TensorFlow中深度前馈网络优化研究及其轴承故障诊断应用

目前在复杂系统的故障诊断中,故障特征与故障类型之间存在较强的非线性关系,且数据量较大,信号处理复杂,诊断效率不高,而深度学习在特征提取与模式识别方面显示出巨大潜力。针对此问题提出基于深度前馈网络的故障诊断模型,将其应用于复杂的轴承故障诊断。该方法直接将原始信号作为模型的输入特征量,然后利用谷歌开源深度学习框架TensorFlow建模,通过相关参数设置、梯度算法优化、正则化处理对网络进行优化设计。构建上万的9种轴承故障类型样本,确保样本多样性,提高网络鲁棒性,最终优化后的模型诊断准确率为98.96%。将该方法与多种传统的机器学习诊断方法进行比较,结果表明该方法能更有效地进行轴承故障诊断,验证了模型的合理性和优越性。

OPTIMIZATION OF DEEP FEEDFORWARD NETWORK IN TENSORFLOW AND ITS APPLICATION OF BEARING FAULT DIAGNOSIS

At present,in the fault diagnosis of complex systems,there is a strong nonlinear relationship between the fault characteristics and the fault type.Due to the big amount of data and the complex signal processing,the efficiency of diagnosis is not high,while the deep learning has shown great potential in feature extraction and pattern recognition.We proposed a fault diagnosis model based on deep feedforward network for this problem,which was applied to complex bearing fault diagnosis.We directly regarded the original signal as the input feature of the model,and then used the Google open source deep learning framework TensorFlow to model and optimize the network through relevant parameters setting,gradient optimization algorithm and regularization processing.Nine kinds of bearing fault samples beyond 10 000,were built to ensure sample diversity,improve network robustness.The final optimized model diagnosis accuracy is 98.96%.This method is compared with many traditional machine learning methods,and the results show that it can diagnose bearing fault more effectively and verify the rationality and superiority of the model.