Search

CN-121980189-A - Paper machine fault detection and diagnosis method based on multi-section sliding window core reconstruction analysis

CN121980189ACN 121980189 ACN121980189 ACN 121980189ACN-121980189-A

Abstract

The invention belongs to the technical field of fault detection and diagnosis in the process of flow industry, and discloses a paper machine fault detection and diagnosis method based on multi-section sliding window nuclear reconstruction analysis, wherein a sample is constructed through a sliding window mechanism, nonlinear dimension reduction and reconstruction error extraction are performed by adopting nuclear principal component analysis, and modeling capacity of a model on complex nonlinear relations is enhanced; the method comprises the steps of carrying out variable magnitude weighted analysis on reconstruction errors, identifying main abnormal variables when faults occur, and realizing high-precision fault source positioning. In order to verify the effectiveness of the method, the method is applied to a typical fault scene of a plurality of working sections of the papermaking process, and experimental results show that the method can accurately detect and diagnose key abnormal variables in various faults, has stronger fault interpretability and diagnosis reliability, and is suitable for abnormal identification and process monitoring of the key variables in a complex industrial process.

Inventors

  • LIU HONGBIN
  • WANG XINYUAN
  • WEI WENGUANG
  • HUANG PENG
  • WAN BING
  • MIAO YANFEN
  • ZHANG FENGSHAN
  • ZHOU JINGPENG
  • LI XIAOLIANG

Assignees

  • 南京林业大学
  • 瞬捷数字科技(山东)有限公司

Dates

Publication Date
20260505
Application Date
20260129

Claims (4)

  1. 1. The paper machine fault detection and diagnosis method based on the multiplex section sliding window core reconstruction analysis is characterized by comprising the following steps of: S1, data preprocessing, namely normalizing operation data by adopting a Z score normalization method, eliminating scale differences caused by different scales, dividing normal data into a training data set X, and taking sample data containing a plurality of different faults as a test set for model performance evaluation and fault detection verification; S2, constructing a sliding window characteristic enhancement model, namely adopting a sliding window mechanism to carry out time sequence reconstruction on the normalized data, constructing a sample sequence containing time dependence information, forming an enhanced input sample with history information, expanding the enhanced input sample into a high-dimensional time sequence vector, and enhancing the dynamic representation capability of the input characteristic; S3, constructing a nonlinear kernel dimension reduction model, namely constructing a kernel principal component analysis model based on a training data set X, introducing a radial basis kernel function to map sliding window feature data into a high-dimensional feature space, extracting main structural information in the high-dimensional feature space, selecting an optimal principal component on the premise of keeping a principal component accumulated feature variance, and realizing feature compression and reconstruction capability in the high-dimensional feature space, respectively constructing two monitoring statistics of reconstruction error square and nonlinear dissimilarity measure by calculating a reconstruction error and normalized deviation of a sample, wherein a calculation formula of the nonlinear dissimilarity measure is as follows: ; Wherein, the Corresponding time index A sliding window vector at which the sliding window is located, Is the length of the sliding window; and S4, constructing a fault detection and diagnosis model, namely inputting test set sample data containing various faults into a trained nonlinear kernel dimension reduction model, calculating corresponding reconstruction error square sum and nonlinear dissimilarity measure in real time, modeling normal data statistic distribution of the training sample by adopting a kernel density estimation method, adaptively and dynamically generating a detection threshold under a given confidence level, calculating average reconstruction error contribution of each variable in a sliding window by utilizing a reconstruction error contribution analysis method, sequencing, extracting key variables of 5 before ranking, and realizing key fault source positioning.
  2. 2. The method for detecting and diagnosing paper machine faults by multi-section sliding window core reconstruction analysis according to claim 1 is characterized in that the normal data statistic sequence is subjected to smooth noise reduction processing through Gaussian filtering.
  3. 3. The method of claim 1, wherein the operation data in step 1 is derived from paper machine production data from a paper mill, and the training data set X comprises normal operation data for each critical section in the paper manufacturing process, and the critical sections include, but are not limited to, a flow section, a wire section, a press section, and a dryer section.
  4. 4. The method for detecting and diagnosing faults in a paper machine based on multi-stage sliding window core reconstruction analysis as claimed in claim 1, wherein the fault sample data set Y is constructed by injecting different typical faults including but not limited to high consistency slurry blockage, vacuum system leakage and exhaust system failure and different types of abnormal modes including but not limited to drift change, periodic change and amplitude change for model performance evaluation and fault detection verification.

Description

Paper machine fault detection and diagnosis method based on multi-section sliding window core reconstruction analysis Technical Field The invention belongs to the technical field of fault detection and diagnosis in the process of flow industry, and particularly relates to a paper machine fault detection and diagnosis method based on multi-section sliding window core reconstruction analysis, which is suitable for intelligent monitoring scenes of multivariable, strong coupling and time-varying characteristics in the papermaking industry. Background In the modern paper industry production process, in order to ensure the stability of paper quality and the reliability of equipment operation, real-time monitoring and intelligent diagnosis of multivariate process data in key process stages are required. However, the papermaking process has the characteristics of multiple variables, strong time sequence, complex coupling among various sections and the like, so that the traditional monitoring method relying on mechanism modeling or static discrimination faces challenges in terms of accuracy and instantaneity. With the continuous improvement of industrial data acquisition capability and intelligent manufacturing level, a fault detection method based on data driving gradually becomes a mainstream development direction. Although the current mainstream method has a certain abnormality identification capability, the current mainstream method still has obvious limitations in coping with nonlinear dynamic characteristics of the papermaking process, historical dependency and cross-section comprehensive modeling. On one hand, part of methods cannot effectively fuse time structure information among samples, so that time variability of a system running state is difficult to sense, on the other hand, the use of static control limits also causes that a model is insufficient in robustness when the model is used for dealing with slight drift or early faults, false alarm or missing report is easy to occur, and in addition, most of the existing methods stay on a global abnormal judgment level, lack of accurate positioning capability on fault source variables, and limit practical engineering application values. Disclosure of Invention The invention provides a paper machine fault detection and diagnosis method based on multi-section sliding window core reconstruction analysis, which aims at the problems in the prior art and comprises the following steps: a paper machine fault detection and diagnosis method based on multi-section sliding window core reconstruction analysis comprises the following steps: S1, data preprocessing, namely normalizing operation data by adopting a Z score normalization method, eliminating scale differences caused by different scales, dividing normal data into a training data set X, and taking sample data containing a plurality of different faults as a test set for model performance evaluation and fault detection verification; S2, constructing a sliding window characteristic enhancement model, namely adopting a sliding window mechanism to carry out time sequence reconstruction on the normalized data, constructing a sample sequence containing time dependence information, forming an enhanced input sample with history information, expanding the enhanced input sample into a high-dimensional time sequence vector, and enhancing the dynamic representation capability of the input characteristic; S3, constructing a nonlinear kernel dimension reduction model, namely constructing a kernel principal component analysis model based on a training data set X, introducing a radial basis kernel function to map sliding window feature data into a high-dimensional feature space, extracting main structural information in the high-dimensional feature space, selecting an optimal principal component on the premise of keeping a principal component accumulated feature variance, and realizing feature compression and reconstruction capability in the high-dimensional feature space, respectively constructing two monitoring statistics of reconstruction error square and nonlinear dissimilarity measure by calculating a reconstruction error and normalized deviation of a sample, wherein a calculation formula of the nonlinear dissimilarity measure is as follows: Wherein, the Corresponding time indexA sliding window vector at which the sliding window is located,Is the length of the sliding window; and S4, constructing a fault detection and diagnosis model, namely inputting test set sample data containing various faults into a trained nonlinear kernel dimension reduction model, calculating corresponding reconstruction error square sum and nonlinear dissimilarity measure in real time, modeling normal data statistic distribution of the training sample by adopting a kernel density estimation method, adaptively and dynamically generating a detection threshold under a given confidence level, calculating average reconstruction error contribution of each