the change of working condition brings strong nonlinearities in surface vibration. To solve these problems, an improved deep residual shrinkage network (IDRSN) is developed for detecting diverse engine faults at various degrees using single channel surface vibration signal. Within IDRSN, a wide convolution kernel is utilized in first convolution layer to capture the long-term fault-related impacts and eliminate the short-time random impact. The residual network module is adopted to enhance the focus the relevant components of vibration signals. Mini-batch training strategy is used to improve the model stability. Meanwhile, Gradient-weighted class activation map is adopted to assess the consistency between the learned knowledge and the fault-related information. The IDRSN is implemented to diagnosing a diesel engine under various faults, faulty degrees and operating speeds. Comparisons with existing models are analyzed in terms of hyper-parameters, training samples, noise resistance, and visualization. Results demonstrate the proposed IDRSN's superior performance on fault diagnosis accuracy, stability, anti-noise performance, and anti-interference performance. An average accuracy rate of 98.38 % was achieved by the proposed IDRSN, in comparison to 96.64 % and 93.56 % achieved by the DRSN and the wide-kernel deep convolutional neural network respectively. These results highlight the proposed IDRSN's superiority in diagnosing multiple faults under various working conditions, offering a low-cost, highly effective, and applicable approach for complex fault diagnosis tasks.
文摘针对滚动轴承退化性能难以评估、寿命状态难以识别的问题,提出一种基于特征噪声能量比(Feature-to-noise energy ratio,FNER)指标及改进深度残差收缩网络(Improved deep residual shrinkage network,IDRSN)的滚动轴承寿命状态识别新方法。首先,将全寿命轴承信号进行希尔伯特(Hilbert)变换和快速傅里叶变换(Fast fourier transform,FFT)得到包络谱,根据故障特征频率及其倍频计算包络谱幅值的特征能量比(Feature energy ratio,FER);然后,根据自相关函数(Autocorrelation function,AF)得到包络信号的总能量,将故障特征能量和噪声能量的比值作为轴承性能退化指标,之后按照FNER指标曲线划分轴承寿命状态和实现样本标签化;随后,使用标签化样本训练引入了密集连接网络的IDRSN得到轴承寿命状态识别模型。为了提高抗干扰能力,将DropBlock层引入第一个大型卷积内核,在全局平均池化之前引入Dropout技术。最后,运用两个滚动轴承全寿命周期数据集验证FNER指标和IDRSN模型的实用性和有效性,结果表明所提方法能更准确地实现滚动轴承寿命状态识别。
文摘滚动轴承是风电机组关键部件,其运行工况复杂,故障类型难以准确识别。针对传统深度神经网络在强噪声环境下特征学习能力不足的问题,提出一种基于稠密连接模块的改进深度残差收缩网络(Deep residual shrinkage network based on dense block,DB-DRSN),实现强噪声、不同负载工况下滚动轴承故障的高效诊断。首先,将添加不同等级噪声的振动信号间隔采样并矩阵化,构建二维灰度图作为输入样本。然后,基于Dense block构造稠密连接残差收缩模块层(Residual shrinkage block unit based on dense block,DB-RSBU),利用Bottleneck层替代残差收缩模块中的卷积隐层,并加入Concat连接,达到对浅层和深层特征的充分利用。在每次稠密连接后通过1×1卷积进行降维,利用注意力模块和软阈值对逐通道特征赋不同阈值并降噪。最后,输入样本经过卷积池化层和DB-RSBU层堆叠的网络得到分类结果。试验表明,DB-DRSN模型在CWRU与PU滚动轴承数据集上不同噪声等级下的平均诊断准确率分别达到99.80%和96.44%,相比其他模型有更高的准确率、更快的收敛速度和更强的抗干扰能力。引入稠密连接核心思想对网络结构的改进可为基于数据驱动的风电机组滚动轴承故障诊断方法提供新思路。
