Abstract:  When the stress and strain of the loose tooth rake, the key device of the drum film recovery machine, is monitored in real time, the obtained stress and strain signal is easy to be interfered by the external environment and it is difficult to identify the back‑up fault from the signal. In order to solve this problem, the strain monitoring position of the pine tooth harrow was determined by ANSYS analysis, and the strain gauge was used to carry out strain monitoring tests on the pine tooth harrow in different working conditions. Based on the monitoring data, a condition recognition method of variational mode decomposition (VMD)-BiLSTM neural network model based on arithmetic optimization algorithm (AOA) was proposed. Firstly, the parameters of k value and penalty factor α of VMD modal component were optimized by AOA. Then, VMD was used for adaptive decomposition of strain signal of pine tooth harrow. Finally, according to Pearson coefficient, the decomposed and reconstructed signals were input into BiLSTM network for feature learning, so as to realize the condition recognition of the pine tooth rake. The results show that the method can accurately recognize 4 kinds of working conditions such as no‑load, normal working conditions, slight back‑up and severe back‑up, and the effect is better than VMD-LSTM, BiLSTM and LSTM neural network models, with the recognition accuracy of more than 99.1%, which effectively improves the recognition accuracy of working conditions of the pine tooth harrow.

Key words: soil loosening device, condition identification, arithmetic optimization algorithm(AOA), variational mode decomposition (VMD), BiLSTM

摘要： 当对滚筒式残膜回收机的关键装置松土齿耙的应力应变进行实时监测时，所获得应力应变信号易受外部环境的干扰，难以从信号中识别壅土故障。针对该问题，通过ANSYS分析确定松土齿耙的应变监测部位，利用应变片对松土齿耙不同工况进行应变监测试验。基于监测数据，提出一种基于算数优化算法（AOA）的变分模态分解（VMD）—双向长短期记忆网络（BiLSTM）神经网络模型工况识别方法。首先，利用AOA对VMD模态分量的k值和惩罚因子α进行参数优化；然后，使用VMD对松土齿耙应变信号进行自适应分解；最后，根据皮尔逊系数将分解并重构后的信号输入BiLSTM网络中进行特征学习，实现松土齿耙的工况识别。结果表明，该方法实现对松土齿耙空载、正常工作、轻度壅土、严重壅土4种工况精准识别，且效果优于VMD-LSTM、BiLSTM、LSTM神经网络模型，识别准确率达到99.1%以上，有效提高松土齿耙工况识别的准确率。

关键词: 松土装置, 工况识别, 算数优化算法, 变分模态分解, 双向长短期记忆网络