Introduction
Electrical motors are critical components in industrial applications, yet their precise control remains a challenge due to non-linear dynamics and environmental factors. Traditional physics-based models often fail to account for parameter drift (e.g., resistance changes due to heat) and sensor noise. This research proposes a data-driven approach to model the relationship between external inputs—voltage and speed—and internal outputs—current and torque.
Proposed Architecture: DiagBiRNN-Skip
The core of the proposed solution is an Encoder-Decoder architecture enhanced with diagonalized recurrent skip connections. Unlike standard skip connections that simply pass features forward, this model incorporates recurrent layers (RNNs) within the skip paths to capture complex temporal patterns in the latent space.
To ensure the model is computationally efficient for industrial hardware, the authors "diagonalized" the RNN weights. By using Hadamard products instead of full matrix multiplications, they significantly reduced parameter counts without compromising predictive performance.
Innovation in Loss Functions
A significant challenge in modeling motor dynamics is the presence of high-frequency transitions. Standard Mean Squared Error (MSE) tends to prioritize smooth, average behavior, leading to poor performance during rapid signal changes. To address this, the authors introduced a Total Variation (TV) weighted loss. This function penalizes the model more heavily for errors in high-variance segments of the signal.
As shown in the evaluation, the TV-weighted MSE consistently yields lower Symmetric Mean Absolute Percentage Error (SMAPE) across various signal complexities compared to standard MSE. This innovation prevents the model from becoming "lazy" and ignoring critical high-frequency dynamics.
Experimental Results and Domain Adaptation
The model was evaluated on both high-frequency simulated data (100 hours) and noisy real-world industrial motor data.
Key performance metrics include:
- Predictive Accuracy: The DiagBiRNN-Skip model achieved a coefficient of determination ($R^2$) of 0.95 on simulated data, vastly outperforming standard CNN and LSTM baselines which often failed to capture the dynamics entirely.
- Error Reduction: It reached a SMAPE of 1.09%, compared to over 6% for the best baseline (Deep CNN).
- Transfer Learning: The model demonstrated robust domain adaptation. By pre-training on simulation data and fine-tuning on just 20% of real-world data, it successfully learned to filter sensor noise while maintaining high accuracy on physical hardware.
Conclusion
The Encoder-Decoder with Recurrent Skip Connections provides a powerful framework for modeling complex industrial dynamics. By combining CNNs for feature extraction, diagonalized RNNs for temporal latent modeling, and a complexity-aware loss function, the researchers have bridged the gap between theoretical simulation and practical real-world application. Future work will focus on real-time inference on microcontrollers and integrating these models into closed-loop control systems.
Summary updated