Coal mill unit fault diagnosis method and system based on multi-layer graph convolutional neural network

Abstract

The application provides a coal mill unit fault diagnosis method and system based on a multi-layer graph convolutional neural network, and the method comprises the following steps: collecting fault data of different working conditions of the coal mill unit by using a sensor; normalizing the sensor data, uniformly intercepting N samples by using a sliding window strategy, and then mapping the fault samples under different working conditions to the same feature space by using a domain adversarial model; constructing a feature matrix X by using the features of the mapped N fault samples, constructing a multi-layer network by using three different indexes, obtaining a corresponding adjacency matrix A, and taking the label of the fault as the label of the node in the multi-layer network; dividing the fault samples into a training set and a test set, inputting the feature matrix X and the different adjacency matrices A into a graph convolutional neural network, outputting a predicted label by convolution, and iteratively training the model by using a cross-entropy loss function. The application solves the technical problems of the prior art, such as not considering the influence of the complex correlation between samples on fault diagnosis, poor diagnosis performance under variable working conditions, and low diagnosis accuracy.

Description

Technical Field

[0001] This invention relates to the field of coal-fired processing equipment maintenance, specifically to a method and system for fault diagnosis of coal mill units based on multilayer graph convolutional neural networks. Background Technology

[0002] Although new energy power generation technologies such as wind power, hydropower, and nuclear power have developed rapidly in recent years, thermal power generation still accounts for more than 70% of the country's total power generation, making it the primary power generation method. In thermal power systems, coal mills are a crucial component, key to improving raw material utilization and the productivity of thermal power plants. Therefore, monitoring and diagnosing the operating status of coal mills in thermal power plants is a prerequisite for ensuring the safe and stable operation of thermal power systems and an important direction for the development of smart grids.

[0003] Coal mill units are core equipment in thermal power plants, used to crush and grind coal into pulverized coal, which is then transported to the furnace by a primary air fan for combustion and power generation. In actual operation, coal mill units are characterized by prolonged high-load operation and complex structures, significantly increasing the probability of malfunctions. A malfunction in the coal mill unit not only affects boiler combustion and reduces the power plant's production efficiency but also threatens the safety of other equipment and production personnel. This makes intelligent fault diagnosis of coal mill units increasingly necessary and urgent. If intelligent fault diagnosis can be performed on the coal mill unit when a fault occurs, accurately locating the location and type of the fault, it has significant theoretical and practical value for the safe operation of the unit and the economic benefits of the thermal power plant.

[0004] Early fault diagnosis methods primarily relied on human experience. While these methods were easy to understand and widely applicable, their accuracy declined and they became susceptible to subjective influences when systems became more complex. With advancements in sensor technology and increased data storage capacity, data-driven fault diagnosis methods have become a research hotspot. As a typical data-driven approach, deep learning-based fault diagnosis research has yielded substantial theoretical and practical results. For example, the existing invention patent document CN106153179A, "Method for Diagnosing Vibration Faults of Medium-Speed ​​Coal Mills," first analyzes the fault characteristics of coal mill vibration, concluding that coal mill current is an important monitoring parameter for coal mill vibration fault analysis. Subsequently, through correlation analysis of historical data, four auxiliary variables are selected: mill outlet air pressure, coal feed rate of the coal feeder, primary air volume at the mill inlet, and mill outlet temperature. These four auxiliary variables are used to predict the current. The method of constructing mill vibration quantity by subtracting the measured current from the predicted current and taking the residual sequence of the coal mill current is used. The residual sequence of the mill current is decomposed into three-level wavelet packet decomposition to obtain the energy ratio of 8 frequency bands. The energy ratio of two types of faults is statistically analyzed to obtain the characteristic quantities of two faults: severe wear of the grinding rollers and foreign objects entering the mill. As can be seen from the specific implementation of the prior art, the coal mill fault diagnosis scheme disclosed in this document uses the pre-processed coal mill inlet and outlet differential pressure, coal feed rate, primary air volume at the mill inlet, and mill outlet temperature as inputs to the coal mill current prediction model. A BP neural network is used to predict a current prediction signal. The difference between the actual coal mill current and the predicted coal mill current is used to characterize the coal mill vibration by obtaining the residual sequence signal of the coal mill current. Another example is the prior art patent application document CN115496188A, "A Coal Mill Fault Early Warning Method Based on Deep Learning Convolutional Neural Network." This method obtains measurement point parameter data of the coal mill's historical normal operating state, selects measurement point parameter data related to coal mill blockage faults as input variables, establishes a CNN-LSTM-Attention fault early warning model, and performs online training. It obtains the residual sequence between the predicted value ym and the actual value y obtained by the CNN-LSTM-Attention fault early warning model, processes it using the sliding window method to obtain the average deviation sequence, and uses the kernel density estimation method to calculate the probability density function of the average deviation sequence. However, when faced with large-scale, non-Euclidean structured fault data under different operating conditions, the neural networks used in the aforementioned existing technologies, as well as conventional CNNs, DBNs, and other models, exhibit poor diagnostic performance or even fail. Furthermore, existing deep learning-based fault diagnosis models typically use only one metric to construct the network, which is not conducive to characterizing the various complex relationships between samples and makes it difficult to aggregate feature information from different types of neighbors, resulting in low fault diagnosis accuracy. In addition, although GCNs can handle non-Euclidean graph data, they cannot be directly used in multi-layer networks.

[0005] In summary, existing technologies do not consider the impact of various complex correlations between samples on fault diagnosis, resulting in poor diagnostic performance and low diagnostic accuracy under varying operating conditions. Summary of the Invention

[0006] The technical problem to be solved by this invention is how to address the issue that existing technologies do not consider the impact of various complex correlations between samples on fault diagnosis, resulting in poor diagnostic performance and low diagnostic accuracy under varying operating conditions.

[0007] This invention solves the above-mentioned technical problems by employing the following technical solution: A fault diagnosis method for coal mill units based on multi-layer graph convolutional neural networks includes:

[0008] S1. Collect fault data of different operating conditions of coal mill units using sensors;

[0009] S2. Normalize the fault data under different operating conditions, use the sliding window strategy to extract the average length data segments to obtain N fault samples under different operating conditions, and use the domain adversarial model to map the fault samples under different operating conditions to the same feature space to obtain the mapped fault samples.

[0010] S3. Use the features of N mapped fault samples to form a feature matrix X. Construct a multi-layer network using no less than two preset indices to obtain an adjacency matrix A. The corresponding nodes in the multi-layer network are the same mapped fault sample, that is, the fault label of the sample is the shared node label of the same node in the multi-layer network, which is used to minimize the loss function to update the model parameters.

[0011] S4. Divide the N mapping fault samples into training set and test set, input the feature matrix X and the difference adjacency matrix A into the graph convolutional neural network to perform inter-layer convolution and intra-layer convolution, output the predicted label of the training set, and use the cross-entropy loss function for iterative training to obtain a multi-layer graph convolutional neural network model.

[0012] S5. Input the test set into the applicable multi-layer graph convolutional neural network model for fault diagnosis, and use the Acc accuracy metric and Macro-F1 metric to measure the classification accuracy of the applicable multi-layer graph convolutional neural network model.

[0013] This invention addresses the fault diagnosis problem of coal mill units under different operating conditions based on domain adversarial mechanisms and multilayer graph convolutional neural networks (MGCN). Domain adversarial mechanisms map fault data from different operating conditions to the same feature space, thus more closely reflecting the actual working scenario of the coal mill unit and improving the robustness and practicality of the model. MGCN can better represent various potential correlations between nodes, and the inter-layer and intra-layer convolutional mechanisms can aggregate information from different neighboring nodes, thereby improving the performance of fault diagnosis.

[0014] In a more specific technical solution, the sensor measurement variables in step S1 include: coal mill current, coal feeder quantity, coal mill inlet air temperature, induced draft fan bearing temperature, primary air fan current, forced draft fan current, and air preheater outlet flue gas temperature.

[0015] The fault diagnosis framework for coal mill units under varying operating conditions proposed in this invention, based on multilayer graph convolutional neural networks (MGCN), can process and analyze large-scale, non-Euclidean structured data under different operating conditions. It solves the shortcomings of existing single-layer graph network methods, such as insufficient extraction of process data features and difficulty in achieving operating condition transitions, and has better fault diagnosis performance and stronger method practicality.

[0016] In a more specific technical solution, step S2 includes:

[0017] S21. Using the maximum-minimum normalization method, process the fault data under different operating conditions to obtain normalized fault data. Then, extract the features of N fault samples under different operating conditions from the normalized fault data through a sliding window w.

[0018] S22. Take one operating condition data from the differential operating condition fault sample as the source domain and the remaining operating condition data as the target domain.

[0019] S23. Based on the source domain and the target domain, use the domain adversarial model to map the fault samples of different operating conditions to the same feature space to obtain the mapped fault samples.

[0020] This invention maps data samples from different operating conditions to the same feature space. Simultaneously, the intra-layer and inter-layer convolutions in the multi-layer graph convolutional neural network (MGCN) effectively extract and aggregate information from different neighboring nodes to improve the model's training performance, thereby enhancing the accuracy and practicality of fault diagnosis. This invention utilizes a domain adversarial strategy to map fault data from different operating conditions to the same feature space, enabling conversion between different operating conditions, making it closer to actual working scenarios, and also improving the model's robustness.

[0021] In a more specific technical solution, step S21 utilizes the following maximum-minimum normalization logic to process fault data under different operating conditions:

[0022]

[0023] Where x represents the original data, x max With x min These are the maximum and minimum values ​​in the original data, respectively.

[0024] In a more specific technical solution, step S3 includes:

[0025] S31. Using the features of N mapped fault samples, construct the feature matrix X∈R of the graph. N×w ;

[0026] S32. Based on the feature matrix, calculate the distance between the current node and other nodes, as well as the similarity between nodes;

[0027] S33. Sort the nodes according to the distance between nodes and the similarity between nodes;

[0028] S34. Select the K nodes with the smallest distance and the K nodes with the largest similarity as the neighbors of the current node, and obtain the adjacency matrix A1∈R corresponding to the first graph. N×N The adjacency matrix A2∈R corresponding to the second graph N×N ;

[0029] S35. Take the previous and next nodes of the current node as neighbor nodes to obtain the adjacency matrix A3∈R of the third graph. N×N This allows us to extract the temporal relationships between nodes.

[0030] In a more specific technical solution, in step S32, the distance S1 between nodes and the similarity S2 between nodes are calculated using the following logic:

[0031]

[0032]

[0033] Here, a and b are the feature vectors of any two nodes.

[0034] In a more specific technical solution, step S4 includes:

[0035] S41. Connect the characteristic matrix X with the difference adjacency matrix A. l (l = 1, 2, 3) are input into a multi-layer graph convolutional network model;

[0036] S42. Utilize the intra-layer and inter-layer convolution mechanisms of the multi-layer graph convolutional network model to aggregate neighbor node information with different structures for nodes, and learn two sets of GCN parameters to decouple intra-layer and inter-layer propagation.

[0037] S42. Sum the feature matrices of the graph network in at least 2 to obtain the features of the aggregate nodes;

[0038] S43. Iterative training is performed using the softmax function and cross-entropy loss function to obtain a suitable multi-layer graph convolutional neural network model.

[0039] This invention employs an intra-layer and inter-layer convolution mechanism that decouples intra-layer and inter-layer propagation by learning two sets of GCN parameters, enabling the model to learn the different importance in the two propagation directions. The multi-layer graph convolutional network (MGCN) model used in this invention constructs a multi-layer network, thereby better representing various correlations between samples. Then, different GCN layers are used to independently propagate node features, and the node representations from different layers are aggregated through intra-layer and inter-layer convolutions, which helps improve diagnostic performance.

[0040] In a more specific technical solution, step S42 uses the following logic to represent neighbor node information for different structures:

[0041]

[0042]

[0043] Where iα represents the i-th node in the α-th layer of the network, intra-k represents the k-th inner convolution, and inter-k represents the k-th inter-layer convolution. Let x represent the features of node i after k+1 convolutions. iα It is the input feature of node i in the graph of layer α.

[0044] In a more specific technical solution, in step S5,

[0045] The Acc indicator can be represented using the following logic:

[0046]

[0047] Where TP is the number of positive samples predicted as positive, FP is the number of negative samples predicted as positive, TN is the number of negative samples predicted as negative, and FN is the number of positive samples predicted as negative.

[0048] The Acc indicator can be represented using the following logic:

[0049]

[0050] Where TP is the number of positive samples predicted as positive, FP is the number of negative samples predicted as positive, TN is the number of negative samples predicted as negative, and FN is the number of positive samples predicted as negative.

[0051] In a more specific technical solution, the coal mill fault diagnosis system based on a multi-layer graph convolutional neural network includes:

[0052] The data acquisition module is used to collect fault data of different operating conditions of the coal mill unit using sensors;

[0053] The domain adversarial module is used to normalize the fault data of different operating conditions. It uses a sliding window strategy to extract data segments of average length to obtain N fault samples of different operating conditions. The domain adversarial model is used to map the fault samples of different operating conditions to the same feature space to obtain the mapped fault samples. The domain adversarial module is connected to the acquisition module.

[0054] The adjacency matrix acquisition module is used to construct a feature matrix X using the features of N mapped fault samples, and then construct a multi-layer network using no less than two preset indices to obtain an adjacency matrix A. Based on this, the multi-layer network node labels of the mapped fault samples are obtained, which are used to update the model parameters by minimizing the loss function. The adjacency matrix acquisition module is connected to the domain adversarial module.

[0055] The model iterative training module is used to divide the N mapped fault samples into training set and test set, input the feature matrix X and the difference adjacency matrix A into the graph convolutional neural network to perform inter-layer convolution and intra-layer convolution, output the predicted label of the training set, and use the cross-entropy loss function to perform iterative training to obtain a multi-layer graph convolutional neural network model. The model iterative training module is connected to the adjacency matrix acquisition module.

[0056] The fault diagnosis and accuracy measurement module is used to input the test set into the applicable multi-layer graph convolutional neural network model for fault diagnosis. The Acc accuracy metric and Macro-F1 metric are used to measure the classification accuracy of the applicable multi-layer graph convolutional neural network model. The fault diagnosis and accuracy measurement module is connected to the model iterative training module.

[0057] Compared with existing technologies, this invention has the following advantages: This invention addresses the fault diagnosis problem of coal mill units under different operating conditions based on domain adversarial mechanisms and multilayer graph convolutional neural networks (MGCN). Domain adversarial mechanisms can map fault data from different operating conditions to the same feature space, thus more closely reflecting the actual working scenario of the coal mill unit and improving the robustness and practicality of the model. Multilayer graph convolutional neural networks (MGCN) can better represent various potential correlations between nodes, and the inter-layer and intra-layer convolutional mechanisms can aggregate information from different neighboring nodes, thereby improving the performance of fault diagnosis.

[0058] The fault diagnosis framework for coal mill units under varying operating conditions proposed in this invention, based on multilayer graph convolutional neural networks (MGCN), can process and analyze large-scale, non-Euclidean structured data under different operating conditions. It solves the shortcomings of existing single-layer graph network methods, such as insufficient extraction of process data features and difficulty in achieving operating condition transitions, and has better fault diagnosis performance and stronger method practicality.

[0059] This invention maps data samples from different operating conditions to the same feature space. Simultaneously, the intra-layer and inter-layer convolutions in MGCN effectively extract and aggregate information from different neighboring nodes to improve model training performance, thereby enhancing the accuracy and practicality of fault diagnosis. This invention utilizes a domain adversarial strategy to map fault data from different operating conditions to the same feature space, enabling conversion between different operating conditions, making it closer to real-world work scenarios, and also improving the model's robustness.

[0060] This invention employs an intra-layer and inter-layer convolution mechanism to decouple intra-layer and inter-layer propagation by learning two sets of GCN parameters, enabling the model to learn the different importance in the two propagation directions. The multi-layer graph convolutional network (MGCN) model used in this invention constructs a multi-layer network, thereby better representing the various correlations between samples. Then, different GCN layers independently propagate node features, and the node representations from different layers are aggregated through intra-layer and inter-layer convolutions, which helps improve diagnostic performance. This invention solves the technical problem in existing technologies that do not consider the impact of multiple complex correlations between samples on fault diagnosis, resulting in poor diagnostic performance and low diagnostic accuracy under varying operating conditions. Attached Figure Description

[0061] Figure 1 This is a schematic diagram of data stream processing for the coal mill fault diagnosis method based on a multi-layer graph convolutional neural network according to Embodiment 1 of the present invention.

[0062] Figure 2 This is a schematic diagram of the basic steps of the coal mill fault diagnosis method based on a multilayer graph convolutional neural network according to Embodiment 1 of the present invention.

[0063] Figure 3 This is a schematic diagram of the structure of the medium-speed coal mill in Embodiment 2 of the present invention;

[0064] Figure 4a This is a graph showing the changes in the test set Acc value and the number of iterations in Embodiment 2 of the present invention;

[0065] Figure 4b This is a graph showing the change of Macro-F1 value versus the number of iterations in Embodiment 2 of the present invention;

[0066] Figure 5 This is a confusion matrix diagram of the test set fault samples in Embodiment 2 of the present invention;

[0067] Figure 6 This is a diagram showing the diagnostic results of various faults in the test set of Embodiment 2 of the present invention. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] Example 1

[0070] like Figure 1 and Figure 2 As shown, the fault diagnosis method for coal mill units based on multi-layer graph convolutional neural networks provided by this invention includes the following basic steps:

[0071] Step S1: Collect fault data of the coal mill unit under different operating conditions using the installed sensors;

[0072] In this embodiment, the sensor measurement variables in step S1 mainly include: pulverizer current, coal feeder quantity, pulverizer inlet air temperature, induced draft fan bearing temperature, primary air fan current, forced draft fan current, and air preheater outlet flue gas temperature, etc.; the fault categories of the pulverizer unit mainly include: A. high differential pressure of the forced draft fan oil station filter, B. abnormal vibration of the primary air fan oil pump, A. large vibration of the pulverizer, B. large vibration of the induced draft fan bearing, C. large vibration of the pulverizer, etc. In this embodiment, the coal feeder quantity determines the operating condition of the pulverizer unit.

[0073] Step S2: Normalize the collected sensor data and use a sliding window strategy to uniformly extract data segments to obtain N samples. Then, map the fault samples under different working conditions to the same feature space through a domain adversarial model.

[0074] In this embodiment, the data in step S2 is processed using the max-min normalization method:

[0075]

[0076] Where x represents the original data, x max With x min The maximum and minimum values ​​in the original data are respectively, and then the features of N data samples are extracted by a sliding window w. Then, one of the working conditions is taken as the source domain and the data of other working conditions is taken as the target domain. Finally, the data are mapped to the same feature space through a domain adversarial model.

[0077] Step S3: Construct a feature matrix X from the features of the N mapped fault samples, and then construct a multi-layer network using three different indices to obtain the corresponding adjacency matrix A. The fault label is the label of the node in the multi-layer network.

[0078] In this embodiment, step S3 constructs a feature matrix X∈R from the mapped N fault sample features. N×w The faulty sample features are node features. Then, the distance and similarity between nodes are calculated using the Euclidean distance formula and the cosine similarity formula, which can be expressed as follows:

[0079]

[0080]

[0081] Here, a and b are the feature vectors of any two nodes. By calculating the distance and similarity from the current node to every other node, sorting each distance and similarity, and then selecting the K nodes with the smallest distance and the K nodes with the largest similarity as the current node's neighbors, the adjacency matrix A1∈R corresponding to the graph can be obtained. N×N and A2∈R N ×N To extract the temporal relationships between nodes, the nodes at the previous and next time steps of a given node are selected as its neighbors, thus obtaining the adjacency matrix A3∈R of the graph. N×N Furthermore, the fault category label is the label of that node.

[0082] Step S4: Divide the N fault samples into a training set and a test set. At the same time, input the feature matrix X and different adjacency matrices A into the graph convolutional neural network, and output the predicted labels of the training set through inter-layer and intra-layer convolution. Finally, use the cross-entropy loss function to iteratively train the model.

[0083] In this embodiment, step S4 compares the obtained feature matrix X with different adjacency matrices A l (l=1,2,3) are input into a multi-layer graph convolutional model. Utilizing the model's intra-layer and inter-layer convolutional mechanisms, information about neighboring nodes with different structures is aggregated for each node. This can be represented as:

[0084]

[0085]

[0086] Where iα represents the i-th node in the α-th layer of the network, intra-k represents the k-th inner convolution, and inter-k represents the k-th inter-layer convolution. Let x represent the features of node i after k+1 convolutions. iα This represents the input feature of node i in the α-th layer graph. Then, the aggregated node features are obtained by summing the feature matrices of the three graph networks:

[0087] H i =sum(Hα ), α=1,2,3

[0088] Where H α These are the feature matrices corresponding to different layers. The intra-layer and inter-layer convolution mechanism decouples intra-layer and inter-layer propagation by learning two sets of GCN parameters, enabling the model to learn the different importance of the two propagation directions. Finally, the softmax function and cross-entropy loss function are used for iterative training of the model.

[0089] Step S5: Input the test set into the pre-trained multi-layer graph convolutional neural network for fault diagnosis, and use the accuracy (Acc) and Macro-F1 metrics to measure the overall classification accuracy of the model.

[0090] In this embodiment, step S5 uses a test set to test the trained model, then outputs the fault class labels for the test set, and finally measures the overall classification accuracy of the model using the Acc metric and the Macro-F1 metric. The Acc metric can be expressed as:

[0091]

[0092] Where TP is the number of positive samples predicted as positive, FP is the number of negative samples predicted as positive, TN is the number of negative samples predicted as negative, and FN is the number of positive samples predicted as negative. In multi-class classification problems, the Acc metric is defined as the ratio of the number of correctly classified samples to the total number of samples.

[0093] In this embodiment, precision and recall are contradictory metrics; generally, high precision tends to be accompanied by low recall, and vice versa. To comprehensively consider these two metrics, the F1 score is used to measure the overall performance of the model in fault diagnosis. The F1 score can be expressed as:

[0094]

[0095]

[0096]

[0097] However, the F1 score metric is only applicable to binary classification problems. As an improved form of the F1 score metric, the Macro-F1 metric first calculates the TP, FP, FN, and TN for each class, then obtains the F1 score value for each class, and finally takes the average of all the F1 score values.

[0098] Example 2

[0099] like Figure 3 As shown, in this embodiment, the medium-speed coal mill includes: a coal chute 1, an outlet gas seal system 2, a discharge valve and multi-outlet device 3, a separator top cover device 4, an inverted cone device 5, a separator body device 6, an inner cone device 7, a spring loading device 8, a grinding roller device 9, a side body device 10, a grinding bowl and impeller device 11, a scraper device 12, and a planetary gear reducer 13. The raw coal in the medium-speed coal mill first falls from the coal chute 1 onto the grinding bowl, where it is ground into coal powder by the grinding roller device 9. The rotating roller throws the coal powder to the edge, where it is carried by hot air to the separator body device 6 for separation. Coarse powder falls from the inner cone back into the grinding bowl for re-grinding, while fine powder is collected from the mill by the airflow through the discharge valve and multi-outlet device 3. In this embodiment, step S1 of embodiment 1 is executed. The coal mill unit mainly consists of coal mills A, B, C, and D, primary air fans A and B, induced draft fans A and B, forced draft fans A and B, and air preheaters A and B. Sensor data is mainly collected at key measuring points of each piece of equipment, including sensor data monitoring 172 variables such as coal mill current, coal feeder quantity, coal mill inlet air temperature, coal mill bearing temperature, induced draft fan bearing temperature, primary air fan current, forced draft fan current, and air preheater outlet flue gas temperature. The operating conditions of the coal mill unit can be roughly divided into 5 operating conditions based on the coal feed rate: operating condition 1 is a coal feed rate of 35-40 t / h, operating condition 2 is a coal feed rate of 40-45 t / h, operating condition 3 is a coal feed rate of 45-50 t / h, operating condition 4 is a coal feed rate of 50-55 t / h, and operating condition 5 is a coal feed rate of 55-60 t / h.

[0100] The data collection period for the coal mill unit spanned 15 months. Normal data was collected every 5 minutes, and fault data was collected every 1 second. A total of 32 types of fault data were collected, including A. High differential pressure of the blower oil station filter, B. Abnormal vibration of the primary air blower oil pump, A. Large vibration of the coal mill, B. Large vibration of the induced draft fan bearing, and C. Large vibration of the coal mill. These data covered 5 operating conditions, and all data came from the actual operation of the coal mill in a power plant in East China.

[0101] In this embodiment, step S2 of embodiment 1 is executed. The 172 sensor features of the fault data are standardized using the max-min normalization method to eliminate the influence of dimensions. Then, a sliding window w (with a value of 1) is used to capture the features of the nodes. Each node feature represents the fault data corresponding to each time point, that is, the dimension of the node feature is 172 dimensions. Next, each operating condition is treated as the source domain and other operating conditions are treated as the target domain. Finally, a domain adversarial model is used to map the samples of the target domain to the same feature space of the source domain.

[0102] In the domain adversarial model, the source domain label is set to 1, the target domain label is set to 0, and the feature dimension of all sample data mapped to the same feature space is set to 172. Then, the domain label loss LOSS_LABEL and the marginal distribution loss LOSS_MMD are used to constrain the degree of confusion between the source and target domains. Therefore, the loss function of the domain adversarial model is:

[0103]

[0104]

[0105] LOSS = LOSS_MMD + LOSS_LABEL

[0106] Where, n s and n t Let x represent the number of samples in the source and target domains, respectively. i Let D be the feature of the i-th sample. s and D t Let D represent the sample spaces of the source and target domains, respectively. c Representation domain discriminator, d i Let represent the actual domain label of sample i, L() be the binary cross-entropy loss function, and Φ() be the nonlinear mapping function that projects the original data onto the Hilbert space. The features of the i-th sample in the source domain, Let LOSS represent the feature of the j-th sample in the target domain, and let LOSS denote the total loss function during domain adversarial processes.

[0107] In this embodiment, step S3 in embodiment 1 is executed, and the feature matrix H∈R of the mapped N fault sample features is constructed. N×172 Then, the distance S1 and similarity S2 between any two nodes are calculated using the Euclidean distance formula and the cosine similarity formula. Next, each distance and similarity is sorted, and the K nodes with the smallest distance and the K nodes with the largest similarity are selected as the neighbor nodes of that node. This yields the adjacency matrix A1∈R of the network. N×N and A2∈R N×N To extract the temporal relationships between nodes, the nodes at the previous and next time steps of a given node are selected as its neighbors, thus obtaining the adjacency matrix A3∈R of the graph. N×N After performing intra-layer convolution operations on three networks with different structures, an adjacency matrix of nodes between layers is constructed. Identical nodes in different layers are connected pairwise, while no edges connect different nodes. This yields the inter-layer adjacency matrix A4∈R. 3N×3N Furthermore, the category label of the faulty sample is the label of that node.

[0108] In this embodiment, step S4 of embodiment 1 is executed, dividing all mapped fault samples into a training set and a test set, with the training set and test set accounting for 70% and 30% of all samples, respectively. The training set is then input into the MGCN. The MGCN model extracts spatiotemporal features from the input data H using two graph convolutional layers (GCNConv), namely the first graph convolutional layer (GCNConv1) and the second graph convolutional layer (GCNConv2), on the three-layer graph. The outputs of the two graph convolutional layers can be represented as H... α (1) and H α (2) :

[0109]

[0110]

[0111]

[0112]

[0113] in, Let represent the feature vector of node i in layer α of the network H. The node feature matrices of the initial three layers of the network are the same. and These are the node feature matrices after the first and second convolutional layers of the α-th layer network, respectively. The dimensions of these two node feature matrices are 256 and 128, respectively. ReLU represents the activation function, and W... α (1) and W α (2) Let be the weight matrix of the two graph convolutional layers in the α-th graph.

[0114] After performing the first graph convolution on each layer of the network, the node features of the three layers are concatenated row by row as follows:

[0115]

[0116] After obtaining the inter-layer feature representation H4 of the nodes, an inter-layer graph convolution operation is performed using the inter-layer adjacency matrix A4 to obtain the output inter-layer feature matrix, which has a dimension of 128.

[0117]

[0118]

[0119] Simultaneously, after performing a second graph convolution on each layer of the network, the node features of the three layers are concatenated row by row to obtain the output intra-layer feature matrix:

[0120]

[0121] Next, the inter-layer feature matrix and the intra-layer feature matrix are summed to obtain the output feature matrix of each node in the three-layer network:

[0122] H out =sum(H intra H inter ),H out ∈R 3N×128

[0123] After obtaining the node feature matrix of the three-layer graph, i.e., H1 is H out The matrix with the first N rows, H2 is H out The matrix from row N+1 to row 2N, H3 is H out The matrix from row 2N+1 to row 3N, by summing the features of the same nodes in different layers of the network, yields the node feature matrix after inter-layer and intra-layer convolutions:

[0124]

[0125] Following the intra- and inter-layer convolutional mechanisms of the MGCN model, two fully connected layers are designed. The first fully connected layer reduces the dimensionality of the extracted features, while the second fully connected layer acts as a classifier in the model, classifying the extracted features. Furthermore, designing two fully connected layers increases the model's non-linear expressive power. Finally, the model is iteratively trained using the predicted and true labels of the samples and the cross-entropy loss function. The trend of the training set loss and the changes in the Acc and Macro-F1 scores on the test set are observed to select the optimal parameters for the model.

[0126] like Figure 4a , Figure 4b , Figure 5 as well as Figure 6As shown, in this embodiment, step S5 of the embodiment is executed, and domain adversarial training is performed with working condition 1 as the source domain to obtain mapped fault samples. The training samples are input into a multi-layer graph convolutional model for model training. It can be observed that the Acc index and Macro-F1 index of the test set show an upward trend with the increase of the number of iterations. In this embodiment, the coal feed rate range of working condition 1 includes 35~40t / h. Then, the overall classification accuracy of the model is measured by the confusion matrix of the classification results of the test set, and the classification accuracy of each type of fault in the model can also be measured by the Acc index and F1 score of each type of fault. Next, domain adversarial training is performed with different working conditions as the source domain, and the results of the Acc index and Macro-F1 index of the MGCN model are compared with those of the GCN and MLP models. In this embodiment, the coal feed rates of working conditions 2-5 are 40~45t / h, 45~50t / h, 50~55t / h, and 55~60t / h, respectively, as shown in Tables 1 and 2.

[0127] Table 1 compares the Acc index of each model mapped to different working conditions.

[0128] MLP model 0.7086 0.8188 0.8155 0.8067 0.7983 GCN model 0.8502 0.9241 0.9143 0.9511 0.9465 MGCN model 0.9293 0.9926 0.9784 0.9966 0.9934

[0129] Table 2 compares the Macro-F1 indices of each model mapped to different working conditions.

[0130] MLP model 0.6916 0.8152 0.8021 0.8034 0.7890 GCN model 0.8485 0.9201 0.9130 0.9500 0.9452 MGCN model 0.9277 0.9926 0.9779 0.9965 0.9936

[0131] The comparison results of the three algorithms under different operating conditions show that the overall Acc score of the MGCN model is about 20% higher than that of the MLP model and about 6% higher than that of the single-layer GCN model. The overall Macro-F1 score also shows a similar improvement. These results demonstrate that the MGCN model of this invention not only has good fault diagnosis performance but also high robustness and practicality.

[0132] In summary, this invention addresses the fault diagnosis problem of coal mill units under different operating conditions based on domain adversarial mechanisms and multilayer graph convolutional neural networks (MGCN). Domain adversarial mechanisms can map fault data from different operating conditions to the same feature space, thus more closely reflecting the actual working scenario of coal mill units and improving the robustness and practicality of the model. MGCN can better represent various potential correlations between nodes, and the inter-layer and intra-layer convolutional mechanisms can aggregate information from different neighboring nodes, thereby improving the performance of fault diagnosis.

[0133] The fault diagnosis framework for coal mill units under varying operating conditions proposed in this invention, based on multilayer graph convolutional neural networks (MGCN), can process and analyze large-scale, non-Euclidean structured data under different operating conditions. It solves the shortcomings of existing single-layer graph network methods, such as insufficient extraction of process data features and difficulty in achieving operating condition transitions, and has better fault diagnosis performance and stronger method practicality.

[0134] This invention maps data samples from different operating conditions to the same feature space. Simultaneously, the intra-layer and inter-layer convolutions in MGCN effectively extract and aggregate information from different neighboring nodes to improve model training performance, thereby enhancing the accuracy and practicality of fault diagnosis. This invention utilizes a domain adversarial strategy to map fault data from different operating conditions to the same feature space, enabling conversion between different operating conditions, making it closer to real-world work scenarios, and also improving the model's robustness.

[0135] This invention employs an intra-layer and inter-layer convolution mechanism to decouple intra-layer and inter-layer propagation by learning two sets of GCN parameters, enabling the model to learn the different importance in the two propagation directions. The multi-layer graph convolutional network (MGCN) model used in this invention constructs a multi-layer network, thereby better representing the various correlations between samples. Then, different GCN layers independently propagate node features, and the node representations from different layers are aggregated through intra-layer and inter-layer convolutions, which helps improve diagnostic performance. This invention solves the technical problem in existing technologies that do not consider the impact of multiple complex correlations between samples on fault diagnosis, resulting in poor diagnostic performance and low diagnostic accuracy under varying operating conditions.

[0136] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A fault diagnosis method for coal mill units based on multi-layer graph convolutional neural networks, characterized in that, The method includes: S1. Collect fault data of different operating conditions of coal mill units using sensors; S2. Normalize the differential operating condition fault data, and use a sliding window strategy to extract average-length data segments to obtain... Using a domain adversarial model, the fault samples with different operating conditions are mapped to the same feature space to obtain mapped fault samples. S3, Utilization The features of the mapped fault samples constitute a feature matrix. A multi-layer network is constructed using at least two predefined metrics to obtain the adjacency matrix. The multi-layer network node labels of the mapped fault samples are obtained to minimize the loss function and update the model parameters. S3 includes: S31, Utilization The feature matrix of the feature map of the mapped fault samples ;in, Indicates a sliding window; S32. Based on the feature matrix, calculate the node distances from the current node to other nodes and the node similarity. S33. Sort the nodes according to the distance between nodes and the similarity between nodes; S34, Take The node with the smallest distance and The nodes with the highest similarity are selected as the neighbors of the current node, and the adjacency matrix corresponding to the first graph is obtained accordingly. Adjacency matrix of the second graph ; S35. Take the previous and next time nodes of the current node as the neighbor nodes to obtain the adjacency matrix corresponding to the third graph. Based on this, the temporal relationships between nodes can be extracted; S4, will The mapped fault samples are divided into a training set and a test set, and the feature matrix is... With difference adjacency matrix The input is fed into a graph convolutional neural network to perform inter-layer convolution and intra-layer convolution, and the predicted labels of the training set are output. Iterative training is performed using the cross-entropy loss function to obtain a multi-layer graph convolutional neural network model. S4 includes: S41, the feature matrix With the difference adjacency matrix The input is fed into a multi-layer graph convolutional network model, where... ; S42. Utilize the intra-layer convolution and inter-layer convolution mechanism of the multi-layer graph convolutional network model to aggregate neighbor node information with different structures for the node, and learn two sets of GCN parameters to decouple intra-layer and inter-layer propagation; The neighbor node information of the different structures is represented by the following logic: in, Indicates the first The first layer in the network 1 node Indicates the first Secondary involution Indicates the first Inter-layer convolution Indicates passage After convolution, the nodes are aggregated. Features It is the first Layer diagram nodes Input features; S42. Summate the feature matrices of at least two graph networks to obtain the aggregate node features; S43. Iterative training is performed using the softmax function and cross-entropy loss function to obtain the applicable multi-layer graph convolutional neural network model. S5. Input the test set into the applicable multi-layer graph convolutional neural network model for fault diagnosis, and use the Acc accuracy index and Macro-F1 index to measure the classification accuracy of the applicable multi-layer graph convolutional neural network model.

2. The method for fault diagnosis of coal mill units based on multi-layer graph convolutional neural networks according to claim 1, characterized in that, The sensor's measurement variables in step S1 include: coal mill current, coal feeder quantity, coal mill inlet air temperature, induced draft fan bearing temperature, primary air fan current, forced draft fan current, and air preheater outlet flue gas temperature.

3. The method for fault diagnosis of coal mill units based on multi-layer graph convolutional neural networks according to claim 1, characterized in that, Step S2 includes: S21. Process the differential operating condition fault data using the max-min normalization method to obtain normalized fault data, and then process the data through a sliding window. Extracted from the normalized fault data Features of each of the aforementioned fault samples under different operating conditions; S22. Take one of the operating condition data from the differential operating condition fault samples as the source domain and the rest of the operating condition data as the target domain. S23. Based on the source domain and the target domain, the differential operating condition fault samples are mapped to the same feature space using a domain adversarial model to obtain the mapped fault samples.

4. The method for fault diagnosis of coal mill units based on multilayer graph convolutional neural networks according to claim 3, characterized in that, In step S21, the differential operating condition fault data is processed using the following maximum-minimum normalization logic: in, The original data, and These are the maximum and minimum values ​​in the original data, respectively.

5. The method for fault diagnosis of coal mill units based on multi-layer graph convolutional neural networks according to claim 1, characterized in that, In step S32, the distance between nodes is calculated using the following logic. and the similarity between the nodes : in, and Let be the feature vector of any two nodes, representing the th node. Each graph node This indicates a sliding window.

6. The method for fault diagnosis of coal mill units based on multi-layer graph convolutional neural networks according to claim 1, characterized in that, In step S5, the Acc index is represented using the following logic: in, It represents the number of positive samples predicted as positive. It represents the number of negative samples predicted as positive. It represents the number of negative samples predicted as negative. It represents the number of positive samples predicted as negative.

7. A coal mill fault diagnosis system based on a multi-layer graph convolutional neural network, used to execute the coal mill fault diagnosis method based on a multi-layer graph convolutional neural network as described in any one of claims 1 to 6, characterized in that, The system includes: The data acquisition module is used to collect fault data of different operating conditions of the coal mill unit using sensors; The domain adversarial module is used to normalize the differential operating condition fault data, and uses a sliding window strategy to extract average-length data segments to obtain... A fault sample with different operating conditions is mapped to the same feature space using a domain adversarial model to obtain a mapped fault sample. The domain adversarial module is connected to the acquisition module. The adjacency matrix acquisition module is used to utilize... The features of the mapped fault samples constitute a feature matrix. Then, a multi-layer network is constructed using at least two predefined metrics to obtain the adjacency matrix. The adjacency matrix acquisition module is connected to the domain adversarial module to obtain the multi-layer network node labels of the mapped fault samples, which are then used to update the model parameters by minimizing the loss function. The model iterative training module is used to... The mapped fault samples are divided into a training set and a test set, and the feature matrix is... With difference adjacency matrix The input is fed into a graph convolutional neural network to perform inter-layer convolution and intra-layer convolution, and the predicted labels of the training set are output. Iterative training is performed using the cross-entropy loss function to obtain a multi-layer graph convolutional neural network model. The model iterative training module is connected to the adjacency matrix acquisition module. The fault diagnosis and accuracy measurement module is used to input the test set into the applicable multi-layer graph convolutional neural network model for fault diagnosis. The classification accuracy of the applicable multi-layer graph convolutional neural network model is measured using the Acc accuracy index and the Macro-F1 index. The fault diagnosis and accuracy measurement module is connected to the model iterative training module.

Images