| Abstract: | 在無農業設施下種植要達到無農藥防治、化學肥料使用是不容易達成目標，以小型人造光源植物工廠栽種蔬菜，無農藥殘留且不受天候影響隨時可栽種可重覆使用不佔空間，因此本研究目的為打造一個居家使用小型植物工廠並達到節能、標準化及兼具安全便利等，提出智慧室內植物工廠系統架構包括四個模組分別是建立、採集數據、傳輸數據和數據分析系統，加入大數據和人工智慧實現最佳植物生長環境條件，為達成上述目的以一次一因子法，利用高頻率發光二極體(Light Emitting Diode, LED)提供不同波長的光源及不同光照週期，確認有利於作物生長光源型態，其中智慧植物工廠必須不斷收集植物工廠的數據並適時更新監測參數，以獲得最佳植物生長條件。本實驗使用萵苣(Lactuca satva L.)品系幼苗，實驗箱中皆使用栽培土填滿培養盤定植供給水，每日定量噴灑於根部(500 ml/日)連續為期30天至採收完成，分別以木製植物栽培箱種植每箱中9株，首先植物箱中照射時數為24小時全日照射設置白光、藍光與紅光不同顏色LED為比較，替代陽光讓作物行光合作用；後續間歇式循環及全照射12小時給予相同紅光測試不同光週期，夜間時段(18時至5時，總照射時數12個小時)、週期2小時(總照射時數6小時)、週期4小時 (總照射時數8小時)，白天(6時至17時)則給予自然日照行光合作用；最後植物箱中額外增加排風系統，以12小時開關通氣循環及24小時進行通氣。三種不同光源全日照連續30天培育期有顯著差異，分別以紅光組(660 nm高功率LED)為最佳，其次為藍光組、白光組為最差，不同光照時間週期試驗中發現，以2小時及4小時、12小時為循環，對萵苣植株重量及高度呈現正成長，但12小時週期植株生長高度低於另兩組生長效益較低，其中提供光照4小時效率最好；植物箱中額外增加排風系統，12小時及24小時試驗其中12小時開關通氣循環對於植株生長最佳。資訊安全對於智慧植物工廠也非常重要，如收集到偽造的數據會影響機器學習性能，因而本研究使用聯邦學習技術針對每個區域內部數據進行訓練，上傳梯度損失和在區塊鏈中產生區塊，監測參數並經過模型訓練驗證及統計確認數據完整性，而不採取直接上傳原始參數的方式，達到保證隱私和數據安全目的。
It is challenging to achieve pesticide-free pest control and minimize the use of chemical fertilizers when planting without agricultural facilities. By cultivating vegetables in small-scale artificial light-based plant factories, it is possible to eliminate pesticide residue and overcome weather limitations. These factories are pesticide-free, unaffected by weather conditions, and allow for continuous cultivation and space efficiency. Therefore, the purpose of this research is to develop a small-scale plant factory for home use that promotes energy efficiency, standardization, safety, and convenience. To achieve these goals, a smart indoor plant factory system architecture is proposed, consisting of four modules: establishment, data collection, data transmission, and data analysis system. By incorporating big data and artificial intelligence, the system aims to create optimal plant growth conditions. To determine the optimal light source conditions for crop growth, a one-factor-at-a-time approach is employed. High-frequency Light Emitting Diodes (LEDs) are utilized to provide different wavelengths of light and operate on different lighting cycles, confirming the light source patterns favorable for crop growth. In the intelligent plant factory, continuous data collection from the plant factory and timely updating of monitoring parameters are essential to obtain the best plant growth conditions.In the M-PFALs experiment, 9 lettuces (Lactuca satva L.) were planted in a plant box. We use potting soil for planting and give daily water (500 ml/day) to harvest for a total of 30 days. It is mainly divided into three experiments, the different color light test, different photoperiod test, and ventilation cycle test. In different color light tests, the plants were illuminated by white, blue, and red LEDs for 24 hours instead of sunlight. Different photoperiod experiments test different night (18:00 to 5:00) cycles in red light, the control group is normal photoperiod, the full-time cycle 12-hour group, the cycle 2-hour group (total 6 hours), the cycle 4-hour group (total 8 hours), the control group was normal photoperiod. Natural light was given to sunlight for photosynthesis from 6:00 to 17:00. Ventilation cycle test, the fan was turned on for 12 hours and turned off for 12 hours in the ventilation group, and the other group was turned on for 24 hours. In different color light tests, the red light group is the best, followed by the blue light group, and the white light group is the worst. Giving red light in different photoperiod experiments was better than the control group, among which the group with a period of 4 hours had the best effect, followed by the group with a period of 2 hours, and the group with a period of 12 hours was the worst. In the ventilation cycle test, the group with fans turned on for 12 hours and turned off for 12 hours is better than the group with fans turned on for 24 hours.Information security is also crucial for smart plant factories. For instance, the collection of falsified data can affect machine learning performance. Therefore, this study employs federated learning techniques to train the data within each region. It involves uploading gradient losses and generating blocks in a blockchain, monitoring parameters, and verifying data integrity through model training, validation, and statistical confirmation. This approach avoids directly uploading raw parameters, ensuring privacy and data security. |
