Predictive Maintenanceand Condition Monitoring with HQNN

The goal is to turn production telemetry and equipment sensor streams into an early, reliable warning of failure or wear, in time to schedule maintenance instead of reacting to it. This page covers two worked cases built on the same method: predicting machine failure from production-line telemetry, and classifying the condition of a hydraulic system's cooler from sensor time series.

Unplanned downtime is costlier than planned service, and early warnings reduce scrap, missed deliveries, and emergency repairs. The classical baselines differ by case — gradient boosting, random forests, and neural networks for tabular telemetry; CNNs for sensor time series — but the quantum approach is the same across both: a hybrid quantum–classical neural network (HQNN) with a shallow variational circuit that can match strong baselines with fewer trainable parameters in compact encodings.

Mechanically, both cases follow the same recipe after classical preprocessing. Each feature controls the rotation angle of one or more single-qubit gates — angle encoding — which maps the classical sample into a quantum state. A stack of variational layers then applies trainable rotations interleaved with entangling gates (CNOT or CZ) that couple qubits, spreading local feature information across multi-qubit correlations. The circuit is measured to produce expectation values (Pauli-Z or Pauli-Y, depending on the case), and that low-dimensional vector feeds a small classical head that makes the final prediction. The whole model — classical layers and quantum gate angles together — is trained end-to-end with gradient-based methods such as the adjoint or parameter-shift rule.

The primary case is a binary classification task: predict impending component failure from production-line telemetry (temperatures, rotational speed, torque, tool wear).

1. Ingest telemetry signals, define warning horizons and failure labels, and build rolling and statistical features with leakage control.
2. Train boosting, random forest, and neural-network baselines with stratified temporal splits and logged hyperparameters.
3. Attach an 5-qubit variational block with depth 1 as a compact head over classical features and train it with a classical optimizer.
4. Report accuracy, precision, recall, ROC-AUC, and PR curves while monitoring class imbalance.
5. Capture seeds, config, environment logs, and a change log for full reruns.

Data can come from industrial or public telemetry sources, and each run documents exact signals, labeling rules, windows, and any synthetic augmentation.

The second case is a multiclass time-series problem: determine the condition of a hydraulic system's cooler from raw process-sensor streams (pressures, volume flows, temperatures) recorded over 60-second duty cycles on a test rig. It's the same predictive-maintenance goal as Case 1 — turn sensor readings into an early, reliable condition signal — applied to a system where the input is a multivariate time series rather than a row of tabular features.

Time series need a different front end than a row of tabular features. Here, a classical LSTM first summarizes each 60-step cycle, and a Parallel Quantum Network (PQN) — several small quantum circuits run side by side rather than one wide circuit — takes the compressed representation from there. Splitting the quantum work across parallel narrow circuits keeps each one shallow and easier to run on near-term hardware, while the parallel blocks together still reach 12 qubits of representational capacity.

For a fault-detection case built on a different quantum method, see steel plate fault detection with QSVM.