CN-115982650-B - Industrial dynamic process robust fault isolation method based on variable decibel leaf inference

CN115982650BCN 115982650 BCN115982650 BCN 115982650BCN-115982650-B

Abstract

The invention discloses an industrial dynamic process robust fault isolation method based on variable decibel leaf inference. The method comprises the steps of collecting data in the process of industrial production under normal working conditions and under fault working conditions through a sensor as a whole data set, establishing respective robust fault isolation models aiming at the data set under the normal working conditions and the fault working conditions, selecting proper prior distribution of each variable in the established robust fault isolation models, solving by using a variable decibel leaf extrapolation method, calculating posterior distribution of each variable, and judging and isolating faults according to a fault indication matrix in a solving result. The method can adapt to industrial dynamic process data with abnormal values or larger noise, can realize accurate fault isolation, and provides effective support for industrial production control behaviors.

Inventors

ZENG JIUSUN
LU CHENG
XU XIAOBIN
YAO LE
LIU YI
SHEN BINGBING

Assignees

杭州师范大学

Dates

Publication Date: 20260508
Application Date: 20221129

Claims (3)

1. A robust fault isolation method for industrial dynamic process based on variable decibels is characterized by comprising the following steps: Step 1, acquiring data in the process of industrial production under normal working conditions and fault working conditions as a whole data set through a sensor; Step 2, establishing respective robust fault isolation models aiming at data sets under normal working conditions and fault working conditions; Step 3, selecting proper prior distribution for each variable in the built robust fault isolation model, then solving by using a variable dB leaf inference method, calculating posterior distribution of each variable, and carrying out fault judgment and isolation according to a fault indication matrix in a solving result; In the step 1, the sensor collects data in the process of industrial production under normal working conditions and fault working conditions as training data and test data respectively, and the training data and the test data are combined and normalized to obtain a data set with zero mean and one unit variance; in the step2, specifically: For training data under normal working conditions, the following robust fault isolation model is established: X={x r } x r ={x i } D ={ } Where X represents training data under normal conditions, X r represents a data vector under normal conditions, X i represents an ith measurement of the data vector under normal conditions, X r , D represents a projection matrix, Is the ith row of the projection matrix D, g represents the dimension-reducing vector after the data vector under the normal working condition is projected to the dimension-reducing space, and s i represents the abnormal value vector under the normal working condition V i represents the difference between the result of multiplying the projection matrix under normal conditions by the dimension-reduction vector and the i-th measured value x i of the data vector; Is the vector of the indicated outliers of the ith measurement x i of the data vector under normal conditions, Representative of Containing outliers, if the contrary Representative of No outliers are included; for training data under fault conditions, the following robust fault isolation model is established: D ={ } ={ } in the formula, Representing training data under normal conditions, Representing the data vector under normal conditions, Representing data vectors under normal conditions Is determined by the measurement value of (c) in the (c), Representing a matrix of fault indications, Representing fault indication matrices Is arranged in the row i of the (a), A dimension reduction vector which represents the projection of the data vector under the fault condition to the dimension reduction space, Indicating abnormal value vector under fault condition Is used for the amplitude of (a) and (b), Representing the difference between the result of multiplying the projection matrix by the dimension reduction vector under the fault condition and the ith measured value x i of the data vector; Is the ith measurement of the data vector under normal conditions Is an indication of an outlier vector of (c), Representative of Containing outliers, if the contrary Representative of No outliers are included; Is a weight matrix; Is a sparse representation of the tag parameter matrix, symbols Representing the hadamard product; In the step 3, only the indicated abnormal value vectors under the normal working condition and the fault working condition are used And The tag parameter matrix Z is set to be Bernoulli-Gaussian prior distribution, and dimension reduction vectors g and under normal working condition and fault working condition are calculated Normal and fault condition indicating outlier vector Amplitude value And indicating an outlier vector Amplitude of (a) of (b) Difference v i and difference under normal and fault conditions The projection matrix D, the weight matrix W are set to gaussian distribution.
2. The robust fault isolation method of industrial dynamic process based on variable dB leaf is characterized in that in the data set, training data and test data are composed of data acquired by different sensors according to time sequence, the data acquired by all the different sensors at the same time form a data vector, and a single data vector in the time t-d-t is expressed as Where m represents the total number of sensors, The number of delay steps is indicated and, Representing the amount of time delay, t representing the current time, Representing t- The data vector acquired at the moment in time, Represents a vector set of m 1.
3. The method for robust fault isolation in an industrial dynamic process based on a variational Bayesian as set forth in claim 1, wherein in said step 3, a fault indication matrix in a solution result is extracted Calculating a fault indication matrix The two norms of each row vector in the (a) are taken as fault variable contribution values, and then judgment is carried out: If the contribution value of the fault variable is greater than 0, the position of the sensor corresponding to the row vector of the contribution value of the fault variable is faulty, and then the sensor is adjusted to solve the fault according to the position of the specific working condition of the fault variable; if the fault variable contribution value is not greater than 0, the position of the sensor corresponding to the row vector of the fault variable contribution value is not faulty and is not processed.

Description

Industrial dynamic process robust fault isolation method based on variable decibel leaf inference Technical Field The invention belongs to an isolation method in the field of process monitoring and fault diagnosis in an industrial control system, and particularly relates to an industrial dynamic process robust fault isolation method based on variable decibel leaf inference. Background With the continuous improvement of the reliability and safety requirements of industrial production processes, the data-driven process monitoring becomes a simple and effective monitoring method and is widely applied. Among the many methods, principal Component Analysis (PCA) -based fault detection is one of the basic methods. Using PCA, data can be projected into principal component space and residual space, and statistics constructed therefrom to monitor industrial processes. The fault variables are then located and isolated by conventional contribution graph methods. Reconstruction methods based on a minimum absolute shrinkage and selection operator (Lasso) are also used for localization and isolation of fault variables. The methods have higher positioning and isolation performance under the condition of small data fluctuation and less interference. Conventional fault isolation models fail when the data contains more outliers or is noisy. In recent years, probability-based fault localization and isolation models have also received much attention, and probability-based methods can effectively incorporate a priori knowledge using a suitable probability distribution, and can handle uncertainty and missing data in industrial processes, such as Probabilistic Principal Component Analysis (PPCA), hidden Markov Models (HMM), and the like, as compared to general methods. Meanwhile, in order to cope with the influence of abnormal values on the results, researchers have proposed to use heavy tail distribution or mixed distribution to improve the robustness of the model, such as student t distribution, mixed gaussian distribution, and the like. Researchers have also found that using the characteristics of the MMV that the sparse variable is unchanged for a short period of time can effectively acquire the timing characteristics of the process data. Therefore, in order to ensure the stable operation of the industrial production process and improve the fault isolation performance, a robust fault isolation method of the industrial dynamic process based on the variational Bayesian inference is developed, and the positioning and the identification of industrial fault variables are realized. Disclosure of Invention In order to solve the technical problems in the background technology, the invention provides an industrial dynamic process robust fault isolation method based on the variational Bayesian inference. The method is suitable for the fault positioning problem of a complex large-scale industrial production system, and has long-term significance for promoting the development of industrial automation and big data technology. The invention researches the data characteristics of abnormal values or more noise contained in the process data, and can meet the robustness requirement in the industrial process. The invention provides effective technical support for the robust fault isolation method of the industrial dynamic process. The technical scheme adopted by the invention is as follows: Step 1, acquiring data in the process of industrial production under normal working conditions and fault working conditions as a whole data set through a sensor; step 2, establishing a robust fault isolation model with the respective window width d according to the data sets under the normal working condition and the fault working condition; And 3, selecting proper prior distribution for each variable in the built robust fault isolation model, solving by using a variable dB leaf inference method, calculating posterior distribution of each variable, and carrying out fault judgment and isolation according to a fault indication matrix in a solving result. In the step 1, the sensor collects data in the process of industrial production under the normal working condition and the fault working condition as training data and test data respectively, and the training data and the test data are combined and normalized to obtain a data set with zero mean and one unit variance. The training data and the test data are used as data sets under normal working conditions and fault working conditions respectively. The data set is characterized in that training data and test data are composed of data acquired by different sensors according to time sequence, the data acquired by all the different sensors at the same time form a data vector, and a single data vector in the time t-d-t is expressed asWhere m represents the total number of sensors contained in the dataset, τ represents the number of delay steps, d represents the amount of time delay, t represents the current time ins