Shaoyang Qiu

Accurate monitoring of centrifugal pump hydraulic characteristics is essential for condition awareness and performance evaluation, while direct installation of flow and pressure sensors may increase system cost and maintenance requirements. This study proposes an edge-deployable intelligent soft measurement framework for estimating pump flow rate and head from readily available operating variables. The proposed MDST-PumpNet integrates multiscale depthwise separable convolution, cross-channel weighted projection transformation (CCWPT)-based feature interaction modeling, and feature fusion to capture nonlinear relationships between operating conditions and hydraulic responses. To enable deployment on resource-constrained devices, the trained model is optimized through attention-oriented knowledge distillation, entropy-based structured pruning, and post-training quantization.Experiments were conducted on a centrifugal circulating water-pump test platform using a group-hold-out evaluation strategy based on pump speed conditions. The proposed soft measurement method achieves flow-rate and head estimation R² values of 0.9990 ± 0.0002 and 0.9984 ± 0.0002, respectively, while reducing the parameter number by 45.31% compared with the baseline 1DCNN model. The optimized model is deployed on multiple computing platforms, including microcontroller unit (MCU), Linux-Arm, x86-64, and neural processing unit (NPU) devices. On an ESP32-S3 platform, the model achieves a mean inference latency of 0.062 ms and an operating power consumption of 0.39 W. A 24-h continuous deployment test further demonstrates stable runtime performance. The proposed framework provides an effective solution for edge-enabled soft measurement of centrifugal pump hydraulic characteristics under the evaluated conditions.