Modeling Of Unsupervised Soft Sensor Based On Modal Partition And Domain Adaptation

As the scale of industrial production expands,the industrial production process becomes more and more complex,and the measurement of some key parameters in the system becomes more important.Traditional measurement technology is mainly based on a new type of process measurement instrument,which implements on-line soft sensor of process parameters in hardware form.However,due to the influence of process mechanism,physical environment,sensor and instrument hardware characteristics,some key parameters of industrial production process can not be measured directly on-line by hardware devices.Soft sensor technology uses the idea of indirect measurement,and establishes the corresponding mathematical model by using the auxiliary data information easily obtained in the process,so as to estimate the difficult-to-test dominant variable.Traditional soft sensor modeling requires that the modeling data and real-time data satisfy the assumption of independent and identical distribution,and the modeling process data comes from a single stable production condition.However,in the actual industrial production process,due to the reorganization of equipment,changes in raw materials and external environment,the process will have multiple stable operating conditions,that is,multiple stable working points in the same production process,and the correlation of variables between different stable working points has different characteristics.Such processes are called multimodal processes.The data distribution between different modes is different,which results in the performance deterioration of traditional soft sensor models.Therefore,finding an appropriate modeling method to adapt to the impact of working conditions is a key issue in multi-working condition soft sensor modeling.Different modal data have different process characteristics and require different models.Modal partitioning is often used in process monitoring and fault diagnosis of multimodal chemical processes.Introducing it to historic data is one of the keys to solve the modeling of multimodal soft sensor.In addition,the current and historical data no longer obey the same probability distribution,resulting in model inaccuracies.In this paper,an unsupervised domain adaptive method is introduced in soft sensor modeling.Historical data is used as the source domain and current data as the target domain.The application of subspace learning in multimodal soft sensor is mainly studied.The main contents and innovations of this paper are as follows:(1)An unsupervised soft sensor modeling method combining modal partition with subspace alignment is presented to solve the problem that a single soft sensor model cannot describe the industrial process with multimodal characteristics,and that the current and historical data do not conform to the same distribution,which results in the inaccuracy of the soft sensor model.Specifically,an offline modal partitioning method based on information entropy is proposed to partition the historic data in order to build different soft-sensing models for historic data with different modes;furthermore,the maximum mean difference distance is used to match the closest historic modes to the current data;then,the neighborhood alignment is proposed to maintain the local manifold structure of the data while preserving the local manifold structure of the data.Align the feature subspaces of the current data and the matched modal data to reduce the distribution differences.Finally,a partial least squares regression prediction model is built using the aligned data to obtain the soft sensor prediction values of the current data.(2)In view of the shortcomings of the off-line modal partitioning method based on information entropy,such as the need to specify the number of clusters and the sliding window threshold,a multi-modal process offline modal partitioning method is proposed,which can calculate the number of modes without specifying the sliding window threshold.Specifically,the process data is divided,the entropy difference between adjacent windows is calculated and sorted,and then the measurement is used.The function obtains the optimal mode partition number r_b,chooses the window with large r_b-1 before the entropy difference as the initial mode change window,and then uses another measure function to directly partition the local data to get the final mode partition result.The results of the modal partition under the numerical simulation experiments and the soft-sensing experiments under the TE simulation process show the effectiveness of the algorithm in this chapter.(3)A geodesic streaming kernel unsupervised soft sensor modeling method based on linear local tangent space arrangement is studied to solve the problem that the common information between historical and current data cannot be effectively used.By extracting the public information of the historical data and the current data,the public information of the two is projected into a manifold subspace to complete the feature migration in the subspace,thereby reducing the distribution differences between the two,and achieving an unsupervised soft sensor modeling.