Research And Application Of Anomaly Detection Algorithm Based On Production Data Of Polysilicon Ingot

With the rapid development of the photovoltaic industry,polysilicon occupies a dominant position in the raw materials of photovoltaic cells by virtue of its cost performance.However,due to its high production cost,the batches of ingredients used in the production process or differences in ingredients may produce a small number of abnormal products.If the early warning is not available,it will cause great waste.During recent years,benefiting from the rapid development of artificial intelligence,anomaly detection has played an important role in the fields of network intrusion detection,industrial fault diagnosis and medical disease detection.Research on abnormal detection of polysilicon ingots is of great significance for realizing timely and accurate early warning of abnormal patterns and reducing unnecessary economic losses.At present,the main method for detecting abnormality of polysilicon ingot is process test.This method not only has high test cost,but also is difficult to operate,which is difficult to implement in detail.Therefore,in response to the application requirements of polysilicon ingot quality abnormality detection,combined with the theory of artificial intelligence,this thesis proposes several improved algorithms from three aspects: dataset construction,feature extraction and recognition model construction,and systematic analysis and research are carried out.These algorithms have obtained good experimental results and provided valuable references for actual production.The main research content of this thesis is summarized as follows:Firstly,the polysilicon ingot dataset is established to train and test the detection models.In the aspect of polysilicon ingot quality prediction,based on the current domestic and foreign polysilicon ingot technology research,with reference to published scientific research papers and patent works,this thesis summarized common ingot quality abnormality detection methods.This thesis analyzed the main factors that affect the quality of the final ingot through field inspections on the production site to understand the production process and data composition.After preprocessing the annual polysilicon ingot production data,this thesis constructed a polysilicon ingot dataset for the research.The polysilicon ingot dataset belongs to the unbalanced data set commonly used in the field of anomaly detection,that is,the amount of abnormal sample data is much smaller than the amount of normal sample data.Two key issues for feature dimensionality reduction and classification recognition of imbalanced data sets.This thesis proposed an anomaly detection model based on the improved diffusion mapping-support vector data description algorithm(DM-SVDD).Firstly,the diffusion mapping(DM)algorithm is used to achieve dimensionality reduction.The eigenvalues and corresponding eigenvectors covering the main structure of the data are taken to make it maintain a stable global relationship in the low-dimensional space;at the same time,for the polysilicon ingot data set,a new nearest neighbor graph is constructed by using two distance measurement formulas of Euclidean distance and Mahalanobis distance to improve the diffusion mapping algorithm;then a new anomaly detection model is constructed by combining the support vector data description algorithm(SVDD).The experimental results show that the new model not only improves the recognition performance of most normal samples,but also has better detection performance for a few abnormal samples than the traditional model,while reducing the time complexity.For the anomaly detection problem of unbalanced data sets,deep learning methods can be used to extract deep features that better express the characteristics of the data,and further improve the classification and recognition performance.Based on this idea,This paper proposes an anomaly detection model based on a composite network: stacked denoising automatic encoders-one class support vector machine(SDAE-OCSVM).Firstly,the dataset is balanced through the SMOTE algorithm;secondly,the stacked denoising automatic encoders(SDAE)performs unsupervised pre-training on the original features of the dataset;thirdly,combined the data labels to supervised fine-tuning by using the back-propagation(BP)algorithm and reconstruct the depth characteristics of the original sample;finally,the one class support vector machine(OCSVM)is used to identify and classify the samples.The experimental results prove that the composite network structure can effectively extract the deep features of the data,fully explore the internal information between the features,and greatly improve the recognition performance of the anomaly detection model.