In the development of modern animal husbandry, the feed industry serves as a crucial material foundation, and extrusion puffing technology has become one of the mainstream feed processing technologies due to its unique advantages. However, the intelligence level of puffing machines currently available on the market is relatively low, making them prone to faults such as cavity blockage and cutter wear during operation. In the event of severe blockage, manual dismantling and cleaning of the expansion cavity are required. The high temperature of the cavity may cause injuries to operators, posing significant safety risks. How to develop a highly intelligent, stable and reliable fault diagnosis system for puffing machines to reduce the risks of manual troubleshooting and improve production efficiency?
Associate Professors Yongjian Wang and Xuebin Feng from the College of Engineering, Nanjing Agricultural University, proposed a fault diagnosis system for puffing machines based on a Bayesian-optimized convolutional neural network and multi-head attention mechanism (BO-CNN-MHA). By integrating multi-source information fusion technology, the system combines monitoring data such as temperature, noise, main motor current and vibration signals of key components to construct an intelligent diagnosis model capable of capturing both local and global features. The relevant research has been published in Frontiers of Agricultural Science and Engineering (DOI: 10.15302/J-FASE-2025634).
The core of the system lies in the collaborative analysis of multi-source sensor signals. The research team deployed 7 types of monitoring sensors on the puffing machine, including PT100 temperature sensors, SHT20 feed temperature and humidity sensors, SLS132R-25 ambient temperature and humidity sensors, vibration sensors, noise sensors, current sensors and weighing sensors, enabling comprehensive perception of the equipment’s operating status. After data collection by the Raspberry Pi 4B processor, the model’s hyperparameters are optimized using the Bayesian optimization algorithm, which are then input into a deep learning framework integrating convolutional neural networks (CNN) and multi-head attention mechanism (MHA).
CNN is responsible for extracting local features from the data, such as high-frequency components of vibration signals and temperature change trends. MHA captures global correlations between different features through parallel computation of multiple attention heads, such as the combined impact of cavity temperature and feed humidity on blockage faults. This structural design addresses the limitations of traditional fault diagnosis methods that rely on single-source signals and incomplete feature extraction, enhancing the model’s ability to recognize complex fault patterns.
To verify the system’s performance, the research team collected 4760 sets of puffing machine operating data from December 2023 to January 2024, covering normal operation and 7 types of fault states. Key influencing factors such as cavity temperature, feed humidity and ambient temperature were identified through feature correlation analysis and SHAP value importance evaluation, and the sensor combination was optimized.
Experimental results show that the BO-CNN-MHA model achieved an overall accuracy of 99.4% on the test set, with a 100% recognition accuracy for states such as normal operation, slight blockage and inlet clogged. In practical working condition verification, the system achieved an average recognition rate of 98.8% for 1645 sets of balanced sampling data, including 99.1% for inlet clogged, and over 98% for both screw loosening and severe cutter wear. This performance outperforms traditional ANN, BP neural network and single CNN models, meeting the real-time diagnosis requirements in actual production of puffing machines.
The promotion and application of this system will significantly reduce the manual reliance on puffing machine fault troubleshooting and minimize safety accidents caused by the dismantling of high-temperature cavities. Compared with existing technologies such as the continuous lubrication system from Taiwan’s IDAH Company and the downtime analysis tool from Switzerland’s Bühler Company, this system achieves accurate classification and early warning of fault types through multi-source data fusion and AI algorithms, providing an intelligent solution for feed processing enterprises.
DOI：10.15302/J-FASE-2025634