齿带式残膜回收机捡拾机构故障监测系统设计与试验

Design and Experiment of Fault Monitoring System for Pickup Mechanism of Tooth‑belt Type Residual Film Recycler

摘要:

针对齿带式残膜回收机在作业过程出现收膜皮带断裂、捡拾钉齿脱落等典型故障问题，设计了一种捡拾机构故障监测系统。以STM32为下位机数据采集核心，采用Qt框架开发上位机监测界面，为增强田间抗干扰能力，利用RS485总线进行数据通信。结合捡拾钉齿周向排布结构，选取脱膜辊转速和相邻钉齿通过传感器时间间隔作为关键监测参数。通过对正常工况下样本数据的筛选与多项式、指数等多类函数拟合对比，正常工况下建立转速与理论时间间隔的二阶指数衰减函数最优模型，引入自然指数函数计算监测值与理论值相对偏差，以增强数据特征差异。将归一化处理后的转速、时间间隔及其相对偏差作为特征向量，输入经超参数优化的多层感知器（Multilayer perceptron，MLP）进行分类识别，并由Softmax函数输出故障状态，整体模型训练识别准确率达98.65%。模型对比试验结果表明MLP性能优于支持向量机与随机森林模型;台架验证试验中，系统平均识别准确率为96.13%。研究结果可为齿带式残膜回收机故障监测系统开发提供理论和技术参考。

Abstract:

Aiming to address typical operational failures such as belt breakage and pickup tooth detachment in film?recovery machines, a fault monitoring system for the pickup mechanism was designed. This system employed an STM32 microcontroller as the data acquisition core and utilized the Qt framework to develop the upper?level monitoring interface. To enhance field interference resistance, RS485 bus communication was adopted for data transmission. Considering the circumferential arrangement of the pick?up teeth, the film?stripping roller speed and the time interval between adjacent teeth passing sensors were selected as key monitoring parameters. Through screening sample data under normal operating conditions and comparing polynomial, exponential, and other function fits, an optimal second?order exponential decay function model was established for the relationship between rotational speed and theoretical time interval values under normal conditions. The natural exponential function was introduced to calculate the relative deviation between monitored and theoretical values, enhancing data feature differentiation. Normalized rotational speed, time interval values, and their relative deviations were used as feature vectors input into a hyperparameter?optimized multilayer perceptron (MLP) for classification. The Softmax function outputted fault states, achieving an overall model training recognition accuracy of 98.65%. Model comparison and bench validation tests demonstrated that the MLP outperformed support vector machines and random forest models. Bench validation achieved an average system recognition accuracy of 96.13%. These findings can provide theoretical and technical references for developing fault monitoring systems for gear?belt residual film recovery machines.