High-frequency Inductive Power Transfer (IPT) is a key technology for wireless charging, offering advantages like increased transmission distance and reduced system size. The performance of these systems is critically dependent on the inductance (L) and quality factor (Q) of the transmitting and receiving coils. Traditional measurement methods using impedance or network analyzers are expensive, bulky, and impractical for sealed products. Simulation-based approaches, while useful, become computationally prohibitive at very high frequencies or for complex coil geometries due to skin and proximity effects.
This paper pioneers a machine learning (ML) solution to this identification problem. By inputting only an image of the coil and its operating frequency into a trained model, the system can rapidly and accurately predict L and Q values. This method is portable, non-invasive, and eliminates the need for costly hardware or disassembly.
This section outlines the fundamental structure of IPT systems and analyzes the critical role of coil parameters.
A typical IPT system consists of an inverter, primary and secondary compensation networks, and loosely coupled transmitter (Ltx) and receiver (Lrx) coils. The inverter generates a high-frequency AC current, which is conditioned by the primary compensation network before flowing through Ltx. Energy is transferred wirelessly to Lrx, then conditioned by the secondary network for delivery to the load.
The inductance L determines the resonant frequency and coupling characteristics, while the quality factor Q, defined as $Q = \frac{\omega L}{R}$, where $\omega$ is the angular frequency and $R$ is the equivalent series resistance, directly impacts system efficiency and the ability to achieve Zero Voltage Switching (ZVS). High Q is essential for minimizing losses, especially in the MHz frequency range. An experimental prototype operating at 6.78 MHz was established to study these effects.
The core innovation is the application of a Convolutional Neural Network (CNN) for visual parameter regression.
A CNN architecture was chosen for its proven efficacy in extracting spatial features from images. The model takes a coil image and the operating frequency as inputs. The image passes through convolutional layers for feature extraction (edges, shapes, turns), followed by pooling layers and fully connected layers that integrate the frequency data to regress the final L and Q values.
A diverse dataset was crucial for model robustness. It included images of coils with and without ferromagnetic cores, coils with excitation wires of varying thicknesses, and coils of different shapes (e.g., spiral, solenoid). This variety ensures the model can generalize across a wide range of real-world coil designs.
Identification Error Rate: 21.6%
This error rate represents the model's performance in predicting L and Q values across the test dataset. While not perfect, it demonstrates a significant proof-of-concept, offering a rapid, low-cost alternative to traditional methods. The error is likely attributable to limitations in dataset size, image resolution, and the inherent complexity of mapping visual features to precise electrical parameters.
Chart Description: While not explicitly detailed in the provided text, a typical results section would include charts such as: 1) A scatter plot of Predicted L vs. Measured L, showing correlation and error distribution. 2) A similar plot for Predicted Q vs. Measured Q. 3) A bar chart comparing the time taken for ML identification versus simulation (e.g., HFSS) or physical measurement, highlighting the speed advantage of the ML approach.
The problem is framed as a supervised regression task. The model learns a function $f$ that maps input features to target parameters:
$[\hat{L}, \hat{Q}] = f(I_{coil}, f_{operation}; \theta)$
where $I_{coil}$ is the coil image tensor, $f_{operation}$ is the operating frequency, $\theta$ represents the trainable parameters (weights and biases) of the CNN, and $\hat{L}, \hat{Q}$ are the predicted values.
The loss function used during training is typically a Mean Squared Error (MSE) or Mean Absolute Error (MAE) between predictions and ground-truth values obtained from traditional measurements:
$\mathcal{L}(\theta) = \frac{1}{N} \sum_{i=1}^{N} \left( (L_i - \hat{L}_i)^2 + \alpha (Q_i - \hat{Q}_i)^2 \right)$
where $N$ is the batch size and $\alpha$ is a weighting factor to balance the scale difference between L and Q.
Non-Code Analysis Framework: Consider a quality control scenario in a wireless charger manufacturing line.
This framework transforms a complex electrical test into a simple visual inspection, drastically reducing testing time and cost.
This paper isn't just about measuring coils; it's a strategic pivot from physics-first to data-first in power electronics design and validation. The authors correctly identify that the bottleneck in high-frequency IPT isn't theoretical understanding but practical parameter extraction. By treating the coil as a visual pattern rather than an electromagnetic boundary-value problem, they bypass the computational tyranny of Maxwell's equations at MHz frequencies. This is reminiscent of how computer vision bypassed explicit feature engineering. The 21.6% error isn't a weakness—it's the price of admission for a paradigm that promises order-of-magnitude reductions in testing time and cost.
The argument is compellingly linear: 1) High-frequency IPT is vital but hard to characterize. 2) Existing tools (analyzers, simulators) are either expensive, slow, or intrusive. 3) Therefore, we need a new, agile method. 4) Machine learning, specifically CNNs proven on ImageNet, offers a path. 5) Here's our proof-of-concept model and dataset. 6) It works with reasonable error. The logic is sound, but the leap from "image" to "inductance" is glossed over. The model is essentially learning a highly non-linear proxy for electromagnetic simulation—a fascinating but black-box approach that would give traditionalists pause.
Strengths: The practicality is undeniable. The method is brilliantly simple in concept—just snap a picture. The use of a diverse dataset (with/without cores, various shapes) shows good foresight for generalization. Aligning with the trend of physics-informed machine learning, they incorporate the operating frequency as a direct input, injecting crucial domain knowledge into the model.
Flaws: The 21.6% error rate, while a start, is far from production-ready for precision applications. The paper is silent on error breakdown—is the error in L or Q? Is it consistent or does it fail catastrophically on certain coil types? The "image" input is vague—what resolution, lighting, angle? As with many ML applications, the model's performance is shackled to its training data. It will likely fail on coil geometries or materials not represented in its dataset, a limitation not faced by fundamental physics simulators like ANSYS HFSS. There's also no discussion of uncertainty quantification—a critical need for engineering decisions.
For researchers: Double down on hybrid models. Don't just use a pure CNN. Use it to predict initial geometry parameters (turn count, diameter), then feed those into a fast, simplified analytical model (e.g., based on Wheeler's formulas) to calculate L and Q. This adds interpretability and physics constraints. For industry: Pilot this for go/no-go quality testing, not for precision design. The cost savings from rapid screening of defective units will justify the investment even with the current error rate. Start building your proprietary dataset of coil images and measured parameters now; that data asset will be more valuable than any single model. Finally, engage with the computer vision community. Techniques from few-shot learning and domain adaptation, as seen in advanced GAN architectures like CycleGAN, could be key to making the system robust to real-world visual variations.
In conclusion, this work is a provocative and necessary step. It doesn't solve the coil identification problem, but it successfully reframes it in a way that opens the door for data-driven acceleration. The future belongs not to the method with the lowest error in a lab, but to the one that delivers "good enough" answers fastest and cheapest on the factory floor. This paper points squarely in that direction.