薄膜阴极极片是新能源汽车中锂电池的关键核心组件，严重的极片缺陷会引发电池短路，导致汽车自燃，威胁消费者的生命和财产。基于机器视觉的锂电池极片缺陷的高效与高精度检测符合国家新能源发展战略，具有研究价值和实际工程意义。目前，极片缺陷的智能检测尚存在缺陷数据获取困难，数据标注成本高的难题，制约了深度学习方法的大规模应用。主要内容如下：首先，针对极片缺陷数据获取困难和标注成本高的问题，提出了通过物理渲染引擎生成极片缺陷的半合成数据方案，可精确指定缺陷生成参数，合成缺陷数据符合均匀分布，避免了如边缘模糊缺陷等样本不均衡导致的长尾分布问题。该方法能够自动生成具有一致性和客观性的标签，显著减少了手动标记大型数据集所需的时间。该方案可生成实际场景中尚未出现的缺陷样本，充分利用真实数据集和合成数据集，解决了深度学习训练方法在极片缺陷数据稀缺时面临的难题。其次，提出基于半合成数据集进行缺陷分割任务的网络预训练方案，无需在训练过程中使用真实数据，采用合成样本上训练的模型应用于真实数据，在手动标记的验证集上实现了平均Dice 得分0.77，平均IOU 达到0.66，验证了方案的有效性。此外，合成数据集上预先训练模型加入真实数据后仅需模型参数微调，显著减少真实场景下模型训练的时间。最后，设计了端到端树提升系统XGBoost 实现了样本级多特征融合的二元分类算法。针对本文分割模型特征图和模板图像的斑点分析值计算特征向量，在多个特征向量上训练XGBoost 模型，通过启发式粒子群优化（PSO）迭代优化寻找最优解，对模型均方误差最小化实现了二元分类。本文提出的方案在真实极片缺陷检测场景测试集上实现了93.4% 的准确率，关键指标查准率、查全率和准确率均显著优于商业软件Aurora Vision Studio，整体实施方案切实有效，易于部署，有望解决新能源行业中锂电池极片缺陷检测的行业共性问题。
Thin-film cathodes, also known as pole pieces, are crucial components of lithium-ionbatteries in new energy vehicles. Defects in pole pieces can result in battery short-circuits,potentially leading to vehicle fires and endangering consumer lives and property. Efficientand precise detection of pole piece defects through machine vision aligns with nationalnew energy development strategies and holds both research and engineering significance.The efficient and effective detection of pole piece defects still faces challenges with regardto acquiring defect samples and high data annotation costs, limiting the large-scale applicationof deep learning methods. This thesis addresses these issues through the followingcontributions:Firstly, a semi-synthetic data generation method is proposed, utilizing a physical renderingengine to tackle the challenges of acquiring pole piece defect data and the highcosts of manual annotation. This approach can accurately specify defect generation parameters,synthesize defect data conforming to a uniform distribution, avoids long-taildistribution problem, and ambiguous annotation of defects with unclear boundaries. Themethod can automatically generate consistent and objective labels, significantly reducingthe time needed for manual annotation of large data sets. The method also generatesdefect samples not yet encountered in real-world scenarios, leveraging both real and syntheticdata sets to overcome the challenges faced by deep learning training methods withlimited pole piece defect data.Secondly, a network pre-training method for defect segmentation tasks based on semisyntheticdata sets is introduced, eliminating the need for real data during the trainingprocess. The model, trained on synthetic samples, is applied to real data and achievesa mean Dice score of 0.77 and a mean IOU of 0.66 on manually annotated validationset, demonstrating the method’s effectiveness. Furthermore, the model pre-trained onsynthetic data sets requires only fine-tuning of model parameters when incorporating realdata, substantially reducing model training time in real-world scenarios.Lastly, an end-to-end tree boosting system, XGBoost, is utilized to implement a multifeaturefusion binary classifier. Feature vectors are calculated based on blob analysisvalues from the segmentation model’s feature map and template image. The XGBoostmodel is trained on multiple feature vectors, and the optimal classifier parameters are found through iterative optimization using heuristic particle swarm optimization (PSO),minimizing the model’s mean square error for binary classification.The proposed method achieves a 93.4% accuracy on the test set, with key performancemetrics such as precision, recall, and accuracy significantly outperforming the commercialsoftware Aurora Vision Studio. The overall implementation scheme is effective, easy todeploy, and has the potential to solve the task of lithium-ion battery pole piece defectdetection in the new energy industry.
