CN-121705810-B - Dynamic and static mixed graph-based fault detection method for dense medium coal separation process

CN121705810BCN 121705810 BCN121705810 BCN 121705810BCN-121705810-B

Abstract

The invention belongs to the technical field of industrial process monitoring and fault diagnosis, and provides a fault detection method for a dense medium coal dressing process based on a dynamic and static mixed graph, which can be integrally called a space-time synchronous attention network (HG-STAN) based on the dynamic and static mixed graph. The method comprises the steps of collecting key variable data in a dense medium coal dressing process, preprocessing, constructing a static diagram based on process priori knowledge, simultaneously learning dynamic association from the data by utilizing a self-attention mechanism, fusing to form a dynamic and static hybrid diagram, constructing space neighborhood information and historical time information of synchronous aggregation nodes of a time-space synchronous diagram attention self-encoder (STGAAE), utilizing a combined loss function training model comprising reconstruction loss and diagram structure sparsity regularization, establishing a multi-level monitoring system of data level statistics and diagram level statistics, and determining various statistic control limits through kernel density estimation to realize fault online detection. The invention obviously improves the accuracy and the sensitivity of fault detection in the dense medium coal separation process.

Inventors

ZHANG XIANGRUI
Gong Zhuoyan
DAI WEI
XIA ZHENXING
NAN JING
LIU XIN
WANG LANHAO

Assignees

中国矿业大学

Dates

Publication Date: 20260512
Application Date: 20260213

Claims (10)

1. The method for detecting the fault in the dense medium coal separation process based on the dynamic and static mixed graph is characterized by comprising the following steps of: S1, data preprocessing and dynamic and static mixed graph construction, namely acquiring process monitoring variable data and priori knowledge aiming at a gravity coal dressing process, carrying out data preprocessing on the process monitoring variable data, obtaining standardized time sequence data by each process variable, generating dynamic time sequence data through a fixed-length sliding window to form standardized dynamic input time sequence data, defining by using the priori knowledge to obtain a static graph adjacent matrix, inputting the preprocessed time sequence data into a spatial self-attention mechanism, obtaining attention weights among variables through calculating an inquiry matrix and a key value matrix, forming an adjacent matrix of a dynamic graph, fusing the static graph and the dynamic graph to obtain a mixed graph serving as a topological basis for subsequent feature extraction; S2, modeling and training a space-time synchronous diagram attention self-encoder, namely establishing a space-time synchronous diagram attention self-encoder STGAAE, wherein STGAAE comprises a space-time synchronous diagram encoder and a space-time synchronous diagram decoder which are stacked by a plurality of layers of space-time synchronous diagram attention modules, taking a mixed diagram adjacent matrix and preprocessed time sequence data as model input, realizing synchronous aggregation and feature extraction of space neighborhood information and historical time information by introducing a masking lower triangular matrix and a time feature extraction matrix in a node updating process, reconstructing input data in a coal re-dressing process, constructing a combined loss function based on a data reconstruction error and a diagram structure sparsity regularization term, training the space-time synchronous diagram attention self-encoder, forcing the model to learn process data key features in a normal working condition of the coal re-dressing process, and constraining a dynamic diagram structure to avoid false correlation, and obtaining a trained monitoring model; s3, constructing multi-level monitoring statistics and off-line modeling, namely respectively constructing data-level statistics and graph-level statistics based on a trained model, wherein the data-level statistics comprise T2 statistics of a feature space and SPE statistics of a residual space, and the graph-level statistics comprise macro topology statistics based on Laplace matrix feature values And microstructure statistics based on dynamic edge count Determining a control limit of each statistic through kernel density estimation; S4, online fault detection, namely acquiring process data in real time and carrying out preprocessing the same as that of an offline stage, inputting the processed data into a trained model, calculating values of four monitoring statistics in real time, and judging that the process is faulty if any one of the values of the monitoring statistics exceeds a corresponding control limit.
2. The method for detecting a fault in a dense medium coal separation process based on a dynamic and static hybrid map as claimed in claim 1, wherein in step S1, the data preprocessing includes abnormal data rejection, missing value filling and maximum and minimum standardization, and the spatio-temporal process data of all process monitoring variables are in the form of data preprocessing X represents training data composed of N normal samples, T represents matrix transposition, Represents the t th normal sample, and the ith column of X is marked as , A time series data vector representing the ith process variable, The time sequence data matrix of the input model at the moment t is obtained by reorganizing the data through a sliding window Where w represents the length of the sliding window.
3. The dynamic and static hybrid diagram-based dense medium coal separation process fault detection method according to claim 1, wherein in step S1, the dense medium coal separation process diagram structure is defined as a directed non-weighted diagram Wherein Is a set of nodes that are configured to communicate, Is a set of directed edges that are directed, Representing adjacency matrix, node Is the ith variable in the dense medium coal separation process, and the node attribute is the normalized time sequence data vector Directed edge Representing variables Sum variable Causal dependencies between.
4. The method for detecting a fault in a dense medium coal separation process based on a dynamic and static hybrid map according to claim 1, wherein in step S1, the construction of the dynamic and static hybrid map structure specifically comprises defining an adjacency matrix of the static map based on priori knowledge Wherein the element is 1 when there is a physical connection or causal relationship between nodes, and 0 otherwise, from a time series data matrix using a self-attention mechanism In-process dynamic graph adjacency matrix : ; Wherein the matrix is queried Key value matrix , And Is two trainable weight matrices of the self-attention mechanism, The static diagram and the dynamic diagram are overlapped and fused to obtain a mixed diagram adjacent matrix : ; Wherein, the Representation fetch And (3) with The edges of two adjacent matrixes are OR-operated, and finally As a topological basis for subsequent spatio-temporal synchronization feature extraction.
5. The method as claimed in claim 1, wherein in step S2, the spatio-temporal synchronous graph attention self-encoder STGAAE is composed of a spatio-temporal synchronous graph attention mechanism embedded graph encoder-graph decoder architecture, and for realizing efficient matrix operation and avoiding future information leakage, spatial attention weights are calculated first, for the hybrid graph Any node of moment And node The input data are respectively And Synchronous attention coefficient Subsequently, by node pairs All neighbors of (a) Is normalized to obtain a normalized attention coefficient Combining the learnable time feature extraction matrix with the masking lower triangular matrix to jointly act on input data to obtain a space-time aggregated matrix expression: ; Wherein the method comprises the steps of As an adjacency matrix for the hybrid map, Is a unitary matrix, is used for adding self-loops for the adjacent matrix of the mixed graph, Is that The input data corresponding to the layer is provided, For the time feature extraction matrix, To mask the lower triangular matrix, a matrix is used for better extracting the characteristics of the data And performing dimension transformation on the updated nodes, and then introducing nonlinear operation through an activation function to enhance the extraction capability of complex high-dimensional features.
6. The method for detecting a fault in a dense medium coal separation process based on a dynamic and static hybrid map as claimed in claim 1, wherein in step S2, the spatio-temporal synchronization map encoder is formed by stacking a plurality of layers of spatio-temporal synchronization map attention modules, and functions to input original spatio-temporal data Mapping to a representation rich in space-time dependencies Wherein For dimension, the spatio-temporal synchronization map decoder is the inverse of the spatio-temporal synchronization map encoder, the goal of which is to slave the characterization To reconstruct the original input data 。
7. The method for detecting a fault in a dense medium coal separation process based on a dynamic and static hybrid map according to claim 1, wherein in step S2, the combined loss function is Lost by reconstruction Sum graph structure sparsity regularization loss The weighting constitution is as follows: ; Wherein the reconstruction loss is used for measuring the output of the space-time synchronous diagram decoder With the original input The difference forces the model to learn the key characteristics of the normal working condition data, and the mean square error is adopted as reconstruction loss: ; To prevent overcomplicating the dynamic graph and introducing false correlations, the graph structure sparsity regularization penalty is increased, penalizing the excessive sum of edge weights in the dynamic adjacency matrix to encourage its sparsity, consistent with the a priori knowledge that the correlation of the real industrial process is usually sparse, The L1 norm of the matrix: ; is a super parameter used for balancing the importance between the reconstruction accuracy and the sparsity of the graph structure, Representing a training batch of the training set, Is of batch size.
8. The method for detecting a fault in a dense medium coal separation process based on a dynamic and static hybrid map as set forth in claim 1, wherein in step S3, the offline modeling comprises training a spatio-temporal synchronization map attention self-encoder STGAAE based on normal training samples, and constructing data-level and map-level monitoring statistics, the data-level monitoring statistics comprising constructing a conventional in a feature space and a residual space And SPE statistics, wherein the feature space statistics at time t The calculation formula is as follows: ; Wherein the method comprises the steps of And Respectively representing the average value of the characteristic vector output by the final hidden layer of the space-time synchronization diagram from the encoder at the time t and the characteristic vector obtained in the off-line modeling stage, Is covariance matrix obtained in final hidden layer offline modeling stage, and t moment residual space statistic The calculation formula is as follows: ; Wherein the method comprises the steps of Is that The original input vector of the moment in time, A reconstructed output for the spatio-temporal synchronization map decoder; the monitoring statistics of the graph layer includes constructing Laplace spectrum monitoring statistics from both macro topology and microstructure layers And dynamic graph structure monitoring statistics For each moment Dynamic diagram structure of (a) First, calculate the corresponding degree matrix And Laplace matrix : ; Subsequently, to the Laplace matrix Decomposing the characteristic values to obtain a group of characteristic values Calculating the variance of the eigenvalues as a comprehensive macroscopic statistic : ; For each moment Dynamic graph adjacency matrix of (2) , Elements of (a) Quantitate nodes And node At the moment of time Dynamically associating all Adding to obtain a statistical dynamic graph adjacent matrix Statistics of the number of all edges ; ; Four statistics ,SPE, And Control limit of (2) , , And The method is obtained by carrying out kernel density estimation on the corresponding statistic sequence of the normal training data under the given significance level.
9. An electronic device comprising a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the dynamic and static hybrid map-based coal re-screening process fault detection method of any one of claims 1-8.
10. A computer readable storage medium, characterized in that a computer program or instructions is stored which, when run on a computer, performs the steps of the dynamic and static hybrid map based method for detecting a fault in a dense media coal separation process as claimed in any one of claims 1 to 8.

Description

Dynamic and static mixed graph-based fault detection method for dense medium coal separation process Technical Field The invention relates to the technical field of industrial process monitoring and fault diagnosis, in particular to a fault detection method for a dense medium coal dressing process based on a dynamic and static mixed graph. Background The dense medium coal separation technology is widely applied due to high-efficiency separation precision, and the system is formed by physically connecting a plurality of key devices such as a mixing barrel, a dense medium cyclone, an arc screen and the like, has highly nonlinear and dynamic coupling processes, and provides a serious challenge for the accuracy and timeliness of fault detection. The existing fault detection method has the following limitations of 1, linear model and isolated modeling, namely the nonlinear characteristics of the dense medium coal separation process are difficult to characterize by the traditional linear method represented by Principal Component Analysis (PCA) and Partial Least Squares (PLS). Although deep learning methods such as self-encoder (AE) and long-short-term memory network (LSTM) promote nonlinear feature extraction capability, sensor variables are generally regarded as isolated time series inputs, topological correlation among variables determined by a process flow is ignored, and therefore, a model lacks physical interpretability and is difficult to locate fault propagation paths. 2. The singleness of the graph structure construction is that although the Graph Neural Network (GNN) can explicitly model the relation among variables, under the heavy medium coal-dressing scene, a static graph constructed only by relying on priori knowledge is difficult to adapt to dynamic changes such as working condition drift, and a dynamic graph constructed only by relying on data correlation is easy to generate false association under a strong noise environment and loses physical significance. The existing method cannot ensure the physical interpretability and the working condition self-adaption capability of the graph structure at the same time. 3. The prior space-time graph neural network mostly adopts a serial (space-first time-second or space-first time-second) architecture to process space-time data, so that the inherent synchronism of space interaction and time evolution in the heavy medium coal separation process is manually split, and complex space-time coupling dependency relationship is difficult to effectively capture. 4. The design of monitoring indexes is incomplete, the existing method mainly depends on statistics of data layers such as T2 and SPE, the characteristics that process faults often accompany system topological structure changes are not fully utilized, and monitoring of structural abnormality of the layer is lacking, so that the detection sensitivity of structural faults is insufficient. Therefore, a fault detection method capable of integrating physical knowledge, dynamically adapting to working conditions, synchronously modeling space-time characteristics and providing multi-level monitoring indexes is needed to improve the reliability and safety of operation in the dense medium coal separation process. Disclosure of Invention The invention aims to provide a dynamic and static hybrid graph-based fault detection method for a dense medium coal separation process, which aims to solve the problems in the background technology. In order to achieve the purpose, the invention provides the technical scheme that the fault detection method for the dense medium coal dressing process based on the dynamic and static mixed graph is hereinafter referred to as HG-STAN (Hybrid Graph based Spatial-Temporal Synchronous Attention Network), and a core model is a space-time synchronous graph attention self-encoder (STGAAE). S1, data preprocessing and dynamic and static mixed graph construction, namely acquiring process monitoring variable data and priori knowledge aiming at a heavy medium coal dressing process, carrying out data preprocessing on the process monitoring variable data, obtaining standardized time sequence data by each process variable, generating dynamic time sequence data through a fixed-length sliding window to form standardized dynamic input time sequence data, using the priori knowledge to define and obtain a static graph adjacent matrix, inputting the preprocessed time sequence data into a spatial self-attention mechanism, obtaining attention weights among variables through calculating an inquiry matrix and a key value matrix, forming an adjacent matrix of a dynamic graph, fusing the static graph and the dynamic graph to obtain a mixed graph, and taking the mixed graph as a topological basis for subsequent feature extraction. And S2, modeling and training a space-time synchronous diagram attention self-encoder, wherein the space-time synchronous diagram attention self-encoder STGAAE is built, the STGAAE compris