Abstract: Owing to the development of artificial intelligence technology, garbage classification methods based on neural networks have become important for solving waste disposal problems and have been widely applied. However, for edge devices, the upgradability of neural networks and the improvement in their accuracy require further research. Hence, this study proposes a waste classification method and system for edge devices. The system applies deep learning methods to garbage classification and identification, combines multistep knowledge distillation technology for class-incremental learning, and uses an adaptive weighted fusion algorithm to fuse image classification results and multisensor data. Specifically, to address the problem of catastrophic forgetting presented by neural networks during incremental learning, this study employs multistep knowledge distillation to perform class-incremental learning on the AlexNet network. This approach forms a teacher–assistant–student model suitable for waste classification, enabling the AlexNet network to maintain its ability to recognize old classes while learning new ones. Simultaneously, all potential distillation paths are evaluated on the CIFAR-100 dataset, and the optimal distillation path is determined. The feasibility and effectiveness of the proposed method in solving catastrophic forgetting are demonstrated by comparing different algorithms in 100 categories of category incremental learning experiments. Additionally, to address the reduced classification accuracy of the convolutional neural network when the features of the garbage image to be classified are concealed or confused, this study uses multisensor data. External weight sensors and metal detectors are connected to an embedded device to determine the weight and metal characteristics of the garbage. This study employs the Q-learning algorithm to perform a weighted fusion of the classification results from the AlexNet network with weight and metal features. The state space, action space, and reward space of the problem are defined, which prevents misclassification caused by ambiguous features and feature confusion in the classification images, thereby improving the classification accuracy. By comparing the proposed algorithm with the AlexNet algorithm and other multisensor garbage classification algorithms, an improvement in the classification accuracy is demonstrated. Finally, waste classification methods are applied to edge devices and designs, and an intelligent waste classification system is developed. Raspberry Pi 4B is selected as the embedded platform for deploying the AlexNet network. The system is equipped with modules such as a camera, metal detector, and weight sensor to obtain multimodal data, and a system workflow is established. To verify the performance of the proposed system, comparative experiments are conducted using the public dataset TrashNet and a self-developed dataset. The experimental results show that the designed classification system achieves an average classification accuracy of 89.7% in the garbage classification task of the self-developed dataset, and that the average time required for a single classification is 130 ms, demonstrating rapid and accurate garbage classification. Compared with an embedded algorithm based on MobileNetV3, the accuracy rate of the proposed system improved by 6.5%. The design and implementation of the system consider actual application scenarios, offer strong practicality and promotional value, and provide a useful approach for the application of garbage classification technology.