Physics-guided deep learning for real-time resonance diagnostics and predictive maintenance in pulsed water jet manufacturing systems

Unlike closure mechanisms for laser or EDM machining, pulsed water jet manufacturing does not currently have any. The reason is timing; nothing tracks the chamber resonance quickly enough. Although FFT analysis requires 30–50 ms, control of a HAPOS (hydrodynamic acoustic pulse oscillatory system) operation up to 450 MPa has to be kept below 10 ms of inference budget. The answer is our 12-layer physics-guided CNN (PG-CNN) which reads the resonance frequency f rez directly from a model’s raw pressure signals (100 kHz, Hann-windowed, N = 4096 samples). Training runs in two stages. First we pre-train on 2400 synthetic spectra from a validated Helmholtz model (Δ f = 6.4% against the five-element HAPOS); we then fine-tune on 129 quality-screened experimental spectra (out of 312 planned acquisition slots) over 5–25 MPa, using discriminative learning rates and early stopping. The dilated convolutions (kernel K = 7, dilations {1, 2, 4}) cover a 1.69 ms receptive field; attention pooling then selects the time segments. Within PG-SAD, the fixed transfer constant k̃ rkk = 9.41·10 − 4 gives way to a learnable function k rkk ( p , g ) under heteroscedastic constraints. On the held-out set ( n = 240) the PG-CNN reaches 98.2% accuracy inside a ± 50 Hz band with MAE = 23.4 Hz ( R 2 = 0.998), while PG-SAD lowers the mean absolute frequency error to 12.8 Hz at F1 = 0.97. After ONNX/TensorRT optimisation, a Jetson Xavier NX delivers 6.2 ms inference (161 FPS at 15 W), well within the τ ≈ 50 ms time constant of plunger pumps. This physics–CNN bridge is what supplies the sensing layer that closed-loop control and predictive maintenance of pulsed water jet cells have been missing.