In this study, we propose a novel intelligent laboratory assistant system designed to revolutionize experimental teaching through a synergistic combination of large language models (LLMs), voice interaction, and state machine control. The proposed system addresses the perennial challenges in hands-on education, including the high demand for instructor time for repetitive guidance, the difficulty of providing immediate and personalized student feedback, and the critical need for real-time safety monitoring. By developing a voice-enabled companion, we aim to transition from static, paper-based manuals to a dynamic, interactive, and intelligent tutoring experience that guides students through experiments step by step, answers contextual questions, and proactively mitigates risks.

The core innovation of this system is its tri-domain adapter architecture, a hierarchical design that harmonizes the broad reasoning capabilities of LLMs with the stringent precision required for experimental instruction. First, a general knowledge adapter establishes the foundational layer by using a locally deployed and fine-tuned DeepSeek model. This adapter is systematically infused with comprehensive electrical engineering expertise, encompassing circuit analysis, analog and digital electronics, and power systems. This injection of domain-specific knowledge ensures the accurate interpretation of complex terminology, such as distinguishing between “series-aiding” and “series-opposing” configurations in mutual inductance experiments, providing a reliable theoretical bedrock. Second, an experimental context adapter dynamically bridges theory and practice. By interfacing with the system’s state machine, it loads standardized procedural workflows corresponding to the student’s selected experiment. This resolves critical contextual ambiguities in which the same component (e.g., “resistor R1”) requires different measurement protocols depending on the experimental setup. It facilitates a “co-piloted” guidance process, replacing static manuals with interactive, step-by-step voice instructions and performing real-time validation of student measurements to prevent error propagation. Third, a real-time safety filter functions as a dedicated risk-mitigation module. It continuously monitors the audio stream for lexical indicators of hazards, such as “unusual smell,” “overheating,” or “sparks,” triggering immediate voice alerts and simultaneous instructor notifications to preempt accidents. Orchestrating these domains yields an intelligent information router. By analyzing a student’s query alongside the current state machine status, the router dynamically arbitrates between the general knowledge base and the specific experimental database. This ensures that every response is both contextually aware and grounded in verified instructional materials, effectively aligning generic AI capabilities with specific educational objectives.

The proposed system was implemented and evaluated in the “Fundamentals of Electronic Technology” laboratory. Experimental results demonstrate that the system performed stably across 120 controlled tests: student operation accuracy reached 97.7% (baseline: 83.3%), the equipment damage rate was reduced to 0.8% (baseline: 4.3%), and the experimental report quality score improved by 29.5 points (on a 100-point scale).

Even with limited initial datasets that were subsequently expanded and enhanced through data augmentation techniques, the proposed system demonstrated a marked improvement in student experiment completion efficiency, a reduction in repetitive queries to instructors, and a notable decrease in simulated safety incidents. This work provides a reusable technological paradigm for intelligent, personalized, and safer laboratory education across various scientific and engineering disciplines.