Semiconductor Design-for-Reliability (DfR) faces a critical knowledge bottleneck. Experimental parameters, failure criteria, and mechanism interpretations remain scattered across publications in inconsistent formats. Such fragmentation hinders systematic reuse and cross-study comparison. To address this challenge, we present an automated framework that extracts reliability data from unstructured technical documents and transforms it into standardized, auditable knowledge for experiment planning. Coordinated AI modules perform data extraction, lifetime and degradation modeling, physical constraint validation, and test condition recommendation. All intermediate results are stored in a shared database for traceability. Local deployment of all AI models ensures data confidentiality suitable for industrial use without reliance on external cloud services. Evaluation on a corpus of 50 IGBT power-cycling publications showed that the extraction module achieved an F1 score of 0.83, which outperformed rule-based methods by 15 percentage points. The largest gains were observed for semantically variable parameters such as failure criteria and Weibull coefficients. For experiment design recommendations, alignment with published test matrices reached up to 62.5%, compared to 0-25% for general-purpose AI models of comparable or greater scale. These results demonstrate that domain-structured knowledge integration, rather than AI model size alone, drives the performance improvement and offers a practical pathway for auditable, privacy-preserving automation in reliability engineering workflows.

Aiming at the reliability bottlenecks of power MOSFETs in such key application fields as new energy vehicles (NEVs), photovoltaic (PV) inverters, and industrial motor drives, this paper proposes an unsupervised anomaly detection method for power MOSFET degradation based on the Long Short-Term Memory Autoencoder (LSTM-AE). Without relying on labeled data or complex hardware modifications, the method can adaptively learn the health patterns and temporal features of devices, addressing the limitations of traditional methods such as difficulty in capturing long-term dependencies and poor noise immunity. Three different test conditions for power cycling tests were designed to collect operational data of power MOSFET devices. On-state resistance (RDS(on)) and body diode voltage drop(VSD) were selected as input features, and the Leave-One-Out Cross-Validation (LOOCV) method was adopted to verify the model performance. Experimental results demonstrate that the method can effectively capture weak degradation features during the stage of insignificant device performance changes, achieving an early warning prior to device failure. Under 3σ principle threshold, its detection performance outperforms the Multilayer Perceptron Autoencoder (MLP-AE) and traditional Mahalanobis Distance (MD) methods, with the optimal sequence size being 64. It achieves a high accuracy in identifying normal states and exhibits high recognition capability for severely degraded states, providing a robust, efficient, and practical solution for the health monitoring of power MOSFETs.

Accurate lifetime prediction of power semiconductor devices is critical for the reliability of power electronic systems. Power cycling tests generate only a small number of failure samples due to their high cost and long duration. This data scarcity makes it difficult to train reliable prediction models. At the same time, modern deep learning models require significant AI expertise to configure and tune, which creates a barrier for reliability engineers. To address these two problems, this paper proposes a domain-constrained multi-agent framework for automated remaining useful life (RUL) prediction of power semiconductor devices. The framework integrates three collaborative agents: a Stats Agent for statistical reliability assessment, a Physics Agent for physical plausibility validation, and a Model Agent for algorithm selection and hyperparameter tuning. A Central Controller manages the workflow and coordinates agent communication. The agents interact through a Propose-Critique-Refine (PCR) mechanism, in which the Model Agent proposes configurations and the other two agents critique them from statistical and physical perspectives. This iterative negotiation eliminates the need for manual parameter tuning. The framework is validated on the NASA IGBT accelerated aging dataset under Leave-One-Device-Out Cross-Validation. Results show that the framework achieves an R2 of 0.914 at a 50% observation ratio and remains above 0.79 even at 30%, demonstrating strong generalization under limited data conditions.