一种结合卷积神经网络(convolution neural networks,简称CNN)和长短时记忆(long short term memory,简称LSTM)神经网络的滚动轴承RUL预测方法。首先,对滚动轴承原始振动信号作快速傅里叶变换(fast Fourier transform,简称FFT);其次,将预处理所得到的频域幅值信号进行归一化处理后,将其作为CNN的输入,然后,再将深层特征输入到LSTM网络中,构建趋势性量化健康指标,同时确定失效阈值;最后,实现轴承寿命预测。
opts = trainingOptions('adam', ...
'MaxEpochs',100, ...
'GradientThreshold',1,...
'ExecutionEnvironment','cpu',...
'InitialLearnRate',0.005, ...
'LearnRateSchedule','piecewise', ...
'LearnRateDropPeriod',50, ... %2个epoch后学习率更新
'LearnRateDropFactor',0.5, ...
'L2Regularization',1e-6,...
'Shuffle','once',... % 时间序列长度
'SequenceLength',k,...
'MiniBatchSize',100,...
'Verbose',1,...
'Plots','training-progress');