基于GAF-DenseNet的旋耕作业质量等级识别模型

为实现基于拖拉机多传感器实测载荷数据的旋耕作业质量准确识别，提出一种基于GAF-DenseNet的拖拉机旋耕作业质量等级识别模型，设计旋耕作业质量等级分级标准，开展旋耕作业田间试验，并进行模型准确性验证和性能分析。该模型通过格拉姆角场（Gramian angular field, GAF）算法，在保留原始载荷序列的时间依赖性的前提下，对时间序列数据进行唯一编码。DenseNet网络对图像阵列中内含的载荷信息进行深层挖掘，通过特征重用、模型压缩等技术环节，在保证特征提取深度的同时，显著提升该网络的运算效率。分析结果表明：过大或过小的重采样滑动窗口大小均会降低模型性能，且格拉姆角差场（Gramian angular difference field, GADF）实验效果强于格拉姆角和场（Gramian angular summation field, GASF），实验数据显示在重采样滑动窗口大小为250且选用格拉姆角差场的条件下，模型性能达到最优。增长率k与模型整体性能呈正相关的趋势，但过大的k值会降低模型的实时性能且对于准确性提升有限，实验场景下将增长率k设为24更能符合实际需求。GAF-DenseNet模型准确率和F1值分别达到96.816%和96.136%，并且在实时性能上具有良好表现，推理时长可低至16s。在与其他智能算法对比分析中，该模型整体性能均优于对照组实验结果。

Quality Indentification Model of Tractor Rotary Tillage Based on GAF-DenseNet

To achieve accurate prediction of rotary tillage quality based on tractor multi-sensor load data, a tractor rotary tillage quality identification model based on GAF-DenseNet was proposed, rotary tillage quality grading standard was designed, and field tests of rotary tillage were carried out, and model accuracy verification and performance analysis were conducted. The Gramian angular field (GAF) algorithm uniquely encoded the time series data while preserving the time dependence of the original load sequence. The DenseNet network deeply mined the load information embedded in the image array, and significantly improved the computing efficiency of this network while ensuring the depth of feature extraction through feature reuse, model compression, and other technical aspects. The analysis results showed that the model performance was reduced by either too large or too small a resampling sliding window size and the experimental effect of Gramian angular difference field (GADF) was stronger than Gramian angular summation field (GASF), and the experimental data showed that the model performance was optimal when the resampling sliding window size was 250 and the GADF algorithm was selected. The growth rate k tended to be positively correlated with the overall performance of the model, but too large a value of k reduced the real-time performance of the model and had limited improvement in accuracy, and the growth rate k was set to 24 in the experimental scenario to better meet the actual demand. The GAF-DenseNet model achieved accuracy and F1 value of 96.816% and 96.136%, respectively. It had good performance in real-time capability, and the interfence time can be as low as 16s. The overall performance of this model was better than the control group analysis results in the comparison tests with other intelligent algorithms.