TL;DR
This paper introduces a novel CNN-BiLSTM-Attention hybrid neural network to predict the aging of automotive wire harnesses. By integrating convolutional neural networks for local feature extraction, bidirectional LSTMs for temporal analysis, and an attention mechanism for weighting important time steps, the model achieves highly accurate predictions of wire harness degradation, outperforming traditional methods.
What Problem Does This Paper Address?
Modern vehicles are increasingly reliant on complex electrical systems, making the health of components like wire harnesses critical for performance and safety. Wire harnesses are susceptible to aging due to various factors like electrical load, heat, and environmental exposure, which can lead to performance degradation and potential failures. Predicting this aging process accurately is challenging due to the multi-physics coupling effects involved. Traditional methods often rely on complex thermal models or are limited in their ability to capture the dynamic, time-varying nature of aging. This paper aims to bridge this gap by developing an intelligent system that can predict wire harness aging using real-time operational data.
Methods
The core of this research is a hybrid neural network architecture: CNN-BiLSTM-Attention.
- Convolutional Neural Networks (CNNs): Used for local feature extraction. Specifically, 1D convolutional layers are employed to capture multi-scale patterns in the input data, effectively identifying local characteristics of the wire harness's operating state.
- Bidirectional Long Short-Term Memory (BiLSTM) Network: This component handles the temporal dynamics of the aging process. BiLSTM processes sequential data by considering both forward and backward dependencies, allowing it to capture a more comprehensive understanding of how the wire harness's state evolves over time.
- Attention Mechanism: Applied to the output of the BiLSTM layer. This mechanism dynamically assigns weights to different historical time steps, allowing the model to focus on the most relevant information for predicting the current aging state and reducing the impact of irrelevant noise.
The model takes voltage, current, and temperature as primary input features, along with derived features like power and impedance, to provide a richer representation of the wire harness's condition.
Datasets
The study utilized data collected from an experimental platform designed to simulate a 12V automotive low-voltage electrical system.
- Accelerated Aging Experiment: Complete automotive wire harnesses were subjected to high-temperature accelerated aging at 140°C for up to 720 hours. This approach was chosen over aging individual material samples to better simulate real-world conditions, including the aging of internal copper wires and connection ends.
- Data Collection: Data, including timestamp, voltage, current, and temperature, were collected at a sampling period of 100 ms. A sliding window of 300 data points was used as input for the LSTM, and a single-step prediction strategy was adopted.
- Dataset Split: The collected data was partitioned into training and validation sets while preserving temporal continuity. A portion of the data was used for testing and comparison with other models.
Key Insights and Ideas
- Dielectric Loss as a Key Indicator: The paper identifies dielectric loss as a crucial, non-destructive indicator of wire harness aging. Its correlation with electrical performance makes it a suitable quantifiable measure for aging degree.
- Synergy of Hybrid Architecture: The combination of CNN, BiLSTM, and Attention is not merely additive; it creates a powerful synergy. CNNs extract local patterns, BiLSTM models temporal evolution, and Attention refines the focus on critical temporal information, leading to a more robust and accurate prediction.
- Importance of Derived Features: Beyond raw sensor readings, derived features like power (voltage * current) and impedance provide deeper insights into the wire harness's condition, particularly its resistance and power loss, which are directly impacted by aging.
- Practical Implications for Vehicle Safety: The ability to predict aging levels with high accuracy opens doors for proactive maintenance. Thresholds for dielectric loss can be defined to trigger diagnostic warnings or recommend timely replacement, significantly enhancing vehicle safety and reliability.
Key Results
The CNN-BiLSTM-Attention model demonstrated superior performance in predicting wire harness aging:
- Low Error Metrics: Achieved a Mean Absolute Error (MAE) of 0.02806.
- Significant Error Reduction: Showed 32.50% and 62.06% error reduction compared to LSTM and Random Forest models, respectively, in terms of MAE.
- Model Stability: Exhibited low performance variations across 10 independent runs (MSE = 0.00292 ± 0.0004, MAE = 0.02806 ± 0.003), indicating robustness.
- Visual Confirmation: Line charts comparing predicted values against actual dielectric loss showed the CNN-BiLSTM-Attention model's predictions were more convergent and had a smaller fluctuation range than those of LSTM and Random Forest.
Limitations and Future Work
- Humidity Influence: The study acknowledges that humidity was not experimentally controlled, and its potential effect on dielectric loss (accelerating ionic conduction) was inferred from prior work. Future research should incorporate humidity sensors.
- Fixed Temperature Training: The current model was trained solely on aging data at a fixed temperature (140°C). Real-world operational conditions involve dynamic temperature variations, which could affect prediction stability.
- Future Directions:
- Conducting multi-temperature joint experiments.
- Integrating humidity sensors to optimize the model.
- Developing intelligent control strategies based on predicted aging degrees to extend harness service life (e.g., by reducing power or adjusting working modes).
- Applying the model to other fields involving connectors and loads to verify its universality.
Summary updated