Abstract
For the deficiency that the traditional single forecast method can not forecast precisely the states of electronic equipments, a combined forecast method based on hidden Markov model (HMM) and least square support vector machine (LS-SVM) is presented. Multi-agent genetic algorithm (MAGA) is used to estimate parameters of HMM in order to overcome the problem that Baum-Welch algorithm is easy to fall into local optimal solution easily. MAGA is used to estimate parameters of LS-SVM. On the base of these, the combined forecast model of electronic equipment states is established. The example results showed the superiority of the combined forecast model at forecast precision, calculation speed and stability.
Introduction
Electronic equipment degradation is a complex process where internal health states are hidden, and only external performance parameters are observable. Predicting these states is critical for proactive maintenance. However, traditional single-model approaches often struggle with precision. Hidden Markov Models (HMMs) are well-suited for state decoding but suffer from training issues where standard algorithms, like Baum-Welch, frequently settle for local optima. Least Square Support Vector Machines (LS-SVMs) are excellent for regression but do not inherently model sequential state transitions. This paper proposes a hybrid "forecast-then-decode" model that leverages the strengths of both, optimized by a Multi-agent Genetic Algorithm (MAGA).
Methodology
The proposed architecture integrates LS-SVM for performance forecasting and HMM for state classification.
Hybrid Architecture
The process begins with data acquisition and preprocessing. LS-SVM is employed to handle the non-linear forecasting of future observation data (performance metrics). These forecasted values are then fed into an HMM to determine the most likely underlying health state.
HMM State Modeling
The equipment states are categorized into five grades: Health, Fine, Notice, Deterioration, and Bad. To reflect the irreversible nature of hardware wear-and-tear, a "left-right" HMM topology is used, where transitions only move forward or stay in the current state.
MAGA Optimization
A core innovation is the use of MAGA to estimate parameters for both HMM and LS-SVM. Unlike standard Genetic Algorithms, MAGA agents interact with their local environment and neighbors, facilitating a more efficient global search and preventing the model from settling for local optima.
Experimental Setup
The model was validated using experimental data from a temperature control amplifier. Four key parameters—T1, T3, and T6 temperature signal voltages, and speed-decreasing warning voltage—were fused into a single "absolute change volume" metric to simplify simulation calculations.
Results and Discussion
The hybrid MAGA-optimized model demonstrated significant improvements over standard methods:
- HMM Optimization: MAGA achieved a logarithmic average likelihood of -4.3818, vastly outperforming the Baum-Welch algorithm (-26.6025) and Ant Colony Algorithm (-10.2467).
- Forecasting Precision: The MAGA-optimized LS-SVM yielded a Mean Absolute Error (MAE) of 0.0203, nearly half that of the standard Adaptive Genetic Algorithm (0.0439).
- Stability: The combined model maintained high calculation speeds and stability across multiple test sequences.
The "division of labor" between LS-SVM (regression) and HMM (state decoding) proved highly effective, allowing each algorithm to operate within its optimal domain.
Conclusion
The combined HMM and LS-SVM approach, optimized via MAGA, provides a robust framework for electronic equipment state forecasting. While highly accurate for near-term predictions, future work will focus on extending the forecast horizon and applying the model to other failure modes, such as capacitor leakage. The ultimate goal is to transition from state forecasting to predicting Remaining Useful Life (RUL).
Summary updated