Esearch And Application Of Deep Learning Neural Network In Soft Sensor Modeling Of Ball Mill Fill Level

Ball mill is one high energy consumption equipment applied widely in electric power, mineral grinding, chemistry, pottery, cement, medical and metallurgy industry. In practical industrial process, because the fill level can’t be measured accurately for the complex environment inside ball mill, it is always kept lower to avoid the safety problem of over load or block mill caused by higher fill level. Consequently, the ball mill can hardly work with high efficiency. Therefore, it’s very important to measure the fill level accurately for improving the efficiency and stability of ball mill.Soft sensor technique is very efficient method to solve the indirectly measured process parameters measurement problem. The main principle of soft sensor is that the estimate of unmeasured parameter is completed depending on the relationship between an instrumental parameter and the unmeasured parameter. In this paper, the vibration signal of ball mill bearing is selected as the instrumental variable. To solve the problems of indirect variable analysis and effective feature extraction, deep belief network is introduced in the modeling process, which is employed to extract effective features from spectrum of vibration signal to represent the fill level. Deep belief network is a deep neural networks and one of classical deep learning methods, which has excellent ability of feature extraction capture the complex nonlinear features of the data. The primary advantages of deep belief network includes:(1) Deep belief network constructs a deep architecture to obtain the high level representations of input data.(2) DBN can obtain optimized initial weights of all hidden layers by the method of greedy layer-wise algorithm based on amount of unlabeled data, and then capture the complex nonlinear features of the data.(3) DBN can fine-tune the weights of the whole network through a few of labeled data and make the network perform preeminent finally. The above-mentioned characteristics of DBN show that DBN is very suitable to extract nonlinear features of vibration signal and it also suitable to solve the issue which is limited by lack of abundant and precise training sample in practical industrial process.In this thesis, we focused on analyze and extract effective features from power spectrum of vibration signal which can be used to represent the ball mill fill level. DBN is formed as a stack of restricted Boltzmann machines (RBM). DBN extracts effective features in two steps:Firstly, train each RBM via the mechanism of layer-by-layer without supervision. After the pre-training, the RBMs are unrolled to create a DBN, which is then fine-tuned using back propagation of error derivatives for optimal network. The layer-by-layer learning algorithm is a very effective way to pre-train the weights of DBN. Through the features transforming layer-by-layer, it makes the feature vector map different feature space and each layer of features captures higher level representation of the data. DBN is a more efficient way than PCA and PLS to progressively reveal low-dimensional, inherent and nonlinear structure of original data.In order to validate the feasibility and effectiveness of the ball mill fill level soft sensor model based on DBN feature extraction. Experiments on a lab-scale ball mill are carried out to simulate the ball mill in industrial fields and record the sample data (including labeled fill levels and corresponding vibration signal data). To analyze the amount vibration signal, firstly, compute the power spectrum density of vibration signal via Welch method. Secondly, adopt feature extraction methods including DBN and the conventional methods PCA and PLS to extract effective features of the high dimensional and collinearity power spectrum. Thirdly, transform the effective features to support vector machine to train the soft sensor model. Finally, use the test set to analyze the results of the soft sensor method. The experimental results show that soft sensor model based on DBN feature extraction is more accuracy and stable than other existing general methods...