Research On Fault Diagnosis Method Of Time Series Data Based On CNN-LSTM

The scale of modern industrial production is huge,the hidden troubles of equipment failures are more and more complicated,and safety accidents pose a greater threat to industrial systems.Establishing a fault diagnosis model can efficiently locate equipment faults and ensure the safe production of the system.The current fault diagnosis methods have developed from the mechanism method for simple systems to the data-driven method for complex systems,but there are still some shortcomings,for example,the characteristics of individual faults are similar and difficult to distinguish;The unbalanced proportion of all kinds of samples in the data affects the training effect of the model.Therefore,this paper takes the industrial production system as the research object,based on the monitored system time series data,and realizes fault diagnosis through deep learning modeling.The main research work is as follows:（1）Considering that the collected data samples are polluted by noise,aiming at the problem that the existing noise reduction algorithms can’t effectively separate the noise when dealing with non-stationary time series data,this paper improves the wavelet threshold denoising method,and puts forward a method of adaptive threshold instead of fixed threshold.In this method,the noise is removed by changing the threshold value layer by layer,and the influence of noise is eliminated more thoroughly while keeping effective information as much as possible.The simulation results show that,after the noise signal is processed by the improved noise reduction method,the noise in the drawn timing diagram is obviously less,which is more similar to the overall shape of the timing diagram of the original signal.（2）Among the data samples sampled by industrial equipment,there is little difference between some faults,so the common fault diagnosis model is difficult to identify.To solve these problems,this paper proposes a fault diagnosis model based on attention mechanism improvement of Long-term circular convolution Network（Attention-CNN-LSTM）,which adds an attention layer in front of the output layer,so that the model can pay attention to the weak differences among different faults.At the same time,the model combines the ability of cyclic network to remember historical state with the ability of convolutional network to extract highdimensional static features,which makes the model can efficiently process a large amount of data and pay attention to the dynamic performance of samples.Experiments show that compared with the traditional Long Short Term memory network（LSTM）method for processing time series data,the average recall rate of the optimized model is increased by16.9%,and the average accuracy rate is increased by 13.45%.It can identify the faults that LSTM model can’t distinguish,and the training efficiency is obviously improved.The convergence speed of the objective function is faster and more stable,which indicates that the optimized modeling method has a good performance in the face of complex time series data.（3）In the real situation of industrial production,the occurrence of faults is small probability,which leads to the scarcity of fault data compared with normal data,making fault diagnosis more difficult.To solve this problem,this paper proposes a fault diagnosis method based on deep convolution generative confrontation network（GAN-CNN-LSTM）.The sequence data is imaged by the Gramian Angular Field（GAF）method,and then complementary samples are synthesized by the generative confrontation network according to the imaged sample data,which alleviates the problem of data imbalance at the data level and improves the fault diagnosis effect of the model.The experiment shows that in the training set where the proportion of small-class samples is more than 20%,adding the synthesized sample expansion can effectively improve the recognition ability of the model for small-class samples.