Data-driven Quality Status Identification Of Air Conditioning Tube Expansion

With the rapid development of Internet of things,these product quality detection methods based on sensor data and machine learning can obtain accurate detection results at a low cost.Due to the complex production environment and frequent switching of products,the quality inspection of tube opening in the production process of air-conditioner condensing tube is faced with difficulties in data acquisition,low detection efficiency and high cost.In view of the above situation,this paper transforms the quality state recognition problem of tube opening in the expanding tube process into a data-driven pattern recognition problem.Then the SOM-BP and the CNN-TL methods are utilized to solve this problem,which can obtain good recognition results on four product data sets with different sample sizes respectively.The main research contents are as follows:Based on the principle of tube expanding process,this paper construct a platform for data collection.In this platform,pressure sensors are used to obtain the pressure of the supporting seats as the inputs of data analysis.In addition,visual sensors are used to recognize the size of tubes as the labels of data analysis.Thus four product datasets of different size are constructed.In this paper,original data are denoised and segmented according to the characteristics of tube expanding process.For the two different models,the multi-dimensional eigenvalues are extracted from the segmented data by different methods.Conventional feature extraction methods are used in the SOM-BP model,including extraction of time-domain,frequencydomain and time-frequency features of pressure data,and feature selection by Pearson correlation analysis.CNN-TL model transforms the pressure data into the wavelet scale spectrum image based on the wavelet change to extract the image features.SOM-BP model is adopted firstly to identify the quality of tube in tube expanding process,which takes into account the advantages of fast convergence speed of SOM neural network and adaptive learning ability of BP network and can obtain good recognition results on our four product datasets.It is worth noting that the recognition results of two small product datasets are not as good as that of the other two large product datasets.In this paper,CNN-TL model is utilized to solve the problem that recognition results in small product datasets are not good enough.We first use the large product datasets to train the CNN model,and then carry out transfer learning in the small product datasets,so as to solve the problems of insufficient training and over-fitting in the small product datasets.Experimental results show that this model can achieve better recognition effect.