基于数据挖掘的电子皮带秤皮带跑偏检测

为提高电子皮带秤连续累计称量精度,针对严重影响精度的电子皮带秤跑偏,采用对皮带秤现有原始传感器的数据挖掘实现跑偏量实时在线检测,以取代传统硬件检测设备。引入流形学习和深度学习,分别提出了基于局部切空间排列(local tangent space alignment,LTSA)+广义回归神经网络(generalized regression neural networks,GRNN)和基于连续深度置信网络(continuous deep belief networks,CDBN)的在线跑偏特征提取模型,再结合极限学习机(extreme learning machine,ELM)以跑偏特征为模型输入进行跑偏量预测。最后通过试验对该文提出的在线跑偏量预测模型的性能进行了验证:LTSA+GRNN+ELM平均跑偏预测精度为93.33%,平均每组预测时间38.29 ms;CDBN+ELM预测精度则高达98.61%,平均每组预测时间1.47 ms。二者预测精度和实时性皆表明能取代传统硬件检测装置,为皮带跑偏检测提供了一种方法,为进一步的皮带秤在线精度补偿和故障预测提供了必要依据。

Detection of belt deviation of belt weigher using data mining

Abstract: At present, belt weigher has been widely used in various transportation and trade occasions of industry and agriculture. Belt deviation is one of the most important indicators of accuracy of belt weigher, and it is also one of the most common faults. In this paper, aiming at the problem of belt deviation, we obtained the real-time online detection of deviation by the data mining based on the existing original sensor data of belt weigher, instead of traditional hardware testing equipment in which CCD, PSD and array phototransistor are usually used as the specialized sensor for detecting deviation. At first, in order to reduce the dimension of existing original data and the complexity of the subsequent detection mode of belt deviation, the online features extraction models of belt deviation based on LTSA (Local Tangent Space Alignment) + GRNN (Generalized Regression Neural Networks), and CDBN(Continuous Deep Belief Networks) were proposed respectively, through introducing manifold learning and deep learning. GRNN was applied to construct the explicit nonlinear mapping from the original data of high dimension to the features of belt deviation extracted by LTSA. CDBN was proposed by introducing CRBM (Continuous Restricted Boltzmann Machine) and combining with the "dropout". Unlike LTSA, CDBN can be used to construct the explicit nonlinear mapping while extracting the deviation features from the original data, which needed more training time. Subsequently, the feature extraction experiments of belt deviation at different flow rates showed that the models based on LTSA+GRNN, and CDBN both had very good feature detection effect which meant that the two features extraction models could effectively reduce the redundancy of the original data while retaining enough features of belt deviation. And the experiments also revealed that, in case of belt deviation, the bigger the flow rate was, the greater the amount of belt deviation was, and vice versa. Further, SVM (Support Vector Machine), ELM (Extreme Learning Machine) and other regression analysis methods were used to build the online prediction models of belt deviation where the deviation features extracted by LTSA+GRNN and CDBN were taken as the input. Finally, the performances of two proposed online detection models of belt deviation based on LTSA+GRNN+ELM and CDBN+ELM respectively were verified through the experiments: the average prediction accuracy of deviation prediction model based on LTSA+GRNN+ELM was 93.33%, while its average prediction time of each group was 38.29 ms and its average training time was 18.91 s; the average prediction accuracy of deviation prediction model based on CDBN was as high as 98.61%, while its average prediction time of each group was as short as 1.47 ms and its average training time was 139.96 s. Besides, the experiments also showed that ELM was more suitable than SVM for the belt deviation, because ELM had almost the same prediction accuracy as SVM while the training and prediction time of ELM was far less than that of SVM. Both the prediction and real-time of the two models mentioned above showed that the two models could be a new approach for online detection of belt deviation and replaced traditional hardware detection device. Moreover, this study provided the necessary basis for the further online precision compensation and fault prediction of belt weigher.