A chemical production line state recognition method and system

Abstract

The application provides a chemical production line state recognition method and system, comprising: determining the proximity of each sensor in the physical space based on the physical space layout of the chemical production line process flow; extracting the signal characteristics of each sensor based on the monitoring signals of each sensor in the current monitoring period; regarding each sensor as a node, and connecting the edges between the nodes based on the proximity of each sensor in the physical space and the similarity of the signal characteristics of each sensor to obtain a graph structure and its adjacency matrix; inputting the adjacency matrix of the graph structure and the node feature matrix composed of the signal characteristics of each sensor into a production line state recognition model to obtain the production line state recognition result corresponding to the current monitoring period. The application creates a graph structure by combining the monitoring signals with the production line process knowledge, and performs graph representation learning to realize knowledge-driven production line state recognition, thereby improving the accuracy of intelligent recognition of the state of the chemical production line.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of chemical production lines, and more particularly relates to a chemical production line state recognition method and system. BACKGROUND

[0002] With the continuous progress of process industry, modern chemical processes have become complex integrated systems involving multiple unit operations. To ensure production safety and quality, the stability of the system must be maintained, and higher requirements are placed on accurate control of the system. The key link of system control is fault detection and diagnosis (FDD, including anomaly detection and fault classification). Therefore, the FDD task in complex chemical processes has become an important research topic.

[0003] Modern chemical process industry relies on a large number of sensors to provide support for state analysis. This has led to an increasing number of applications of data-driven FDD models, and methods based on deep learning (DL) have achieved good results due to their excellent nonlinear learning capabilities, such as convolutional neural networks (CNN), long short-term memory networks (LSTM), and stacked autoencoders (SAE). However, pure data-driven models lack prior knowledge in the field, which limits the accuracy of production line state recognition. SUMMARY

[0004] In view of the defects of the prior art, the purpose of the present application is to provide a chemical production line state recognition method and system, which aims to solve the problem of low recognition accuracy of the existing chemical production line state recognition method.

[0005] To achieve the above-mentioned purpose, in a first aspect, the present application provides a chemical production line state recognition method, comprising:

[0006] Step S101, based on the physical space layout of the chemical production line process, determining the proximity of each sensor in the physical space;

[0007] Step S102, based on the monitoring signals of each sensor in the current monitoring period, extracting the signal features of each sensor;

[0008] Step S103, taking each sensor as a node, and based on the proximity of each sensor in the physical space and the similarity of the signal features of each sensor, connecting the edges between the nodes to obtain the graph structure and its adjacency matrix;

[0009] Step S104, inputting the adjacency matrix of the graph structure and the node feature matrix composed of the signal features of each sensor into a production line state recognition model to obtain the production line state recognition result corresponding to the current monitoring period; the production line state recognition model is trained based on the adjacency matrix of the sample graph structure, the sample node feature matrix, and the corresponding production line state label.

[0010] In an optional example, step S102 specifically includes:

[0011] Based on the monitoring signals of each sensor during the current monitoring period, the initial signal characteristics of each sensor are extracted;

[0012] Based on the reaction mechanism of the chemical production line process, the importance of each sensor is determined;

[0013] The signal characteristics of each sensor are determined based on the initial signal characteristics of each sensor and the weights assigned according to the importance of each sensor.

[0014] In an optional example, the initial signal characteristics of each sensor are obtained specifically based on the following steps:

[0015] The monitoring signals of each sensor during the current monitoring period are normalized and noise-reduced to obtain the pre-processed signals of each sensor.

[0016] Calculate the contribution of the preprocessed signal of each sensor, and use the contribution of the preprocessed signal of each sensor as the initial signal feature of each sensor.

[0017] In an optional example, the adjacency matrix of the graph structure is determined based on the following steps:

[0018] Based on the proximity of each sensor in physical space, each sensor is connected to other sensors that are close to it in physical space by connecting them with edges. The first adjacency matrix of the graph structure is determined based on the edge connection relationship between the nodes.

[0019] Based on the similarity of the signal features of each sensor, each sensor is connected to its nodes with those of similar signal features by edge connections. The second adjacency matrix of the graph structure is determined based on the similarity of the connected nodes.

[0020] Based on the first adjacency matrix and the second adjacency matrix, determine the adjacency matrix of the graph structure.

[0021] In an optional example, the production line status recognition model includes graph convolutional layers, graph pooling layers, fully connected layers, and a classifier constructed based on Chebyshev multinomials.

[0022] Secondly, the present invention provides a chemical production line status identification system, comprising:

[0023] The spatial proximity determination module is used to determine the physical spatial proximity of each sensor based on the physical spatial layout of the chemical production line process flow.

[0024] The node feature embedding module is used to extract the signal features of each sensor based on the monitoring signals of each sensor during the current monitoring period;

[0025] The graph structure representation module is used to treat each sensor as a node and connect the nodes by edges based on the proximity of each sensor in physical space and the similarity of the signal characteristics of each sensor, so as to obtain the graph structure and its adjacency matrix.

[0026] The production line status recognition module is used to input the adjacency matrix of the graph structure and the node feature matrix composed of the signal features of each sensor into the production line status recognition model to obtain the production line status recognition result corresponding to the current monitoring period. The production line status recognition model is trained based on the adjacency matrix of the sample graph structure, the sample node feature matrix, and the corresponding production line status label.

[0027] In an optional example, the node feature embedding module is specifically used for:

[0028] Based on the monitoring signals of each sensor during the current monitoring period, the initial signal characteristics of each sensor are extracted;

[0029] Based on the reaction mechanism of the chemical production line process, the importance of each sensor is determined;

[0030] The signal characteristics of each sensor are determined based on the initial signal characteristics of each sensor and the weights assigned according to the importance of each sensor.

[0031] In an optional example, the initial signal features of each sensor in the node feature embedding module are obtained based on the following steps:

[0032] The monitoring signals of each sensor during the current monitoring period are normalized and noise-reduced to obtain the pre-processed signals of each sensor.

[0033] Calculate the contribution of the preprocessed signal of each sensor, and use the contribution of the preprocessed signal of each sensor as the initial signal feature of each sensor.

[0034] In an optional example, the adjacency matrix of the graph structure in the graph structure representation module is determined based on the following steps:

[0035] Based on the proximity of each sensor in physical space, each sensor is connected to other sensors that are close to it in physical space by connecting them with edges. The first adjacency matrix of the graph structure is determined based on the edge connection relationship between the nodes.

[0036] Based on the similarity of the signal features of each sensor, each sensor is connected to its nodes with those of similar signal features by edge connections. The second adjacency matrix of the graph structure is determined based on the similarity of the connected nodes.

[0037] Based on the first adjacency matrix and the second adjacency matrix, determine the adjacency matrix of the graph structure.

[0038] In an optional example, the production line status recognition model in the production line status recognition module includes a graph convolutional layer, a graph pooling layer, a fully connected layer, and a classifier, all constructed based on Chebyshev multinomials.

[0039] In summary, the technical solutions conceived by this invention have the following beneficial effects compared with the prior art:

[0040] This invention provides a method and system for identifying the status of a chemical production line. By considering the physical spatial layout of sensors involved in the chemical production line process as graph nodes, and connecting nodes with spatially adjacent sensor nodes using edges, multiple heterogeneous sensor signals representing the same critical process can be jointly analyzed. The similarity between the signal features of the graph nodes is calculated, and a chemical production line status graph structure is constructed based on spatial proximity, enhancing the internal correlation between multi-source heterogeneous sensor monitoring data in the same critical process. The adjacency matrix and node feature matrix of the graph structure are input into the production line status identification model to mine high-order feature information characterizing the production line status. This invention achieves knowledge-driven production line status identification by creating a graph structure by combining monitoring signals with production line process knowledge and performing graph representation learning, thereby improving the accuracy of intelligent identification of chemical production line status. Attached Figure Description

[0041] Figure 1 This is one of the flowcharts illustrating the chemical production line status identification method provided in this embodiment of the invention;

[0042] Figure 2 This is a schematic diagram of the production line status recognition model provided in an embodiment of the present invention;

[0043] Figure 3 This is a second flowchart illustrating the chemical production line status identification method provided in this embodiment of the invention.

[0044] Figure 4 This is a schematic diagram illustrating the connection of nodes in the spatial layout of production line sensors provided in an embodiment of the present invention.

[0045] Figure 5 This is an experimental result diagram of the selection of the number of aggregated neighbor nodes in the spatial weighted graph provided in the embodiment of the present invention;

[0046] Figure 6 This is a graph showing the experimental results of order selection for Chebyshev convolution polynomials provided in an embodiment of the present invention;

[0047] Figure 7This is a comparison chart of the accuracy of chemical production line status identification between the method provided in this embodiment and existing methods;

[0048] Figure 8 This is an architecture diagram of the chemical production line status identification system provided in an embodiment of the present invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0050] Combining prior knowledge-based methods with deep learning techniques offers an effective approach to solving the aforementioned problems. However, manually selecting prior knowledge-based constraints weakens model convenience. Furthermore, utilizing the selected expert knowledge makes it difficult to effectively extract system features, and these rules often fail when the system state changes. Graph theory-based methods offer an effective solution to these problems.

[0051] In the chemical process industry, prior knowledge can be fused into graph structures. Graph models are used to learn the inherent prior knowledge of objects, enabling them to more accurately represent reality. However, these methods have limitations. Specifically, they tend to convert graph structures into fault vectors rather than directly mining graph information, leading to the loss of graph information and insufficient fault representation. In light of this, graph convolutional networks (GCNs) have been proposed, extracting signal features from discrete spatial structures through convolution operations on graphs.

[0052] However, these graphs are typically constructed from measured signals, rarely considering the underlying process knowledge of the system. Therefore, integrating process knowledge into genetic neural network models is a significant challenge. To address this, this invention utilizes the process flow to construct a graph, treating sensors as nodes and capturing the interrelationships between system signals through edge connections. However, constructing a graph using direct connections between all process units fails to accurately establish the internal relationships between different key processes and contains redundant information. Within the process flow, strong interrelationships exist between different sensors at the same key step, while sensor information from different steps generally exhibits weak correlations.

[0053] Therefore, how to use graph-structured data to enhance the internal connections between sensors in the same process and weaken the connections between sensors in different processes is worthy of further research.

[0054] To address this issue, the present invention provides a method for identifying the status of a chemical production line. Figure 1This is one of the flowcharts illustrating the chemical production line status identification method provided in this embodiment of the invention, such as... Figure 1 As shown, the method includes the following steps:

[0055] Step S101: Based on the physical spatial layout of the chemical production line process, determine the physical proximity of each sensor.

[0056] Step S102: Extract the signal features of each sensor based on the monitoring signals of each sensor during the current monitoring period;

[0057] Step S103: Treat each sensor as a node, and connect the nodes by edges based on the proximity of each sensor in physical space and the similarity of the signal characteristics of each sensor, to obtain the graph structure and its adjacency matrix.

[0058] Step S104: Input the adjacency matrix of the graph structure and the node feature matrix composed of the signal features of each sensor into the production line status recognition model to obtain the production line status recognition result corresponding to the current monitoring period; the production line status recognition model is trained based on the adjacency matrix of the sample graph structure, the sample node feature matrix, and the corresponding production line status label.

[0059] Specifically, the physical spatial layout of sensors in the chemical production line process can be directly mapped and transformed to a spatial rectangular coordinate system to calculate the spatial coordinates and relative positions of each sensor, thereby measuring the process correlation between multiple sensors. The production line sensors are regarded as nodes, and the K-nearest neighbor method is used to connect the nearby nodes according to the physical spatial proximity of the sensors, thereby introducing the spatial characteristics of sensor deployment.

[0060] On the other hand, after acquiring the monitoring signals from each sensor during the current monitoring period, the signal features of each sensor can be extracted, which serve as the node features of each node. Based on the similarity of the node features, nearby nodes are connected. Finally, combining the spatial characteristics of the sensor deployment and the similarity of the sensor signals, a graph structure is constructed, and its adjacency matrix is ​​obtained. By including the signal features of each node in the graph, the heterogeneous sensor measurement signals for each monitoring period in the spatial layout of the production line process flow are transformed into a graph structure corresponding to each monitoring period, thus transforming the chemical production line status identification problem into a graph classification problem.

[0061] Here, the current monitoring period can be determined based on the time accuracy required by the current diagnostic task. For example, if the current diagnostic task requires monitoring the production status of the production line once per hour, then the current monitoring period is one hour. It can be understood that if the sampling interval of the sensor is once every three minutes, then the monitoring signal of the current monitoring period will include the monitoring signals corresponding to 20 time points.

[0062] The adjacency matrix of the graph structure and the node feature matrix composed of signal features from each sensor are input into the production line status recognition model to obtain the production line status recognition result corresponding to the current monitoring period. The production line status recognition model is built based on a graph convolutional neural network and is obtained through supervised training based on the adjacency matrix of the sample graph structure, the sample node feature matrix, and the corresponding production line status labels. Here, the production line status recognition result can include whether a fault occurred on the production line during the monitoring period and the type of fault. The sample graph structure is constructed using historical monitoring data samples.

[0063] It should be noted that the production status of a chemical production line, including whether a fault has occurred and the different types of faults, is reflected in the signals monitored by sensors in the production line. Based on this, production line status identification can be performed. Existing GCN-based methods typically extract information equally from each node without distinguishing the importance of different units in the system. However, this invention considers that sensors for the same critical process often have similar physical spatial layouts in a chemical production line. For example, a critical process in the middle of the process may be monitored by three sensors: temperature, pressure, and flow. These three sensors describe the same critical process, but they are analyzed from different signal perspectives. Joint analysis of these sensors can greatly improve the ability of sensor signals to provide status information. Therefore, sensors describing the same process can be purposefully aggregated using K-nearest neighbors, and graph convolutional neural networks can aggregate information from connected nodes, thereby strengthening the internal connections of the same reaction process and jointly analyzing the production status of the chemical production line from the perspective of heterogeneous signals.

[0064] The method provided in this invention treats sensors as graph nodes based on the physical spatial layout of sensors involved in the chemical production line process. Each node is connected to its spatially adjacent sensor nodes with edges, thereby enabling joint analysis of multiple heterogeneous sensor signals representing the same critical process. The method calculates the similarity between the signal features of the graph nodes and constructs a chemical production line state graph structure based on spatial proximity, enhancing the internal correlation between multi-source heterogeneous sensor monitoring data in the same critical process. The adjacency matrix and node feature matrix of the graph structure are input into a production line state recognition model to mine high-order feature information characterizing the production line state. This invention achieves knowledge-driven production line state recognition by creating a graph structure by combining monitoring signals with production line process knowledge and performing graph representation learning, thus improving the accuracy of intelligent recognition of chemical production line states.

[0065] Based on the above embodiments, and considering that in practical applications, key chemical processes contain main monitoring units and auxiliary monitoring units of varying importance, there is an urgent need to study how to assign different weights to different types of sensors according to the physical characteristics of process industry systems. Therefore, this invention proposes a production line status identification method based on a spatially weighted graph representation of chemical production line processes. Specifically, the reaction mechanism derived from the process flow serves as the basis for weight allocation of different sensors, and the weighted features serve as the signal characteristics of the nodes; furthermore, the edge connection coefficients between nodes also need to be determined in conjunction with the similarity between the weighted features. Compared to an unweighted graph, the graph structure provided by this invention, namely the spatially weighted graph, provides a more accurate fault representation.

[0066] Step S102 specifically includes:

[0067] Based on the monitoring signals of each sensor during the current monitoring period, the initial signal characteristics of each sensor are extracted;

[0068] Based on the reaction mechanism of the chemical production line process, the importance of each sensor is determined;

[0069] The signal characteristics of each sensor are determined based on the initial signal characteristics of each sensor and the weights assigned according to the importance of each sensor.

[0070] It should be noted that, based on the reaction mechanism, sensors representing different production line states are assigned corresponding weights according to their importance; that is, different priorities are assigned to the signals monitored by different sensors. Key types of sensors are assigned higher weights to amplify their effectiveness, while auxiliary sensors should be assigned lower weights to weaken their influence. The product of the node weight and the initial signal characteristics is the final signal characteristic of each node, i.e., the node-weighted characteristic.

[0071] The entire chemical production line process performs specific production tasks, and the reactions are fixed. For example, in an exothermic reaction, the importance of each sensor can be: the temperature sensor is the most important, the pressure sensor is secondary, and the flow sensor is the least important; in a liquefaction reaction, the flow sensor is the most important.

[0072] In chemical production line processes, different sensors at the same critical step are closely interconnected, while sensor information from different steps generally has weak correlation. Therefore, by assigning different weights to different types of sensors based on the physical characteristics of the process industry system, determining the signal characteristics of nodes and the edge coefficients between nodes, and constructing a spatial weighted graph, the aim is to enhance the internal connections between sensors at the same critical step, weaken the connections between sensors in different processes, and improve the fault representation capability of critical nodes.

[0073] This invention incorporates prior knowledge from the process industry, including physical spatial layout and reaction mechanisms. First, it establishes connections between sensors describing the same critical process based on the production line layout for joint analysis. Second, it assigns different weights to different sensors based on reaction mechanism knowledge, enabling more targeted processing of sensor signals. Existing research treats sensors as equally important (i.e., all sensors have the same priority in monitoring indicators) or performs dimensionality reduction, relying on purely data-driven methods. This invention, however, integrates the physical spatial layout and reaction mechanisms of the production line, better aligning with the trends and conditions of the production line itself, resulting in more accurate identification.

[0074] Based on any of the above embodiments, considering that the signals monitored by heterogeneous sensors cannot be directly added, the initial signal characteristics of each sensor in the embodiments of the present invention are obtained based on the following steps:

[0075] The monitoring signals of each sensor during the current monitoring period are normalized and noise-reduced to obtain the pre-processed signals of each sensor.

[0076] Calculate the contribution of the preprocessed signal of each sensor, and use the contribution of the preprocessed signal of each sensor as the initial signal feature of each sensor.

[0077] Preferably, normalization methods can include Min-Max normalization, logarithmic normalization, etc. Noise reduction can be achieved using techniques such as principal component analysis (PCA) and kernel principal component analysis (KPCA) to project the original signal into a high-dimensional feature space for noise reduction. Considering the varying contributions of multi-source heterogeneous signals from the production line to production line status identification, XGBoost is used to calculate the contribution of each sensor's preprocessed signal to the production line status classification. The contribution of the sensor signal components is then used to replace the original signal, serving as a new node feature. XGBoost has the same input and output dimensions.

[0078] Based on any of the above embodiments, the adjacency matrix of the graph structure is determined based on the following steps:

[0079] Based on the proximity of each sensor in physical space, each sensor is connected to other sensors that are close to it in physical space by connecting them with edges. The first adjacency matrix of the graph structure is determined based on the edge connection relationship between the nodes.

[0080] Based on the similarity of the signal features of each sensor, each sensor is connected to its nodes with those of similar signal features by edge connections. The second adjacency matrix of the graph structure is determined based on the similarity of the connected nodes.

[0081] Based on the first adjacency matrix and the second adjacency matrix, determine the adjacency matrix of the graph structure.

[0082] Preferably, the distance between each node in physical space is calculated using Euclidean distance metric, and the k nearest neighbors algorithm is used to find the k nearest neighbors of each node. Each node is connected to its k nearest neighbors, and connected nodes are represented by 1 in the first adjacency matrix, while unconnected nodes are represented by 0 in the first adjacency matrix.

[0083] The cosine similarity function is used to measure the connection strength, i.e., similarity, between the multidimensional features of nodes, and serves as the edge connection coefficient. The K-nearest neighbor algorithm is used to find the k' nearest neighbors of each node, and each node is connected to its k' nearest neighbors. Connected nodes are represented by similarity in the second adjacency matrix, while unconnected nodes are represented by 0 in the first adjacency matrix. This differentiates the strength of the connection between heterogeneous sensors and enhances the correlation between important nodes.

[0084] By fusing spatial edge connectivity relationships and edge connectivity coefficients, the final graph structure, i.e., the spatially weighted graph, is constructed, and the adjacency matrix of the graph structure is obtained. Specifically, the fusion method can be to add corresponding elements from the first adjacency matrix and the second adjacency matrix.

[0085] Furthermore, the number of neighboring nodes, k, aggregated at the physical space level is an important parameter. An appropriate k value can include nodes from the same process without aggregating nodes that are too far apart. To this end, this invention conducted multiple experiments with different k values ​​based on sample monitoring signals and selected the optimal k value based on the final classification accuracy. A larger k value can lead to the aggregation of irrelevant sensors, ultimately reducing classification accuracy and increasing time consumption.

[0086] Based on any of the above embodiments, the production line status recognition model includes a graph convolutional layer, a graph pooling layer, a fully connected layer, and a classifier constructed based on Chebyshev polynomials.

[0087] Preferably, Figure 2 This is a schematic diagram of the production line status recognition model provided in an embodiment of the present invention, as shown below. Figure 2 As shown, the production line status recognition model can contain three graph convolutional layers constructed based on Chebyshev multinomials. The classifier can be a softmax classifier, and the parameters of the softmax classifier can be fixed during training.

[0088] This invention proposes a graph learning model for chemical production line status identification. It combines monitoring signals with production line process knowledge to create a graph structure and performs graph representation learning to achieve knowledge-driven production line status identification. Specifically, a spatially weighted graph is constructed based on the spatial dependencies in the sensor layout of the chemical production line and the signal correlations determined by the reaction mechanism. On this basis, a graph convolutional neural network model is used to learn both graph data structure information and node attribute information, considering the structural correlations between nodes, to obtain the final graph node feature representation, i.e., the production line status representation vector. This vector is then fed into a softmax classifier to output the production line status identification result.

[0089] Based on any of the above embodiments Figure 3 This is a second flowchart illustrating the chemical production line status identification method provided in this embodiment of the invention, as shown below. Figure 3 As shown, the method includes the following steps:

[0090] Step S1: Based on the physical spatial layout of the sensors in the chemical production line process, directly map and transform them to a spatial rectangular coordinate system to calculate the spatial coordinates and relative positions of each sensor, thereby measuring the process correlation between multiple sensors; the coordinate mapping process can be completed at the edge, and the two-dimensional spatial coordinate system positions of the sensors are transmitted to the cloud for graph data construction; the cloud treats the production line sensors as nodes, and connects each node according to the physical proximity of the sensors through the K-Nearest Neighbors (KNN) algorithm;

[0091] Specifically, the sensors to which the data belongs are considered nodes. Based on this, 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 then connected to its k nearest neighbors. The Euclidean distance calculation formula is as follows:

[0092]

[0093] In the formula, x i y i These are nodes v i The x-coordinate and y-coordinate in a rectangular coordinate system.

[0094] Step S2: Apply a sliding window to segment each single sensor sequence to obtain samples for the monitoring period. Perform noise reduction preprocessing on the segmented sample data. Use the XGBoost algorithm to identify the contribution of each sensor in the noise reduction dataset and establish it as the node feature embedding of the production line spatial weighted graph.

[0095] Specifically, using the n-dimensional important feature matrix obtained from monitoring, it is divided into n sequences, each corresponding to one of the n sensors. A sliding window of length m is used to divide each single sensor sequence into data segments of length m, where m is determined based on the time precision and sampling interval required for the current diagnostic task. The features of these nodes are included in the graph. The mean-filling method is used to align the inconsistent data intervals in the original monitoring data. When the last segment of the chemical production line monitoring sample is less than m in length, the mean-filling method is used to fill its length to m. Each single sensor sequence ultimately yields h segmented monitoring samples.

[0096] The monitoring samples from the chemical production line are normalized using the Min-Max normalization method. Further, Principal Component Analysis (PCA) denoising is performed to obtain a denoised dataset. Based on this, XGBoost is used to identify the contribution of each signal in the denoised dataset to the label classification. This process can be completed at the edge, and the contribution data is then transmitted to the cloud. The cloud then replaces the original signal with the signal component contribution as a new node feature L.

[0097] Furthermore, the Min-Max normalization formula is as follows:

[0098]

[0099] In the formula, x i nor This is the output after Min-Max normalization, where X is the original input data. max Let x be the maximum value of the elements in X. min Let X be the minimum value of the elements in X, and (max-min) be the data span.

[0100] Step S3: Based on the different importance of the multi-source heterogeneous sensor signals deployed in each process link to characterize the state of the chemical production line, assign corresponding weights to the sensors and obtain weighted node features; then determine the edge connection coefficients, further integrate the spatial edge connection relationships and edge connection coefficients, and finally construct a spatial weighted graph;

[0101] Specifically, based on the constructed spatial map, and according to the chemical reaction mechanism, sensors representing different production line states are assigned corresponding weights according to their importance, using w i This indicates that key sensor types should be assigned higher weights to amplify their effectiveness, while auxiliary sensors should be assigned lower weights to mitigate their impact. The product of node weights and contribution characteristics is the node weighted characteristic.

[0102] Construct the node weight vector W = (w1, w2, ..., w n ), w i∈[0,1], and based on this, the weighted feature matrix F = (f1,f2,…,f...) is obtained by multiplying the node features L with the node weight vector. n ), f i =l i ×w i Taking the node connection metric function represented by cosine similarity as an example, the cosine similarity function is used to measure the connection strength between the multi-dimensional features of nodes as the edge connection coefficient. The K-nearest neighbor algorithm is used to find the k' nearest neighbors of each node. Each node and its k' nearest neighbors are connected to obtain the edge connection of the spatial weighted graph. The spatial edge connection relationship and the edge connection coefficient are fused to construct the spatial weighted graph.

[0103] Furthermore, the cosine similarity formula is as follows:

[0104]

[0105] Furthermore, the mathematical representation of a graph structure is G = {V, E, A, F}. Where V = {V...} i} represents a set of nodes, indicating the sensors on the production line at a specific measurement moment; E is represented by an array of edges connecting the nodes; A = {a i,j Let} be the adjacency matrix, and the adjacency matrix is ​​a sparse matrix, where when a i,j = 0 indicates that there is no edge connection between node i and node j; otherwise, it indicates that there is an edge connection between the nodes. F is the weighted characteristic matrix.

[0106] Step S4: Establish a spatial weighted graph-driven production line status recognition model, and use historical online monitoring data of the production line to train the model to obtain a trained production line status recognition model;

[0107] Specifically, a production line status identification model based on a spatial weighted graph representation of the chemical production line process flow is established, and the model is trained using a sample graph structure constructed from historical monitoring data samples to obtain a trained production line status identification model.

[0108] The production line status recognition model based on the chemical production line process flow is a spatial weighted graph learning model. The model consists of three ChebGCN layers, one GraphPool layer, and one fully connected layer. After learning the original spatial weighted graph through three Chebyshev convolutions, max pooling is used to obtain the spatial representation vector of the production line status for the corresponding monitoring period. After learning, the fully connected layer outputs a production line status representation vector with a number of labels. The softmax classifier model takes as input the production line status representation vector and the probability of the status label at the production line detection time as outputs, respectively, ultimately calculating the chemical production line status recognition result.

[0109] The loss function of the production line status identification model is the cross-entropy cost function. The status identification results include: whether a fault occurred during the monitoring period and the type of fault.

[0110] Step S5: Perform the above transformation process on the real-time online monitoring data of the production line to obtain a spatially weighted map of the real-time online data. Input the obtained real-time online map data of the production line into the trained production line status recognition model for classification and recognition to obtain the real-time production line status during the online monitoring period.

[0111] This invention discloses a method for identifying the state of a chemical production line based on a weighted graph representation of the process flow. The method includes: first, treating sensors as graph nodes according to the physical spatial layout of sensors involved in the chemical production line process flow; second, classifying heterogeneous sensors in the process flow into different importance levels based on the reaction mechanisms of each process step, assigning different weight coefficients to the sensor nodes, and calculating the weighted features of the graph nodes; further, calculating the similarity between the weighted features of the graph nodes to determine the node edge connection coefficients, and constructing a weighted graph of the chemical production line state space to enhance the internal correlation between multi-source heterogeneous sensor monitoring data in the same key process flow; finally, inputting the constructed weighted graph of the chemical production line state space into a graph convolutional neural network model to mine high-order feature information representing the characteristics of the production line state. This invention improves the intelligent identification level of chemical production line states by integrating knowledge of the chemical production line process flow domain and combining it with a weighted graph representation learning method.

[0112] Experiment 1: Validation

[0113] The Tennessee Eastman Process (TEP) is a simulation model of an actual chemical process. Due to its close resemblance to actual production, it is widely used in health monitoring and condition identification model evaluation in process industries. The process includes 41 different monitoring devices, encompassing 22 system monitoring variables such as temperature, pressure, and flow rate, and 19 product detection variables such as acidity, purity, and proportions. The monitoring variables generated by the TEP model exhibit nonlinear, dynamic, and strongly coupled characteristics. Furthermore, the model operation does not include the start-up and shutdown phases of the complex system. In this study, the simulation model was set to mode_1 (baseline), and all 12 manual variables were initialized. Based on this, 10 production states were simulated, including the normal state and 9 different process disturbances, representing typical failures that occur in reality. Each simulation lasted 48 hours, with a sampling interval of 3 minutes, generating a total of 960 data points. It is important to note that reasonable upper and lower limits for operation were set for each device in the simulation model. Once certain limits are reached, the reaction will stop to protect this complex system. Of all the disturbance types described, interruption of the feed flow (A) triggers the shutdown of the complex system. All samples from fault states and normal states were mixed to obtain a total of 9600 data points. 70% of the samples were randomly selected as the training set, and 30% as the test set. The fault types of the Tennessee Eastman process are shown in Table 1.

[0114] Table 1

[0115]

[0116] The specific steps are as follows:

[0117] (1) Data preprocessing and feature extraction

[0118] The production line data obtained from the segmentation were normalized. Principal Component Analysis (PCA) was then applied to the normalized data to obtain a denoised dataset, with an output dimension of 22. Based on this, XGBoost was used to identify the importance of each signal in the label classification. Then, the importance, or contribution, of each signal component was used to replace the original signal as a new feature.

[0119] (2) Sensor spatial layout in production line

[0120] The sensors to which the production line data belongs are considered as nodes. Based on the physical spatial layout of the production line sensors, they are mapped to a Cartesian coordinate system to obtain the spatial coordinates and relative positional relationships of each node. 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 then connected to its k nearest neighbors, incorporating the spatial characteristics of the sensors. A schematic diagram of the node edge connection determination based on the spatial layout of the production line sensors provided in this embodiment is shown below. Figure 4As shown, the physical space layout includes flow rate, pressure, and temperature sensors. Sensors 1-13 in the figure represent flow rate sensors.

[0121] (3) Construction of spatial weighted graph structure

[0122] Sensors are treated as nodes, and weights are assigned according to their importance. A spatially weighted graph is constructed by multiplying the weight vector with the original node vector features. Cosine similarity is used to measure the connection strength between the multidimensional features of nodes, which is the edge connection coefficient. Each node is connected to its k nearest neighbors, fusing the spatial edge connection relationships and edge connection coefficients. The connections between weighted nodes are determined by the process flow reaction mechanism and can reflect the product production status.

[0123] (4) Production line status identification model construction reaction mechanism

[0124] A chemical production line status identification model based on a process flow spatial weighted graph representation was constructed and trained. Specifically, the chemical production line status identification model consists of three ChebGCN convolutional layers and one GraphPool layer. The GraphPool uses max pooling output. The model was trained under supervised conditions, with the fault state corresponding to each monitoring period serving as the data label for that period. Considering the absence of data imbalance among samples, the accuracy metric was used to evaluate the model's diagnostic precision, calculated using the following formula:

[0125]

[0126] (5) Production line status identification

[0127] The production line monitoring signal data obtained online is processed through steps (1), (2), and (3) to obtain a spatially weighted graph, which is then substituted into the production line status identification model constructed in step (4). Based on the structural correlation between nodes, the graph data structure information and node attribute information are learned simultaneously to obtain the final fault feature representation, and the status label corresponding to the data at each time point is output.

[0128] Experiment 2: Hyperparameter Settings

[0129] The number of neighboring nodes, k, aggregated in a spatially weighted graph is an important parameter depending on the spatial layout of different process industry systems. An appropriate value of k can include nodes from the same process without aggregating nodes that are too far apart. To this end, this invention conducted 10 experiments with different orders of k (k = 2–7) and compared the average results. The experimental results of selecting the number of neighboring nodes k in the spatially weighted graph provided in this embodiment are shown in the figure below. Figure 5As shown, classification accuracy improves when k increases from 2 to 3, but performance degrades when k increases from 3 to 7. This observation indicates that setting the number of sensors in the critical steps to 3 yields the best results, as larger k values ​​lead to the clustering of irrelevant sensors, ultimately reducing classification accuracy. Specifically, the accuracy is highest at k=3, reaching 93.14%. Therefore, the parameter k for the spatially weighted graph is set to 3 in this experiment. It is worth noting that larger k values ​​result in greater time consumption, which is essentially the cost of constructing the graph space structure to improve model accuracy.

[0130] The parameter T is the number of terms in the Chebyshev polynomial, which determines the aggregation of (T-1) hop neighbor information in the ChebGCN network. A value that is too small will limit the network's ability to extract information from graph-structured data. At the same time, increasing the value of T will significantly increase the amount of information, which will consume a lot of resources. Therefore, an appropriate value of T can balance performance and computational efficiency. When T=1, the network learns only from the features of the nodes themselves, at which point the network degenerates into a regular convolutional neural network. To determine the optimal number of terms T for the model, 10 batches of experiments were conducted for hyperparameter selection. The experimental results of selecting the order of the Chebyshev convolutional polynomial T provided in this embodiment of the invention are shown in the figure below. Figure 6 As shown, when T is between 2 and 4, the diagnostic accuracy increases to 94.32%. When T = 5, the accuracy decreases because the fifth-order adjacent points contain information about different process indications. When T continues to increase to 6 and 7, the fluctuation of the model's diagnostic accuracy converges somewhat, but the accuracy does not improve significantly. Considering that increasing T would lead to a large consumption of computational resources, this experiment chose T = 4.

[0131] Experiment 3: Comparison with other chemical production line status identification methods

[0132] To verify the effectiveness of this invention, the Spatial Weighted Graph (SWG) method of this invention was compared with five classic production line status recognition methods, including Principal Component Analysis + Linear Discriminant Analysis (LDA+PCA), Convolutional Neural Network (CNN), Graph Convolutional Network (GCN), Chebyshev Graph Convolutional Network (ChebGCN), and Graph Attention Network (GAT). The accuracy comparison of the chemical production line status recognition method provided in this embodiment with existing methods is shown in the figure below. Figure 7As shown, traditional statistical learning methods (LDA+PCA) struggle to learn the characteristics of high-dimensional nonlinear systems, resulting in a diagnostic accuracy as low as 63.24%. Deep learning-based methods (CNN) achieve a better diagnostic accuracy (84.23%), indicating that deep learning-based methods have better nonlinear fitting capabilities. Furthermore, the graph convolutional network (GCN) model achieves an even better diagnostic accuracy of approximately 87.76%. This suggests that signal features can be enhanced by converting signals into graphical data. This invention (SWG) proposes a chemical production line condition identification model based on a spatially weighted graph representation of the process flow, improving upon traditional graph-based deep learning methods. It extracts key points from prior knowledge, enhancing the fault representation capability of the graph, achieving the highest diagnostic accuracy (94.53%) on the TEP dataset. It is noteworthy that the node weights in the GAT network are learned from the original graph and are independent of the graph's structure. However, the results show that node weights generated by incorporating prior knowledge are more effective than simply learning node weights from monitoring data.

[0133] Based on any of the above embodiments, the present invention provides a chemical production line status identification system. Figure 8 This is an architecture diagram of the chemical production line status identification system provided in an embodiment of the present invention, such as... Figure 8 As shown, the system includes:

[0134] The spatial proximity determination module 810 is used to determine the physical spatial proximity of each sensor based on the physical spatial layout of the chemical production line process flow.

[0135] The node feature embedding module 820 is used to extract the signal features of each sensor based on the monitoring signals of each sensor during the current monitoring period;

[0136] The graph structure representation module 830 is used to treat each sensor as a node and connect the nodes by edges based on the proximity of each sensor in physical space and the similarity of the signal characteristics of each sensor, so as to obtain the graph structure and its adjacency matrix.

[0137] The production line status recognition module 840 is used to input the adjacency matrix of the graph structure and the node feature matrix composed of the signal features of each sensor into the production line status recognition model to obtain the production line status recognition result corresponding to the current monitoring period. The production line status recognition model is trained based on the adjacency matrix of the sample graph structure, the sample node feature matrix, and the corresponding production line status label.

[0138] It is understood that the detailed functional implementation of each of the above modules can be found in the description of the aforementioned method embodiments, and will not be repeated here.

[0139] In addition, this invention provides another chemical production line status identification device, which includes: a memory and a processor;

[0140] The memory is used to store computer programs;

[0141] The processor is configured to implement the methods described in the above embodiments when executing the computer program.

[0142] In addition, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the methods described in the above embodiments.

[0143] Based on the methods in the above embodiments, this embodiment of the invention provides a computer program product that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0144] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying the status of a chemical production line, characterized in that, include: Step S101: Based on the physical spatial layout of the chemical production line process, determine the physical proximity of each sensor. Step S102: Extract the signal features of each sensor based on the monitoring signals of each sensor during the current monitoring period; Step S103: Treat each sensor as a node, and connect the nodes by edges based on the proximity of each sensor in physical space and the similarity of the signal characteristics of each sensor, to obtain the graph structure and its adjacency matrix. Step S104: Input the adjacency matrix of the graph structure and the node feature matrix composed of the signal features of each sensor into the production line status recognition model to obtain the production line status recognition result corresponding to the current monitoring period; the production line status recognition model is trained based on the adjacency matrix of the sample graph structure, the sample node feature matrix, and the corresponding production line status label. Specifically, step S102 includes: Based on the monitoring signals of each sensor during the current monitoring period, and based on each monitoring signal, the preprocessed signal of each sensor is determined; Calculate the contribution of the preprocessed signal of each sensor, and use the contribution of the preprocessed signal of each sensor as the initial signal feature of each sensor. Based on the reaction mechanism of the chemical production line process, the importance of each sensor is determined; The signal characteristics of each sensor are determined based on the initial signal characteristics of each sensor and the weights assigned according to the importance of each sensor.

2. The method according to claim 1, characterized in that, The preprocessed signals of each sensor are obtained based on the following steps: The monitoring signals are normalized and noise reduced to obtain the preprocessed signals of each sensor.

3. The method according to claim 1, characterized in that, The adjacency matrix of the graph structure is determined based on the following steps: Based on the proximity of each sensor in physical space, each sensor is connected to other sensors that are close to it in physical space by connecting them with edges. The first adjacency matrix of the graph structure is determined based on the edge connection relationship between the nodes. Based on the similarity of the signal features of each sensor, each sensor is connected to its nodes with those of similar signal features by edge connections. The second adjacency matrix of the graph structure is determined based on the similarity of the connected nodes. Based on the first adjacency matrix and the second adjacency matrix, determine the adjacency matrix of the graph structure.

4. The method according to any one of claims 1 to 3, characterized in that, The production line status recognition model includes graph convolutional layers, graph pooling layers, fully connected layers, and a classifier, all constructed based on Chebyshev multinomials.

5. A chemical production line status identification system, characterized in that, include: The spatial proximity determination module is used to determine the physical spatial proximity of each sensor based on the physical spatial layout of the chemical production line process flow. The node feature embedding module is used to extract the signal features of each sensor based on the monitoring signals of each sensor during the current monitoring period; The graph structure representation module is used to treat each sensor as a node and connect the nodes by edges based on the proximity of each sensor in physical space and the similarity of the signal characteristics of each sensor, so as to obtain the graph structure and its adjacency matrix. The production line status recognition module is used to input the adjacency matrix of the graph structure and the node feature matrix composed of the signal features of each sensor into the production line status recognition model to obtain the production line status recognition result corresponding to the current monitoring period. The production line status recognition model is trained based on the adjacency matrix of the sample graph structure, the sample node feature matrix, and the corresponding production line status label. The node feature embedding module is specifically used to determine the preprocessed signal of each sensor based on the monitoring signal of each sensor in the current monitoring period; calculate the contribution of the preprocessed signal of each sensor and use the contribution of the preprocessed signal of each sensor as the initial signal feature of each sensor; determine the importance of each sensor based on the reaction mechanism of the chemical production line process; and determine the signal feature of each sensor based on the initial signal feature of each sensor and the weight assigned according to the importance of each sensor.

6. The system according to claim 5, characterized in that, The preprocessed signals of each sensor in the node feature embedding module are obtained based on the following steps: The monitoring signals are normalized and noise reduced to obtain the preprocessed signals of each sensor.

7. The system according to claim 5, characterized in that, The adjacency matrix of the graph structure in the graph structure representation module is determined based on the following steps: Based on the proximity of each sensor in physical space, each sensor is connected to other sensors that are close to it in physical space by connecting them with edges. The first adjacency matrix of the graph structure is determined based on the edge connection relationship between the nodes. Based on the similarity of the signal features of each sensor, each sensor is connected to its nodes with those of similar signal features by edge connections. The second adjacency matrix of the graph structure is determined based on the similarity of the connected nodes. Based on the first adjacency matrix and the second adjacency matrix, determine the adjacency matrix of the graph structure.

8. The system according to any one of claims 5 to 7, characterized in that, The production line status recognition module includes a production line status recognition model comprising a graph convolutional layer, a graph pooling layer, a fully connected layer, and a classifier, all constructed based on Chebyshev multinomials.

Images