In recent years

deep learning technology has shown great potential in bearing fault diagnosis based on vibration signals.However

in the fault diagnosis method based on deep learning

the traditional single network topology feature extraction has weak discrimination and low noise robustness

and the accuracy of fault diagnosis is not high.In addition

most of the current research methods have a low fault recognition rate in a variable load environment.In response to the above problems

this paper proposes an improved neural network end-to-end fault diagnosis model.The model combines convolutional neural networks (CNN) and the attention long short-term memory (ALSTM) based on the attention mechanism

and uses ALSTM to capture long-distance correlations in time series data

Effectively suppress the high frequency noise in the input signal.At the same time

a multi-scale and attention mechanism is introduced to broaden the range of the convolution kernel to capture high and low frequency features

and highlight the key features of the fault. After testing on multiple data sets

and comparing with existing methods

experiments show that the method in this paper has significant performance in accuracy

noise robustness

and fault recognition rate under variable load conditions.