The action power curve of a ZDJ9 switch machine reflects the combined electrical response of the driving motor, the transmission load and the switching process. To address the scarcity of high-confidence maintenance labels, the storage of monitoring data as interface screenshots, and the difficulty of whole-curve discrimination in distinguishing local response deviations from global fluctuations, this paper proposes a power-response state deviation detection method based on stage-wise residuals and boundary perturbation stability. Each action is divided into three power response segments (initial, middle and terminal); a reference curve is constructed from normal samples of the same device and movement direction; local deviations are characterized by stage-wise residuals; and boundary perturbation routes samples whose segmentation is sensitive to boundary changes into review. Experiments on 310 power curves show that the method achieves a screening F1 of 0.968 and produces no reliable false alarms among the 50 normal test samples (0/50, 95% CI: [0%, 7.1%]), outperforming the differential-peak baseline with a reliable false alarm rate of 12%, while providing reliable outputs for 87.27% of the samples.

ZDJ9 switch machine; power curve; stage-wise residual; boundary perturbation stability; review-oriented output; state deviation detection