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

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.