Multiple open-switch fault diagnosis in four-level ANPC inverters using adaptive CNN with reinforced attention

Dyan Puspita Apsari; Dong-Choon Lee

Multiple open-switch fault diagnosis in four-level ANPC inverters using adaptive CNN with reinforced attention

Dyan Puspita Apsari

Dong-Choon Lee

Vol. 26, No. 4, pp. 945-961, Apr. 2026

10.1007/s43236-026-01287-3

Deep Learning

Induction motor

Multilevel inverters

Multiple switch faults

Open-switch fault

Abstract

This paper proposes a fault diagnosis framework for multiple open-switch faults in four-level active neutral-point-clamped (ANPC) inverters under limited labeled data conditions. In multilevel inverter systems, the number of possible fault combinations increases exponentially with the number of switches. The proposed method leverages one-phase voltage signals with half-bridge (HB) reconstruction to improve observability and reduce data requirements. An adaptive convolutional neural network with reinforced attention optimization (ACNN-RAO) is developed, where depthwise convolutions minimize the parameters and a reinforcement agent dynamically adjusts attention weights to enhance classification accuracy. For online implementation, three parallel models are deployed to classify single, double, and triple open-switch faults. Experimental results demonstrate 99.6% diagnostic accuracy with 96% fewer labeled data, and fault identification within 1.5 fundamental cycles.

Statistics

Show / Hide Statistics

Statistics (Past 3 Years)
Multiple requests among the same browser session are counted as one view. If you mouse over a chart, the values of data points will be shown.

Cite this article

[IEEE Style]

D. P. Apsari and D. Lee, "Multiple open-switch fault diagnosis in four-level ANPC inverters using adaptive CNN with reinforced attention," Journal of Power Electronics, vol. 26, no. 4, pp. 945-961, 2026. DOI: 10.1007/s43236-026-01287-3.

[ACM Style]

Dyan Puspita Apsari and Dong-Choon Lee. 2026. Multiple open-switch fault diagnosis in four-level ANPC inverters using adaptive CNN with reinforced attention. Journal of Power Electronics, 26, 4, (2026), 945-961. DOI: 10.1007/s43236-026-01287-3.