Minimizing Data Transmission Volume in Smart Farm Networks
As part of the research requirements for a Modeling Multimedia Systems course I took in the spring semester of 2024, I undertook a project to apply and improve existing methods for Minimizing Data Transmission Volume in Smart Farm Networks. The objective was to find a relevant research paper in the field of multimedia, understand its methodology, and contribute to it through enhancements. I selected the paper titled "K-predictions based data reduction approach in WSN for Smart Agriculture" by Christian Salim and Nathalie Mitton, which proposed a machine learning data reduction (MLDR) method to optimize data transmissions in wireless sensor networks (WSN).
In my project, I focused on improving the predictive accuracy of the model at the sink level, which is critical for reducing unnecessary data transmissions. In this approach, sensor nodes initially transmit data to the sink, and after a trend is identified, the sink predicts future data. However, accuracy at the sink level is essential: when predictions are inaccurate, the sink halts its predictions, leading to frequent data transmissions until a new trend is learned, which increases energy consumption and inefficiency.
I targeted improvements in the prediction of two key parameters, temperature and humidity, as they are vital in smart agriculture systems. I used the hybrid model from a research paper by Sabat et al, which combined a Vector Autoregressive (VAR) model and a Gated Recurrent Unit (GRU) network to capture both linear and non-linear data patterns. To enhance the accuracy further, I introduced attention mechanism cells to the GRU model. This addition allowed the model to focus on the most relevant aspects of the time-series data, significantly boosting prediction accuracy.
I trained my model using a collection of climate data from Bengaluru, a city in India, and applied two methods: the MLDR method suggested by the paper's authors and my modified VAR-GRU hybrid model with attention mechanisms. I then compared the results to evaluate the improvement in prediction accuracy.
The results showed that my modified model achieved a 20% improvement in temperature prediction accuracy and an 11% improvement in humidity prediction accuracy compared to the original method. These improvements reduced the frequency of data transmissions and increased the overall energy efficiency of the network. This could potentially help the system predict for longer periods without the need for unnecessary data transmissions between nodes and the sink, optimizing the performance of smart farm systems.
References:
-
Salim, Christian & Mitton, Nathalie. (2021). K-Predictions Based Data Reduction Approach in WSN for Smart Agriculture. Computing. 10.1007/s00607-020-00864-z.
-
Sabat, Naba & Nayak, Rashmiranjan & Pati, Umesh & Das, Santos. (2023). A Deep Learning-Based Vector Autoregressive-Gated Recurrent Unit Hybrid Model for Long-Term Forecasting of Weather Parameters for Smart Farms. 10.4018/978-1-6684-8516-3.ch009.