Research On Bearing Sub-health Recognition Algorithm Based On LSTM Fusion Multi-scale Convolution

Rolling bearings are an indispensable part of modern manufacturing,and are widely used in most mechanical equipment.However,due to long-term high-speed and overload and other harsh working conditions,the bearings are easily in a sub-healthy state.In order to ensure the reliability and safety of the equipment,it is necessary to monitor the health of the bearings.Therefore,many experts and scholars attach great importance to bearing sub-health recognition and carry out in-depth research.Aiming at the situation that the traditional bearing sub-health recognition method relies too much on the relevant experience of the experimenter in the feature extraction process,and the final diagnosis effect is not good,this paper proposes an improved LSTM fusion multi-scale convolution sub-health recognition algorithm.On the one hand,since traditional convolutional neural networks only use a singlesize convolution kernel in each convolutional layer,when feature extraction is performed on bearing sub-health data,if the size of the convolution kernel is not selected properly,it may not be able to accurately locate The location of the shock feature in the domain signal may not cover a complete shock vibration period in the signal,so a multi-scale convolutional neural network is proposed to enhance the input classification information.Divide two channels in the multi-scale convolutional layer,each channel uses different size convolution kernels to convolve the input data,and then continue the convolution and pooling operations on the obtained feature vectors,and then obtain the two channels The result of using element-wise product for feature fusion,and then input to the fully connected layer to summarize and the output layer to give the final sub-health recognition results.On the other hand,for CNN,the main consideration is the spatial correlation of features,which can describe local spatial features.Since the measured vibration signal of bearing sub-health recognition is sequence data,the impact of the timing information existing between the data on the diagnosis accuracy has to be considered,so the features obtained after multi-scale convolution fusion are selected to be input into the improved LSTM network to extract the timing information.At the same time,after multi-scale convolution and pooling,the dimensionality of the data can be greatly reduced,and the long-term dependency problem of LSTM due to the long input data sequence can also be alleviated.The improved LSTM network adopts a new unit structure,which merges the three "gates" in the original LSTM structure into an updated "gate",simplifies the network structure and reduces the number of network parameters,thereby shortening the running time of the model and ensuring accuracy At the same time,it enables the model to perform sub-health recognition faster.Finally,the bearing data set of Case Western Reserve University(CWRU)is used to experimentally verify the improved LSTM fusion multi-scale convolution sub-health recognition algorithm proposed in this paper.It can be obtained from the experimental result data and the results of the comparison experiment.This paper proposes The algorithm can effectively identify the sub-health state of the bearing,which verifies the feasibility of the improved algorithm in this paper.