IMS-FACNN(Improved Multi-Scale Convolution Neural Network integrated with a Feature Attention Mecha

Introduce a 2020 Thesis of Shanghai University of science and Technology , Improved multiscale convolution network with attention mechanism .

Catalog

1. introduction

2. Network architecture

3. Model details

3.1 Improved multi-scale coarse-grained processing

3.2 Characteristic attention layer

3.3 The overall architecture

5. experiment

6. Conclusion

---

1. introduction

Traditional diagnosis methods based on signal processing , Such as EMD,VMD,ELMD etc. , Based on time-frequency analysis , Popular learning , Or sparse representation . The features extracted by the above methods will be regarded as shallow （shallow） Learning machine （SVM,Logistic Regression） The input of . However , The upper limit of these learning methods depends heavily on the quality of the extracted features . more importantly , Due to the limitation of diagnostic ability , These shallow learning models will lead to poor generalization ability under complex working conditions . There are three main limitations ：1. The design of feature extraction and classification, two processes that affect the performance of diagnosis, is tested independently （ The compatibility between feature extraction and classifier is not considered ）.2. Feature extraction requires expert knowledge and signal processing skills , It takes time and effort .3. The existing diagnostic methods are so unique that they are difficult to be applied to other industrial fields . That is, the generalization is not good .

Deep learning can overcome the above shortcomings , Excellent performance in classification and prediction tasks . The advantages of deep learning ：1. Nonlinear high-dimensional representation ability .2. Directly connecting data does not require feature extraction .3. Strong learning ability and generalization .

introduce CNN, Less computing resources due to weight sharing and sparse connections . The sensor signal is generally 1D The signal , There will be some 1D Signal turn 2D Image as 2DCNN Input method （ spectrum . Wavelet spectrum ）. However ,2D Image as CNN Model input Unable to learn the natural vibration characteristics in the signal . therefore 1DCNN Is used to extract more detailed features . also 【25】 prove , The original signal is used as CNN The input trained model has better generalization ability and robustness under load conditions . Vibration signals usually contain multi-scale features , But the depth CNN It is difficult to capture the inherent multi-scale features 【27】. Small batch It is more suitable for processing images and larger batch Better for handling 1D Timing information .【30】 prove 1D The original signal is used as CNN Input is more effective than 2D Images . When more sensitive information is extracted , The accuracy of the model is higher and more stable .

therefore , The author puts forward IMS-CACNN, Feature the attention layer to adapt to the weight of learning , There is no need to search for the optimal time scale . The main contributions are as follows ：1. Put forward IMS-CACNN Obtain vibration characteristics of different scales .2. Feature attention mechanism is introduced to deal with the multi-scale features learned from adaptive fusion .3. The optimal mini-batch.

2. Network architecture

3. Model details

3.1 Improved multi-scale coarse-grained processing

In the research of traditional multiscale diagnosis , Due to the use of non overlapping windows to divide the time scale , Some inherent characteristic information will be ignored . therefore , This paper uses overlapping window sliding to extract signals .

3.2 Characteristic attention layer

The feature attention mechanism can adaptively score the features learned at different scales and assign their weights .

3.3 The overall architecture

The features of sub layers with different scales are fused through the attention mechanism , Then enter the classification layer to diagnose the fault category .

5. experiment

5.1CWRU On dataset Anti noise performance

 5.2CWRU On dataset Migration performance

5.3 Mixed condition （0hp-3hp） Compare the results of different models

 

 In order to reflect the effectiveness and superiority of the proposed IMS-FACNN model, a scenario of mixed data with different noise is set, where training set includes 0HP,1HP, 2HP and 3HP clean bearing vibration signals, the testing set includes 0HP,1HP, 2HP and 3HP bearing vibration signals with different SNR.

In order to prove the proposed IMS-FACNN The efficiency and superiority of the model , In the compound working condition without noise （0hp-3hp） Train the model . Add noise during the test , That is to say, the test set is a composite working condition + Noise of different decibels .

6. Conclusion

Put forward IMS-FACNN, Directly act on the original vibration signal . Continuous slip sampling , Ensure multi-scale . Introduce attention mechanism .

1. Due to the introduction of improved multi-scale coarse granularity （ Coarse texture ）（coarse-grained） step , It can extract more effective information and has anti-interference （anti-interference）.

2. It can distinguish fault category and fault degree .

3. Multi scene （multiple scenario）, Under load , Compared to other methods （MS-CNN,MC-CNN） With higher accuracy .

4. For the fan data in the real scene ,IMS-FACNN The model achieves good diagnostic accuracy .

5. The “2nd scale” plays a leading role instead of “1st scale” when distinguishing the outer race bearing fault.

The classification results of learning features from different scales show ： The features with larger weight given by attention mechanism have more obvious performance on classification . Although some features with small weight have little influence on the classification results , But the features of different scales also complement each other .