Intelligent Manufacturing Quality Inspection Technology Based on Deep Learning Algorithms: Exploring a Path to Enhance New Productivity
TL;DR
The results indicate that self-supervised pretraining and multi-level feature fusion effectively improve feature representation in complex manufacturing scenarios and, at the same time, computational cost is reduced without sacrificing detection accuracy.
Abstract
Traditional quality inspection methods in intelligent manufacturing often suffer from low efficiency, limited accuracy, and poor generalization to real manufacturing environments. This study aims to develop an intelligent quality inspection method with high accuracy and robust generalization capability. A deep learning model that integrates Convolutional Neural Networks (CNNs) with selfsupervised learning is proposed. CNNs are used to extract key visual features from manufacturing processes. Self-supervised learning is applied for pretraining on unlabeled data to capture latent process representations. In addition, a multi-layer architecture is designed to jointly model spatial and temporal features. An ensemble learning strategy is further introduced to fuse multimodal information and enhance overall detection performance. Experiments were conducted on three quality inspection tasks: printed circuit board solder joints, ceramic substrates, and flexible packaging. The proposed method achieved detection accuracies of 98.1%, 97.3%, and 95.8%, respectively. At a mean average precision (mAP) of 0.5, the convergence speed improved by an average of 33.6%. The inference latency on the same hardware platform was approximately 22 ms. After INT8 quantization, the performance retention reached 98.9%. The results indicate that self-supervised pretraining and multi-level feature fusion effectively improve feature representation in complex manufacturing scenarios. At the same time, computational cost is reduced without sacrificing detection accuracy. This enhances the practicality of the method in real industrial environments. The proposed quality inspection method combining CNNs and self-supervised learning outperforms ResNet50 and traditional detection algorithms in terms of accuracy, stability, and computational efficiency. It provides a practical and reliable solution for intelligent manufacturing quality inspection.