A semi-supervised transferable LSTM with feature evaluation for fault diagnosis of rotating machinery

Tang, Zhi; Lin, Bo; Liu, Xiaofeng; Wei, Daiping

doi:10.1007/s10489-021-02504-1

Cited by 23 publications

(10 citation statements)

References 38 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…where h is the smoothing parameter and k : ð Þ is the Gaussian kernel function. Similarly, the probability density distribution function q x ð Þ of the sample point Y is obtained, and it is brought into formulas (11) and (12) to obtain the KLD value between a sample point and the center point in the same distribution. Some current studies have shown that due to the constant probability density function of the center point in the same distribution, KLD, which has better computational efficiency, is more suitable for the calculation of distribution differences than Wasserstein distance.…”

Section: Incremental Learning Networkmentioning

confidence: 99%

See 1 more Smart Citation

Incremental learning BiLSTM based on dynamic proportional adjustment mechanism and experience replay for quantitative detection of blade crack propagation

Shen

Song

et al. 2023

Structural Health Monitoring

View full text Add to dashboard Cite

In the traditional quantitative detection model for blade cracks in centrifugal fan, it is assumed that the data distribution is fixed or stable. However, the new data brought by the crack propagation would break the stable distribution, thereby disturbing the old data, and resulting in a decrease in the detection performance of the model. To overcome catastrophic forgetting and reduce the extra computational cost of retaining intact old data, a quantitative detection method based on incremental learning bidirectional long short-term memory (BiLSTM) with dynamic proportional adjustment mechanism and experience replay for blade crack propagation is proposed. First, a basic BiLSTM model is constructed by inputting the data of cracks with a length of 0–5 mm. Second, the fully connected layer features in the model are selected for t-distributed stochastic neighbor embedding (t-SNE) dimensional reduction, and the Kullback–Leibler divergence is used as an indicator of feature distribution evaluating the representative old data. Third, a dynamic proportional adjustment mechanism for the old data retention proportion is constructed according to the feature distribution index and the model detection accuracy. Finally, the data of the crack with a length of 6–10 mm are gradually input to proceed with the incremental learning of the model. Verified by the measured data of the centrifugal fan, the model can adjust the retained number of old crack length data dynamically, and import new crack length data for incremental learning, making it characterized by high detection accuracy, stability, and plasticity for the quantitative detection of crack length propagation in blades.

show abstract

Section: Incremental Learning Networkmentioning

confidence: 99%

“…This mechanism enables the network to store and transmit information for a long time. The calculation formula of LSTM from input value to output value is given below 12,13 :…”

Section: Review Of Related Algorithmsmentioning

confidence: 99%

Incremental learning BiLSTM based on dynamic proportional adjustment mechanism and experience replay for quantitative detection of blade crack propagation

Shen

Song

et al. 2023

Structural Health Monitoring

View full text Add to dashboard Cite

show abstract

“…Better results can be obtained by extracting the time domain characteristics from the collected vibration signals and inputting them into the LSTM neural network for fault detection. To reflect the most comprehensive gear fault situation, this paper extracts 15 time-domain indexes based on the comprehensive consideration of literature [7] , which are mean value, standard deviation, absolute mean value, variance, root amplitude, maximum value, minimum value, root mean value, peak value, margin index, waveform index, pulse index, peak value, skewness and kurtosis. The detailed process of the fault detection model based on IPSO-LSTM proposed is shown in Figure1.…”

Section: Fault Detection Model Based On Ipso-lstmmentioning

confidence: 99%

Gear fault detection method based on improved PSO and LSTM

Zhang

zhang

Shen

2023

Third International Conference on Mechanical, Electronics, and Electrical and Automation Control (METMS 2023)

View full text Add to dashboard Cite

Gear is an important part of power transmission between machines. Long Term and short term memory network (LSTM) is adopted as the basic model for gear fault detection. Aiming at the problem that the model hyperparameters of LSTM are difficult to be determined manually, by using the improved particle swarm optimization (IPSO) to find the optimal LSTM superparameters automatically. Optimize the learning factor and inertia weight in PSO to enhance particle optimization ability, so as to build IPSO-LSTM model with better detection ability. In this paper, the gear fault data set published by the University of Connecticut was used to verify the detection model, and the fault detection accuracy of the gear fault data set reached 98.7%. Compared with the unimproved PSO-LSTM model, LSTM and CNN, the accuracy rate is increased by 2.9%,7.4% and 10% respectively.

show abstract

“…At this time, 1) the loss generated by large-scale samples is large; 2) the loss generated by large-scale samples accounts for a large proportion. The researchers [23] proposed the amount function in the binary classification task, as shown in formula (6) (6) Where reduces the contribution of large-scale samples to the loss function. Compared with the cross-entropy loss, the loss of large-scale samples is reduced by , and the loss of effective samples is expanded, thereby highlighting the performance of small-scale sample accuracy.…”

Section: Focal Loss Functionmentioning

confidence: 99%

Bearing fault diagnosis under class unbalanced data based on deep learning

2022

AJETS

View full text Add to dashboard Cite

There is a huge difference in the number of operating samples and failure samples in industrial production. If the diagnostic model is trained through deep learning under an unbalanced data set, it will make the model recognize the faulty samples as normal samples. Aiming at this problem, an adaptive focus loss function mechanism is proposed. It can avoid over-learning large-scale samples during smallbatch imbalance training. At the same time, to improve the generalization ability of fault samples, a pretraining data enhancement mechanism is proposed. By using the rolling element bearing fault diagnosis data set, the effectiveness of the two mechanisms is verified. The two mechanisms can not only complete training tasks under unbalanced data and improve generalization ability.

show abstract

A semi-supervised transferable LSTM with feature evaluation for fault diagnosis of rotating machinery

Cited by 23 publications

References 38 publications

Incremental learning BiLSTM based on dynamic proportional adjustment mechanism and experience replay for quantitative detection of blade crack propagation

Incremental learning BiLSTM based on dynamic proportional adjustment mechanism and experience replay for quantitative detection of blade crack propagation

Gear fault detection method based on improved PSO and LSTM

Bearing fault diagnosis under class unbalanced data based on deep learning

Contact Info

Product

Resources

About