A double-channel coal chemical fault diagnosis method based on graph space-time fusion modeling
By combining graph-temporal fusion modeling with GCN and LSTM, a knowledge graph was constructed and a dual-supervised training network was designed. This solved the problem of fault identification in the coal-to-ethylene glycol process, enabling rapid and accurate fault diagnosis and classification, and improving production safety and efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING TECH UNIV
- Filing Date
- 2026-04-17
- Publication Date
- 2026-07-07
AI Technical Summary
Existing technologies are insufficient to quickly and accurately identify faults in the complex coal-to-ethylene glycol process, which may lead to the paralysis of the production system or cause catastrophic consequences. There is a lack of effective early warning and fault diagnosis methods.
A graph-based spatiotemporal fusion modeling approach is adopted, combining graph convolutional neural networks (GCN) and long short-term memory neural networks (LSTM). Spatiotemporal features are extracted by constructing a knowledge graph, and a dual-supervised training network is designed for fault diagnosis, including anomaly classification and fault diagnosis classification.
It enables a systematic description and fault characteristic representation of coal chemical processes, allowing for rapid fault identification and accurate classification, reducing equipment downtime, and improving production safety and efficiency.
Smart Images
Figure CN122045782B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of chemical engineering, fault prediction and optimization, and specifically, but not limited to, a dual-channel coal chemical fault diagnosis method based on graph spatiotemporal fusion modeling. Background Technology
[0002] In the anomaly management of the coal-to-ethylene glycol process, fault detection and diagnosis play a central role. The technological processes and control systems involved in chemical production are intricate, with close interactions between various variables. Even minor faults can trigger a chain reaction, leading to the paralysis of the entire production system and potentially catastrophic consequences. Therefore, implementing accurate early warning and rapid, effective fault diagnosis is crucial for ensuring the safe and stable operation of chemical production. Fault detection and diagnosis are not only key technologies for ensuring the safe and efficient operation of the coal-to-ethylene glycol process, but also have significant theoretical and practical value for improving the level of intelligent manufacturing in the chemical industry.
[0003] With the increasing complexity of industrial systems, fault diagnosis in modern coal-to-ethylene glycol processes faces new challenges. There is an urgent need to develop new methods capable of rapidly and accurately identifying faults under numerous monitored variables and dynamic interactions. The focus of fault detection and diagnosis should be on identifying anomalies in the overall process, rather than being limited to alarms for single variables. This is crucial for monitoring safety status, responding to potential faults, and predicting risks in coal-to-ethylene glycol processes.
[0004] Specifically, fault detection aims to determine whether a system is in a faulty state, while fault diagnosis, after detecting a fault, further determines the nature of the fault, including its root cause, location, scale, timing, and impact. The main tasks of process fault diagnosis include fault detection, fault identification, fault isolation, and fault recovery. In practical applications, this process can be viewed as a multi-classification problem, relying on sensor signals to accurately diagnose various potential system faults. By adopting efficient fault detection and diagnosis strategies, equipment downtime can be significantly reduced, factory operational safety can be improved, and production costs can be effectively lowered. Summary of the Invention
[0005] To address the problems in existing technologies, this invention proposes a dual-channel coal chemical fault diagnosis method based on graph spatiotemporal fusion modeling. It combines graph convolutional neural networks with long short-term memory neural networks and incorporates fully connected layers into the framework to obtain the spatiotemporal features of coal-to-ethylene glycol process data. This method employs a dual-supervised training strategy to learn the fused features and classifies process data for anomalies, ultimately learning specific fault categories.
[0006] The technical solution to achieve the purpose of this invention is as follows:
[0007] A dual-channel coal chemical fault diagnosis method based on graph spatiotemporal fusion modeling includes the following:
[0008] S1. Establish a steady-state model of the entire coal-to-ethylene glycol process, and obtain a dynamic model of the coal-to-ethylene glycol process through the steady-state model; introduce faults by artificially changing the output mode of the controller, and obtain a fault time series dataset of the coal-to-ethylene glycol process.
[0009] S2. Construct a knowledge graph based on prior knowledge of the coal-to-ethylene glycol process, and design entities in a hierarchical manner, including physical entities. and variable entities The types of relationships are divided into physical connections, state associations, and material composition;
[0010] S3. Based on knowledge graphs, a graph feature extraction module is constructed using a graph convolutional neural network (GCN) to extract high-level feature representations that integrate graph structure information and node attribute information. Then, a long short-term memory neural network (LSTM) is used to encode the features temporally as temporal features to capture dependencies in time series data. Simultaneously, spatial features are extracted from the coal-to-ethylene glycol process fault time series dataset using a fully connected layer network. Finally, spatial and temporal features are fused.
[0011] S4. Design a dual-supervised training network by combining two multilayer perceptrons, including an anomaly classifier and a fault diagnosis classifier. The anomaly classifier learns whether the input data is abnormal by using the original fault time series data, and the fused features are used as the input of the fault diagnosis classifier to finally obtain an accurate fault classification result.
[0012] Preferably, the fault simulation of the coal-to-ethylene glycol process in S1 is as follows:
[0013] S1-1 The coal-to-ethylene glycol process includes coal-to-syngas production, syngas purification and separation, oxalate synthesis, oxalate hydrogenation to ethylene glycol, and product refining steps. A steady-state model of the entire process is constructed.
[0014] S1-2. Manually introduce faults by changing the controller output to obtain a fault dataset for the coal-to-ethylene glycol process. The controllers include a flow controller, a temperature controller, a pressure controller, and a level controller.
[0015] Preferably, in S1-2, the fault is a failure of the PCR tower condenser, a failure of the PCR tower reflux pump, a failure of the MSC tower reflux pump, or a sudden increase in NO feed flow rate.
[0016] Preferably, the prior knowledge of the coal-to-ethylene glycol process in S2 is used to construct a knowledge graph, specifically as follows:
[0017] S2-1, Prior Knowledge Graph, represented as ,in It is an entity set. It is a set of relations. Indicates the form of The triple represents the head entity. Tail entity and relationships The connections between them; the entity set can be formally represented as:
[0018]
[0019] in, Represents a set of device entities. Represents the set of flowing entities. Represents a set of component entities. Represents a collection of variable entities;
[0020] S2-2. Relationships based on prior knowledge graphs are established through the interaction of head and tail entities, expressed as various types of relationships. Inherent connections exist between different physical entities, and variable entities represent the state representation relationships with their corresponding physical entities. The set of relationships is represented as follows:
[0021]
[0022] in, This indicates that the device outputs a specific stream; This indicates that a stream flows into a certain device; Indicates the components contained in the stream; The variable represents the operating status of the equipment or flow stream; Used to express the semantics of reaction and transformation direction;
[0023] S2-3. For the mapping from knowledge graph to process data, the set of process data collection points is as follows: The variable entity set is The mapping function from the measurement point to the variable entity is: , Indicates data collection points In the knowledge graph, the corresponding variable entities are attached to device entities or flow entities through State relationships to represent the operational state of the corresponding physical objects; represented as:
[0024]
[0025] in, Indicates a device entity or a flow entity. Represents variable entities, Represents a set of State relations;
[0026] S2-4. Achieve continuous characterization of the operating state while maintaining the stability of the process topology, based on the observation sequence of any variable entity. for:
[0027]
[0028] in, Indicates the sampling time. Represents the observed value of a variable at that moment; observation sequence The state attributes of the variable nodes are stored to depict the evolution of the variable over time.
[0029] S2-5. Data needs to be standardized before being written into the graph. The mean and standard deviation of the variables under normal operating conditions are as follows: and Variable standardization Represented as:
[0030]
[0031] in, A very small constant is introduced to prevent the denominator from being zero; through standardization transformation, different variables can be characterized on a unified scale to represent their degree of deviation from normal operating conditions, providing a basis for anomaly intensity calculation and subsequent state aggregation; in addition to the original observation value and standardized value, the variable node also stores derived state information including deviation intensity, anomaly label, duration, trend of change, and sliding window statistics.
[0032] Preferably, the graph feature extraction module, time encoding module, and spatial feature extraction module in S3 are as follows:
[0033] S3-1, In the graph feature extraction module, based on the observation sequence... The node feature matrix is constructed using the input features. Adjacency matrix constructed based on prior knowledge Combined, as input to GCN, where Indicates the number of nodes. This represents the feature dimension of each node; the initial node feature matrix represents the original feature input. Then the first The node feature matrix of layer +1 is represented as follows:
[0034]
[0035] in, Indicates the first The node feature matrix of the layer, Indicates the first Layer weight matrix; The degree matrix representing the nodes; Activation function; global graph features are extracted by stacking L layers of convolutions. ;
[0036] S3-2, Graph features output by graph convolution Convert to time series input; in the time dimension, graph embedding sequence. ;in, Indicates a time step. This represents the number of nodes in the current graph structure. This represents the feature dimension of each node. Indicates the first The graph embedding matrix corresponding to each time step;
[0037] Graph Embedding Sequences Perform time encoding: Use a multi-layer LSTM model to capture temporal dependencies:
[0038]
[0039] in, The hidden state at the current time step. It represents the dimension of the LSTM hidden state, indicating the graph features after temporal encoding; Indicates the input gate. Represents the Gate of Oblivion Indicates the output gate. Indicates the first The cell state at each time step. Indicates the first The hidden state at each time step and These are the learnable weight matrix and bias vector of the input gate, respectively. and Here are the learnable weight matrix and bias vector for the forget gate. and Here are the learnable weight matrix and bias vector for the forget gate. and The learnable weight matrix and bias vector for the forget gate; the final output of the temporal encoding module is: ,in This represents the hidden state vector at time step t, which is the feature representation output by the LSTM model after time encoding.
[0040] S3-3, Time-series data of the coal-to-ethylene glycol process under different operating conditions are represented as follows: ,in For time steps; The number of chemical equipment; Representing the feature dimensions of each device; spatial feature encoding for each time step data:
[0041]
[0042] in, Indicates the first The spatial input matrix corresponding to each time step This represents the bias vector in the spatial feature encoding module; The weight matrix is a learnable matrix. For output dimensions; Spatial embedding representation for each time step; Full-time-step output: ;
[0043] S3-4, Combining spatial characteristics yes Device embedding representation, time features yes Time step representation in dimensionality; through spatial features Perform dimensionality reduction so that its output is a global embedding for each time step: Then the global embedding of the time step The set of: Then, the temporal and spatial features are weighted and fused: ,in These are learnable parameters.
[0044] The preferred dual-supervised training network structure in S4 is as follows:
[0045] S4-1. For the anomaly classifier, input chemical engineering data and output a binary classification result indicating whether it is an anomaly. The predicted label of the model is quantified using the Cross Entropy Loss function. and truth labels Error between, predicted label :
[0046] in, This represents the feature representation input into the anomaly classifier. This represents the multilayer perceptron model corresponding to the anomaly classifier.
[0047] The expression for the cross-entropy loss function is shown below:
[0048]
[0049]
[0050] The loss function for the anomaly classifier is shown in the following formula:
[0051]
[0052] in express The true category, Indicates the output value Predicted probability after applying the sigmoid activation function;
[0053] S4-2, The fault diagnosis classifier learns spatial features and abnormal data to further classify specific fault categories. The output of the fault diagnosis classifier is... :
[0054]
[0055] in, This represents the fused feature obtained by fusing the spatial features extracted by GCN with the temporal features extracted by LSTM.
[0056] Supervised training is performed using the Focal Loss function:
[0057]
[0058]
[0059] in, This represents the loss function value of the fault diagnosis classification channel. It is an adjustable hyperparameter; This represents the model's predicted output;
[0060] S4-3. For the fault diagnosis model, four metrics are used to evaluate the model performance: accuracy (ACC), loss value, recall, and F1 score. The expression for accuracy (ACC) is: The Recall expression is: The F1 score expression is: Where TP is the number of samples predicted as positive and actually positive; FP is the number of samples predicted as positive but actually negative; FN is the number of samples predicted as negative but actually positive; and TN is the number of samples predicted as negative and actually negative.
[0061] The present invention has the following beneficial effects:
[0062] (1) Based on the prior knowledge of the coal-to-ethylene glycol process, this invention constructs a knowledge graph that can uniformly express the relationships between equipment, streams, components and variables, thereby realizing a systematic description of the structural information of complex coal chemical processes.
[0063] (2) This invention uses graph convolutional neural networks to extract features from knowledge graphs, which can characterize the topological relationships between process objects; combined with long short-term memory networks to encode graph features in time, it can further characterize the temporal dependency information in the fault evolution process.
[0064] (3) In addition to graph structure features and time features, the present invention further introduces a spatial feature extraction and fusion mechanism, which can jointly model multi-source operation information in coal chemical process, thereby forming a more complete fault feature representation.
[0065] (4) The present invention sets up an anomaly classifier and a fault diagnosis classifier to form a dual-supervised training network, which can first identify whether the sample is abnormal, and then make a specific fault category judgment on the abnormal sample, thereby realizing a hierarchical output that combines anomaly detection and fault diagnosis.
[0066] (5) This invention is applicable to fault diagnosis scenarios in coal chemical processes with many variables, strong coupling, and obvious dynamic evolution, and can meet the application requirements of fault monitoring and classification identification in coal-to-ethylene glycol processes. Attached Figure Description
[0067] Figure 1 This is a diagram showing the overall structure of the fault diagnosis method.
[0068] Figure 2 This is a schematic diagram of the process of constructing a priori knowledge graph.
[0069] Figure 3 This is a schematic diagram of spatiotemporal feature fusion.
[0070] Figure 4 This is a graph showing the prediction results.
[0071] Figure 5 This is a schematic diagram of a confusion matrix. Detailed Implementation
[0072] To further understand the present invention, preferred embodiments of the present invention are described below in conjunction with examples. However, it should be understood that these descriptions are only for further illustrating the features and advantages of the present invention, and not for limiting the scope of the claims of the present invention.
[0073] The description in this section pertains to only a few typical embodiments, and the present invention is not limited to the scope of the embodiments described. Combinations of different embodiments, substitution of some technical features in different embodiments, and substitution of similar or identical prior art with some technical features in the embodiments are also within the scope of the description and protection of the present invention.
[0074] Taking the coal-to-ethylene glycol (CTEG) process as an example, this paper predicts the types of faults in the coal-to-ethylene glycol process online, such as... Figure 1As shown, the overall framework for a fault prediction method in the coal-to-ethylene glycol process based on graph convolutional neural networks is constructed. Graph features, temporal features, and spatial features of the coal-to-ethylene glycol process data are extracted and fused.
[0075] The fault diagnosis method of this technical solution is used to predict faults in the coal-to-ethylene glycol process. The specific steps include:
[0076] S1. Establish a steady-state model of the entire coal-to-ethylene glycol process, and obtain a dynamic model of the coal-to-ethylene glycol process through the steady-state model; introduce faults by artificially changing the output mode of the controller, and obtain a fault time series dataset of the coal-to-ethylene glycol process.
[0077] S1-1. The coal-to-ethylene glycol process includes steps such as coal-to-syngas production, syngas purification and separation, oxalate synthesis, oxalate hydrogenation to ethylene glycol, and product refining. A steady-state model of the entire process is constructed. The global property method is PR-BM, the property method for syngas purification and separation is PSRK, the property method for oxalate synthesis and hydrogenation is NRTL, and the property method for the distillation unit is NRTL-RK. The unit modules are as follows: the gasifier uses the RGibbs module; the desulfurization or decarbonization process in syngas purification uses the RadFrace tower module; the water-gas shift converter uses the RStoic module; the oxalate synthesis reactor uses the RPlug module; the oxalate hydrogenation reactor uses the RCSTR module; and the distillation tower uses the RadFrace module.
[0078] S1-2. Using Aspen dynamic simulation technology, faults are manually introduced by changing the controller output to obtain a fault dataset for the coal-to-ethylene glycol process. The controllers include a flow controller, a temperature controller, a pressure controller, and a level controller.
[0079] Optionally, in S1-2, the faults mentioned can be: tower PCR condenser failure, tower PCR reflux pump pressure loss, tower MSC reflux pump pressure loss, and NO feed flow rate surge, as shown in Table 1.
[0080] Table 1 Simulated Fault Types
[0081]
[0082] The fault was simulated and fault data was collected, as shown in Table 2. A dynamic dataset containing process parameters such as liquid level, temperature, pressure, and composition was obtained, and monitoring points were set up to collect data.
[0083] Table 2 Monitoring Points for Coal-to-Ethylene Glycol Process
[0084]
[0085] S2. Construct a knowledge graph using prior knowledge of the coal-to-ethylene glycol process, such as... Figure 2 As shown, importing physical entities, variable entities, and various types of relations into Neo4j yields a complete knowledge graph, which is represented as follows: The entity is designed in layers, including physical entities. and variable entities The types of relationships are divided into physical connections, state associations, and material composition.
[0086] S2-1, The process of constructing a priori knowledge graph is as follows: Figure 2 As shown, data entities are obtained through industrial process data, and physical entities are obtained through technological processes. The prior knowledge graph is represented as follows: ,in It is an entity set. It is a set of relations. Indicates the form of The triple represents the head entity. Tail entity and relationships The connections between them; therefore, the entity set can be formally represented as:
[0087]
[0088] in, Represents a set of device entities. Represents the set of flowing entities. Represents a set of component entities. Represents a collection of variable entities.
[0089] S2-2. Relationships based on prior knowledge graphs are established through the interaction of head and tail entities, expressed as various types of relationships. Different physical entities have inherent connections, and variable entities typically have state representation relationships with their corresponding physical entities. The set of relationships can be represented as:
[0090]
[0091] in, This indicates that the device outputs a specific stream; This indicates that a stream flows into a certain device; Indicates the components contained in the stream; The variable represents the operating status of the equipment or flow stream; Used to express the semantics of reaction and transformation direction.
[0092] S2-3. For the mapping from knowledge graph to process data, let the set of process data collection points be... The variable entity set is Then the mapping function from the measurement point to the variable entity can be defined as: ,in, Indicates data collection points In the knowledge graph, these are the corresponding variable entities. These variable entities are attached to device entities or flow entities through State relationships to represent the operational state of the corresponding physical objects. Formally, this can be represented as:
[0093]
[0094] in, Indicates a device entity or a flow entity. Represents variable entities, Represents a set of State relationships.
[0095] S2-4. Achieve continuous characterization of the operating state while maintaining the stability of the process topology. Let the observation sequence of any variable entity be:
[0096]
[0097] in, Indicates the sampling time. This represents the observed value of the variable at that moment. This sequence is stored as the state attribute of the variable node and is used to characterize the evolution of the variable over time.
[0098] S2-5. Data needs to be standardized before being written into the graph. Let the mean and standard deviation of the variable under normal operating conditions be respectively... and Its standardized representation is:
[0099]
[0100] in, A very small constant is introduced to prevent the denominator from being zero. Through standardization, different variables can be characterized on a uniform scale to represent their degree of deviation from normal operating conditions, providing a basis for anomaly intensity calculation and subsequent state aggregation. In addition to the original observations and standardized values, variable nodes can also store derived state information such as deviation intensity, anomaly label, duration, trend, and sliding window statistics. Let the variable at time... The deviation intensity is When it exceeds the threshold When this happens, an exception indicator can be defined as: This allows variable nodes to simultaneously carry multiple layers of information, including the original state, the standardized state, and the abnormal state.
[0101] S3. Based on knowledge graphs, a graph feature extraction module is constructed using a graph convolutional neural network (GCN) to extract high-level feature representations that integrate graph structure information and node attribute information. Then, a long short-term memory (LSTM) neural network is used to temporally encode the features as temporal features to capture dependencies in time-series data. Simultaneously, spatial features are extracted from the coal-to-ethylene glycol process fault time-series dataset using a fully connected layer network. For example... Figure 3 As shown, spatial features and temporal features are finally fused.
[0102] S3-1, In the graph feature extraction module, based on the observation sequence... Used as input features to construct the node feature matrix Adjacency matrix constructed based on prior knowledge Combined, as input to GCN, where Indicates the number of nodes. This represents the feature dimension of each node; the initial node feature matrix represents the original feature input. Then the first The node feature matrix of layer +1 can be represented as:
[0103]
[0104] in, Indicates the first The node feature matrix of the layer, Indicates the first Layer weight matrix; The degree matrix representing the nodes; Activation function; global graph features are extracted by stacking L layers of convolutions. ;
[0105] S3-2, Graph features output by graph convolution Convert to time series input; in the time dimension, graph embedding sequence. ;in, Indicates a time step. This represents the number of nodes in the current graph structure. This represents the feature dimension of each node. Indicates the first The graph embedding matrix corresponding to each time step;
[0106] Graph Embedding Sequences Perform time encoding: Use a multi-layer LSTM model to capture temporal dependencies:
[0107]
[0108] in, The hidden state at the current time step. It represents the dimension of the LSTM hidden state, indicating the graph features after temporal encoding; Indicates the input gate. Represents the Gate of Oblivion Indicates the output gate. Indicates the first The cell state at each time step. Indicates the first The hidden state at each time step and These are the learnable weight matrix and bias vector of the input gate, respectively. and Here are the learnable weight matrix and bias vector for the forget gate. and Here are the learnable weight matrix and bias vector for the forget gate. and The learnable weight matrix and bias vector for the forget gate; the final output of the temporal encoding module is: ,in It represents the hidden state vector at a specific time step and is the feature representation output by the LSTM model after time encoding.
[0109] S3-3, Time-series data of the coal-to-ethylene glycol process under different operating conditions are represented as follows: ,in For time steps; The number of chemical equipment; Representing the feature dimensions of each device; spatial feature encoding for each time step data:
[0110]
[0111] in, Indicates the first The spatial input matrix corresponding to each time step This represents the bias vector in the spatial feature encoding module; The weight matrix is a learnable matrix. For output dimensions; Spatial embedding representation for each time step; Full-time-step output: ;
[0112] S3-4, Combining spatial characteristics yes Device embedding representation, time features yes Time step representation in dimensionality; through spatial features Perform dimensionality reduction so that its output is a global embedding for each time step: Then the global embedding of the time step The set of: Then, the temporal and spatial features are weighted and fused: ,in These are learnable parameters.
[0113] S4. Design a dual-supervised training network by combining two multilayer perceptrons, including an anomaly classifier and a fault diagnosis classifier. The anomaly classifier learns whether the input data is abnormal by using the original fault time series data, and the fused features are used as the input of the fault diagnosis classifier to finally obtain an accurate fault classification result.
[0114] S4-1. For the anomaly classifier, input chemical engineering data and output a binary classification result indicating whether it is an anomaly. The predicted label of the model is quantified using the Cross Entropy Loss function. and truth labels The error between them. Predicted label. :
[0115] in, This represents the feature representation input into the anomaly classifier. The multilayer perceptron model corresponding to the anomaly classifier.
[0116] The expression for the cross-entropy loss function is shown below:
[0117]
[0118]
[0119] The loss function for the anomaly classifier is shown in the following formula:
[0120]
[0121] in express The true category, Indicates the output value Predicted probabilities after passing through the sigmoid activation function.
[0122] S4-2, The fault diagnosis classifier learns spatial features and abnormal data to further classify specific fault categories. The output of the fault diagnosis classifier is... .
[0123]
[0124] in, This represents the fused feature obtained by fusing the spatial features extracted by GCN with the temporal features extracted by LSTM.
[0125] Supervised training is performed using the Focal Loss function:
[0126]
[0127]
[0128] in, This represents the loss function value of the fault diagnosis classification channel. It is an adjustable hyperparameter; This represents the model's predicted output.
[0129] S4-3. For the fault diagnosis model, four metrics are used to evaluate model performance: accuracy (ACC), loss value, recall, and F1 score. The expression for accuracy (ACC) is: The recall rate expression is: The F1 score expression is: Where TP is the number of samples predicted as positive and actually positive; FP is the number of samples predicted as positive but actually negative; FN is the number of samples predicted as negative but actually positive; and TN is the number of samples predicted as negative and actually negative.
[0130] The proposed method is used for fault diagnosis in the coal-to-ethylene glycol process. The proposed network is built on a graph convolutional neural network to enhance the learning ability of graph features. LSTM and fully connected neural networks are used to extract the temporal and spatial features of the fault data, respectively, to enhance the network's spatial feature learning ability. Furthermore, a dual-supervised training network is constructed using two multilayer perceptron networks to improve the overall accuracy and reliability of the diagnosis. The results are as follows: Figure 4 As shown, with the progress of training, the model accuracy gradually improves and tends to stabilize, the loss value continues to decrease and remains at a low level, while the recall and F1 score remain at a high level overall. These results indicate that the method has good convergence and stability, can accurately identify different fault types, has fewer missed detections of abnormal samples, and demonstrates good diagnostic performance on a real coal-to-ethylene glycol fault dataset. Figure 5 The fault confusion matrix results show that the fault categories are mainly concentrated on the diagonal, indicating that the model has a high classification and discrimination ability; however, there is still some confusion between a few categories.
[0131] Table 3 Various Fault Indicators
[0132]
[0133] As shown in Table 3, the model exhibits excellent performance and generalization ability on both the training and test sets. It ultimately achieves a high accuracy of 98.01%, a recall of 97.99%, and an F1 score on the test set, indicating high prediction accuracy and few missed targets. The training and testing metrics are very close (the test loss is 0.0216, even slightly lower than the training loss of 0.0350), showing no overfitting. The model demonstrates strong discriminative ability across most categories, except for the heat exchanger temperature +5... It has slight shortcomings in class and dehydration tower heat load fluctuation classes. Its prediction confidence is extremely high, with a Top-3 accuracy of 99.69%, meaning that even if the prediction is wrong, the correct class will almost always be among its top three options. Overall, this is a highly reliable, high-performance model that reaches practical application level.
[0134] The description and application of the present invention herein are illustrative and not intended to limit the scope of the invention to the embodiments described above. The effects or advantages described in the specification may not be apparent in actual experimental cases due to uncertainties in specific conditions or other factors, and such descriptions are not intended to limit the scope of the invention. Variations and modifications to the embodiments disclosed herein are possible, and various substitutions and equivalents of the components in the embodiments are well known to those skilled in the art. It should be understood by those skilled in the art that the invention can be implemented in other forms, structures, arrangements, proportions, and with other components, materials, and parts without departing from the spirit or essential characteristics of the invention. Other variations and modifications can be made to the embodiments disclosed herein without departing from the scope and spirit of the invention.
Claims
1. A dual-channel coal chemical fault diagnosis method based on graph spatiotemporal fusion modeling, characterized in that, Including the following: S1. Establish a steady-state model of the entire coal-to-ethylene glycol process, and obtain a dynamic model of the coal-to-ethylene glycol process through the steady-state model; introduce faults by artificially changing the output mode of the controller, and obtain a fault time series dataset of the coal-to-ethylene glycol process. S2. Construct a knowledge graph based on prior knowledge of the coal-to-ethylene glycol process, and design entities in a hierarchical manner, including physical entities. and variable entities The types of relationships are divided into physical connections, state associations, and material composition; S3. Based on knowledge graphs, a graph feature extraction module is constructed using graph convolutional neural network (GCN) to extract high-level feature representations that integrate graph structure information and node attribute information. Then, a Long Short-Term Memory (LSTM) neural network is used to encode the features in time as temporal features to capture the dependencies in the time series data. At the same time, a fully connected layer network is used to extract spatial features from the coal-to-ethylene glycol process fault time series dataset. Finally, the spatial features and temporal features are fused. S4. Design a dual-supervised training network, combining two multilayer perceptrons, including an anomaly classifier and a fault diagnosis classifier. The anomaly classifier learns whether the original fault time-series data contains anomalies, and the fused features are used as input to the fault diagnosis classifier to obtain accurate fault classification results. The specific structure of the dual-supervised training network in S4 is as follows: S4-1. For the anomaly classifier, input chemical engineering data and output a binary classification result indicating whether it is an anomaly. The predicted label of the model is quantified using the Cross Entropy Loss function. and truth labels Error between, predicted label : in, This represents the feature representation input into the anomaly classifier. This represents the multilayer perceptron model corresponding to the anomaly classifier. The expression for the cross-entropy loss function is shown below: The loss function for the anomaly classifier is shown in the following formula: in express The true category, Indicates the output value Predicted probability after applying the sigmoid activation function; S4-2, The fault diagnosis classifier learns spatial features and abnormal data to further classify specific fault categories. The output of the fault diagnosis classifier is... : in, This represents the fused feature obtained by fusing the spatial features extracted by GCN with the temporal features extracted by LSTM. Supervised training is performed using the Focal Loss function: in, This represents the loss function value of the fault diagnosis classification channel. It is an adjustable hyperparameter; This represents the model's predicted output; S4-3. For the fault diagnosis model, four metrics are used to evaluate the model performance: accuracy (ACC), loss value, recall, and F1 score. The expression for accuracy (ACC) is: The Recall expression is: The F1 score expression is: Where TP is the number of samples predicted as positive and actually positive; FP is the number of samples predicted as positive but actually negative; FN is the number of samples predicted as negative but actually positive; and TN is the number of samples predicted as negative and actually negative.
2. The method according to claim 1, characterized in that, The specific fault simulation of the coal-to-ethylene glycol process in S1 is as follows: S1-1 The coal-to-ethylene glycol process includes coal-to-syngas production, syngas purification and separation, oxalate synthesis, oxalate hydrogenation to ethylene glycol, and product refining steps. A steady-state model of the entire process is constructed. S1-2. Manually introduce faults by changing the controller output to obtain a fault dataset for the coal-to-ethylene glycol process. The controllers include a flow controller, a temperature controller, a pressure controller, and a level controller.
3. The method according to claim 2, characterized in that, In S1-2, the faults mentioned are: failure of the PCR condenser in the tower, failure of the PCR reflux pump in the tower, failure of the MSC reflux pump in the tower, or a sudden increase in NO feed flow rate.
4. The method according to claim 1, characterized in that, The prior knowledge of the coal-to-ethylene glycol process in S2 is used to construct a knowledge graph, specifically as follows: S2-1, Prior Knowledge Graph, represented as ,in It is an entity set. It is a set of relations. Indicates the form of The triple represents the head entity. Tail entity and relationships The connection between them; formally representing the entity set as: in, Represents a set of device entities. Represents the set of flowing entities. Represents a set of component entities. Represents a collection of variable entities; S2-2. Relationships based on prior knowledge graphs are established through the interaction of head and tail entities, expressed as various types of relationships. Inherent connections exist between different physical entities, and variable entities represent the state representation relationships with their corresponding physical entities. The set of relationships is represented as follows: in, This indicates that the device outputs a specific stream; This indicates that a stream flows into a certain device; Indicates the components contained in the stream; The variable represents the operating status of the equipment or flow stream; Used to express the semantics of reaction and transformation directions; S2-3. For the mapping from knowledge graph to process data, the set of process data collection points is as follows: The variable entity set is The mapping function from the measurement point to the variable entity is: , Indicates data collection points In the knowledge graph, the corresponding variable entities are attached to device entities or flow entities through State relationships to represent the operational state of the corresponding physical objects; represented as: in, Indicates a device entity or a flow entity. Represents variable entities, Represents a set of State relations; S2-4. Achieve continuous characterization of the operating state while maintaining the stability of the process topology, based on the observation sequence of any variable entity. for: in, Indicates the sampling time. Represents the observed value of a variable at that moment; observation sequence The state attributes of the variable nodes are stored to depict the evolution of the variable over time. S2-5. Data needs to be standardized before being written into the graph. The mean and standard deviation of the variables under normal operating conditions are as follows: and Variable standardization Represented as: in, A very small constant is introduced to prevent the denominator from being zero; through standardization transformation, different variables can be characterized on a unified scale to represent their degree of deviation from normal operating conditions, providing a basis for anomaly intensity calculation and subsequent state aggregation; in addition to the original observation value and standardized value, the variable node also stores derived state information including deviation intensity, anomaly label, duration, trend of change, and sliding window statistics.
5. The method according to claim 1, characterized in that, The graph feature extraction module, temporal encoding module, and spatial feature extraction module in S3 are as follows: S3-1, In the graph feature extraction module, the observation sequence The node feature matrix is constructed using the input features. Adjacency matrix constructed based on prior knowledge The combination, after which it is used as input to GCN, wherein... Indicates the number of nodes. It is the feature dimension of each node; the initial node feature matrix. This represents the original feature input: Then the first The node feature matrix of layer +1 is represented as follows: in, Indicates the first The node feature matrix of the layer Indicates the first Layer weight matrix; The degree matrix representing the nodes; Activation function; global graph features are extracted by stacking L layers of convolutions. ; S3-2, Graph features output by graph convolution Convert to time series input; in the time dimension, graph embedding sequence. ;in, Indicates a time step. This represents the number of nodes in the current graph structure. This represents the feature dimension of each node. Indicates the first The graph embedding matrix corresponding to each time step; Graph Embedding Sequences Perform time encoding: Use a multi-layer LSTM model to capture temporal dependencies: in, The hidden state at the current time step. It represents the dimension of the LSTM hidden state, indicating the graph features after temporal encoding; Indicates the input gate. Represents the Gate of Oblivion Indicates the output gate. Indicates the first The cell state at each time step. Indicates the first The hidden state at each time step and These are the learnable weight matrix and bias vector of the input gate, respectively. and Here are the learnable weight matrix and bias vector for the forget gate. and Here are the learnable weight matrix and bias vector for the forget gate. and The learnable weight matrix and bias vector for the forget gate; the final output of the temporal encoding module is: temporal features. ,in This represents the hidden state vector at time step t, which is the feature representation output by the LSTM model after time encoding. S3-3, Time-series data of the coal-to-ethylene glycol process under different operating conditions are represented as follows: ,in For time steps; The number of chemical equipment; Representing the feature dimensions of each device; spatial feature encoding for each time step data: in, Indicates the first The spatial input matrix corresponding to each time step This represents the bias vector in the spatial feature encoding module; The weight matrix is a learnable matrix. For output dimensions; Spatial embedding representation for each time step; Full-time-step output: Spatial features ; S3-4, Combining spatial characteristics yes Device embedding representation, time features yes Time step representation in dimensionality; through spatial features Perform dimensionality reduction so that its output is a global embedding for each time step: Then the global embedding of the time step The set of: Then, the temporal and spatial features are weighted and fused: ,in These are learnable parameters. This is a weighted fusion feature.