基于ELM神经网络的果品冷链乙烯监测校准模型与验证

冷链环境监测对于维持易腐果品品质安全至关重要,乙烯是关键监测要素之一。然而,现有果品冷链乙烯监测设备较少考虑与温湿度之间互作作用,影响监测精度和应用效果。该研究提出了一种基于ELM神经网络的乙烯监测校准模型并在多要素监测设备中验证。首先,以乙烯电化学传感器固有电压信号为基础,综合考虑温湿度变化影响,构建ELM神经网络乙烯监测校准模型,实现乙烯监测自适应校准;其次,以乙烯校准模型为核心,集成相关传感器、微控制器等模块,引入LoRa技术,开发果品冷链环境多要素监测设备;最后,以监测设备为载体,进行ELM模型离线测试和实际场景多要素监测性能验证。结果表明,该模型乙烯校准均方根误差达0.30 μL/L,平均训练耗时0.062 5 s,有效提高了动态环境下乙烯自适应监测性能;同时,该设备在冷链实际多要素环境中温度、相对湿度、乙烯浓度监测均方根误差达0.46 ℃,1.65%,1.11 μL/L,可以满足果品冷链环境多要素监测精度需求。研究成果对于精准控制冷链环境、准确预测果品品质有指导意义。

Ethylene monitoring calibration model and verification for fruit cold chain based on ELM neural network

China has been the largest producer of fruits and vegetables in the world, particularly with an output of over 1 billion tons in 2020. However, the annual loss rate of fruits and vegetables has been up to 15% (only 5% in developed countries). The cold chain can effectively maintain the quality of perishable products for food safety, thereby reducing the process losses under the low-temperature environment. The cold chain environmental monitoring has also been essential to predict the fruit quality, thus regulating the low-temperature environment. Furthermore, ethylene is one of the key monitoring elements to maintain the quality and safety of perishable fruits. However, the existing ethylene monitoring device in a fruit cold chain rarely considers the intelligent interaction with temperature and humidity, leading to low monitoring accuracy and application. Therefore, this research aims to propose an ethylene monitoring calibration model using Extreme Learning Machine (ELM) neural network, and then to verify in a multi-element monitoring device. The temperature, relative humidity, and electrochemical ethylene sensors were integrated to realize the multi-element perception under the cold chain environment. In addition to the inherent voltage signal, the temperature and humidity data of the electrochemical ethylene sensor were introduced as the inputs to construct an ELM neural network ethylene calibration model with a higher learning speed and stronger generalization, thereby enhancing the accuracy and applicability of ethylene online monitoring under a dynamic environment. The ELM neural network ethylene calibration model was integrated into the fruit cold chain environmental multi-element monitoring device, further to calibrate the ethylene sensor data online and verify the environmental multi-element monitoring performance. Meanwhile, a Long Range Radio (LoRa) technology was used to ensure the long-distance, low-cost data transmission in the complete fruit cold chain, including pre-cooling, cold storage, cold chain transportation, and sales trains. The ELM neural network ethylene calibration model was trained at the temperatures of 0, 2, 4, 6, 8 and 12 ℃, compared with the BP model. Further, the developed multi-element monitoring device was used to verify the multi-element monitoring performance in the actual cold storage environment of the fruits. The results showed that the calibration Root Mean Square Error (RMSE) of the ELM neural network ethylene calibration model reached 0.30 μL/L, and the average training time for five training sessions of the ELM model was 0.062 5 s, indicating the better adaptive monitoring performance of ethylene in a dynamic environment. The monitoring RMSE values of temperature element, humidity, and ethylene were 0.46 ℃, 1.65%, and 1.11 μL/L, respectively, fully meeting the actual demand on the accurate monitoring of multi-element for the fruit cold chain environment. The finding can offer a great contribution to accurately conntrolling the cold chain environment, and then predicting the fruit quality, particularly for the effective decision-making on cold chain management.