Acoustic-Based Condition Recognition for Pumped Storage Units Using a Hierarchical Cascaded CNN and MHA-LSTM Model

As an important regulating resource in power systems, pumped storage units frequently switch operating conditions due to peak shaving and frequency regulation, making the condition transitions complex. Traditional methods struggle to achieve high-precision classification. This paper proposes a hierarchical cascade deep learning model based on noise signals, which integrates a convolutional neural network (CNN) with a multi-head attention long short-term memory network (MHA-LSTM) to address the differentiated recognition of steady-state and transitional conditions. The CNN efficiently extracts multi-scale spatial features from sound spectrograms, enabling fast classification under steady-state conditions. The MHA-LSTM combines attention mechanisms with time-series modeling. This enhances its ability to capture long-range dependencies in the signals. And it significantly improves classification accuracy in ambiguous boundaries and transitional scenarios. Testing on 3413 noise samples shows that the proposed method achieves an overall accuracy of 92.22%, with steady-state condition recognition exceeding 98%, and recall and F1 score above 90% for major categories. Compared with other approaches, this model provides a high-precision classification tool for unit health monitoring, supporting the intelligent operation and maintenance of power plants.