Aiming to further improve the prediction accuracy of intelligent diagnosis of vibration signals in the cavitation state of centrifugal pumps, a combination of optimization algorithms and long short term memory networks was proposed, i. e. , the method of variational mode decomposition ( VMD) of BiLSTM. This algorithm was used to identify various cavitation states in centrifugal pumps. An algorithm was introduced to optimize the BiLSTM model, thereby improving the prediction accuracy of time series data. In order to improve the prediction accuracy, mirage swarm optimization (MSO), sparrow search algorithm (SSA) and Pelican optimization algorithm ( POA) were adopted. The POA algorithms were compared and parameter optimization was carried out, obtaining different prediction models. For the vibration data of the cavitation experiment, the original vibration signal was firstly decomposed into multiple intrinsic mode function (IMF) components through VMD. As environmental pressure decreased, cavitation would gradually intensify. Acceleration sensors were placed at five positions to collect vibration signals. These measurement points were as follows: the outlet flange, the radial direction on the upper side of the pump body, the axial direction of the pump body, the radial direction on the right side of the pump body, and the pump feet. The collected vibration acceleration signal was subjected to the fast Fourier transform (FFT) to obtain the time-domain signal at the rated speed of 1450r/min, and the time-domain signal was converted into a frequency-domain signal based on the fast Fourier transform (FFT). Spectral analysis indicated that the vibration energy was mainly concentrated between the axis passing frequency (APF) and the blade passing frequency (BPF). Among them, the shaft frequency APF was 24.2 Hz, and the blade passing frequency BPF was 145.0Hz. Then, based on these components, features were constructed, and deep learning algorithms were used to accurately identify and predict the cavitation state. Based on the vibration data of centrifugal pumps under three different working conditions (no cavitation, primary cavitation, and severe cavitation), the experimental results showed that the prediction model fitting of VMD-SSA-BiLSTM after training was higher than that of LSTM, BiLSTM, VMD – BiLSTM, VMD-POA-BiLSTM, and VMD-MSO-BiLSTM. The research result can provide a reliable tool for identifying the multi-cavitation state of centrifugal pumps.