Normal production processes will be substantially impacted by industrial devices in abnormal working conditions. Anomaly sound detection (ASD) model can monitor the working condition of devices by the non-contact and non-invasive way. When new device data is introduced, traditional ASD models are trained using data from all devices, to accommodate every device. However, in real-world settings, the kinds and amounts of devices are constantly changing, which raises difficulties for the current ASD models. This paper proposes a teacher-student incremental learning method for ASD models, aiming to solve ASD model scalability problem. In this paradigm, teacher model has knowledge of all the old devices. The objective of student model is to learn new device knowledge, while avoiding the forgetting of old device knowledge. When student model learns new device data, teacher model transfers the acoustic feature knowledge of old devices to student model via knowledge distillation. Furthermore, the imbalance between old and new knowledge causes challenges, such as knowledge forgetting or lower learning efficiency for student model. This paper presents a dual-teacher-student (DTS) model to solve the problem of knowledge imbalance. Different teacher models for new and old devices in DTS, directing student model to accomplish continuous and deep integration of knowledge. Evaluation for proposed method on the DCASE 2020 Task2 dataset. The results show, the proposed method outperforms other methods in terms of learning capability and robustness during the incremental learning process. Analysis of significance test on the experimental results demonstrates that the method outperforms other methods statistically.