Causal aggregation and grouping fault root cause tracing method for typical equipment of gas turbine

Abstract

The application discloses a kind of typical equipment's causal aggregation and grouping fault root cause tracing method for gas turbine.The method constructs the causal aggregation and grouping model consisting of sparse causal inference module, node grouping module, obtains the position information of fault measuring point in the entire causal graph using the vector similarity calculation of space-time dislocation, obtains the relative position information between fault measuring point by the design of grouping penalty term training model using the idea of cutout clustering, comprehensively measures the causal similarity between fault measuring point from two angles of global and local.Using grouping result, the causal relationship in group and between groups is sparsified, and a more simplified causal structure is obtained.The application can construct two-level causal representation between variables in group and between groups, and trace the fault root cause of gas turbine equipment according to the fault root cause variables indicated in the causal graph.The method directly represents the fault propagation path of gas turbine equipment, and provides an efficient and reliable new idea for fault root cause tracing decision of typical equipment of gas turbine.

Description

Technical Field

[0001] This invention belongs to the field of intelligent fault root cause tracing of large-scale distributed gas turbine equipment. In particular, it proposes a method for grouping fault measurement points by aggregating causal relationships. It proposes a causal similarity metric index to group fault measurement points with similar causal relationships under a causal framework, effectively realizing the sparsity of the causal graph network. On this basis, it extracts the causal relationships between variables and constructs a two-level causal graph structure between variables within and between groups. According to the causal relationships in the two-level causal graph, it indicates the fault propagation path, characterizes the root cause of the fault, and realizes the tracing of the fault source of gas turbine equipment. Background Technology

[0002] With the deepening of Industry 4.0, the operation and maintenance of intelligent equipment has received widespread attention in the industrial sector. Generally, data-driven intelligent equipment operation and maintenance consists of three stages: First, fault diagnosis methods are used to alarm for faults occurring in industrial processes. Fault isolation methods are then used to analyze monitoring samples that have experienced faults, separating faulty and healthy measurement points within the samples. Next, based on this, causal relationships between variables are extracted from known time-series data of the faulty measurement points. A variable causal graph structure is constructed based on these relationships. This topology graph indicates the system fault propagation path and visually represents the source of the system fault. Finally, these fault tracing and diagnosis results are saved, and based on these results and recommendations, specific areas of the industrial process system are maintained, effectively extending the equipment's lifespan. Root cause analysis methods primarily play a role in the second stage of intelligent equipment operation and maintenance, providing reference opinions for maintenance personnel.

[0003] Root cause diagnosis research has been applied to various industrial processes and factory equipment, such as hot-rolled steel and cigarette cutting, but it has not yet been effectively applied to typical gas turbine equipment. As a complex industrial system, typical gas turbine equipment contains multiple sub-components and numerous measuring points. During operation, due to the connection of material and information flows, faults propagate through multiple sub-components within this complex system, causing anomalies in monitoring data across a wide range of measuring points. Furthermore, the presence of control loops and other system structures leads to severe equipment coupling and mutual influence between components, resulting in interactions between fault measuring points and increasing the difficulty of extracting causal relationships. In summary, the main problems with current root cause tracing methods for typical gas turbine equipment, such as compressors and turbines, are the large number of fault measuring points, the vast number of causal relationships to be determined, the exponentially increasing number of possible causal graph structures generated by existing methods, and the overly complex causal networks that fail to identify fault propagation paths and root variables.

[0004] In recent years, to address the challenges of numerous fault measurement points and complex causal network structures in root cause analysis of large-scale complex industrial systems, many data-driven root cause analysis methods employ Lasso regularization constraints for variable selection to achieve sparsity of causal relationships. However, these methods fail to consider the inherent characteristics of fault propagation, where faults typically propagate first within subsystems of a complex system through causal connections, and then to adjacent subsystems, with the complex system exhibiting a hierarchical structure during propagation. Furthermore, these methods often suffer from over-parameterization due to the nonlinear and strongly coupled nature of complex industrial processes, failing to achieve the desired sparsity. Therefore, some methods first group the set of fault measurement points in the complex system into subsets based on equipment operating mechanisms or measurement point data characteristics, then infer causal relationships between measurement points within subsets and between subsets themselves, thereby diagnosing fault propagation paths and characterizing root cause variables. However, these methods do not directly utilize causal relationships to group fault measurement points, often failing to guarantee consistency between the grouping results and the causal relationship grouping results, hindering the extraction of accurate causal relationships. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a method for causal aggregation and grouping of root cause tracing for typical gas turbine equipment.

[0006] The objective of this invention is achieved through the following technical solution:

[0007] A method for causal aggregation and grouping of root cause tracing for typical gas turbine equipment includes the following steps:

[0008] (1) Obtain time-series data of typical equipment fault measurement points of the gas turbine, determine the time delay hyperparameter L of the time-series data, and generate past time-series data X from the time-series data based on the time delay hyperparameter. ＜t =[x 1,＜t ,x 2,＜t ,...,x D,＜t ] and current time series data X t =[x 1t ,x 2t ,...,x Dt ]; where D is the number of fault measurement points, T is the total number of time-series data samples of typical gas turbine equipment fault measurement points, and x it The current time series data of the i-th fault measurement point is composed of the last TL samples from the time series data of the i-th fault measurement point. Each sample is the fault measurement point data at the current time. i,＜t The past time series data represents the i-th fault measurement point. The past time series data corresponds one-to-one with the current time series data, satisfying that the fault measurement point data at the current moment corresponds to the fault measurement point data at the previous L moments.

[0009] (2) Construct a causal aggregation and grouping model, which includes a sparse causal inference module and a node grouping module; wherein, the sparse causal inference module consists of D convolutional neural networks corresponding one-to-one with fault measurement points and a causal graph weighted adjacency matrix generation unit, and each convolutional neural network is used to generate a causal graph weighted adjacency matrix based on past time series data X. ＜t Predict the current time series data x of the corresponding fault measurement point it The causal graph weighted adjacency matrix generation unit is used to quantify and calculate the Granger causal relationship between fault measurement points based on the connection weights between the first hidden layer and the input layer of D convolutional neural networks, and generate a causal graph weighted adjacency matrix; the node grouping module includes a node feature extractor, a group feature extractor, and a grouping unit. The node feature extractor is used to extract features from past time series data X. ＜t Extract the temporal features of each fault measurement point and concatenate them with the corresponding location embedding vector to obtain the temporal features of each node; the group feature extractor is used to extract the temporal features from the current time series data X. t The feature vectors of the current time and corresponding time-delay data are extracted and then concatenated with the corresponding position embedding vectors to reduce the spatial dimension and obtain the group features; the grouping unit is used to construct a similarity matrix based on the node temporal features and group features and to group the nodes to obtain the node grouping results;

[0010] (3) Transfer the past time series data X obtained in step (1) ＜t And the current time series data X t The input is fed into the causal aggregation and grouping model, which is trained by minimizing the loss function. After training, the causal aggregation and grouping model outputs the final causal graph weighted adjacency matrix and node grouping results. The loss function includes the error loss between the output and the ground truth of each convolutional neural network, a grouping regularization term, a grouping penalty term, an inter-group loop structure penalty term, and an intra-group loop structure penalty term. The grouping regularization term standardizes the grouping results, on the one hand promoting the node grouping module to classify faulty test points into defined groups, and on the other hand preventing the erroneous situation of faulty test points being grouped alone. The grouping penalty term L is designed using the graph tiling concept. g This allows the node grouping module to group variables according to the strength of the causal relationship between them, and designs an inter-group loop structure penalty term L. a1 Intra-group loop structure penalty term L a2 This is used to train the sparse causal extraction module, reducing the loop structure in the topological graph formed by causal relationships;

[0011] (4) Based on the final weighted adjacency matrix of the causal graph and the node grouping results, determine the causal relationship between fault measurement points and the causal relationship between fault measurement point subsets within the group, and complete the root cause tracing of the fault.

[0012] Furthermore, the time-series data of typical gas turbine equipment fault measurement points are normalized data.

[0013] Furthermore, the causal graph weighted adjacency matrix generation unit calculates the Granger causal relationship between fault measurement points based on the connection weights between the first hidden layer and the input layer of D convolutional neural networks, and generates a causal graph weighted adjacency matrix. The specific calculation method for each element of the causal graph weighted adjacency matrix is ​​as follows:

[0014]

[0015] Among them, a ij Let A be the element in the i-th row and j-th column of the weighted adjacency matrix A of the causal graph. In the convolutional neural network corresponding to the i-th fault test point, is the weight vector connecting the past time-series data input node of the j-th fault test point to the first hidden layer node z, where z is the number of hidden layer nodes in the first hidden layer; norm represents the second norm of the matrix.

[0016] Furthermore, the node feature extractor consists of a first convolutional layer, an activation function layer, and a splicing layer, each with an independent convolutional process for each measurement point, connected in sequence.

[0017] Specifically, including:

[0018] (1) Using past time series data X ＜t As input, features are extracted from the first convolutional layer and then passed through the GLU activation function layer to obtain the feature vector.

[0019] (2) The concatenation layer superimposes the position embedding vectors of the temporal data onto the extracted feature vectors to obtain the final temporal features H of each node:

[0020] H = GLU(f1(X) ＜t ;Θ1))+concat(Pos(pos,i))

[0021] The position embedding Pos(pos,i) is obtained by calculating the function Pos.

[0022] Furthermore, the group feature extractor consists of a second convolutional layer, an activation function layer, a splicing layer, and a third convolutional layer connected in sequence.

[0023] Specifically, including:

[0024] (1) Using the current time series data X t As input, the current time series data X is left-filled. t This ensures that the time dimension of the input remains unchanged during the convolution process. The features extracted by the second convolutional layer are then processed by the GLU activation function layer to obtain the feature vector.

[0025] (2) The splicing layer superimposes the position embedding vector of the time series data onto the extracted feature vector.

[0026] (3) The concatenated vectors are passed through a third convolutional layer, which keeps the time dimension of the features unchanged while reducing the spatial dimension, reducing the number of vectors from D to K, to obtain the group features F:

[0027] F = f(GLU(f2(X)) t ;Θ2))+concat(Pos(pos,i)))

[0028] Furthermore, the position embedding vector is obtained by calculating the function Pos.

[0029] Furthermore, the grouping regularization term L o Grouping penalty item L g Inter-group loop structure penalty term L a1 With the group's loop structure penalty term L a2 Specifically:

[0030]

[0031]

[0032]

[0033]

[0034] Where S represents the similarity matrix, I K This represents an identity matrix with dimension K, where K represents the number of groups, i.e., the number of group features. ||*|| F Let F denote the F-norm, Tr(*) denotes finding the trace of a matrix, matrix D is the degree matrix of the weighted adjacency matrix A of the causal graph, and I denotes the identity matrix. and Let A be the normalized inter-group adjacency matrix and the normalized intra-group adjacency matrix, respectively, where k∈[1,K]; k Let D be the k-th submatrix reconstructed from the weighted adjacency matrix A of the causal graph based on the grouping results. in Let D be the in-degree matrix of the weighted adjacency matrix A in a causal graph. out Let A be the out-degree matrix of the weighted adjacency matrix A in a causal graph.

[0035] Furthermore, the grouping unit constructs a similarity matrix based on node temporal features and group features, and then groups the nodes to obtain the grouping results, specifically as follows:

[0036] The similarity S between the temporal feature vector of each node and the temporal feature vector of the group is calculated using the vector dot product. ij Construct a similarity matrix:

[0037]

[0038] Among them, h i Let f represent the temporal feature vector of the i-th node. j Let K represent the time-series feature vector of the j-th group, and K represent the number of group features.

[0039] For each fault measurement point, the similarity between the node temporal feature vector of the fault measurement point and the temporal feature vectors of all groups is compared, and the fault measurement point is assigned to the corresponding group with the highest similarity to obtain the node grouping result.

[0040] Furthermore, the loss function includes the error loss between the output and the true value of each convolutional neural network, expressed as mean squared error L. mse Specifically:

[0041]

[0042] in and x it These represent the predicted value and the actual value of the t-th data point in the current time series data of the i-th fault measurement point, respectively.

[0043] Furthermore, step (4) specifically includes:

[0044] Based on the final weighted adjacency matrix of the causal graph and the node grouping results, determine the causal relationship between the fault measurement points: if element a of the weighted adjacency matrix of the causal graph... ij If the value is greater than the average value of the matrix elements, then the j-th fault measurement point is considered to have a causal effect on the i-th fault measurement point.

[0045] Determine the causal relationships between fault measurement point subsets within a group: Calculate the weighted adjacency matrix A of the inter-group causal graph. g =S T AS, there exists a threshold δ, if a g,ij -a g,ji >δ, i≠j, a g,ij A represents g The element value in the i-th row and j-th column indicates that the j-th fault measurement point subset has a causal effect on the i-th fault measurement point subset, if a g,ji -a g,ij If >δ, i≠j, then the i-th measurement point subset has a causal effect on the j-th measurement point subset.

[0046] Furthermore, step (4) also includes:

[0047] Based on the causal relationships between fault measurement points and between fault measurement point subsets within a group, a two-level causal graph is constructed. The root node G of the inter-group level graph is found; this root node represents a measurement point subset graph. Then, the root node v of graph G is searched; the fault measurement point represented by this root node is the root dependent variable of the fault.

[0048] Furthermore, the time-series data includes fault measurement point samples collected at different times. Fault measurement points may include typical process parameters of gas turbine power generation, such as gas turbine natural gas volume flow rate, gas turbine natural gas mass flow rate, compressor inlet temperature, compressor outlet pressure, compressor outlet temperature, compressor bearing temperature, compressor thrust bearing generator end temperature, compressor thrust bearing gas turbine end temperature, compressor bearing vibration, compressor side main shaft vibration, compressor inlet pressure differential, gas turbine exhaust average temperature, gas turbine side main shaft vibration, combustion chamber pressure differential, gas turbine cooling air regulating valve position, gas turbine humming, gas turbine speed, gas turbine power, cooling air pressure in the second-stage stationary vane ring chamber, and cooling air pressure in the third-stage stationary vane ring chamber.

[0049] Compared with existing technologies, the beneficial effects of this invention are: This invention provides a new grouping approach for root cause analysis of typical gas turbine equipment failures. Addressing the problems of existing methods, such as the potential over-parameterization of all measurement points through direct sparsification, and the inability to guarantee interpretability of grouping results when extracting causal relationships based on mechanism or time-series data similarity, this invention proposes a causal relationship similarity index based on two complementary understandings: from a local perspective, the stronger the causal relationship between two variables, the higher their similarity; from a global perspective, the more similar the causal structures of two variables relative to other variables, the higher their similarity. For the local perspective, this invention utilizes the concept of graph tangency to design a grouping penalty term L. g To ensure the strength of causal relationships within a group is as high as possible, this invention designs a similarity calculation method based on time misalignment from a global perspective. This method uses node temporal features containing past time information and group temporal features containing current time information to transform vector similarity calculation into temporal prediction similarity comparison, thereby obtaining similarity comparison results of causal structures within the Granger causality framework. Based on the causal relationship similarity index, fault measurement points are grouped to extract sparser and more interpretable causal relationships, constructing a two-level causal graph structure within and between groups. This intuitively represents the fault propagation path and root cause of the fault, thereby achieving intelligent tracing of the root cause of gas turbine equipment faults. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating the process of constructing a causal learning model to obtain a two-level causal graph within and between groups, as described in this invention.

[0051] Figure 2 This is a diagram of the model framework constructed according to the present invention;

[0052] Figure 3 These are the fault measurement point numbers and measurement point names of the gas turbine equipment described in this invention;

[0053] Figure 4 This is a concept diagram of grouping variables according to causal similarity as proposed in this invention;

[0054] Figure 5 This is a specific embodiment of the invention that constructs a two-level cause-effect graph to trace the root causes of gas turbine equipment failures;

[0055] Figure 6 This is a comparison chart of the root causes of gas turbine equipment failures traced using the Neural-Granger method in a specific embodiment of the present invention;

[0056] Figure 7 This is a comparison chart of the results of the method for extracting causal relationships after grouping variables using DTW-Kmeans clustering in a specific embodiment of the present invention to trace the root causes of gas turbine equipment failures. Detailed Implementation

[0057] The purpose of this invention is to address the phenomenon that typical gas turbine equipment faults first propagate within sub-components and then to adjacent sub-components. Considering that directly sparsifying all measurement points may lead to over-parameterization, and that extracting causal relationships by grouping according to mechanism or time series data similarity cannot guarantee the interpretability of the grouping results, this invention proposes a causal relationship similarity index. By aggregating similar causal relationships, fault measurement points are grouped, and a sparse two-level causal graph structure within and between groups is extracted. First, the raw data of the monitoring samples from the fault measurement points are preprocessed to eliminate convergence difficulties caused by differences in data values. Then, a causal aggregation and grouping model is constructed, including building a sparse causal extraction module and applying neural networks to extract Granger causal relationships between variables. A node grouping module is constructed to extract features from the fault measurement point data. A time misalignment design is used to transform vector similarity calculation into causal relationship similarity calculation. A tangent loss is designed to group fault measurement points with strong causal relationships into one group. Next, a two-level acyclic constraint within and between groups is designed to reduce the loop structure in the causal graph and enhance its interpretability. Finally, the preprocessed time-series data of the fault measurement points is fed into the constructed module for training. The causal relationships between variables are quantitatively judged through network parameter calculation, and a two-level causal graph within and between groups is constructed to trace the root cause of gas turbine equipment failures.

[0058] The present invention will be further described below with reference to the accompanying drawings and specific examples.

[0059] The operation of a gas turbine combined cycle generator set is a highly complex industrial process characterized by time-varying and dynamic changes. The gas turbine unit consists of three core components: the compressor, the combustion chamber, and the gas turbine. The compressor and turbine are typically multi-stage axial-flow designs. Its working principle is as follows: Air enters through the gas turbine inlet, and the high-speed rotation of the multi-stage rotor blades in the compressor accelerates and pressurizes the airflow, which is then sent to the combustion chamber. Fuel (gaseous or liquid fuel) is then injected into the combustion chamber and mixed with the high-temperature compressed air, undergoing combustion at a constant pressure. The resulting high-temperature, high-pressure flue gas expands after combustion and enters the turbine region, driving the turbine blades to rotate at high speed until it is discharged from the outlet as exhaust gas. This exhaust gas can be reused, for example, by feeding it into a waste heat boiler in a gas turbine generator set to form a combined cycle. For the gas turbine, changes in unit load cause changes in turbine power, resulting in continuous changes in the unit's operating state, exhibiting typical large-range transient characteristics. This invention addresses the problem of complex and difficult-to-identify causal graphs generated by severely coupled gas turbine equipment. It proposes a method for causal aggregation and grouping of typical gas turbine equipment to trace the root causes of faults. Figure 1 The diagram shown is a flowchart of the process of constructing a model to obtain a causal graph according to the present invention. Figure 2 The diagram shown is a model framework diagram of the present invention. The method of the present invention includes the following steps:

[0060] (1) Obtain fault measurement point time series data under fault conditions of gas turbine equipment, determine the time delay over-parameter of time series data, and generate past time series data and current time series data from time series data based on the time delay over-parameter;

[0061] The fault measurement point time series data includes T samples from D measurement points, represented as a two-dimensional data matrix. In this embodiment, 720 fault data samples of a Siemens SGT5-4000F gas turbine were collected from a power plant in Zhejiang Province. The selected case involved an abnormal cooling air pressure alarm in the gas turbine compressor. There were 10 fault measurement points, such as... Figure 3 As shown, the parameters include: gas turbine natural gas volume flow rate, gas turbine natural gas mass flow rate, compressor inlet temperature, compressor outlet pressure, gas turbine exhaust average temperature, combustion chamber pressure difference, gas turbine power, compressor bearing temperature, cooling air pressure of the second-stage stationary vane ring chamber, and cooling air pressure of the third-stage stationary vane ring chamber.

[0062] Generally, normalizing the original fault measurement point time series data eliminates the convergence difficulties caused by data numerical differences; specifically: standard deviation is used to normalize the original data X=[x1,x2,...,x...]. D For each variable dimension, the mean is removed and the variance is normalized to obtain the preprocessed result. The specific formula for normalization is as follows:

[0063]

[0064] Where, μ i and σ i The original time series data x of measurement point i are respectively i The mean and variance of.

[0065] Determine the time delay hyperparameter L of the time series data, and then use the time delay hyperparameter to analyze the preprocessed data X. * Two sets of data are generated, namely, past time series data X. ＜t =[x 1,＜t ,x 2,＜t ,...,x D,＜t ] and current time series data X t =[x 1t ,x 2t ,...,x Dt ].in This represents the current time series data of the i-th fault measurement point, derived from the time series data of measurement point i. The data consists of the last TL samples, each representing the fault measurement point data at the current moment. The past time series data represents the i-th fault measurement point. The past time series data corresponds one-to-one with the current time series data, satisfying that one fault measurement point data at the current moment corresponds to the fault measurement point data at the previous L moments. In the embodiment, the time delay hyperparameter L = 5.

[0066] (2) Construct a causal aggregation and grouping model, which includes a sparse causal inference module and a node grouping module; specifically, it includes the following sub-steps:

[0067] (2.1) Construct a sparse causal inference module to predict current time series data using past time series data and calculate the weighted adjacency matrix A of the causal graph. The sparse causal inference module consists of D convolutional neural networks corresponding one-to-one with fault measurement points and a causal graph weighted adjacency matrix generation unit. The convolutional neural network is a convolutional neural network with multiple hidden layers. Past time series data is input into the network, and it can output the predicted value of the current time series data. The specific formula is as follows:

[0068] x it =g i (x 1,＜t ,x 2,＜t ,...x D,＜t )+e i

[0069] Among them, g i Let x represent the prediction function corresponding to the convolutional neural network. itIt is the current time series data of the i-th fault measurement point, e i To account for the error between the predicted value and the true value, this convolutional neural network uses past time series data from all measurement points to predict the current time series data of a certain measurement point.

[0070] D parallel convolutional neural networks are used to predict the current time series data of the corresponding fault measurement points.

[0071] The causal graph weighted adjacency matrix generation unit is passed through the first hidden layer h of the convolutional neural network. 1 The Granger causality relationship between fault measurement points is calculated using the connection weights W of the input layer. The specific calculation method is as follows:

[0072]

[0073] Among them, a ij The element in the i-th row and j-th column of the weighted adjacency matrix A of the causal graph is used to quantify the causal relationship between the j-th fault measurement point and the i-th fault measurement point. The meaning is that in the convolutional neural network corresponding to the i-th fault measurement point, z is the weight vector connecting the past time series data input node of the j-th fault measurement point to the first hidden layer node 1, and so on, where z is the weight vector connecting the past time series data input node of the j-th fault measurement point to the first hidden layer node h. 1 The number of hidden layer nodes included. `norm` represents the 2-norm of a matrix; the expression calculates the F-norm of a matrix concatenated from z such weight vectors.

[0074] (2.2) Construct a node grouping module to group fault measurement points. The preprocessed time-series data is input, trained, and outputs the node grouping result S. The node grouping module includes a node feature extractor, a group feature extractor, and a grouping unit; where:

[0075] (2.2.1) Node feature extractor, from past time series data X ＜t The temporal features of each node are extracted, and these temporal features can represent the fault measurement points in the causal graph. In this embodiment, the feature extractor uses a strategy that makes the first convolutional layer f1(X) independent for each measurement point. ＜t ; Θ1) Extract features, where Θ1 represents the parameters of the first convolutional layer, and then pass through a gated linear unit (GLU) as the activation function layer to obtain the feature vector result. The expression of the activation function layer is as follows:

[0076]

[0077] Where σ(·) is the sigmoid function. This represents element-wise matrix multiplication, where W, V, b, and c are linear operation parameters of the input x, i.e., the output of the first convolutional layer f1. The position embedding vector of the time-series data is then superimposed on this. The position embedding is calculated using the function Pos, with the specific formula as follows:

[0078]

[0079] Where pos is the time sequence number of the time series data, and i is the measurement point code used to calculate the measurement point number. This indicates rounding down. The position embeddings calculated by the above formula are arranged into position embedding vectors according to their corresponding temporal order. For the same measurement point encoding, the sine calculation result comes first, followed by the cosine calculation result. This process is denoted as concat(Pos(pos,i)). The concatenation layer superimposes the position embedding vectors with the feature vectors from the first step to obtain the final temporal features H = [h1, h2, ..., h...] for each node. D The forward propagation process of the node feature extractor network is as follows:

[0080] H = GLU(f1(X) ＜t ;Θ1))+concat(Pos(pos,i))

[0081] (2.2.2) Group feature extractor, from the current time series data X t The group time-series features are extracted. These group time-series features represent virtual measurement points that indicate the absolute location of fault measurement points in the cause-effect graph. In this embodiment, the feature extractor first fills the current time-series data X with left padding. t The input to the second convolutional layer ensures that the time dimension of the output data after convolution is consistent with the current time series data X. t To maintain consistency, and to extract only the information of the current time and corresponding time delay data during the convolution process, feature vectors are obtained through the GLU activation function layer. Then, the position embedding vectors are obtained using the method described in step (4.1). The concatenation layer concatenates the feature vectors with the position embedding vectors to obtain D feature vectors containing both current time data and position information. Finally, these D feature vectors pass through a third convolution layer similar to the above convolution process, maintaining the feature vector dimension and reducing the number of feature vectors to K, where K represents the number of groups, to obtain the final group features F = [f1, f2, ..., f...]. K The forward propagation process of the group feature extractor network is as follows:

[0082] F = f(GLU(f2(X)) t ;Θ2))+concat(Pos(pos,i)))

[0083] In the formula, f2(*) represents the second convolutional layer, Θ2 is the parameter of the group feature extractor, and f(*) represents the third convolutional layer.

[0084] (2.2.3) The grouping unit is used to construct a similarity matrix based on the node temporal features and group features, and to obtain the node grouping results. Specifically, the similarity between the node temporal feature vector and the group temporal feature vector is calculated using the vector dot product. The specific calculation formula is as follows:

[0085]

[0086] The softmax function normalizes the calculated similarity, ensuring that the sum of the similarities between each node's features and all group features equals 1. The specific calculation formula is as follows:

[0087]

[0088] Refers to the above formula

[0089] Since node time-series features contain information from the past, and group features contain information from the current time, the calculation of their similarity actually measures the predictive ability of faulty measurement points for virtual measurement points. ij A larger value indicates better predictability of the faulty measurement point for a given virtual measurement point. By comparing the predictability of the faulty measurement point for virtual measurement points in each group, the faulty measurement point is assigned to the group with the best predictability (highest similarity). Based on the grouping results, the faulty measurement points within each group form a subset of the measurement points for that group. For example... Figure 4 As shown, the grouping results indicate that faulty measurement points within the same subset of measurement points all exhibit the best predictive ability for the same virtual measurement point. This suggests that these faulty measurement points have similar causal structures in the causal graph within the Granger causality framework, and that the measurement points are located close to each other in the causal graph from a global perspective. This can be further demonstrated by using grouping regularization terms. Standardize the grouping results, where S represents the similarity matrix, derived from the similarity S... ij Composition, I K This represents an identity matrix with dimension K, where K represents the number of groups, ||*||. F The F-norm is used to facilitate the node grouping module in assigning faulty measurement points to specific groups, while also preventing the erroneous situation where faulty measurement points are grouped separately.

[0090] (3) Transfer the past time series data X obtained in step (1) ＜t And the current time series data X tThe input is fed into the causal aggregation and grouping model, and a loss function penalty term is designed to complete the model training process. The loss function includes the loss of the sparse causal inference module, a grouping regularization term, a grouping penalty term, an inter-group loop structure penalty term, and an intra-group loop structure penalty term. Specifically, the grouping regularization term standardizes the grouping results, on the one hand, promoting the node grouping module to assign faulty test points to the designated groups, and on the other hand, preventing the erroneous situation where faulty test points are grouped alone. The grouping penalty term L is designed using the graph tiling concept. g This allows the node grouping module to group variables according to the strength of the causal relationship between them, and designs an inter-group loop structure penalty term L. a1 Intra-group loop structure penalty term L a2 This is used to train the sparse causal extraction module, reducing the loop structure in the topological graph formed by causal relationships; specifically:

[0091] The grouping penalty item L g This means that, based on the grouping results, the edges in the causal graph are divided into intra-group causal edges and inter-group causal edges, and the inter-group causal edges represent the proportion of causal strength in the overall causal strength of the graph.

[0092]

[0093] Where Tr(*) represents finding the trace of a matrix, Tr(S T AS) represents the causal strength within a group after grouping, Tr(S) T DS) represents the causal strength of the entire graph after grouping, and matrix D represents the degree matrix of the weighted adjacency matrix A of the causal graph. Summing by columns to obtain the out-degree matrix or summing by rows to obtain the in-degree matrix has the same effect. In the example, the out-degree matrix is ​​used to calculate the loss.

[0094] Inter-group loop structure penalty term L a1 With the group's loop structure penalty term L a2 Reduce the loop structure in the generated causal graph by suppressing the higher-order self-reachability of nodes:

[0095]

[0096]

[0097] Where I represents the identity matrix, and Let A and K be the normalized inter-group adjacency matrix and the normalized intra-group adjacency matrix, respectively, where k represents the group and its value ranges from [1, K]. k The k-th submatrix reconstructed from the weighted adjacency matrix A of the causal graph based on the grouping results is composed of elements in the weighted adjacency matrix A of the causal graph representing the connection relationships of the test points within the k-th group. This indicates that elements within the matrix are multiplied together. Let A be the in-degree matrix of the weighted adjacency matrix A in a causal graph. Let D be the out-degree matrix of the weighted adjacency matrix A in the causal graph. In this example, D equals 10, representing that there are 10 measurement points in the causal graph.

[0098] The loss function is constructed by weighting and summing the loss of the sparse causal inference module, the grouping regularization term, and the aforementioned penalty term. Simultaneously, the sparse causal inference module and the node grouping module are trained, enabling the model to extract sparse causal relationships. In this embodiment, the loss function of the sparse causal inference module is the mean square error between the current time-series data of the measurement point and the true value, specifically:

[0099]

[0100] in and x it These represent the predicted value and the actual value of the t-th data point in the current time series data of the i-th fault measurement point, respectively.

[0101] (4) The trained model can output a weighted adjacency matrix A of the causal graph and node grouping results. The method for determining causal relationships between measurement points and between subsets of measurement points uses a threshold to determine whether a causal relationship exists. Specifically:

[0102] Methods for determining the causal relationship between measurement points: If Here, sum(*) represents the sum of the matrix element values. That is, if the weighted adjacency matrix element of the causal graph is greater than the average value of the matrix elements, then the j-th fault measurement point is considered to have a causal effect on the i-th fault measurement point.

[0103] Method for determining causal relationships between subsets of measurement points in different groups: Calculate the weighted adjacency matrix A of the inter-group causal graph. g =S T AS, there exists a threshold δ, if a g,ij -a g,ji >δ, i≠j, a g,ij A represents g The element value in the i-th row and j-th column indicates that the j-th subset of measurement points has a causal effect on the i-th subset of measurement points, if a g,ji -a g,ij If δ > 0 and i ≠ j, then the i-th measurement point subset has a causal effect on the j-th measurement point subset. In this embodiment, the threshold δ is taken as an experimental empirical value of 0.05.

[0104] Once the causal relationships between measurement points and between subsets of measurement points are determined, a two-level causal graph within and between groups can be constructed using these causal relationships. This topological network can indicate the propagation path of the fault between measurement points when the gas turbine fails in the embodiment, and can intuitively characterize the root cause measurement point of the fault.

[0105] To prevent damage to the gas turbine's hot-channel components due to overheating, a cooling air system is installed to supply cold air into the turbine to cool components such as the stator blades, rotor blades, and impeller. In this embodiment, the fault warning system frequently triggers alarms regarding the cooling air pressure (X30, X32) of the second and third stage stator blades. On one hand, due to the low gas input, abnormal pressure difference in the combustion chamber, and low average temperature inside the turbine, the unit is operating at a low load. To maintain stable exhaust temperature, the controller reduces the pressure of the cooling air inside the gas turbine. On the other hand, the second and third stage stator blades of the gas turbine are connected to the compressor, which supplies cooling air; therefore, the compressor outlet pressure directly affects the cooling air pressure of the turbine stator blades. Furthermore, due to a malfunction in the cooling system, the exhaust temperature of the gas turbine will also be abnormal, affecting the compressor equipment through the mechanism of exhaust gas recycling. From the above fault cause analysis, it can be seen that fault measurement points X1 (natural gas volume flow rate), X2 (natural gas mass flow rate), and X4 (compressor outlet pressure) are the root cause variables of the fault.

[0106] Figure 5 The diagram illustrates a two-level causal graph network within and between gas turbine equipment groups constructed in a specific embodiment of the present invention. The results clearly demonstrate the propagation path of the fault and the root cause variables of the fault. The leftmost part of the diagram shows the causal relationship mapping between measurement points, with the dark area representing the causal effect of the corresponding column measurement point on the corresponding row measurement point. The middle part shows the grouping results of the nodes and the causal relationship mapping results between groups. The right side shows the causal graph structure based on the causal relationship mapping.

[0107] As a comparison method Figure 6 The image shows the results of constructing a causal graph network using the Neural-Granger method. The left side of the image shows the causal relationship mapping between measurement points, with dark areas indicating that the measurement points in the corresponding column have a causal effect on the measurement points in the corresponding row. The right side shows the structure of the causal graph drawn based on the causal relationship mapping. Through analysis and searching, it can be found that this method incorrectly identifies X29 (gas turbine power) and X7 (compressor bearing temperature) as root cause variables of the fault.

[0108] As a comparison method Figure 7 The image shows the results of intra-group and inter-group causal graphs constructed after grouping variables using the DTW-Kmeans clustering method and extracting Granger causal relationships within and between groups. The leftmost part of the image shows the causal mapping between measurement points; the darker areas represent the causal effect of corresponding column measurement points on corresponding row measurement points. The middle part shows the grouping results of the nodes and the inter-group causal mapping results. The right side shows the causal graph structure based on the causal mapping. Analysis and searching reveal that this method incorrectly identified X29 (gas turbine power) as the root cause variable of the fault.

[0109] The reason why the comparative methods all misclassified gas turbine power as the root cause variable is that gas turbine power does indeed affect equipment operation, thus affecting many measuring points. However, equipment failures do not originate at the gas turbine power measuring point, but rather propagate through fuel flow and compressor outlet pressure. Therefore, this invention can more accurately trace the root cause of the failure.

[0110] This invention addresses the problem of the large number of fault measurement points in typical gas turbine equipment by proposing a causal relationship similarity index. Starting from two complementary perspectives—global and local—it aggregates the causal relationships between variables, thereby achieving grouping of fault measurement points. Using the grouping results, a two-level causal graph structure (within and between groups) is constructed, intuitively representing the fault propagation path, tracing the root cause of the fault, and intelligently upgrading the fault repair of gas turbine equipment.

[0111] It should be understood that the present invention is not limited to the root cause tracing process of cooling air pressure failure of the second and third stage stator blades of the gas turbine as described in the above specific embodiment. Those skilled in the art can make equivalent modifications or substitutions without departing from the spirit of the present invention, and these equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A method for causal aggregation and grouping of root cause tracing for typical gas turbine equipment, characterized in that, Includes the following steps: (1) Obtain time-series data of typical equipment fault measurement points of the gas turbine, determine the time delay hyperparameter L of the time-series data, and generate past time-series data X from the time-series data based on the time delay hyperparameter. ＜t =[x 1,＜t ,x 2,＜t ,...,x D,＜t ] and current time series data X t =[x 1t ,x 2t ,...,x Dt ]; Where D is the number of fault measurement points, x it The current time series data of the i-th fault measurement point is composed of the last TL samples from the time series data of the i-th fault measurement point, where T is the total number of samples in the time series data of the typical equipment fault measurement points of the gas turbine. Each sample is the fault measurement point data at the current moment. i,＜t The past time series data represents the i-th fault measurement point. The past time series data corresponds one-to-one with the current time series data, satisfying that the fault measurement point data at the current moment corresponds to the fault measurement point data at the previous L moments. (2) Construct a causal aggregation and grouping model, which includes a sparse causal inference module and a node grouping module; wherein, the sparse causal inference module consists of D convolutional neural networks corresponding one-to-one with fault measurement points and a causal graph weighted adjacency matrix generation unit, and each convolutional neural network is used to generate a causal graph weighted adjacency matrix based on past time series data X. ＜t Predict the current time series data x of the corresponding fault measurement point it The causal graph weighted adjacency matrix generation unit is used to quantify and calculate the Granger causal relationship between fault measurement points based on the connection weights between the first hidden layer and the input layer of D convolutional neural networks, and generate a causal graph weighted adjacency matrix; the node grouping module includes a node feature extractor, a group feature extractor, and a grouping unit. The node feature extractor is used to extract features from past time series data X. ＜t Extract the temporal features of each fault measurement point and concatenate them with the corresponding location embedding vector to obtain the temporal features of each node; the group feature extractor is used to extract the temporal features from the current time series data X. t The feature vectors of the current time and corresponding time-delay data are extracted and then concatenated with the corresponding position embedding vectors to reduce the spatial dimension and obtain the group features; the grouping unit is used to construct a similarity matrix based on the node temporal features and group features and to group the nodes to obtain the node grouping results; (3) Transfer the past time series data X obtained in step (1) ＜t And the current time series data X t The input is fed into the causal aggregation and grouping model, which is trained by minimizing the loss function. After training, the causal aggregation and grouping model outputs the final causal graph weighted adjacency matrix and node grouping results. The loss function includes the error loss between the output and the ground truth of each convolutional neural network, a grouping regularization term, a grouping penalty term, an inter-group loop structure penalty term, and an intra-group loop structure penalty term. (4) Based on the final weighted adjacency matrix of the causal graph and the node grouping results, determine the causal relationship between fault measurement points and the causal relationship between fault measurement point subsets within the group, and complete the root cause tracing of the fault.

2. The method according to claim 1, characterized in that, The causal graph weighted adjacency matrix generation unit calculates the Granger causal relationship between fault measurement points based on the connection weights between the first hidden layer and the input layer of D convolutional neural networks, and generates a causal graph weighted adjacency matrix. The specific calculation method for each element of the causal graph weighted adjacency matrix is ​​as follows: Among them, a ij Let A be the element in the i-th row and j-th column of the weighted adjacency matrix A of the causal graph. In the convolutional neural network corresponding to the i-th fault test point, is the weight vector connecting the past time-series data input node of the j-th fault test point to the first hidden layer node z, where z is the number of hidden layer nodes in the first hidden layer; norm represents the second norm of the matrix.

3. The method according to claim 1, characterized in that, The node feature extractor consists of a first convolutional layer, an activation function layer, and a splicing layer, each with an independent convolutional process for each measurement point, connected in sequence.

4. The method according to claim 1, characterized in that, The group feature extractor consists of a second convolutional layer, an activation function layer, a splicing layer, and a third convolutional layer connected in sequence.

5. The method according to claim 1, characterized in that, The position embedding vector is obtained by calculating the function Pos.

6. The method according to claim 1, characterized in that, The grouping regularization term L o Grouping penalty item L g Inter-group loop structure penalty term L a1 With the group's loop structure penalty term L a2 Specifically: Where S represents the similarity matrix, I K This represents an identity matrix with dimension K, where K represents the number of groups, i.e., the number of group features. ||*|| F Let F denote the F-norm, Tr(*) denotes finding the trace of a matrix, matrix D is the degree matrix of the weighted adjacency matrix A of the causal graph, and I denotes the identity matrix. and Let A be the normalized inter-group adjacency matrix and the normalized intra-group adjacency matrix, respectively, where k∈[1,K]; k Let D be the k-th submatrix reconstructed from the weighted adjacency matrix A of the causal graph based on the grouping results. in Let D be the in-degree matrix of the weighted adjacency matrix A in a causal graph. out Let A be the out-degree matrix of the weighted adjacency matrix A in a causal graph.

7. The method according to claim 1, characterized in that, The grouping unit constructs a similarity matrix based on node temporal features and group features, and then groups the nodes to obtain the grouping results. Specifically: The similarity S between the temporal feature vector of each node and the temporal feature vector of the group is calculated using the vector dot product. ij Construct a similarity matrix: Among them, h i Let f represent the temporal feature vector of the i-th node. j Let K represent the time-series feature vector of the j-th group, and K represent the number of group features. For each fault measurement point, the similarity between the node temporal feature vector of the fault measurement point and the temporal feature vectors of all groups is compared, and the fault measurement point is assigned to the corresponding group with the highest similarity to obtain the node grouping result.

8. The method according to claim 1, characterized in that, The loss function includes the error loss between the output of each convolutional neural network and the true value, using the mean squared error L. mse Specifically: in and x it These represent the predicted value and the actual value of the t-th data point in the current time series data of the i-th fault measurement point, respectively.

9. The method according to claim 1, characterized in that, Step (4) specifically involves: Based on the final weighted adjacency matrix of the causal graph and the node grouping results, determine the causal relationship between the fault measurement points: if element a of the weighted adjacency matrix of the causal graph... ij If the value is greater than the average value of the matrix elements, then the j-th fault measurement point is considered to have a causal effect on the i-th fault measurement point. Determine the causal relationships between fault measurement point subsets within a group: Calculate the weighted adjacency matrix A of the inter-group causal graph. g =S T AS, there exists a threshold δ, if a g,ij -a g,ji >δ, i≠j, a g,ij A represents g The element value in the i-th row and j-th column indicates that the j-th fault measurement point subset has a causal effect on the i-th fault measurement point subset, if a g,ji -a g,ij If >δ, i≠j, then the i-th measurement point subset has a causal effect on the j-th measurement point subset.

10. The method according to claim 1, characterized in that, Step (4) further includes: Based on the causal relationships between fault measurement points and between fault measurement point subsets within a group, a two-level causal graph is constructed.

Images