A team led by Qing Wei at the College of Engineering, China Agricultural University, has investigated the transformative potential of artificial neural networks (ANNs) in the drying of agricultural products. Their findings, published in Frontiers of Agricultural Science and Engineering, address significant challenges faced by the industry.
The drying of agricultural products is essential for maintaining food safety and enhancing value. Fresh produce, including grains, fruits, and vegetables, is susceptible to spoilage from excessive moisture. In China, inadequate drying practices result in losses of approximately 21 million tons of grain annually. Conventional drying methods, which often rely on manual experience and simplistic models, can lead to problems such as over-drying, which diminishes quality, or insufficient drying, which encourages mold growth, all while consuming high amounts of energy.
The introduction of artificial intelligence raises the question of whether ANNs can revolutionize this traditional field. The research conducted by Wei and colleagues highlights the innovative use of neural networks in agricultural drying, providing solutions to existing industry issues.
Traditional drying techniques often depend on mathematical formulas or physical equations that struggle to accommodate the nonlinear dynamics of drying processes. For example, during hot air drying, the interplay of temperature, air velocity, and humidity can complicate moisture removal, frequently resulting in uneven drying. Conversely, neural networks, which mimic the connectivity of human brain neurons, can learn from extensive experimental data to make accurate predictions.
In practical applications, studies indicate that neural networks significantly enhance moisture ratio predictions. For instance, in experiments with peppermint drying, conventional models produced substantial prediction errors, whereas neural network approaches achieved a coefficient of determination (R²) of 0.998, closely aligning with actual measurements. Furthermore, when applied to drying shiitake mushrooms, hawthorns, and other products, these networks can track moisture changes in real time, maintaining low prediction error margins. This capability enables drying equipment to adaptively adjust parameters, thereby avoiding issues like cracking and nutrient degradation from over-drying.
Beyond moisture removal, drying techniques also aim to preserve product quality. Traditional drying methods can result in undesirable effects such as browning in fruits and vegetables and the loss of vitamins due to poor temperature management. By integrating diverse data sources, including image recognition and sensor information, neural networks facilitate comprehensive quality control. In the case of kiwifruit drying, researchers developed a color prediction model using neural networks that optimizes temperature settings based on color changes, thereby enhancing the retention of nutritional elements.
Energy consumption poses another significant challenge in agricultural drying. Neural networks can optimize energy use and efficiency by refining drying processes. In trials involving infrared drying of garlic, researchers employing neural network models were able to reduce drying time by 6.5% and cut energy consumption by 36%.
The combination of neural networks with traditional control systems is reshaping production methodologies. For example, integrating neural networks with PID controllers allows for real-time adjustments of temperature and humidity within drying chambers, improving precision and stabilizing product quality compared to conventional manual controls.
In tobacco curing, the deep learning-based TobaccoNet model autonomously calibrates temperature and humidity settings based on images of tobacco leaves, resulting in a mere 1.62% prediction error while significantly diminishing the need for manual intervention.
Despite these advancements, challenges remain in the application of neural networks for agricultural product drying. Many existing models require substantial amounts of labeled data, and limited experimental data for newer agricultural products can hinder broader implementation. Additionally, the opaque nature of deep neural networks raises concerns about their reliability among some producers. The research team emphasizes the importance of future work in developing more streamlined and interpretable models, as well as integrating Internet of Things (IoT) technologies for real-time data collection and model updates.
