Comparative study on the accuracy and robustness of piezoelectric valve flow prediction models based on deep learning

Abstract

To enhance the dependability and autonomy of the μg/s piezoelectric flow control device in long-term on-orbit operations, the on-orbit adaptability and robustness of various deep learning prediction models are compared and analyzed. Multidimensional time series data such as excitation voltage, spool displacement, inlet pressure, and valve body temperature were acquired and normalized. The accuracy of flow prediction was assessed using three models: standard Long Short-Term Memory (Standard LSTM) network, Attention LSTM, and convolutional neural network LSTM (CNN-LSTM). The models’ robustness was evaluated using several sensor degradation scenarios. CNN-LSTM has the best flow prediction ability. Compared with the Standard LSTM, the CNN-LSTM reduces absolute error (MAE) by 34.59 μg/s and relative error (MRE) by 2.82% when four-dimensional parameters are supplied (baseline state). The displacement of the valve spool is the key characteristic. When provided with a unit displacement signal, both Attention-LSTM and CNN-LSTM exhibit excellent flow prediction capabilities. The prediction error of the three LSTM models increases significantly when this feature is removed. Following the removal of the displacement feature using the CNN-LSTM model, the MRE and MAE exceed the baseline by 69.65% and 406.55 μg/s, respectively. Furthermore, the CNN-LSTM model demonstrates superior robustness in the scenario of forward-filled missing data, whereas the Attention-LSTM exhibits outstanding robustness in the presence of random noise. However, when confronted with sensitivity deterioration and signal drift, the prediction accuracy and robustness of all three models experience a significant decline.