基于宽度卷积神经网络的异常农情数据检测方法

Objective: In order to accurately and effectively detect data anomalies from agricultural internet of things, one detection method of abnormal agricultural data based on broad convolution neural network (BCNN) was proposed for providing a reference for achieving high-quality data collection in agricultural internet of things. Method: Firstly, the standardized agricultural data were encoded as polar coordinates, and then they were divided into subsets by sliding window mechanism. Subsequently, the data were reconstructed as matrix format. Finally, the BCNN was designed and trained for conducting anomaly detection. The experiment was conducted using the data monitored in the culturing farm environment. Result: The sliding window mechanism could improve the detection ability and reduce the time consumption. The accuracy and F1 score of the designed BCNN in the datasets of air temperature and humidity, soil temperature and humidity were more than 97.5% and 0.985 respectively, which on average outperformed SVM, RF and CNN with the increase of 1.69%, 2.76%, 3.05% and 0.009 3, 0.0149, 0.0163, respectively. In particular, while handling the air and soil temperature and humidity data with high fluctuation, the gain in accuracy and F1 score ranged 3.61%-5.98% and 0.018 8-0.031 0, respectively. In addition, the proposed BCNN model has less time consumption of anomaly detection, only 1/6 to 1/7 of classical CNN model, and as well as with less hyperparameter. Conclusion: The proposed data preprocessing (data coding, subset partition and reconstruction) method and BCNN model exhibit better performance on abnormal agricultural data.