Stall warning for compressors based on wavelet features and multi-scale convolutional recurrent encoder–decoder

Due to the complexities of compressors and the influence of varied operational factors, a gradual decline in their performance status is inevitable, ultimately leading to compressor stalls. Compressor stalls can inflict substantial damage, thus, it is imperative to detect anomalies promptly and issue early warnings as soon as initial signs of reduced performance or suboptimal operation become apparent. Existing techniques commonly anticipate the onset of compressor stall by detecting spike inception; however, they have not successfully prevented stall effectively. The reason is that once spike inception is identified, it tends to quickly evolve into extensive stall cells within the span of just a few rotations. This article proposes a novel method for early warning of compressor stall, utilizing wavelet features and a Multi-Scale Convolutional Recurrent Encoder–Decoder (MSCRED). The approach extracts wavelet features from the compressor data and fuses information from sensors placed at three critical locations, feeding this data into the Convolutional Long Short-Term Memory (ConvLSTM) network. Furthermore, it utilizes the auto-thresholding technique to establish a pre-stall threshold. This method effectively analyzes precursor signs of compressor stalls, thereby significantly improving the timing of stall warnings for the compressor. The experimental outcomes demonstrate that the MSCRED method excels in the early warning of compressor stall compared to conventional approaches. By conducting a quantitative assessment of monotonicity, robustness, and correlation, the superior performance of the MSCRED method across different operating conditions has been validated.