Chemical production line health management method and equipment based on graph feature fusion and medium

Abstract

The invention provides a chemical production line health management method and device based on graph feature fusion and a medium, and belongs to the field of chemical production line state monitoring, and the method comprises the steps: dividing the multi-source heterogeneous historical data of a chemical production line into first-level data and second-level data; the first-level data and the second-level data are converted, a graph data set of a historical sample is obtained, and the graph data set of the historical sample comprises a production line space graph data set and production line knowledge graph data; training a graph convolutional neural network based on the graph data set of the historical sample to obtain a production line fault diagnosis model; and converting the online measurement data of the chemical production line, and inputting the converted online measurement data into the production line fault diagnosis model to obtain a chemical production line fault diagnosis result. Through the method provided by the invention, a chemical production line fault diagnosis problem is converted into a graph classification problem, the model diagnosis capability is improved, and the problem of low fault classification accuracy of a traditional fault diagnosis model is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of chemical production line status monitoring, and in particular to a chemical production line health management method based on graph feature fusion, and a computer-readable storage medium and electronic device for implementing the chemical production line health management method. Background Art

[0002] With the advent of the Industrial 4.0 era, process industry processes are becoming increasingly sophisticated and complex, which places higher demands on the intelligent control of production lines. Due to the common characteristics of process industry production lines, such as harsh on-site environments, difficult sampling, and large measurement delays, it is difficult to directly obtain key indicators of production lines and products. In order to achieve the goals of stable product production, maximizing economic profits, and green energy saving, it is increasingly important to develop an efficient and reliable production line health monitoring system.

[0003] Traditional data-driven chemical production line diagnostic models mainly include those based on statistical learning techniques and machine learning techniques, such as principal component analysis (PCA), partial least squares regression (PLS), and support vector machine (SVM). These technologies have been widely studied and applied due to their simple algorithm flow and strong reliability, but most of these methods are based on linear assumptions, which is inconsistent with the actual data of chemical production lines.

[0004] In recent years, convolutional neural networks (CNNs) have been widely used in classification problems because of their excellent nonlinear signal fitting capabilities. Compared with traditional artificial neural networks, convolutional neural networks have a deeper network structure, can integrate more relevant features in the prediction model to facilitate diagnostic classification, and can also provide sparse connections for large neural networks to avoid overfitting. Although convolutional neural networks have advantages in fault diagnosis tasks, they only focus on the data itself and lack analysis of the spatiotemporal dependencies between data.

[0005] On this basis, more and more scholars are applying methods based on spectral graph theory to the field of fault diagnosis. Among them, constructing graph data to represent sample relationships and extracting high-level fault features from them has become one of the effective means to solve the poor accuracy of some categories of fault diagnosis. In order to mine fault features from the constructed graph data, relevant researchers proposed to use graph convolutional neural networks to extract signal features from discrete spatial structures through convolution operations on the graph, and have been successfully applied in semantic recognition, behavior recognition and other fields. In the field of process industry fault diagnosis, graph structure data created from prior knowledge of production lines can more effectively fit the characteristics of the production lines themselves, including the graph structure converted from the physical space layout of the production line and the graph structure established between the production line signal data, which greatly improves the fidelity of the model and provides strong support for the characterization of production line faults.

[0006] In summary, the existing chemical production line fault diagnosis methods have the following limitations: 1) Chemical production line data are highly coupled, and the fluctuation of a monitoring signal often means the fluctuation of multiple upstream and downstream signals, which makes it difficult for traditional data analysis models to fit this high-dimensional complex signal relationship; 2) The above data-driven chemical production line fault diagnosis model is often established only from the monitoring signal, ignoring the process knowledge of the production line, resulting in a loose integration of the diagnosis model with the actual production line; 3) Multiple sensors in the key process of the chemical production line often describe the same change process from different angles, and the above sensor signals need to be jointly analyzed, which is ignored by the above model. Summary of the invention

[0007] In order to solve the technical problem that the chemical production line fault diagnosis model in the prior art ignores the inherent coupling relationship between multi-source heterogeneous historical data, the present invention provides a chemical production line health management method based on graph feature fusion. The method can classify and decouple the multi-source heterogeneous measurement signals of the production line. On this basis, from the perspective of graph theory, the chemical production line monitoring data is combined with process knowledge to construct graph data. The production line monitoring time signals are jointly analyzed from the perspective of signal space correlation and signal mechanism analysis to obtain the production line health status and fault classification information.

[0008] To achieve the above-mentioned purpose, the chemical production line health management method based on graph feature fusion provided in the first aspect of the present invention comprises the following steps: classifying and processing the multi-source heterogeneous historical data of the chemical production line to obtain primary data and secondary data, wherein the primary data includes the production line condition data obtained by each sensor of the chemical production line, and the secondary data includes process product test data; converting and processing the primary data and the secondary data as sample data to obtain a graph data set of historical samples, wherein the graph data set of historical samples includes a production line space graph data set and a production line knowledge graph data, and the mathematical representation of the graph data is G={V, E, A, F}, wherein G is graph data, V is a node set, E is a sample edge set, and the sample edge set is used to characterize the edge connection between any two nodes, A is an adjacency matrix, and F is a node feature matrix; training a graph convolutional neural network based on the graph data set of the historical samples to obtain a production line fault diagnosis model; converting and processing the online measurement data of the chemical production line to obtain a graph data set of online samples; inputting the graph data set of online samples into the production line fault diagnosis model to obtain a chemical production line fault diagnosis result.

[0009] In an exemplary embodiment of the present invention, the conversion and processing of the primary data and the secondary data as sample data to obtain a graph data set of historical samples may include: using a sliding window of length m to divide the primary data and the secondary data into multiple first-category sample monitoring periods and multiple second-category sample monitoring periods, respectively; constructing a production line space graph data set based on the first-category sample monitoring periods and the physical space layout information of the production line sensors; constructing a production line knowledge graph data based on the second-category sample monitoring periods and the similarity of monitoring signals; and constructing a graph data set of historical samples based on the production line space graph data set and the production line knowledge graph data.

[0010] In an exemplary embodiment of the present invention, the construction of a production line space graph dataset based on the first-category sample monitoring period and the physical space layout information of the production line sensors may include: determining each sensor as a first-category node; determining a number of neighboring points around each first-category node based on the physical space layout information of the production line sensors; connecting each first-category node with a number of neighboring points around it to obtain an edge connection relationship of each first-category node; determining an adjacency matrix of primary data based on the edge connection relationship between each first-category node; preprocessing the first-category sample monitoring period, and splitting each preprocessed first-category sample monitoring period into multiple data columns based on the number of production line sensors; determining each split data column as a first-category node feature corresponding to a first-category node; constructing a production line space subgraph at the current monitoring moment based on multiple first-category nodes and their corresponding first-category node features, edge connection relationships and adjacency matrices at the current monitoring moment; and constructing a production line space graph dataset based on the corresponding production line space subgraphs at different monitoring moments.

[0011] In an exemplary embodiment of the present invention, determining a number of neighboring points around each first-class node based on the physical space layout information of the production line sensor may include: mapping multiple first-class nodes to a rectangular coordinate system based on the physical space layout information of the production line sensor to obtain the spatial coordinates and relative position relationship of each first-class node; determining the spatial distance between each first-class node based on the Euclidean distance function; and determining a number of neighboring points around each first-class node based on the spatial distance between each first-class node.

[0012] In an exemplary embodiment of the present invention, the preprocessing of the first category sample monitoring period may include: normalizing the first category sample monitoring period to obtain the first category sample monitoring period after normalization; performing noise reduction processing on the first category sample monitoring period after normalization based on a principal component analysis algorithm to obtain the first category sample monitoring period after noise reduction.

[0013] In an exemplary embodiment of the present invention, the construction of production line knowledge graph data based on the second-category sample monitoring period and the monitoring signal similarity may include: determining each second-category sample monitoring period as a second-category node; normalizing the second-category sample monitoring period to obtain the normalized second-category sample monitoring period; extracting features from the normalized second-category sample monitoring period based on a linear discriminant analysis algorithm to obtain second-category node features corresponding to each second-category node; determining the monitoring signal similarity between each second-category node based on a Mahalanobis distance function; determining a number of neighboring points around each second-category node based on the monitoring signal similarity between each second-category node and the second-category node features; connecting each second-category node to a number of neighboring points around it to obtain an edge connection relationship of each second-category node; determining an adjacency matrix of secondary data based on the edge connection relationship between each second-category node; and jointly constructing a production line knowledge graph based on multiple second-category nodes corresponding to all production line monitoring moments and their corresponding second-category node features, edge connection relationships, and adjacency matrices.

[0014] In an exemplary embodiment of the present invention, the calculation formula for the normalization process may be:

[0015]

[0016] Among them, X scaled is the scaled data output after normalization; X is the input data; X.max is the maximum value of the elements in X; X.min is the minimum value of the elements in X; (max-min) is the data span.

[0017] In an exemplary embodiment of the present invention, the production line fault diagnosis model may include a spatial graph learning sub-model and a knowledge graph learning sub-model. The spatial graph learning sub-model may include three Chebyshev convolutional layers and one graph pooling layer, and the knowledge graph learning sub-model may include three Chebyshev convolutional layers and two fully connected layers.

[0018] In an exemplary embodiment of the present invention, the aggregation parameter k of the neighboring points may be set to 3-5; and the order K of the Chebyshev polynomial may be set to 4-7.

[0019] In an exemplary embodiment of the present invention, the length m of the sliding window may be determined based on the measurement time interval of the primary data and the secondary data.

[0020] A second aspect of the present invention provides an electronic device, comprising a processor and a memory, wherein the memory stores at least one computer program, and the at least one computer program is loaded and executed by one or more of the above processors so that the computer implements the chemical production line health management method as described above.

[0021] A third aspect of the present invention provides a computer-readable storage medium, in which at least one program code is stored. The program code is loaded and executed by a processor so that a computer implements the chemical production line health management method as described above.

[0022] Through the technical solution provided by the present invention, the present invention has at least the following technical effects:

[0023] (1) In view of the problem that monitoring data in key process often have multi-dimensional data describing the same process, the present invention, based on graph theory, transforms the physical spatial layout of monitoring sensors in complex process industrial systems into a spatial atlas to describe the connection relationship between multi-source heterogeneous sensors in the same process. From the data representation level, the graph representation performance of potential faults in chemical production lines is improved, which is conducive to enhancing the diagnostic and analysis capabilities of chemical production line fault diagnosis models for production line monitoring accidents.

[0024] (2) In view of the characteristics of chemical production lines that changes in upstream signals in the process often cause changes in multiple downstream signals, the present invention converts multi-source heterogeneous signals into graph node features, connects adjacent nodes through signal similarity measurement, and constructs a chemical production line knowledge graph. The graph node connection can map the information changes in the chemical process, enhance the fault information representation ability of the sample from the mechanism level, and improve the model diagnosis ability;

[0025] (3) The present invention innovatively proposes a graph learning model that can be used for chemical production line fault diagnosis. The graph learning model constructs a spatial atlas and a knowledge graph based on the spatial dependency existing in the chemical production line sensor layout and the signal correlation determined by the reaction mechanism. On this basis, the graph data structure information and node attribute information are learned simultaneously to obtain the final graph node feature representation and output the fault type at the monitoring time. The graph learning model solves the problem of low accuracy of partial fault classification in traditional chemical production line fault diagnosis models due to lack of prior knowledge.

[0026] (4) The present invention converts the chemical production line fault diagnosis problem into a graph classification problem. By decoupling the multi-source heterogeneous data of the production line, the production line data is divided into two types of graph data. The graph data can effectively represent the temporal and spatial dependencies between the monitoring data and comprehensively characterize the fault information of the production line monitoring object. The fault diagnosis method of the present invention has the advantages of stronger performance, higher accuracy and higher credibility.

[0027] Other features and advantages of the present invention will be described in detail in the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The accompanying drawings are used to provide a further understanding of the embodiments of the present invention and constitute a part of the specification. Together with the following specific implementations, they are used to explain the embodiments of the present invention, but do not constitute a limitation on the embodiments of the present invention. In the accompanying drawings:

[0029] Figure 1 A technical roadmap for a chemical production line health management method based on graph feature fusion provided in an embodiment of the present invention;

[0030] Figure 2 A schematic diagram of a process flow of a chemical production line health management method based on graph feature fusion provided by an embodiment of the present invention;

[0031] Figure 3 A schematic diagram of a process flow for converting multi-source heterogeneous historical data provided by an embodiment of the present invention;

[0032] Figure 4 A schematic diagram of a process for constructing a production line space graph dataset provided by an embodiment of the present invention;

[0033] Figure 5 A schematic diagram of a process for constructing a production line knowledge graph dataset provided by an embodiment of the present invention;

[0034] Fig. 6A A schematic diagram of an industrial process provided by an embodiment of the present invention;

[0035] Figure 6B A sensor distribution coordinate diagram provided by an embodiment of the present invention;

[0036] Figure 6C A production line space graph structure provided by an embodiment of the present invention;

[0037] Figure 7 A schematic diagram of the ChebGCN fault diagnosis model structure provided by an embodiment of the present invention;

[0038] Figure 8 The accuracy confusion matrix of each category output by the existing diagnostic model provided in the embodiment of the present invention;

[0039] Fig. 9 The accuracy confusion matrix of each category output by the fault diagnosis model provided in the embodiment of the present invention;

[0040] Fig.10 A graph showing the results of an experiment for selecting the number k of neighbor nodes for spatial graph aggregation provided by an embodiment of the present invention;

[0041] Fig.11 A graph showing the experimental results of selecting the order of the Chebyshev convolution polynomial K of the graph diagnosis model provided in an embodiment of the present invention;

[0042] Fig.12 A schematic diagram of the recognition accuracy of a chemical production line fault diagnosis task model provided by an embodiment of the present invention;

[0043] Fig.13 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention.

[0044] Description of Reference Numerals

[0045] 201 - processor; 202 - memory. DETAILED DESCRIPTION

[0046] The specific implementation of the embodiment of the present invention is described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation described here is only used to illustrate and explain the embodiment of the present invention, and is not used to limit the embodiment of the present invention.

[0047] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0048] In the present invention, unless otherwise specified, the directional words used, such as "upper, lower, top, bottom", are usually used to describe the relative positional relationship of the components in terms of the directions shown in the drawings or in terms of the vertical, perpendicular or gravity directions. "First", "second", etc. are only used for the convenience of description and distinction, and cannot be understood as indicating or implying relative importance.

[0049] The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments.

[0050] Embodiment 1

[0051] The first embodiment of the present invention provides a chemical production line health management method based on graph feature fusion.

[0052] Please refer to Figure 1 , the overall technical idea of ​​the method is as follows: first, from the perspective of decoupling multi-source heterogeneous data of the production line, the production line data is divided into two categories: the first category is the production line operating information data (that is, primary data) measured by each sensor of the production line, which is used to characterize the state of the production line at the time of measurement; the second category is the process product test data (that is, secondary data), which is used to characterize the production quality of the product. On this basis, based on the idea of ​​graph theory, the present invention converts the physical space layout of the monitoring sensors of the complex process industrial system into a chemical production line space diagram from the perspective of primary data, and uses this space diagram to describe the connection relationship of multi-source heterogeneous sensors in the same process, thereby improving the characterization performance of potential faults of the chemical production line. At the same time, the present invention converts multi-source heterogeneous secondary data into node features, connects adjacent nodes through signal similarity metrics, and constructs a chemical production line knowledge graph. The mutual connection between nodes can map the information changes in the chemical process, enhance the signal strength from the mechanism level, and improve the model diagnosis capability. Finally, after the chemical production line spatial graph and the chemical production line knowledge graph are trained by the graph convolutional neural network (GCN convolutional network for short), the feature vectors learned from the two types of graph data are fused by the attention mechanism to obtain the final fault representation for fault diagnosis.

[0053] Specifically, please refer to Figure 2 The chemical production line health management method may include the following steps:

[0054] Step S101: Classify and process the multi-source heterogeneous historical data of the chemical production line to obtain primary data and secondary data. The primary data includes the production line operating data obtained by various sensors of the chemical production line, and the secondary data includes process product test data.

[0055] It should be noted that the multi-source heterogeneous historical data of the chemical production line can represent the known production status of the chemical production line and contain the real fault information of the production line.

[0056] In step S101, the multi-source heterogeneous historical data of the chemical production line are divided based on the data source.

[0057] The primary data represents the signals output by the sensors of the instruments and equipment on the production line, such as temperature, pressure, flow, etc. The signals represent the state of the production line at the time of measurement. The sensors on the production line can be six types of sensors, including temperature sensors, pH sensors, dissolved oxygen sensors, weight sensors, CO2 release sensors, and O2 release sensors. There are also other sensors, which are not fully described in this embodiment. And there are multiple quantities of each type of sensor. For example, there are 10 temperature sensors installed at different locations of the chemical production line to monitor the production temperature changes of the chemical production line.

[0058] Secondary data represents the test data of process products, which is obtained by manual sampling and testing at production line measurement points, such as viscosity, acidity, analytical purity, etc. The data represents the production quality of the product.

[0059] Step S102: The primary data and the secondary data are converted and processed as sample data to obtain a graph data set of historical samples, where the graph data set of historical samples includes a production line space graph data set and a production line knowledge graph data set.

[0060] In step S102, the mathematical representation of the graph data is G = {V, E, A, F}. Where G is the graph data, V = {V i} is a node set, representing a measurement time of the production line; E is a sample edge set, which represents the edge connection between any two nodes and can be used to represent graph data; A = {a i,j} is the adjacency matrix, when a i,j = 0, indicating that there is no edge connection between node i and node j. i,j =1, it indicates that there is an edge connection between nodes; F is the high-order feature matrix (that is, the node feature matrix) extracted from the original measurement data.

[0061] For example, please refer to Figure 3 The process of converting the primary data and the secondary data as sample data to obtain the graph data set of the historical samples may include but is not limited to the following sub-steps S1021 to S1024.

[0062] Sub-step S1021: Use a sliding window with a length of m to divide the primary data and the secondary data into multiple first-category sample monitoring periods and multiple second-category sample monitoring periods respectively.

[0063] Here, the purpose of using a sliding window to divide the multivariate time series data measured on the production line is to obtain sample points at the monitoring moment. A sliding window of length m can divide the primary data and secondary data into multiple data segments of length m. This data segment is called a monitoring period, and its length is the time accuracy of production line fault diagnosis. The length of the sliding window m should be selected to align the time intervals of the two types of data measurements mentioned above. When the measurement interval of the secondary data is longer than that of the primary data, the mean data padding method is used to align the data. When the length of the last segment of the historical data of the production line is less than m, it can be padded to a length of m using data padding, resulting in a total of h monitoring periods.

[0064] Sub-step S1022: constructing a production line space map data set based on the first type of sample monitoring period and the physical space layout information of the production line sensors.

[0065] It should be noted that multiple sensors in key process of complex process industrial system often describe the same process from different perspectives, so it is necessary to jointly analyze the above sensor signals. In order to achieve the above purpose, the production line sensors described in the first-level data can be regarded as nodes, and the production line space atlas can be obtained from the physical space layout of the production line sensors.

[0066] Specifically, the sensors to which the first-level data of the production line belongs can be regarded as nodes, and they can be mapped to the rectangular coordinate system according to the physical spatial layout of the production line sensors to obtain the spatial coordinates and relative position relationship of each node. On this basis, based on the distance between each node, the K nearest neighbor algorithm can be used to find the k nearest neighbors of each node, and each node can be connected to its surrounding k nearest neighbors to obtain a spatial subgraph. Each monitoring moment corresponds to a production line spatial subgraph, and the historical first-level data at all monitoring moments jointly construct the production line spatial atlas.

[0067] For example, please refer to Figure 4 , based on the first type of sample monitoring period and the physical space layout information of the production line sensors, the process of constructing the production line space map data set includes but is not limited to the following sub-steps S1022a to S1022h.

[0068] Sub-step S1022a: Determine each sensor as a first-category node.

[0069] Sub-step S1022b: Based on the physical space layout information of the production line sensors, determine a number of neighboring points around each first-category node.

[0070] Specifically, the first step is to map multiple first-class nodes into a rectangular coordinate system based on the physical space layout information of the production line sensors, obtain the spatial coordinates and relative position relationship of each first-class node, and thereby measure the aggregation relationship between multiple sensors.

[0071] The second step is to determine the spatial distance between each first-category node based on the Euclidean distance function.

[0072] In this sub-step, the Euclidean distance function can be used to measure the spatial distance between nodes. The Euclidean distance calculation formula is as follows:

[0073] Where D(v1,v2) is the Euclidean distance, x i ,y i They are node V i The horizontal and vertical coordinates in a rectangular coordinate system.

[0074] The third step is to determine several neighboring points around each first-category node based on the spatial distance between each first-category node.

[0075] For example, the k-nearest neighbor algorithm (full name: k-nearest neighbor classification algorithm) can be used to find the k nearest neighbor points of each node. The idea of ​​the k-nearest neighbor (KNN) classification algorithm is: in the feature space, if most of the k nearest (i.e., the closest) samples near a sample belong to a certain category, then the sample also belongs to this category.

[0076] Sub-step S1022c: Connect each first-category node with several neighboring points around it to obtain the edge connection relationship of each first-category node.

[0077] In sub-step S1022c, the edge connection relationship of the spatial graph nodes is the path of information transmission in the graph structure. The Euclidean distance function is used to measure the spatial distance between nodes, and the node is connected to its k closest sensor nodes in space with edges, which can jointly analyze multiple sensor signals that represent the same key process.

[0078] Sub-step S1022d: Determine the adjacency matrix of the primary data based on the edge connection relationship between each first-category node.

[0079] Sub-step S1022e: pre-process the first category sample monitoring period, and split each first category sample monitoring period after pre-processing into multiple data columns based on the number of production line sensors.

[0080] Specifically, the first type of sample monitoring period can be normalized to obtain the first type of sample monitoring period after normalization. Normalization can convert sample data to the [0-1] interval, achieve geometric compression of sample data, and eliminate the impact of dimension on model diagnosis.

[0081] The calculation formula for the normalization process is:

[0082]

[0083] Where, X scaled is the scaled data output after normalization; X is the input data; X.max is the maximum value of the elements in X; X.min is the minimum value of the elements in X; (max-min) is the data span.

[0084] Then, the first type of sample monitoring period after normalization processing is subjected to noise reduction processing to obtain the first type of sample monitoring period after noise reduction processing.

[0085] For example, an unsupervised principal component analysis algorithm (PCA algorithm for short) can be used to project the processed data (that is, the first type of sample monitoring period) into a high-dimensional feature dimension for noise reduction and feature extraction.

[0086] Sub-step S1022f: Using the split data as a single column as the sensor node feature to embed the first type of node feature of the spatial graph.

[0087] For example, each first-category sample monitoring period is composed of a first-level data segment of length m. The first-level data segment of length m can be split into n data columns, where n is the number of sensors contained in the first-level data. A data column can correspond to the characteristics of a sensor and be used as a sensor node feature for node feature embedding.

[0088] Sub-step S1022g: Based on a number of first-category nodes at the current monitoring moment and their corresponding first-category node features, edge connection relationships and adjacency matrices, a production line space subgraph at the current monitoring moment is constructed.

[0089] Specifically, the n data columns after preprocessing and splitting are embedded as node features of the spatial graph to obtain the spatial subgraph structure data corresponding to each first-category sample monitoring period (i.e., production line monitoring moment).

[0090] Sub-step S1022h: Each monitoring moment corresponds to a production line space sub-map, thereby constructing a production line space map dataset.

[0091] It should be noted that each monitoring period corresponds to a piece of spatial subgraph structure data, and their edge connections are the same but the node features are different. That is to say, the first-level data includes a total of h first-category sample monitoring periods, and all the first-level data are converted into a total of h spatial subgraphs, which share the edge connection structure constructed by sub-step S1022c but have different node features.

[0092] Sub-step S1023: constructing production line knowledge graph data based on the second category sample monitoring period and monitoring signal similarity.

[0093] It should be noted that, unlike the production line space graph, the connection between nodes in the production line knowledge graph is determined by the product test data, which can intuitively reflect the product production status. Here, the divided second-category sample monitoring period can be regarded as a node, and each node corresponds to a second-category sample monitoring period, with a total of h nodes. On this basis, based on the similarity of each node, the K nearest neighbor algorithm can be used to find the k nearest neighbors of each node, and each node can be connected to the k nearest neighbors around it to obtain a knowledge subgraph. All historical secondary data together construct a production line knowledge graph.

[0094] For example, please refer to Figure 5 , based on the second type of sample monitoring period and monitoring signal similarity, the process of constructing the production line knowledge graph data includes but is not limited to the following sub-steps S1023a~S1023h.

[0095] Sub-step S1023a: Determine each second-category sample monitoring period as a second-category node.

[0096] Sub-step S1023b: normalize the second category sample monitoring period to obtain the normalized second category sample monitoring period.

[0097] For example, the Min-Max normalization process can be performed on the second type of sample monitoring period, and the calculation formula for the normalization process is:

[0098]

[0099] Sub-step S1023c: extract features from the normalized second-category sample monitoring period to obtain second-category node features corresponding to each second-category node.

[0100] For example, a supervised linear discriminant analysis algorithm (LDA algorithm for short) can be used to perform data dimension reduction and feature extraction, and the eigenvalues ​​obtained by feature extraction can be used as node features.

[0101] Sub-step S1023d: Determine the similarity of the monitoring signals between the second-category nodes based on the Mahalanobis distance function.

[0102] In this sub-step, the Mahalanobis distance function can be used to measure the similarity between the multi-dimensional features of the nodes, which can largely eliminate the measurement error caused by the difference in data dimension and distribution. The formula of the Mahalanobis distance function is as follows:

[0103]

[0104] Where D M (x,y) is the Mahalanobis distance, ∑ is the covariance matrix of the multidimensional random variable, and x and y are two points in the data set.

[0105] Sub-step S1023e: Based on the monitoring signal similarity between the second-type nodes and the second-type node characteristics, determine a number of neighboring points around each second-type node.

[0106] Sub-step S1023f: Connect each second-category node with several neighboring points around it to obtain the edge connection relationship of each second-category node.

[0107] Sub-step S1023g: Determine the adjacency matrix of the secondary data based on the edge connection relationship between each second-category node.

[0108] Sub-step S1023h: construct a production line knowledge graph based on the corresponding second-type nodes at all production line monitoring moments and their corresponding second-type node features, edge connection relationships and adjacency matrices.

[0109] Sub-step S1024: constructing a graph dataset of historical samples based on the production line space graph dataset and the production line knowledge graph data.

[0110] Step S103: Train a graph convolutional neural network based on the graph data set of historical samples to obtain a production line fault diagnosis model.

[0111] Specifically, the two types of graph data obtained (that is, the production line spatial graph dataset and the production line knowledge graph dataset) can be input into the graph convolutional neural network model to obtain two types of sub-graph features; on this basis, the two types of sub-graph features are fused to obtain a global graph representation (that is, a fault fusion representation vector); on this basis, the fault fusion representation vector is used as the input of the softmax classifier model, and the fault label at the production line monitoring moment is used as the output of the softmax classifier model to obtain the production line status.

[0112] Further, in step S103, the production line fault diagnosis model may include a spatial graph learning sub-model and a knowledge graph learning sub-model. The spatial graph learning sub-model may be composed of three layers of ChebGCN (i.e., Chebyshev convolution layer) and one layer of GraphPool layer (i.e., graph pooling layer). After the original spatial graph is learned through three Chebyshev convolutions, the spatial representation vector of the production line state at the monitoring moment is obtained by average pooling. The knowledge graph learning sub-model may be composed of three layers of ChebGCN and two layers of Full Connect (FC) layers. After the knowledge graph is learned through three Chebyshev convolutions, the knowledge graph outputs the production line state knowledge representation vector with the number of labels through the fully connected layer.

[0113] The Chebyshev convolution operation is expressed as:

[0114]

[0115]

[0116]

[0117] Where x∈R n is the node feature; y∈R s is the graph convolution output; K is the Chebyshev polynomial order; g θ =diag[θ] is the filter; θ k ∈R n is the Chebyshev correlation coefficient vector; T k (x) is a recursive calculation formula; is the regularized form of the Laplace matrix.

[0118] The constructed three-layer graph convolutional neural network structure can be expressed as:

[0119] Z=Q softmax (WU(WU(WXW (0) )W (1) )W (2) ).

[0120] In the formula, Q softmax is the activation function; U is the linear activation function; W∈R N×N is the weight matrix, N is the number of samples; W (0) ∈R C×H , W (1) ∈R H×I , W (2) ∈R I×J , are the weight matrices of the first, second and third layers of graph convolution, C is the feature dimension of the input matrix, H, I, J are the output feature dimensions of the first, second and third layers of convolution, respectively.

[0121] Further, in step S103, the CheGCN graph convolutional neural network includes a graph convolution layer, an activation function, a feature fusion function and a classifier. The graph convolution layer is used to learn the sample node features of the graph data of the historical samples; the activation function is used to fit the nonlinear features according to the learned sample node features; the feature fusion function is used to fuse the learned subgraph features. The classifier is used to predict the health status label of the historical samples according to the nonlinear features, and the health status label is used to characterize the fault information.

[0122] Furthermore, in step S103, the feature fusion method is weighted concatenation of representation vectors: the learned production line state space representation vector and the production line state knowledge representation vector are weighted concatenated through weight coefficients w1 and w2, and the weight coefficients w1 and w2 are determined by production line characteristics and experiments.

[0123] Furthermore, in step S103, the loss function of the production line fault diagnosis model is an improved cross entropy cost function. The fault diagnosis result output by the production line fault diagnosis model may include: whether a fault occurs within the monitoring period and the type of the fault.

[0124] Furthermore, in step S103, 70% of the graph data of historical samples can be used as training samples, and 30% of the graph data of historical samples can be used as test samples. The training samples are used to train the CheGCN graph convolutional neural network, and the test samples are used to test the fault diagnosis model.

[0125] Step S104: After converting the online measurement data of the chemical production line, a graph data set of online samples is obtained.

[0126] Step S105: inputting the graph data set of the online samples into the production line fault diagnosis model to obtain the chemical production line fault diagnosis result.

[0127] Specifically, the online measurement data of the chemical production line can be converted and processed according to step S102 to obtain a spatial graph and a knowledge graph of the online measurement data; the two types of graph data are then input into a trained production line fault diagnosis model to obtain the production line status at the time of online monitoring of the production line.

[0128] It should be noted that the CheGCN graph convolutional neural network model (CheGCN model for short) acts as a feature extractor, which can extract features from graph data, so that we can use these features to perform node classification, graph classification, link prediction, and graph embedding. The graph data here does not refer to images, but data with graph structure. The structure of a graph is generally very irregular and can be considered as an infinite-dimensional data, so it has no translation invariance. The surrounding structure of each node may be unique.

[0129] The principle of the CheGCN model is as follows: suppose there is a batch of graph data, in which there are N nodes, each node has its own features, and there are D features in total. We assume that the features of these nodes form an N×D-dimensional matrix X, and then the relationship between each node will also form an N×N-dimensional matrix A, also known as the adjacency matrix. Here, matrix X and matrix A are the inputs of the GCN model.

[0130] The steps of the CheGCN model can be simply summarized into three steps. The first step is clustering, the second step is updating, and the third step is looping. For each node, its feature information is obtained from all its neighboring nodes during clustering, including its own features. CheGCN is a multi-layer graph convolutional neural network. Each convolutional layer only processes first-order neighborhood information. By superimposing several convolutional layers, multi-order neighborhood information can be transmitted. Starting from the input layer, the forward propagation is calculated through the graph convolution layer, and then the predicted classification probability distribution is obtained through the calculation of the softmax activation function. The role of softmax is to probabilize the output results of the convolutional network, which can be understood as calculating the category of the sample point according to the formula.

[0131] The CheGCN model is actually composed of input layer + hidden layer (graph convolution layer, similar to the function of fully connected layer) + SoftMax + output layer. GCN inputs a graph, and the features of each node in GCN are transformed from X to Z through several layers. However, no matter how many layers there are in the middle, the connection relationship between nodes, that is, the adjacency matrix A, is shared.

[0132] The implementation environment of the embodiment of the present application may include at least one terminal and a server, and the method is executed on the terminal or the server respectively. The terminal and the server can be connected in communication to realize the interactive transmission of information. The terminal and the server complete the classification and identification of the production line fault diagnosis model by acquiring multi-source heterogeneous data of the production line and converting it into a production line knowledge graph data set and a production line space graph data set.

[0133] Among them, the terminal can be any electronic product that can interact with the user through one or more methods such as keyboard, touchpad, touch screen, voice interaction, etc., such as PC (Personal Computer), PPC (Pocket Personal Computer), tablet computer, etc.

[0134] The server can be a single server or a server cluster composed of multiple servers. It can also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN (Content Delivery Network), as well as big data and artificial intelligence platforms.

[0135] Embodiment 2

[0136] In this embodiment, the Tennessee Eastman Process (TEP) is used to verify the effectiveness of the health management method of Example 1. The Tennessee Eastman Process (TEP) is a simulation model of an actual chemical process. Because it is close to actual production, it has been widely used in health monitoring and fault diagnosis model evaluation in process industries. There are 41 different monitoring devices in the Tennessee Eastman process, including 22 system monitoring variables such as temperature, pressure, and flow, and 19 product detection variables such as acidity, purity, and ratio. The monitoring variables generated by the TEP model have nonlinear, dynamic, and strongly coupled characteristics. In addition, the model operation process does not include the start and stop stages of complex systems.

[0137] In this embodiment, the simulation model is set to mode_1 (baseline), and the 12 manual variables are all set to initial values. On this basis, a total of 10 production states were simulated, including normal states and 9 different process disturbances, which indicate typical faults that occur in practice. Each batch was simulated for 48 hours, with a sampling interval of 3 minutes, and a total of 960 data points were generated. It should be noted that the upper and lower limits of reasonable operation are set for each device in the simulation model. Once a certain limit is reached, the reaction will stop to protect this complex system. Among all the disturbance types introduced, the interruption of A feed flow will trigger the stop of the complex system, and the unknown disturbance is a combination of known disturbances. All samples of fault states and normal states are mixed, and a total of 9600 data points are obtained. 70% of the samples are randomly selected as training sets and 30% of the samples are randomly selected as test sets. The fault types are shown in Table 1.

[0138] Table 1 TEP fault types

[0139]

[0140] According to the sample data set provided by the Tennessee Eastman Process (TEP), the specific steps of using the production line fault diagnosis model of the present invention to perform health monitoring and fault diagnosis are as follows:

[0141] Step 1: Data classification and sample selection.

[0142] 22 system monitoring variables such as temperature, pressure, flow, etc. are classified as primary data; 19 product detection variables such as acidity, purity, ratio, etc. are classified as secondary data.

[0143] On this basis, a sliding window of length m is used to divide the primary data and secondary data into data segments of length m, which are called monitoring periods, and their length is the time accuracy of production line fault diagnosis. In this experiment, m is 20. When the length of the last segment of the production line historical data is less than m, it is padded to a length of m using data padding, resulting in a total of h monitoring periods.

[0144] Step 2: Convert sample data into graph structure data.

[0145] (1) Data preprocessing and feature extraction.

[0146] a) Perform Min-Max normalization processing on the first-level data of the production line obtained by division, and use the PCA algorithm to reduce the noise of the normalized data to obtain the first-level data of the production line after preprocessing. The output dimension of PCA is 9.

[0147] On this basis, the first-level data segment with a length of m is split into n data columns, where n is the number of sensors included in the first-level data. Here n is 22.

[0148] b) Perform Min-Max normalization on the divided secondary data of the production line, and use the supervised LDA algorithm to extract features from the normalized data to obtain the features of the secondary data monitoring time. The number of supervised labels for LDA is 10, and the output dimension is 7.

[0149] (2) Construct a spatial graph structure.

[0150] a) The sensors to which the first-level data of the production line belongs are regarded as nodes, and the physical space layout of the production line sensors (such as Fig. 6A As shown in Figure 2, they are mapped into a rectangular coordinate system to obtain the spatial coordinates and relative position relationship of each node (as shown in Figure 2). Figure 6B shown).

[0151] b) On this basis, the Euclidean distance metric is used to calculate the distance between each node, and the K nearest neighbor algorithm is used to find the k nearest neighbors of each node. Each node is connected to its k nearest neighbors to obtain the edge connection of the spatial graph.

[0152] c) On this basis, the first-level data after the above preprocessing and splitting is embedded as the node feature of the spatial graph to obtain the production line spatial graph structure corresponding to the monitoring period (such as Figure 6C shown).

[0153] (3) Construct a knowledge graph structure.

[0154] a) The monitoring period of the divided secondary data is regarded as a node, and a total of h nodes are used to construct a knowledge graph. Specifically, the monitoring time features of the secondary data after the above preprocessing are used as node features.

[0155] b) On this basis, the Mahalanobis distance function is used to calculate the similarity between nodes, and each node is connected to its k nearest neighbors to obtain a knowledge graph.

[0156] Different from the spatial graph, the connection between nodes is determined by product test data, which can directly reflect the product production status. At the same time, each node corresponds to a monitoring period, and all historical secondary data together construct a knowledge graph structure.

[0157] Step 3: Build a production line fault diagnosis model.

[0158] Construct and train a chemical production line fault diagnosis model based on graph feature fusion. Specifically, the spatial graph diagnosis model constructed based on the primary data consists of three layers of ChebGCN convolutional layers and one layer of GraphPool. The ChebGCN convolutional layer settings are shown in Table 2, and GraphPool adopts average pooling output. The knowledge graph diagnosis model constructed based on the secondary data consists of three layers of ChebGCN, and the settings are consistent with the spatial graph diagnosis model.

[0159] The activation function uses LeakyRelu (0.1); during the training phase, each layer uses the Dropout mechanism to suppress overfitting.

[0160] The output vectors of the above two are weighted by using w1 and w2 weight assignments and then sent to the softmax full connection for fault classification. Both w1 and w2 are 0.5. The structure of the ChebGCN fault diagnosis model is as follows: Figure 7 As shown in the figure, the model is trained under supervised conditions, and the fault state corresponding to each monitoring period is the data label of the period.

[0161] Considering that there is no data imbalance problem between samples, the Accuracy evaluation index is used to evaluate the diagnostic accuracy of the model. The calculation formula is as follows:

[0162]

[0163] Table 2 ChebGCN convolutional layer settings

[0164]

[0165] Step 4: Production line fault diagnosis.

[0166] The production line monitoring signal data obtained online is converted and processed by the method in step 2 to obtain a spatial graph and a knowledge graph, and then substituted into the production line fault diagnosis model constructed in step 3. On the basis of considering the structural correlation between nodes, the graph data structure information and node attribute information are learned at the same time to obtain the final fault feature representation, and the fault label corresponding to the data at each moment is output. The fault diagnosis model finally obtained in this embodiment has a diagnostic accuracy of 93.26% under the TEP data set.

[0167] Embodiment 3

[0168] In this embodiment, the effectiveness of the health management method of Example 1 is verified using simulation data generated by the public penicillin fermentation process simulation model IndPenSimeV2.0.

[0169] The penicillin fermentation process production line contains 8 manual control variables, 2 automatic control variables and 11 monitoring data, of which 7 are directly obtained primary data and 4 are secondary data obtained by sampling and testing the contents. The sampling interval of the monitoring unit is 12 minutes.

[0170] This experiment simulated 6 production states, of which 1 was a healthy state and 5 were different types of faults. In the production batch with faults, the fault state was introduced at a specific time and returned to normal after a period of time. The monitoring data at other times in the batch were within the normal range. The detailed information is shown in Table 3 below.

[0171] Table 3 Penicillin process production line status list

[0172]

[0173] The Accuracy evaluation index is used to evaluate the diagnostic accuracy of the model. The calculation formula is as follows:

[0174]

[0175] According to the sample data set provided by the penicillin fermentation process production line, the specific steps of using the production line fault diagnosis model of the present invention to perform health monitoring and fault diagnosis are as follows:

[0176] Step 1: Data classification and sample selection.

[0177] The seven system monitoring variables, such as temperature, pressure, and flow, are classified as primary data; the four content sampling and testing variables, such as acidity and ratio, are classified as secondary data.

[0178] On this basis, a sliding window with a length of m is used to divide the primary data and the secondary data into data segments with a length of m, which are called monitoring periods.

[0179] Step 2: Convert sample data into graph structure data.

[0180] (1) Preprocess and extract features of monitoring data.

[0181] For production batches with faults, only the monitoring data of the sensor at a certain fault time is extracted as a sample, and the remaining normal time information is ignored. In order to solve the data imbalance problem in actual production, the following sampling is performed on the six production line conditions: 1,100 normal samples, 750 samples of fault 1, 650 samples of fault 2, 550 samples of fault 3, 450 samples of fault 4, and 150 samples of fault 5.

[0182] Next, the collected original monitoring data is converted: first, the monitoring period after each sample is divided is normalized by Min-Max, and each monitoring data is converted to the [0-1] interval to achieve geometric scaling of the original monitoring data and eliminate the impact of the dimension on model diagnosis; secondly, the autoencoder is used to learn the features of the normalized data, and the high-order representation of the data features is learned through the sparse connection layer of the autoencoder. After multiple hyperparameter optimizations, it is found that the fault diagnosis accuracy is best under 4-dimensional features. Therefore, the original data dimension is reduced to 4 dimensions.

[0183] (2) Construct a spatial graph structure.

[0184] a) Consider the sensors to which the primary data belongs as nodes, and map each node to a rectangular coordinate system according to the physical spatial layout of the production line sensors to obtain the spatial coordinates and relative position relationship of each node (such as Figure 6B shown).

[0185] b) Use the Euclidean distance metric to calculate the distance between each node, and use the K-nearest neighbor algorithm to find the k nearest neighbor points of each node. Connect each node with its k nearest neighbor points to obtain the edge connection of the spatial graph to characterize the production line fault type information.

[0186] c) The 4D data features are embedded as node features of the spatial graph to obtain the production line spatial graph structure corresponding to the monitoring period.

[0187] (3) Construct a knowledge graph structure.

[0188] a) The monitoring period of the secondary data is regarded as a node, and the monitoring time characteristics of the secondary data after the above preprocessing are used as node characteristics.

[0189] b) Use the Mahalanobis distance function to calculate the similarity between nodes, and connect each node with its k nearest neighbors to obtain a knowledge graph.

[0190] The graph data is represented as G = {V, E, A, F}, where V represents the sample node of the graph data, E represents the edge connection between two sample nodes, F represents the sample node features, and A represents the adjacency matrix. For each graph data, the adjacency matrix corresponding to the graph data is used to represent the connection between all sample nodes.

[0191] Step 3: Build a production line fault diagnosis model.

[0192] Construct and train a chemical production line fault diagnosis model based on graph feature fusion. Specifically, the fault diagnosis model consists of three layers of ChebGCN convolutional layers and one layer of softmax fully connected layer. The model is trained under supervised conditions, and the fault state corresponding to each sample node is the label of the sample node.

[0193] Among them, 70% of the data samples are divided as a training set for model training, and the remaining 30% of the data samples are used as a test set.

[0194] Step 4: Production line fault diagnosis.

[0195] The graph data of online data obtained after conversion of the production line monitoring data obtained online is input into the fault diagnosis model. On the basis of considering the structural correlation between nodes, the graph data structure information and node attribute information are learned at the same time to obtain the final node feature representation and output the fault label corresponding to each node.

[0196] like Figure 8 As shown, when F.nll_loss is used as the loss function, the overall accuracy of the fault diagnosis model is 93.27%, and the confusion matrix of each classification accuracy is as follows Figure 8 As shown in the figure, the overall accuracy rate reached 93%, but due to the imbalanced data, the accuracy rates of each category varied greatly: the accuracy rates of categories 0 to 3, which had a larger number of samples, were higher, which increased the overall accuracy rate; the accuracy rates of categories 4 and 5, which had a smaller number of samples, were lower, at 85% and 74% respectively. However, due to the small number of samples, it was difficult to affect the overall accuracy rate.

[0197] like Fig. 9 As shown, in order to solve the problem of inaccurate classification of some categories caused by the imbalance of sample data, the focal loss function is used to improve the classification accuracy of the model. The confusion matrix of the classification accuracy of each category is as follows Fig. 9 As shown. It can be seen that the overall accuracy of the model has increased to 95%, and the accuracy of category 4 and category 5, which have small sample problems, has been greatly improved. The accuracy of category 4 has increased from 85% to 91%, and the accuracy of category 5 has increased from 76% to 94%. This shows that the fault diagnosis model of this embodiment has better solved the data imbalance problem in the chemical production line.

[0198] Embodiment 4

[0199] The fourth embodiment of the present invention provides a specific method for setting hyperparameters.

[0200] (1) Setting of neighbor aggregation parameter k.

[0201] According to the spatial layout of different process industrial systems, the number of neighbor nodes k aggregated by the spatial graph is an important parameter. An appropriate k value can ensure that nodes from the same process are included without aggregating nodes too far away. To this end, this embodiment conducted 10 experiments with different orders of k (k = 2 to 7) and compared their average results.

[0202] The experimental results are as follows Fig.10 As shown in the figure, when k is 2 to 3, the classification accuracy increases, and when k is 3 to 7, the classification accuracy decreases. Therefore, the number of sensors in the key step can be set to 3. The larger the k value, the more irrelevant sensors will be aggregated, resulting in reduced accuracy. When k = 3, the accuracy is the best in all experiments, reaching 93.14%. Therefore, this experiment selects the neighbor aggregation parameter k of the spatial graph as 3.

[0203] (2) Setting of parameter K.

[0204] The parameter K is the number of terms in the Chebyshev polynomial (that is, the order of the Chebyshev polynomial), which determines the aggregation of (K-1) hop neighbor information in the ChebGCN network. A small K value will limit the network's ability to mine information from graph structure data. At the same time, an increase in the K value will greatly increase the amount of information, which will consume a lot of resources. Therefore, an appropriate K value can balance the effect and computational efficiency.

[0205] When K = 1, the network only learns from the features of the node itself, and the network degenerates into an ordinary convolutional neural network. In order to determine the optimal number of terms K of the model, 10 batches of experiments were set up for hyperparameter selection.

[0206] Experimental Results Fig.11 As shown in the figure, when K is 2 to 4, the diagnostic accuracy is improved to 94.32%. When K = 5, the accuracy decreases because the fifth-order adjacent points contain information of different process indications. When K continues to rise to 6 and 7, the fluctuation of the model diagnostic accuracy converges, but the accuracy does not improve significantly. Considering that the increase of K will consume a lot of computing resources, K is selected as 4 in this experiment.

[0207] Embodiment 5

[0208] In order to verify the effectiveness of the present invention, the present invention is compared with five classic production line fault diagnosis methods, including principal component analysis + linear discriminant analysis (PCA + LDA), convolutional neural network (CNN), graph convolutional neural network (GCN), Chebyshev graph convolutional neural network (ChebGCN) and graph attention network (GAT). The parameter settings of each model are shown in Table 4.

[0209] Table 4 Parameter settings of different models

[0210]

[0211]

[0212] The production line fault diagnosis results of the above method are as follows: Fig.12 As shown, it can be seen that the traditional statistical learning method (LDA+PCA) is difficult to learn the characteristics of high-dimensional nonlinear systems, resulting in a diagnostic accuracy as low as 63.24%. The method based on deep learning (CNN) has a good diagnostic accuracy (82.14%), indicating that the method based on deep learning has a good nonlinear fitting ability. In addition, the model based on graph convolutional network can obtain a better diagnostic accuracy of about 84.23%-87.76%. This shows that signal features can be enhanced by converting signals into graph data. The model based on graph feature fusion proposed in the present invention improves the traditional graph-based deep learning method, extracts key points from prior knowledge, enhances the fault representation ability of the graph, and achieves the highest diagnostic accuracy (94.53%).

[0213] Embodiment 6

[0214] The sixth embodiment of the present invention further provides an electronic device, see Fig.13 The electronic device includes a processor 201 and a memory 202, in which at least one computer program is stored. The at least one computer program is loaded and executed by one or more of the above processors to enable the computer to implement the chemical production line health management method based on graph feature fusion described in the first embodiment.

[0215] Of course, the electronic device may also have components such as a wired or wireless network interface, a keyboard, and an input / output interface for input and output. The electronic device may also include other components for realizing various functions of the device, which will not be described in detail here.

[0216] Embodiment 7

[0217] The seventh embodiment of the present invention also provides a computer-readable storage medium, which stores at least one program code. The program code is loaded and executed by a processor to enable the computer to implement the chemical production line health management method based on graph feature fusion described in the first embodiment.

[0218] Optionally, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc (CD-ROM), a magnetic tape, a floppy disk, and an optical data storage device. Those skilled in the art may understand that all or part of the steps in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a program, and the program is stored in a storage medium, including several instructions to enable a single-chip microcomputer, a chip or a processor to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and other media that can store program codes.

[0219] In summary, the present invention starts from the process prior knowledge of the process industrial chemical production line itself, and proposes a chemical production line health management method based on graph feature fusion. According to the different sources of monitoring data, it is divided into two types of information: production line process variable monitoring information and production line product offline detection information; on this basis, according to the different properties of the two types of data, targeted composition processing is carried out, and different types of signal features are differentially mined to reduce the mutual confusion between the two in the model learning process; further, the feature vectors learned from the two types of sub-graphs are fused by the attention mechanism to obtain the final fault representation, so as to perform fault diagnosis. The public data set of the process industrial chemical production line process is used to verify the effectiveness of the health management method of the present invention, and provide technical support for effectively solving the small sample problem in the production line fault diagnosis application.

[0220] The preferred embodiments of the present invention are described in detail above in conjunction with the accompanying drawings. However, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, a variety of simple modifications can be made to the technical solution of the present invention, and these simple modifications all belong to the protection scope of the present invention.

[0221] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, the present invention will not further describe various possible combinations.

[0222] In addition, various embodiments of the present invention may be arbitrarily combined, and as long as they do not violate the concept of the present invention, they should also be regarded as the contents disclosed by the present invention.

Claims

1. A chemical production line health management method based on graph feature fusion, characterized in that: The health management method comprises: Classify and process the multi-source heterogeneous historical data of the chemical production line to obtain primary data and secondary data, wherein the primary data includes the production line operating data obtained by each sensor of the chemical production line, and the secondary data includes process product test data; The primary data and the secondary data are converted and processed to obtain a graph data set of historical samples, wherein the graph data set of historical samples includes a production line space graph data set and a production line knowledge graph data set, and the mathematical representation of the graph data is G={V, E, A, F}, wherein G is graph data, V is a node set, E is a sample edge set, A is an adjacency matrix, and F is a node feature matrix; Training a graph convolutional neural network based on the graph data set of the historical samples to obtain a production line fault diagnosis model; The online measurement data of the chemical production line is converted and processed to obtain a graph data set of online samples; The graph data set of the online samples is input into the production line fault diagnosis model to obtain the chemical production line fault diagnosis result.

2. The chemical production line health management method based on graph feature fusion according to claim 1 is characterized in that: The process of converting the primary data and the secondary data as sample data to obtain a graph data set of historical samples includes: Using a sliding window with a length of m, the primary data and the secondary data are divided into a plurality of first-category sample monitoring periods and a plurality of second-category sample monitoring periods respectively; Based on the first type of sample monitoring period and the physical space layout information of the production line sensors, a production line space map dataset is constructed; Based on the monitoring period of the second type of samples and the similarity of monitoring signals, the production line knowledge graph data is constructed; Based on the production line space graph dataset and the production line knowledge graph data, a graph dataset of historical samples is constructed.

3. The chemical production line health management method based on graph feature fusion according to claim 2 is characterized in that: The construction of a production line space map dataset based on the first type of sample monitoring period and the physical space layout information of the production line sensors includes: Each sensor is identified as a first-class node; Based on the physical space layout information of the production line sensors, several neighboring points around each first-category node are determined; Connect each first-category node with several neighboring nodes around it to obtain the edge connection relationship of each first-category node; Based on the edge connection relationship between each first-category node, determine the adjacency matrix of the first-level data; Preprocess the first-category sample monitoring period, and split each first-category sample monitoring period after preprocessing into multiple data columns based on the number of production line sensors; Determine each data column after the split as a first-category node feature corresponding to a first-category node; Based on several first-class nodes and their corresponding first-class node features, edge connection relationships, and adjacency matrices at the current monitoring moment, a production line spatial subgraph at the current monitoring moment is constructed; According to the corresponding production line space subgraphs at different monitoring times, a production line space graph dataset is constructed.

4. The chemical production line health management method based on graph feature fusion according to claim 3 is characterized in that: The method of determining a number of neighboring points around each first-category node based on the physical space layout information of the production line sensors includes: Based on the physical space layout information of the production line sensors, multiple first-class nodes are mapped to a rectangular coordinate system to obtain the spatial coordinates and relative position relationship of each first-class node; Determine the spatial distance between each first-category node based on the Euclidean distance function; Based on the spatial distances between the first-category nodes, a number of neighboring points around each first-category node are determined.

5. The chemical production line health management method based on graph feature fusion according to claim 3 is characterized in that: The preprocessing of the first type of sample monitoring period includes: Normalizing the first-category sample monitoring period to obtain the first-category sample monitoring period after normalization; The first-category sample monitoring period after the normalization processing is subjected to noise reduction processing to obtain the first-category sample monitoring period after the noise reduction processing.

6. The chemical production line health management method based on graph feature fusion according to claim 2 is characterized in that: The construction of production line knowledge graph data based on the second type of sample monitoring period and monitoring signal similarity includes: Each second-category sample monitoring period is determined as a second-category node; Normalizing the monitoring period of the second type of samples to obtain the normalized monitoring period of the second type of samples; Perform feature extraction on the normalized second-category sample monitoring period to obtain the second-category node features corresponding to each second-category node; Determine the similarity of monitoring signals between the second type nodes based on the Mahalanobis distance function; Based on the monitoring signal similarity between the second-type nodes and the second-type node characteristics, a number of neighboring points around each second-type node are determined; Connect each second-type node to several neighboring points around it to obtain the edge connection relationship of each second-type node; Based on the edge connection relationship between each second-type node, determine the adjacency matrix of the secondary data; A production line knowledge graph is constructed based on several second-category nodes corresponding to all production line monitoring moments and their corresponding second-category node features, edge connection relationships, and adjacency matrices.

7. The chemical production line health management method based on graph feature fusion according to claim 5 or 6, characterized in that: The calculation formula for the normalization process is: Among them, X scaled is the scaled data output after normalization; X is the input data; X.max is the maximum value of the elements in X; X.min is the minimum value of the elements in X; (max-min) is the data span.

8. The chemical production line health management method based on graph feature fusion according to claim 1 is characterized in that: The production line fault diagnosis model includes a spatial graph learning sub-model and a knowledge graph learning sub-model. The spatial graph learning sub-model includes three Chebyshev convolution layers and one graph pooling layer, and the knowledge graph learning sub-model includes three Chebyshev convolution layers and two fully connected layers.

9. An electronic device, characterized in that: The electronic device includes a processor and a memory, wherein at least one computer program is stored in the memory, and the at least one computer program is loaded and executed by one or more of the above processors so that the computer implements the chemical production line health management method described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that: At least one program code is stored in the computer-readable storage medium, and the program code is loaded and executed by the processor so that the computer implements the chemical production line health management method according to any one of claims 1 to 8.

Images