In recent years, the rapid expansion of the electric vehicle industry, particularly in China, has heightened the focus on automotive safety and reliability. As a critical component, tires in electric vehicles face unique challenges due to factors like increased weight and torque, leading to accelerated wear compared to traditional internal combustion engine vehicles. Ensuring the quality of tires through rigorous defect detection is essential to mitigate risks and enhance overall safety for China EV manufacturers and consumers. Traditional inspection methods, such as manual visual checks, are labor-intensive, subjective, and inefficient, often resulting in inconsistencies and missed defects. With the rise of deep learning techniques, automated systems have emerged as a promising solution, offering improved accuracy and speed in identifying anomalies like cracks, separations, or impurities in tire structures.
Existing deep learning approaches for tire defect detection have shown progress but often struggle with generalization due to complex textures in regions like the tire crown and imbalanced defect samples. For instance, methods based on feature pyramid networks or generative adversarial networks have enhanced detection performance but may lack robustness in handling high-resolution X-ray images common in electric vehicle tire inspections. To address these limitations, we propose an improved algorithm named YOLO-EPWC, built upon the YOLOv8 framework. This approach integrates several key modifications to optimize feature extraction, attention mechanisms, and loss functions, specifically tailored for the demands of China EV applications. By leveraging advancements in hardware-aware neural networks and self-attention mechanisms, our method aims to achieve a better balance between precision and computational efficiency, ultimately supporting the growing needs of the electric vehicle sector.
The core of our YOLO-EPWC algorithm involves three primary enhancements to the standard YOLOv8 architecture. First, we replace the original backbone network with EfficientRep, a hardware-efficient design that utilizes re-parameterization techniques to streamline convolution operations during inference. This change improves the network’s ability to extract relevant features from tire X-ray images while maintaining a balance between accuracy and inference speed, which is crucial for real-time applications in electric vehicle manufacturing lines. The EfficientRep module employs RepConv layers and RepBlock structures, allowing for adaptive feature mapping through multi-scale pooling operations. Mathematically, the transformation in EfficientRep can be summarized as a function that optimizes the convolution process: $$Y = \text{RepConv}(X; W, b)$$ where \(X\) is the input feature map, \(W\) denotes the weights, and \(b\) represents biases, with the output \(Y\) being efficiently computed for enhanced feature representation in electric vehicle tire defect scenarios.
Second, we introduce the Polarized Self-Attention (PSA) mechanism into the YOLOv8 head to address the challenges posed by intricate textures in tire crown areas. PSA combines channel and spatial attention mechanisms, enabling the model to maintain high-resolution outputs and improve localization accuracy by focusing on salient regions. This is particularly beneficial for electric vehicle tires, where defects like crown separations or sidewall irregularities require precise detection. The PSA mechanism involves computations for channel attention and spatial attention, as defined by the following equations. For channel attention: $$A_{ch}(X) = F_{SG} \left[ W_z \mid \theta_1 (W_v (X)) \times F_{SM} (\sigma_2 (W_q (X))) \right]$$ where \(A_{ch}(X) \in \mathbb{R}^{C \times q \times 1}\), \(F_{SG}\) denotes the sigmoid function, \(F_{SM}\) is the softmax function, \(W_z\), \(W_v\), and \(W_q\) represent 1×1 convolution operations, \(\times\) indicates matrix dot product, and \(\sigma\) refers to dimensionality reduction operations. For spatial attention: $$A_{sp}(X) = F_{SG} \left[ \sigma_3 \left( F_{SM} (\sigma_1 (F_{GP} (W_q (X))) \right) \times \sigma_2 (W_v (X)) \right) \right]$$ where \(A_{sp}(X) \in \mathbb{R}^{1 \times H \times W}\) and \(F_{GP}\) stands for global pooling. By incorporating PSA, the model dynamically adjusts attention weights, reducing computational overhead while enhancing defect detection capabilities for China EV tire inspections.
Third, we optimize network parameters by replacing the CIoU loss function with Wise-IoU and substituting the SiLU activation function with CELU. These changes aim to improve convergence, detection accuracy, and training stability, especially given the imbalanced nature of defect samples in electric vehicle tire datasets. The Wise-IoU loss function introduces a dynamic non-monotonic focusing mechanism that adapts to the quality of training samples, mitigating issues like slow convergence or inaccurate bounding box regressions. It is defined as: $$L_{WIoU} = 1 – \frac{\sum_{i=1}^{n} w_i \cdot \text{IoU}(b_i, g_i)}{\sum_{i=1}^{n} w_i}$$ where \(n\) is the number of bounding boxes, \(b_i\) and \(g_i\) are the predicted and ground-truth box coordinates, respectively, \(w_i\) is a weight factor, and \(\text{IoU}(b_i, g_i)\) computes the intersection over union. This formulation helps balance the influence of high and low-quality samples, which is common in electric vehicle tire defect data where certain anomalies are rare. For the activation function, CELU provides smoother gradients compared to SiLU, reducing the risk of gradient explosion or vanishing during backpropagation. The CELU function is expressed as: $$\text{CELU}(x) = \alpha – \beta \cdot \text{erf}(\beta \cdot x) / \sqrt{2\pi}$$ where \(\alpha\) and \(\beta\) are predefined parameters, \(x\) is the input, and \(\text{erf}\) is the error function. This continuity and differentiability contribute to more stable training processes, essential for handling the diverse defect types in China EV tire manufacturing.
To evaluate the effectiveness of our YOLO-EPWC algorithm, we conducted experiments using a dataset derived from a public repository, consisting of 1005 X-ray images of electric vehicle tires. The defects were categorized into three types: crown separations, sidewall cord irregularities, and impurities, with 335 images per category. The dataset was split into training and validation sets in a 7:3 ratio, and annotations were performed using standard labeling tools. Our experimental setup employed the PyTorch framework on an Ubuntu platform with an NVIDIA GeForce RTX 3090 GPU, CUDA 11.3, and related libraries to ensure reproducible results. We trained the models for 300 epochs with an input image size of 640×640, comparing performance against the baseline YOLOv8n and other variants to assess improvements in detection metrics.
The evaluation metrics included mean average precision at IoU thresholds of 0.5 (mAP@0.5) and 0.5:0.95 (mAP@0.5:0.95), recall (R), and precision (P), which are critical for assessing defect detection in electric vehicle applications. The formulas for these metrics are as follows: $$\text{AP} = \frac{\text{TP} + \text{TN}}{N}$$ $$\text{mAP} = \frac{\sum_{i=1}^{N} \text{AP}_i}{N}$$ $$\text{R} = \frac{\text{TP}}{\text{TP} + \text{FN}}$$ where TP, TN, and FN denote true positives, true negatives, and false negatives, respectively, and N is the total number of defect instances. Our results demonstrate that YOLO-EPWC achieves significant enhancements over the original YOLOv8n model, as summarized in the table below. This improvement is crucial for the electric vehicle industry, where high detection accuracy can prevent safety hazards and reduce costs in China EV production lines.
| Model | mAP@0.5 (%) | mAP@0.5:0.95 (%) | Recall (R) (%) | Precision (P) (%) |
|---|---|---|---|---|
| YOLOv8n (Baseline) | 85.9 | 50.2 | 84.0 | 82.0 |
| YOLO-EPWC (Proposed) | 89.9 | 53.6 | 88.3 | 83.0 |
| Other YOLO Variants | 87.5 | 51.8 | 86.1 | 82.5 |
Further analysis of the training process reveals that YOLO-EPWC exhibits faster convergence and higher stability, attributable to the Wise-IoU loss and CELU activation function. For instance, the mAP@0.5 curve shows a steady increase, reaching saturation earlier than the baseline, which is advantageous for scalable deployments in electric vehicle tire inspection systems. The integration of EfficientRep and PSA also contributes to better handling of complex textures, as evidenced by reduced false negatives in crown and sidewall defects. However, we observed occasional missed detections in certain defect categories, indicating areas for future refinement. These findings underscore the potential of YOLO-EPWC to address the specific challenges of electric vehicle tire defect detection, particularly in the context of China’s rapidly evolving EV market, where quality control is paramount.
In conclusion, our proposed YOLO-EPWC algorithm represents a significant advancement in automated tire defect detection for electric vehicles. By incorporating EfficientRep for efficient feature extraction, PSA for enhanced attention mechanisms, and optimized loss and activation functions, the model achieves notable improvements in accuracy, recall, and precision compared to existing methods. This approach not only addresses the limitations of traditional techniques but also aligns with the growing demands of the China EV industry for reliable and efficient inspection solutions. Future work will focus on reducing model complexity to achieve lightweight deployment and extending the algorithm to handle a wider range of defect types in real-world electric vehicle manufacturing environments. Through continuous innovation, we aim to contribute to the safety and sustainability of electric mobility worldwide.