Industrial Process Fault Detection Based On Monitoring Data Analysis

With the acceleration of China’s industrialization process,industrial production equipment and systems continue to develop towards scale,refinement and complexity.The growing security demand and the rapid development of industrial processes restrict and influence each other.On the one hand,the strong coupling of industrial production process makes the fault occur in a wider range and with higher frequency.On the other hand,the scale and complexity of production equipment will lead to higher concealment and stronger destructive power of faults.If the fault can not be found and handled in time,it will bring huge economic losses and even cause casualties.As an indispensable part of the whole industrial process monitoring,fault detection is an important means to ensure the safe and reliable operation of industrial processes and the efficient and stable production of products.The industrial process itself is a typical multivariable system with a complex structure,many production processes,and interrelated links.In the industrial big data context,a large number of industrial process operation monitoring data are collected,which reflect the operation status of industrial processes.However,the operation monitoring data of industrial processes often have characteristics,such as missing data,uneven data distribution,and multiprocess data correlation,which lead to the difficulty of detecting the fault of the complex industrial system directly.The key variables that reflect the characteristics or status of industrial process systems are generally hidden and cannot be observed directly.Data analysis methods can extract and summarize the collected monitoring data.And summarize and refine the information hidden in the data to obtain the internal laws of the research object.Based on this,this study investigates how to more effectively extract the core latent variables representing the running state of the system.The corresponding fault detection methods are proposed.The lack of industrial process data leads to the insufficient expressiveness of monitoring data for faults.The probabilistic principal component analysis is the most widely used probabilistic latent variable model to deal with missing data.The missing data are estimated by alternating iterations of the maximum expectation algorithm combined with the natural advantages of the probabilistic latent variable model.At the same time,considering that probabilistic principal component analysis cannot effectively utilize the label information in historical data and the limitations of traditional linear discriminant analysis in dimensionality reduction,probabilistic discriminant analysis is introduced.In this method,the monitoring data are mapped to two hidden variable spaces within and between classes,and then the fault type is determined by calculating the logarithm probability ratio of samples.The proposed method not only realizes missing data reconstruction but also effectively improves the class separability of implicit variable space.A process fault detection method based on a hybrid probabilistic latent variable model under the constraint of quality indicators is proposed to solve the problematic measurement of quality indicators in industrial process operation and the insufficient expressiveness of latent variables of operation monitoring data to quality-related faults.The hybrid model can automatically determine the dimension of the hidden variable space by giving different Gaussian distributions to each column of the projection matrix.This model is combined with the teacher-student network framework.The variational mixed factor analysis is regarded as the teacher model and the neural network as the student model.The low-dimensional latent variable obtained from the teacher model is used to supervise and train the student model so that the network structure contains quality-related information.The method proposed extracted the quality-related fault information and improved the fault detection performance,and realize quality-related fault detection.Under the multi-process data association of industrial processes,the traditional blocking method ignores the correlation between subblocks,resulting in the insufficient expressiveness of the independent blocking methods to faults.Therefore,a process fault detection method of distributed variational self-coder under multi-process data association is proposed.Considering the limited modeling ability of shallow models for complex problems,a deep latent variable model variational autoencoder is introduced to extract deeper features from the data.Considering the noise distribution of the data and the information of adjacent subblocks,a distributed variational autoencoder model is proposed to overcome the disadvantage that the traditional block method cannot effectively utilize the information of adjacent subblocks.The proposed method has a high detection rate in the case of complete data and missing data,and can quickly locate the fault locations that occur in local subblocks.A process fault detection method based on weighted kernel global-local projection is proposed to address the insufficient preservation of sparse area information by traditional local preservation methods due to the uneven distribution of industrial process data.Considering that the local preservation method ignores the data distribution characteristics when saving the neighbor relationship,the cam weighted distance is introduced to the kernel global-local preservation projection algorithm,so that the adjacent points can be automatically selected in accordance with the probability distribution of the surrounding points.Considering that the latent variables of different dimensions have different sensitivity to different faults,a function of the statistic change rate is constructed to realize the adaptive selection of fault-sensitive principal components.Obtaining a suitable parameter group is difficult because many parameters are found in the proposed model,and the genetic algorithm is used to optimize the model parameters.The experimental results show that the proposed algorithm can improve the fault detection performance of the traditional kernel global-local projection.