Motor test benches in university laboratories frequently operate unattended, especially during overnight fatigue tests. Under abnormal conditions such as overcurrent or locked rotor, conventional protection mechanisms, typically an overcurrent relay with a local buzzer and a cryptic fault code, exhibit three limitations: the buzzer is inaudible outside the closed laboratory; the fault code conveys no semantic information; and the immediate power cutoff destroys the fault scene, precluding post-incident waveform analysis. Although large language models have shown promise in industrial fault diagnosis, most existing solutions rely on public cloud APIs that require transmitting raw equipment data over the Internet, posing severe data security risks where current waveforms, temperature profiles, and proprietary control parameters constitute unpublished intellectual property. In this study, we propose an intelligent early warning system that delivers remote, human-readable diagnostic reports to researchers’ mobile devices while keeping all raw data strictly within an air-gapped intranet.

A three-layer decoupled architecture termed “edge sensing–intranet inference–cross-gateway pushing” is proposed, with data flowing strictly upward and no upper-layer node permitted to initiate downward connections. The edge-sensing layer employs a Raspberry Pi 4B with an RS485 CAN HAT, communicating with the variable-frequency drive via Modbus RTU and polling three-phase current RMS values and heatsink temperature at 50 ms intervals. Samples are maintained in a ring buffer corresponding to a 10 s sliding window. A dual-threshold anomaly trigger algorithm with a consecutive over-limit counterbalances sensitivity against noise immunity, and the counter threshold is jointly calibrated with the drive’s hardware overcurrent protection delay to prevent missed detections and false alarms. The intranet inference layer runs on an air-gapped GPU workstation with an NVIDIA RTX 4070, hosting the Ollama engine loaded with a GGUF-quantized Qwen2.5-14B-Instruct model and the open-source agent framework OpenClaw. Upon receiving an anomaly frame, OpenClaw retrieves fault waveforms, 30-day baseline statistics, and maintenance records from a local SQLite database; assembles them into a structured prompt with semantic pre-digestion; and invokes the model to generate a report containing a phenomenon description, root-cause analysis, and recommended actions. The cross-gateway pushing layer uses a lightweight mini-PC with dual NICs. Unidirectional data transfer is enforced through independent routing tables, iptables firewall whitelists, Linux namespace and seccomp sandboxing, regex-based content validation, and outbound audit logging, emulating industrial-grade isolation without dedicated hardware.

Experiments on a 1.5 kW three-phase asynchronous motor bench used two fault conditions: locked rotor simulated by a sudden magnetic powder brake torque step and heat dissipation degradation simulated by obstructing the cooling fan inlet. Over ten trials per fault type, the mean end-to-end latencies from anomaly trigger to DingTalk notification were 8.3 and 9.1 s for locked rotor and heat dissipation degradation, respectively, with model inference accounting for approximately 65% of the delay. Under concurrent stress testing with three simultaneous bench anomalies, peak GPU memory utilization reached only 61% of 12 GB. Blind evaluation by two senior engineers across 20 reports yielded average scores of 4.6/5.0 for phenomenon description accuracy, 4.2/5.0 for root-cause analysis reasonableness, and 4.5/5.0 for recommendation practicality.

The proposed architecture delivers semantic-level diagnostic reports to remote researchers within approximately 8–9 s while ensuring that no raw data traverse the external network, satisfying both intelligent early warning and intranet data compliance requirements at an incremental hardware cost below 2000 CNY. Future work will explore LoRA fine-tuning on domain-specific fault samples, multi-sensor fusion, and continuous batching inference to enhance diagnostic accuracy and scalability.