On-line prediction method and electronic equipment for lubricating oil life of gas turbine

Abstract

The present invention provides an online prediction method for the lubricating oil life of a gas turbine, comprising: obtaining an oil consumption node diagram of a target gas turbine and rated oil parameters of the lubricating oil used by the target gas turbine, wherein the oil consumption node diagram includes oil consumption nodes and connecting edges between oil consumption nodes, and the connecting edges have connection values ​​determined according to the oil circuit transmission volume; pruning the system node diagram at the node level and the data level through a trained first spatiotemporal graph convolutional network to obtain an intermediate node diagram, wherein the first spatiotemporal graph convolutional network is trained to identify the operating state of the oil consumption node through a first training data set; inputting the intermediate node diagram and the rated oil parameters into a trained second spatiotemporal graph convolutional network for prediction processing to obtain a prediction result of the lubricating oil life of the gas turbine, wherein the second spatiotemporal graph convolutional network is trained to predict the lubricating oil life through a second training data set. The present invention improves the accuracy of oil product life detection of the gas turbine lubricating oil.

Description

Technical Field

[0001] The invention relates to the field of intelligent monitoring of gas turbines, and in particular to an online prediction method and electronic equipment for lubricating oil life of a gas turbine. Background Art

[0002] A gas turbine is a device that uses high-temperature, high-pressure gas generated by fuel combustion to drive the turbine to rotate, thereby driving the generator to generate electricity. The lubricating oil system of a gas turbine is an important auxiliary system used to provide lubrication, cooling and cleaning for various components of the gas turbine. During the operation of the gas turbine, the lubricating oil is extracted from the oil tank, filtered by the filter, powered by the lubricating oil pump, and transported to various components that need lubrication through the lubricating oil pipeline. The lubricating oil forms an oil film on the surface of the component, which plays a role in lubrication, cooling and cleaning. When the gas turbine is operating normally, the lubricating oil system needs to maintain a stable oil pressure and flow to ensure the normal operation of the gas turbine. In order to ensure the normal operation of the gas turbine, it is necessary to evaluate the service life of the lubricating oil, so as to carry out corresponding maintenance and maintenance according to the evaluation results. Generally speaking, for the service life evaluation of the lubricating oil of the gas turbine, it is necessary to conduct offline (shutdown) testing of the gas turbine, and test the viscosity, acid value, oxidation degree and other indicators of the oil to evaluate its quality. High-quality lubricating oil can maintain good performance for a longer time, and the higher the quality, the longer the service life. However, since the gas turbine needs to be shut down to take oil samples, the lubricating oil in the gas turbine cannot be tested at any time. Moreover, oil testing is a static testing method, which cannot comprehensively consider the oil degradation caused by various abnormalities during operation, making the accuracy of the test results only applicable to static oil life prediction. Since gas turbines work for a long time and are not easily shut down, using static oil evaluation results to guide the maintenance and care of gas turbines will result in large errors. Therefore, the existing oil life detection of gas turbine lubricating oil has the problem of low accuracy. Summary of the invention

[0003] The embodiment of the present invention provides an online prediction method for the lubricating oil life of a gas turbine, aiming to solve the problem of low accuracy in the oil life detection of the existing gas turbine lubricating oil. The oil consumption node graph of the target gas turbine is pruned by a first spatiotemporal graph convolutional network to obtain an intermediate node graph that is strongly related to the operating state. The intermediate node graph and the rated oil parameters of the lubricating oil used by the target gas turbine are processed by a second graph convolutional network to obtain the prediction result of the lubricating oil life of the gas turbine. Since the lubricating oil does not need to be sampled, the gas turbine does not need to be shut down for oil sampling. Through the pruning process, the information that is strongly related to the operating state is retained, so that the second graph convolutional network can use the information that is strongly related to the operating state and the rated oil parameters to predict the lubricating oil life of the gas turbine during operation, thereby improving the accuracy of the oil life detection of the gas turbine lubricating oil.

[0004] In a first aspect, an embodiment of the present invention provides an online prediction method for lubricating oil life of a gas turbine, the method comprising:

[0005] Obtaining an oil consumption node graph of a target gas turbine and rated oil parameters of lubricating oil used by the target gas turbine, wherein the oil consumption node graph includes oil consumption nodes and connecting edges between the oil consumption nodes, wherein the connecting edges have connection values ​​determined according to oil line transmission volume, and the oil consumption nodes include corresponding historical operating condition data, historical temperature data, historical cleaning data, and historical external environment data;

[0006] The system node graph is pruned at the node level and the data level by using a trained first spatiotemporal graph convolutional network to obtain an intermediate node graph, wherein the pruning process includes node pruning and data pruning, and the first spatiotemporal graph convolutional network is trained by performing operation status recognition of oil consumption nodes using a first training data set;

[0007] The intermediate node graph and the rated oil parameters are input into a trained second spatiotemporal graph convolutional network for prediction processing to obtain a prediction result of the lubricant oil life of the gas turbine. The second spatiotemporal graph convolutional network is trained for lubricant oil life prediction using a second training data set.

[0008] In a second aspect, an embodiment of the present invention provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the steps of the online prediction method for the lubricating oil life of a gas turbine provided in an embodiment of the present invention are implemented.

[0009] In an embodiment of the present invention, an oil consumption node graph of a target gas turbine and rated oil parameters of the lubricating oil used by the target gas turbine are obtained, wherein the oil consumption node graph includes oil consumption nodes and connecting edges between the oil consumption nodes, wherein the connecting edges have connection values ​​determined according to the oil circuit transmission volume, and the oil consumption nodes include corresponding historical operating condition data, historical temperature data, historical cleaning data, and historical external environment data; the system node graph is pruned at the node level and the data level by a trained first spatiotemporal graph convolutional network to obtain an intermediate node graph, wherein the pruning process includes node pruning and data pruning, and the first spatiotemporal graph convolutional network is trained to identify the operating status of the oil consumption node using a first training data set; the intermediate node graph and the rated oil parameters are input into a trained second spatiotemporal graph convolutional network for prediction processing to obtain a prediction result of the lubricating oil life of the gas turbine, and the second spatiotemporal graph convolutional network is trained to predict the lubricating oil life using a second training data set. The oil consumption node graph of the target gas turbine is pruned through the first spatiotemporal graph convolutional network to obtain an intermediate node graph that is strongly correlated with the operating status. The intermediate node graph and the rated oil parameters of the lubricating oil used by the target gas turbine are processed through the second graph convolutional network to obtain the lubricating oil life prediction result of the gas turbine. Since there is no need to sample the lubricating oil, there is no need to shut down the gas turbine for oil extraction. Through pruning, the information that is strongly correlated with the operating status is retained, so that the second graph convolutional network can use the information that is strongly correlated with the operating status and the rated oil parameters to predict the lubricating oil life of the gas turbine during operation, thereby improving the accuracy of oil life detection of the gas turbine lubricating oil. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0011] Figure 1 It is a flow chart of an online prediction method for lubricating oil life of a gas turbine provided by an embodiment of the present invention;

[0012] Figure 2 It is a structural schematic diagram of a first to-be-trained graph total area network provided by an embodiment of the present invention;

[0013] Figure 3 is a schematic diagram of the structure of a second graph convolutional network to be trained provided by an embodiment of the present invention;

[0014] Figure 4It is a structural schematic diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0015] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0016] like Figure 1 As shown, Figure 1 The method flow chart of an online prediction method of lubricating oil life of a gas turbine provided by an embodiment of the present invention. The online prediction method of lubricating oil life of a gas turbine comprises the following steps:

[0017] 101. Obtain an oil usage node diagram of a target gas turbine and rated oil parameters of lubricating oil used by the target gas turbine.

[0018] In an embodiment of the present invention, the oil consumption node graph includes oil consumption nodes and connecting edges between the oil consumption nodes, the connecting edges have connection values ​​determined according to the oil transmission volume, and the oil consumption nodes include corresponding historical operating condition data, historical temperature data, historical cleaning data and historical external environment data.

[0019] The above-mentioned oil-using nodes can be functional nodes in the transmission system, hydraulic control system, and control system that need to use lubricating oil, and specifically can be bearing seats, bearing shafts, gear boxes, lubricating oil pumps, seals, etc. Each functional node corresponds to an oil-using node, and the functional nodes are connected through oil circuits, and lubricating oil is then transported to the functional nodes through the oil circuits. The connection value between the two oil-using nodes is determined according to the oil circuit transmission volume between the two functional nodes. The larger the oil circuit transmission volume, the larger the connection value, and vice versa, the smaller the connection value.

[0020] The historical operating condition data may include the operating time, load, speed, stroke, etc. of the oil consumption node in the gas turbine. For example, if an oil consumption node is often in a high load state in the past operation, this information may affect the prediction of the lubricant life. The above historical operating condition data can be obtained based on the operation log.

[0021] The above historical temperature data may include the operating temperature of the oil node in the gas turbine, the temperature of the lubricating oil, etc. If the lubricating oil of the oil node is often in a high temperature environment in the past operation, then this information may affect the prediction of the lubricating oil life, and can be collected by setting a temperature sensor at the oil node.

[0022] The above historical cleaning data may include data such as cleaning time and cleaning method of the oil-using nodes in the gas turbine. If the lubricating oil of the oil-using nodes is often contaminated in the past operation, then this information may affect the prediction of the lubricating oil life. The historical cleaning data can be obtained based on the maintenance records of the maintenance personnel.

[0023] The above historical external environment data may include the ambient temperature, humidity, pollutant concentration, etc. of the oil consumption node in the gas turbine. If the oil consumption node was often in a bad environment in the past operation, this information may affect the prediction of the lubricating oil life.

[0024] 102. Prune the system node graph at the node level and the data level through the trained first spatiotemporal graph convolutional network to obtain an intermediate node graph.

[0025] In an embodiment of the present invention, the first spatiotemporal convolutional network can be a network composed of a graph convolutional network GNN and a long short-term memory network LSMT. The pruning process includes node pruning and data pruning, and the first spatiotemporal graph convolutional network is trained by performing operating status recognition of oil consumption nodes through a first training data set.

[0026] Specifically, the node pruning can be understood as deleting certain oil consumption nodes that have little or negligible impact on the operating status. The data pruning can be understood as trimming certain data with a longer time period to obtain shorter time series data that is still representative.

[0027] By pruning the oil consumption node graph of the target gas turbine through the first spatiotemporal graph convolutional network, the oil consumption nodes and historical data can be screened and processed to obtain a more concise and representative intermediate node graph.

[0028] 103. Input the intermediate node graph and the rated oil parameters into the trained second spatiotemporal graph convolutional network for prediction processing to obtain a prediction result of the lubricating oil life of the gas turbine.

[0029] In an embodiment of the present invention, the second spatiotemporal convolutional network may be a network based on a graph convolutional network GNN and a long short-term memory network LSMT. The second spatiotemporal graph convolutional network is obtained by performing lubricant life prediction training on a second training data set.

[0030] Specifically, the rated oil parameters and the intermediate node graph can be input into the trained second spatiotemporal graph convolutional network for prediction processing to obtain the prediction result of the lubricant life of the gas turbine. The lubricant life prediction result can include a predicted life value, such as 398 days, indicating that the lubricant in the target gas turbine can be used for 398 days.

[0031] After obtaining the prediction result of the lubricating oil life of the target gas turbine, a maintenance and servicing plan for the target gas turbine can be formulated according to the prediction result of the lubricating oil life.

[0032] In an embodiment of the present invention, an oil consumption node graph of a target gas turbine and rated oil parameters of the lubricating oil used by the target gas turbine are obtained, wherein the oil consumption node graph includes oil consumption nodes and connecting edges between the oil consumption nodes, wherein the connecting edges have connection values ​​determined according to the oil circuit transmission volume, and the oil consumption nodes include corresponding historical operating condition data, historical temperature data, historical cleaning data, and historical external environment data; the system node graph is pruned at the node level and the data level by a trained first spatiotemporal graph convolutional network to obtain an intermediate node graph, wherein the pruning process includes node pruning and data pruning, and the first spatiotemporal graph convolutional network is trained to identify the operating status of the oil consumption node using a first training data set; the intermediate node graph and the rated oil parameters are input into a trained second spatiotemporal graph convolutional network for prediction processing to obtain a prediction result of the lubricating oil life of the gas turbine, and the second spatiotemporal graph convolutional network is trained to predict the lubricating oil life using a second training data set. The oil consumption node graph of the target gas turbine is pruned through the first spatiotemporal graph convolutional network to obtain an intermediate node graph that is strongly correlated with the operating status. The intermediate node graph and the rated oil parameters of the lubricating oil used by the target gas turbine are processed through the second graph convolutional network to obtain the lubricating oil life prediction result of the gas turbine. Since there is no need to sample the lubricating oil, there is no need to shut down the gas turbine for oil extraction. Through pruning, the information that is strongly correlated with the operating status is retained, so that the second graph convolutional network can use the information that is strongly correlated with the operating status and the rated oil parameters to predict the lubricating oil life of the gas turbine during operation, thereby improving the accuracy of oil life detection of the gas turbine lubricating oil.

[0033] Optionally, the acquiring of the oil consumption node diagram of the target gas turbine comprises: determining a first oil consumption node diagram according to the main oil circuit structure of the target gas turbine, the first oil consumption node diagram comprising a first oil consumption node and a first connecting edge between the first oil consumption nodes, the first connecting edge having a first connection value determined according to the oil circuit transmission volume; determining a second oil consumption node diagram according to the auxiliary oil circuit structure of the target gas turbine, the second oil consumption node diagram comprising a second oil consumption node and a second connecting edge between the second oil consumption nodes, the second connecting edge having a second connection value determined according to the oil circuit transmission volume; determining the second oil consumption node diagram according to the auxiliary oil circuit structure of the target gas turbine, the second oil consumption node diagram comprising a second oil consumption node and a second connecting edge between the second oil consumption nodes, the second connecting edge having a second connection value determined according to the oil circuit transmission volume; determining the second oil consumption node diagram according to the target gas turbine; The emergency oil circuit structure of the gas turbine determines a third oil consumption node graph, wherein the third oil consumption node graph includes a third oil consumption node and a third connecting edge between the third oil consumption nodes, and the third connecting edge has a third connection value determined according to the oil circuit transmission volume; the first oil consumption node graph, the second oil consumption node graph and the third oil consumption node graph are merged to obtain an initial oil consumption node graph of the target gas turbine; historical operating condition data, historical temperature data, historical cleaning data and historical external environment data corresponding to the oil consumption nodes are added to the initial oil consumption node graph to obtain the oil consumption node graph of the target gas turbine.

[0034] In an embodiment of the present invention, the first oil usage node diagram, the second oil usage node diagram and the third oil usage node diagram can be determined respectively according to the main oil circuit structure, the auxiliary oil circuit structure and the emergency oil circuit structure, and then the first oil usage node diagram, the second oil usage node diagram and the third oil usage node diagram are merged, and the historical data corresponding to the oil usage nodes are added to obtain the oil usage node diagram of the target gas turbine.

[0035] The embodiment of the invention takes into account the different oil circuit structures of the gas turbine and obtains the connection relationship and oil circuit transmission volume of each oil consumption node, so as to more comprehensively and accurately represent the oil supply system of the target gas turbine. By fusing the first oil consumption node graph, the second oil consumption node graph and the third oil consumption node graph, a more comprehensive oil consumption node graph can be obtained, which includes the oil consumption nodes and connection edges of each oil circuit structure.

[0036] Adding historical data corresponding to the oil usage node can provide richer information for evaluating the life and performance of lubricants. These historical data include historical operating data, historical temperature data, historical cleaning data, and historical external environment data, which can reflect the use of lubricants in gas turbines under different operating conditions, provide more comprehensive information for evaluating the life of lubricants, and thus improve the accuracy of life prediction.

[0037] Optionally, the first oil consumption node graph, the second oil consumption node graph and the third oil consumption node graph are merged to obtain the initial oil consumption node graph of the target gas turbine, including: merging the nodes belonging to the same oil consumption point among the first oil consumption node, the second oil consumption node and the third oil consumption node to obtain merged oil consumption nodes; merging the edges with the same oil consumption point among the first connecting edge, the second connecting edge and the third connecting edge to obtain merged connecting edges, and weightedly calculating the first connection value, the second connection value and the third connection value according to a preset weight value to obtain the connection value of the merged connecting edge; based on the merged oil consumption nodes, the merged connecting edges and the connection values ​​of the merged connecting edges, constructing the initial oil consumption node graph of the target gas turbine.

[0038] In the embodiment of the present invention, first, the nodes belonging to the same oil consumption point among the first oil consumption node, the second oil consumption node and the third oil consumption node are merged to obtain a merged oil consumption node, so as to ensure that the oil consumption node remains unique in the merged graph.

[0039] Next, the edges with the same oil usage point among the first connection edge, the second connection edge, and the third connection edge are merged to obtain a merged connection edge. The connection edge merging process can ensure that the connection edge information remains consistent in the fused graph.

[0040] After merging the connection edges, the first connection value, the second connection value, and the third connection value are weighted according to the preset weight value to obtain the connection value of the merged connection edge. This step can ensure that in the fused graph, the weight values ​​of the connection edges of different oil circuit structures can also be reasonably merged together. Different oil circuit structures correspond to different weight values. Specifically, for a connection value of a connection edge, it can be obtained by weighted calculation using the following formula:

[0041] d i,j =δd (i,j)1 +εd (i,j)2 +γd (i,j)3

[0042] Among them, d i,j is the connection value of the merged connection edge, δ is the connection edge d in the first oil consumption node graph (i,j)1 The weight value of ε is the connecting edge d in the second oil consumption node graph. (i,j)2 The weight value of γ is the connection edge d in the first oil consumption node graph. (i,j)3 The weight value is δ+ε+γ=1.

[0043] Finally, based on the merged oil consumption nodes, merged connection edges, and connection values ​​of the merged connection edges, the initial oil consumption node graph of the target gas turbine is constructed. This ensures that the fused graph can fully represent the oil supply system structure and connection relationship of the target gas turbine.

[0044] The embodiment of the present invention can merge the first oil usage node diagram, the second oil usage node diagram and the third oil usage node diagram into a complete oil usage node diagram for subsequent lubricant life prediction and analysis. This fusion process can improve the accuracy and reliability of the data and provide better support for subsequent prediction and analysis.

[0045] Optionally, before the trained first spatiotemporal graph convolutional network is used to prune the system node graph at the node level and the data level to obtain the intermediate node graph, the method further includes: obtaining a first graph convolutional network to be trained and a first training data set, the first graph convolutional network to be trained including a first output and a second output, the first output being a predicted value of the operating state of the oil-using node, the second output being a pruned node graph, the first training data set including a first sample node graph and true values ​​of the operating states of the oil-using nodes corresponding to the first sample node graph; inputting the first sample node graph into the first graph convolutional network to be trained for processing to obtain a predicted value of the operating state of the first sample node graph and a pruned node graph of the first sample node graph; The operating state loss value between the operating state prediction value of the oil-using node of the first sample node graph and the operating state true value of the oil-using node of the first sample node graph is calculated by a preset first loss function, and the node graph loss value between the first sample node graph and the pruned node graph of the first sample node graph is calculated by a preset second loss function, and a total loss value is determined based on the operating state loss value and the node graph loss value; based on the total loss value, the model parameters of the first graph convolution network to be trained are adjusted by a back propagation algorithm, and the parameter adjustment process is iterated until the total loss value reaches a preset condition, and the training is stopped to obtain a trained first spatiotemporal graph convolution network, and the trained first spatiotemporal graph convolution network only retains the second output.

[0046] In an embodiment of the present invention, the first graph convolution network to be trained can be a network composed of a graph convolution network GNN and a long short-term memory network LSMT. The true value of the operating state of the oil-using node of the first sample node graph can be understood as the real operating state of each oil-using node under the condition of the first sample node graph. The above operating state can be quantified as a score value between normal and abnormal, for example, it can be quantified as a score value between 0 and 1. The higher the score value, the closer it is to normal, and the lower the value, the closer it is to abnormal. The above operating state true value can be obtained by annotation by experts.

[0047] The embodiment of the present invention uses the prediction of the operating status to assist in the pruning of the oil consumption node graph, so that the remaining nodes in the intermediate nodes after pruning also include the main information that can predict the operating status.

[0048] The first loss function can be an existing loss function, such as a square error loss function or a cross entropy loss function. The second loss function can be a residual loss function or a square error loss function. It should be noted that for the operating state loss value, the training expectation is the minimum, so that the first spatiotemporal graph convolutional network can accurately predict the operating state of the oil consumption node graph. For the node graph loss value, the training expectation is the maximum, so that the oil consumption node graph can be pruned to the greatest extent, while retaining the most concise information and reducing the impact on the prediction accuracy of the operating state. When the graph convolutional network to be trained outputs the smallest-scale pruned node graph, and the operating state prediction value is very close to the operating state true value, it can be determined that the model training is relatively successful. Specifically, when the total loss value reaches the preset value, the training can be stopped, the first output can be deleted, and only the second output can be retained to obtain the trained first spatiotemporal graph convolutional network.

[0049] The training process enables the first spatiotemporal graph convolutional network to learn how to perform pruning based on the input node graph to obtain an intermediate node graph with fewer nodes. This pruning process can remove redundant nodes and connecting edges, improving the efficiency of subsequent predictions. At the same time, parameter adjustment through the back-propagation algorithm can optimize the performance of the network and improve the accuracy and reliability of the pruning process.

[0050] Optional, please refer to Figure 2 , Figure 2 is a schematic diagram of the structure of a first to-be-trained graph total area network provided by an embodiment of the present invention, such as Figure 2As shown, the first graph convolution network to be trained includes a first spatial convolution layer to be trained, a first temporal convolution layer to be trained, a spatial mask layer to be trained, a temporal mask layer to be trained, a first feature fusion layer, a first output layer to be trained and a second output layer to be trained. The output of the first spatial convolution layer to be trained is respectively connected to the input of the spatial mask layer to be trained and the input of the first temporal convolution layer to be trained, the output of the spatial mask layer to be trained is respectively connected to the input of the first training temporal convolution layer and the first feature fusion layer, the output of the temporal convolution layer to be trained is respectively connected to the input of the first feature fusion layer and the temporal mask layer to be trained, the output of the temporal mask layer to be trained is connected to the input of the first feature fusion layer, the output of the first feature fusion layer is respectively connected to the input of the first output layer and the second output layer, the first output layer is used to output the running state prediction value of the oil consumption node, the second output layer is used to output the pruned node graph, the first sample node graph is input into the first graph convolution network to be trained for processing, and the running state prediction value of the first sample node graph and the first sample node graph are obtained. The pruned node graph of the first sample node graph includes: inputting the first sample node graph into the first graph convolution network to be trained, performing spatial feature extraction on the first sample node graph through the first spatial convolution layer to be trained, and obtaining the spatial feature of the first sample node graph; performing masking processing on the spatial feature of the first sample node graph in the spatial dimension through the spatial mask layer to be trained, and obtaining the spatial mask feature; performing temporal feature extraction on the spatial feature and the spatial mask feature through the first temporal convolution layer to be trained, and obtaining the spatiotemporal feature of the first sample node graph; performing masking processing on the spatiotemporal feature of the first sample node graph in the temporal dimension through the temporal mask layer to be trained, and obtaining the temporal mask feature; performing feature fusion on the spatial mask feature, the temporal mask feature and the spatiotemporal feature, and obtaining the first fused feature; performing full connection processing on the first fused feature through the first output layer, and obtaining the running state prediction value of the first sample node graph; and performing deconvolution processing on the first fused feature through the second output layer, and obtaining the pruned node graph of the first sample node graph.

[0051] In an embodiment of the present invention, the first sample node graph can be input into the first graph convolution network to be trained, and the spatial features of the first sample node graph can be extracted by the first spatial convolution layer to be trained to obtain the spatial features of the first sample node graph. The spatial features of the first sample node graph are masked in the spatial dimension by the spatial mask layer to be trained to obtain the spatial mask features. The spatial features and the spatial mask features are extracted by the first temporal convolution layer to be trained to obtain the spatiotemporal features of the first sample node graph. The spatiotemporal features of the first sample node graph are masked in the temporal dimension by the temporal mask layer to be trained to obtain the temporal mask features. The spatial mask features, the temporal mask features and the spatiotemporal features are feature fused to obtain the first fused features. The first fused features are fully connected by the first output layer to obtain the running state prediction value of the first sample node graph, and the first fused features are deconvolved by the second output layer to obtain the pruned node graph of the first sample node graph. The nodes in the first sample node graph are masked by the spatial mask layer, which is equivalent to deleting the corresponding masked nodes. Therefore, during the training process, the spatial mask layer learns how to mask the nodes without affecting the accuracy of the running status prediction value. The spatial features are masked in the time dimension by the temporal mask layer, which is equivalent to clipping the feature information of the corresponding time period without affecting the accuracy of the running status prediction value, so that the pruned node graph still retains the information that can accurately predict the running status, discards the redundant information, and avoids information interference.

[0052] This network structure can effectively extract useful features from the node graph and predict the operating status of the oil-using nodes. At the same time, through the pruning process in the network, redundant nodes and connecting edges can be removed, as well as redundant historical data, to improve the efficiency of subsequent predictions.

[0053] Optionally, the operation state loss value between the operation state prediction value of the oil consumption node of the first sample node graph and the operation state true value of the oil consumption node of the first sample node graph is calculated by a preset first loss function, and the node graph loss value between the first sample node graph and the pruned node graph of the first sample node graph is calculated by a preset second loss function, including: calculating the operation state loss value between the operation state prediction value of the oil consumption node of the first sample node graph and the operation state true value of the oil consumption node of the first sample node graph by a preset first loss function, and the first loss function is:

[0054]

[0055] Where Nc is the number of nodes using oil in the first sample node graph, f(x i ) is the i-th oil-using node x in the first sample node graph iThe predicted value of the running status, y i is the i-th oil-using node x in the first sample node graph i The true value of the running state, y j is the true value of the operating status of the jth oil-consuming node in the first sample node graph, d i,j is the sum of the connection values ​​of at least one edge between the i-th oil consumption node and the j-th oil consumption node. It can be seen that the operating state loss value L 1 The smaller it is, the smaller the predicted value and true value of the operating status of the same oil-consuming node are, and the difference between the predicted value of the operating status of an oil-consuming node and the true value of the operating status of other oil-consuming nodes is determined according to the distance between the two oil-consuming nodes. The longer the distance is, the more connection edges are required to connect, so the total connection value is larger; the closer the distance is, the fewer connection edges are required to connect, so the total connection value is smaller.

[0056] The node graph loss value between the first sample node graph and the pruned node graph of the first sample node graph is calculated by a preset second loss function, where the second loss function is:

[0057]

[0058] Wherein, M is the total number of data dimensions corresponding to the historical operating condition data, historical temperature data, historical cleaning data and historical external environment data in the oil consumption node of the first sample node, and x i,k is the i-th oil-using node x in the first sample node graph i The k-th dimension data in G(x i,k ) is the i-th oil-using node x in the pruned node graph of the first sample node graph i The pruned data of the k-th dimension in N G is the first sample node graph |G(x i,k )-x i,k |≠0, the number of oil consumption nodes, G(d i,j ) is the i-th oil-using node x in the pruning node graph of the first sample node i The sum of the connection values ​​of at least one connecting edge to the jth oil-using node. It can be seen that the node graph loss value L 2 The larger it is, the greater the difference in the number of oil consumption nodes and the amount of historical data between the pruned node graph and the first sample node graph. That is, the smaller the pruned node graph is, the greater the residual error with the first sample node graph is.

[0059] Optionally, determining the total loss value based on the running state loss value and the node graph loss value includes: calculating the total loss value between the running state loss value and the node graph loss value by using a total loss function, and the total loss function is:

[0060] L t =(1-α)·L 1 -α·L 2

[0061] Here, α ranges from 0 to 1 and increases with the number of iterations.

[0062] In the embodiment of the present invention, the training expectation for the running state loss value is the minimum, so that the first spatiotemporal graph convolutional network can accurately predict the running state of the oil consumption node graph. And for the node graph loss value, the training expectation is the maximum, so that the oil consumption node graph can be pruned to the greatest extent, while retaining the most concise information and reducing the impact on the prediction accuracy of the running state. It can be seen that the total loss value L t The smaller the L 1 The smaller the L 2 The larger the total loss value, the greater the total loss value. Therefore, through the total loss function, the first graph convolutional network to be trained can be trained with the minimum total loss value as the optimization goal.

[0063] Optionally, before inputting the intermediate node graph and the rated oil parameters into the trained second spatiotemporal graph convolutional network for prediction processing to obtain the prediction result of the lubricant life of the gas turbine, the method includes: obtaining a second graph convolutional network to be trained and a second training data set, the second graph convolutional network to be trained includes a third output, the third output is a predicted value of the lubricant life, the second training data set includes a second sample node graph, a sample rated oil parameter and a true value of the lubricant life of the second sample node graph, and the second sample node graph is pruned by the trained first spatiotemporal graph convolutional network Obtain; input the second sample node graph and the sample rated oil parameters into the second graph convolution network to be trained for processing to obtain the predicted value of lubricant life of the second sample node graph; calculate the life loss value between the predicted value of lubricant life of the second sample node graph and the true value of lubricant life of the second sample node graph through a preset third loss function; based on the life loss value, adjust the model parameters of the second graph convolution network to be trained through a back propagation algorithm, and iterate the parameter adjustment process until the life loss value reaches a preset condition, stop training, and obtain a trained second spatiotemporal graph convolution network.

[0064] In an embodiment of the present invention, the second graph convolutional network to be trained may be a network composed of a graph convolutional network GNN and a long short-term memory network LSMT. The true value of the lubricating oil life of the second sample node graph may be understood as the actual service life of the lubricating oil in the target gas turbine under the condition of the second sample node graph, and the true value of the lubricating oil life may be days.

[0065] A second graph convolution network to be trained and a second training data set are obtained. The second graph convolution network to be trained includes a third output, which is a predicted value of lubricant life. The second training data set includes a second sample node graph, sample rated oil parameters, and a true value of lubricant life of the second sample node graph. The second sample node graph is obtained by pruning the trained first spatiotemporal graph convolution network.

[0066] Specifically, the above true value can be determined based on historical data. For example, if the actual expiration date of the lubricant life is September 1, 2018, the historical data before September 1, 2017 can be obtained to construct a first sample node graph, and the first sample node graph is pruned through the trained first spatiotemporal graph convolutional network to obtain a second sample node graph, and the number of days between September 1, 2017 and September 1, 2018 is used as the true value of the lubricant life in the second sample node graph.

[0067] Next, the second sample node graph and the sample rated oil parameters are input into the second graph convolution network to be trained for processing to obtain the predicted value of the lubricant life of the second sample node graph. The life loss value between the predicted value of the lubricant life of the second sample node graph and the true value of the lubricant life of the second sample node graph is calculated by a preset third loss function. The above-mentioned sample rated oil parameters serve as auxiliary information, so that the second graph convolution network to be trained can learn the life prediction of various types of oil under different operating conditions. The third loss function can be an existing loss function, such as a square difference loss function or a cross entropy loss function.

[0068] Then, based on the life loss value, the model parameters of the second graph convolutional network to be trained are adjusted through the back propagation algorithm. This process is iterated until the life loss value reaches the preset condition and the training is stopped to obtain the trained second spatiotemporal graph convolutional network. For the life loss value, the training expectation is the minimum, so that the second spatiotemporal graph total area network can accurately predict the life of the lubricant.

[0069] Through the above training process, the second spatiotemporal graph convolutional network can learn how to predict the life of lubricating oil based on the input intermediate node graph and rated oil parameters. Parameter adjustment through the back propagation algorithm can optimize the performance of the network and improve the accuracy and reliability of life prediction.

[0070] Optional, please refer to Figure 3 , Figure 3 is a schematic diagram of the structure of a second graph convolutional network to be trained provided by an embodiment of the present invention, such as Figure 3As shown, the second graph convolution network to be trained includes a main input layer to be trained, an auxiliary input layer to be trained, a second spatial convolution layer to be trained, a second time convolution layer to be trained, a second feature fusion layer, and a third output layer to be trained. The outputs of the main input layer to be trained and the auxiliary input layer to be trained are connected to the input of the second feature fusion layer at the same time, the output of the second feature fusion layer is connected to the input of the second spatial convolution layer to be trained, the output of the second spatial convolution layer to be trained is connected to the input of the second time convolution layer to be trained, the output of the second time convolution layer to be trained is connected to the input of the third output layer to be trained, the input of the third output layer to be trained is used to output the predicted value of lubricant life, the second sample node graph and the sample rated oil parameters are input into the second graph convolution network to be trained for processing, and the lubricant life of the second sample node graph is obtained. The method comprises: inputting the second sample node graph into the main input layer to be trained for processing to obtain the main input feature, and inputting the sample rated oil parameter into the auxiliary input layer to be trained for processing to obtain the auxiliary input feature; fusing the main input feature with the auxiliary input feature through the second feature fusion layer to obtain the fused input feature; inputting the fused input feature into the second graph convolution network to be trained, performing spatial feature extraction on the fused input feature through the second spatial convolution layer to be trained to obtain the fused spatial feature of the second sample node graph; performing temporal feature extraction on the fused spatial feature through the second temporal convolution layer to be trained to obtain the second spatiotemporal feature of the second sample node graph; performing full connection processing on the second spatiotemporal feature through the third output layer to obtain the lubricant life prediction value of the second sample node graph.

[0071] In an embodiment of the present invention, the main input layer to be trained may be a 1×1 convolution kernel, and the auxiliary input layer to be trained may be an encoder based on Transformer, through which the sample rated oil parameters are encoded into features with computer semantics.

[0072] Specifically, the second sample node graph is input into the main input layer to be trained and convolution is performed through a 1×1 convolution kernel to obtain the main input feature. The sample rated oil parameters are input into the auxiliary input layer to be trained and encoded through Transformer to obtain the auxiliary input feature. The main input feature and the auxiliary input feature are fused through the second feature fusion layer to obtain the fused input feature. The fused input feature is input into the second graph convolution network to be trained, and the spatial feature of the fused input feature is extracted through the second spatial convolution layer to be trained to obtain the fused spatial feature of the second sample node graph. The temporal feature of the fused spatial feature is extracted through the second temporal convolution layer to be trained to obtain the second spatiotemporal feature of the second sample node graph. The second spatiotemporal feature is fully connected through the third output layer to obtain the predicted value of the lubricant life of the second sample node graph.

[0073] By training the second graph convolution model to be trained, useful features can be effectively extracted from the intermediate node graph and rated oil parameters, and the lubricant life can be predicted. Parameter adjustment through the back propagation algorithm can optimize the performance of the network and improve the accuracy and reliability of life prediction.

[0074] See also Figure 4 , Figure 4 is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention, such as Figure 4 As shown, it includes: a memory 402, a processor 401, and a computer program of an online prediction method for lubricating oil life of a gas turbine stored in the memory 402 and executable on the processor 401, wherein:

[0075] The processor 401 is used to call the computer program stored in the memory 402 to perform the following steps:

[0076] Obtaining an oil consumption node graph of a target gas turbine and rated oil parameters of lubricating oil used by the target gas turbine, wherein the oil consumption node graph includes oil consumption nodes and connecting edges between the oil consumption nodes, wherein the connecting edges have connection values ​​determined according to oil line transmission volume, and the oil consumption nodes include corresponding historical operating condition data, historical temperature data, historical cleaning data, and historical external environment data;

[0077] The system node graph is pruned at the node level and the data level by using a trained first spatiotemporal graph convolutional network to obtain an intermediate node graph, wherein the pruning process includes node pruning and data pruning, and the first spatiotemporal graph convolutional network is trained by performing operation status recognition of oil consumption nodes using a first training data set;

[0078] The intermediate node graph and the rated oil parameters are input into a trained second spatiotemporal graph convolutional network for prediction processing to obtain a prediction result of the lubricant oil life of the gas turbine. The second spatiotemporal graph convolutional network is trained for lubricant oil life prediction using a second training data set.

[0079] Optionally, the acquiring of the oil consumption node diagram of the target gas turbine executed by the processor 401 includes:

[0080] Determine a first oil consumption node graph according to the main oil circuit structure of the target gas turbine, wherein the first oil consumption node graph includes first oil consumption nodes and first connection edges between the first oil consumption nodes, and the first connection edges have first connection values ​​determined according to the oil circuit transmission volume;

[0081] determining a second oil consumption node graph according to the auxiliary oil circuit structure of the target gas turbine, wherein the second oil consumption node graph includes second oil consumption nodes and second connection edges between the second oil consumption nodes, and the second connection edges have second connection values ​​determined according to the oil circuit transmission volume;

[0082] Determine a third oil consumption node diagram according to the emergency oil circuit structure of the target gas turbine, wherein the third oil consumption node diagram includes third oil consumption nodes and third connection edges between the third oil consumption nodes, and the third connection edges have third connection values ​​determined according to the oil circuit transmission volume;

[0083] The first oil consumption node graph, the second oil consumption node graph and the third oil consumption node graph are merged to obtain an initial oil consumption node graph of the target gas turbine;

[0084] The historical operating condition data, historical temperature data, historical cleaning data and historical external environment data corresponding to the oil consumption nodes are added to the initial oil consumption node diagram to obtain the oil consumption node diagram of the target gas turbine.

[0085] Optionally, the processor 401 performs the step of fusing the first oil consumption node graph, the second oil consumption node graph, and the third oil consumption node graph to obtain an initial oil consumption node graph of the target gas turbine, including:

[0086] Merging nodes belonging to the same oil consumption point among the first oil consumption node, the second oil consumption node and the third oil consumption node to obtain a merged oil consumption node;

[0087] Merging the edges having the same oil consumption point among the first connection edge, the second connection edge, and the third connection edge to obtain a merged connection edge, and performing weighted calculation on the first connection value, the second connection value, and the third connection value according to a preset weight value to obtain a connection value of the merged connection edge;

[0088] Based on the merged oil consumption nodes, the merged connection edges and the connection values ​​of the merged connection edges, an initial oil consumption node graph of the target gas turbine is constructed.

[0089] Optionally, before pruning the system node graph at the node level and the data level by using the trained first spatiotemporal graph convolutional network to obtain an intermediate node graph, the method executed by the processor 401 further includes:

[0090] Obtaining a first graph convolutional network to be trained and a first training data set, wherein the first graph convolutional network to be trained includes a first output and a second output, wherein the first output is a predicted value of an operating state of an oil-consuming node, and the second output is a pruned node graph, and the first training data set includes a first sample node graph and true values ​​of operating states of oil-consuming nodes corresponding to the first sample node graph;

[0091] Inputting the first sample node graph into the first to-be-trained graph convolutional network for processing to obtain a running state prediction value of the first sample node graph and a pruned node graph of the first sample node graph;

[0092] Calculate the operating state loss value between the operating state prediction value of the oil consumption node of the first sample node graph and the operating state true value of the oil consumption node of the first sample node graph by using a preset first loss function, and calculate the node graph loss value between the first sample node graph and the pruned node graph of the first sample node graph by using a preset second loss function, and determine the total loss value based on the operating state loss value and the node graph loss value;

[0093] Based on the total loss value, the model parameters of the first graph convolutional network to be trained are adjusted by a back-propagation algorithm, and the parameter adjustment process is iterated until the total loss value reaches a preset condition, and the training is stopped to obtain a trained first spatiotemporal graph convolutional network, and the trained first spatiotemporal graph convolutional network only retains the second output.

[0094] Optionally, the first graph convolution network to be trained includes a first spatial convolution layer to be trained, a first temporal convolution layer to be trained, a spatial mask layer to be trained, a temporal mask layer to be trained, a first feature fusion layer, a first output layer to be trained, and a second output layer to be trained. The output of the first spatial convolution layer to be trained is respectively connected to the input of the spatial mask layer to be trained and the first temporal convolution layer to be trained, the output of the spatial mask layer to be trained is respectively connected to the input of the first training temporal convolution layer and the first feature fusion layer, the output of the temporal convolution layer to be trained is respectively connected to the input of the first feature fusion layer and the temporal mask layer to be trained, the output of the temporal mask layer to be trained is connected to the input of the first feature fusion layer, the output of the first feature fusion layer is respectively connected to the input of the first output layer and the second output layer, the first output layer is used to output the running state prediction value of the oil consumption node, and the second output layer is used to output the pruned node graph. The processor 401 performs the step of inputting the first sample node graph into the first graph convolution network to be trained for processing to obtain the running state prediction value of the first sample node graph and the pruned node graph of the first sample node graph, including:

[0095] Inputting the first sample node graph into the first graph convolution network to be trained, performing spatial feature extraction on the first sample node graph through the first spatial convolution layer to be trained, and obtaining the spatial features of the first sample node graph;

[0096] Masking the spatial features of the first sample node graph in the spatial dimension by using the spatial mask layer to be trained to obtain spatial mask features;

[0097] Performing temporal feature extraction on the spatial features and the spatial mask features through the first to-be-trained temporal convolutional layer to obtain spatiotemporal features of the first sample node graph;

[0098] Masking the spatiotemporal features of the first sample node graph in the time dimension through the time mask layer to be trained to obtain a time mask feature;

[0099] Fusing the spatial mask feature, the temporal mask feature, and the spatiotemporal feature to obtain a first fused feature;

[0100] The first fused feature is fully connected through the first output layer to obtain a running state prediction value of the first sample node graph, and the first fused feature is deconvolved through the second output layer to obtain a pruned node graph of the first sample node graph.

[0101] Optionally, the processor 401 performs the step of calculating the operating state loss value between the operating state prediction value of the oil consumption node of the first sample node graph and the operating state true value of the oil consumption node of the first sample node graph by using a preset first loss function, and calculating the node graph loss value between the first sample node graph and the pruned node graph of the first sample node graph by using a preset second loss function, including:

[0102] The operating state loss value between the operating state prediction value of the oil consumption node of the first sample node graph and the operating state true value of the oil consumption node of the first sample node graph is calculated by a preset first loss function, and the first loss function is:

[0103]

[0104] Where Nc is the number of nodes using oil in the first sample node graph, f(x i ) is the i-th oil-using node x in the first sample node graph i The predicted value of the running status, y i is the i-th oil-using node x in the first sample node graph i The true value of the running state, y j is the true value of the operating status of the jth oil-consuming node in the first sample node graph, d i,j is the sum of the connection values ​​of at least one connecting edge between the i-th oil consumption node and the j-th oil consumption node;

[0105] The node graph loss value between the first sample node graph and the pruned node graph of the first sample node graph is calculated by a preset second loss function, where the second loss function is:

[0106]

[0107] Wherein, M is the total number of data dimensions corresponding to the historical operating condition data, historical temperature data, historical cleaning data and historical external environment data in the oil consumption node of the first sample node, and x i,k is the i-th oil-using node x in the first sample node graph i The k-th dimension data in G(x i,k ) is the i-th oil-using node x in the pruned node graph of the first sample node graph i The pruned data of the k-th dimension in N G is the first sample node graph |G(x i,k )-x i,k |≠0, the number of oil consumption nodes, G(d i,j ) is the i-th oil-using node x in the pruning node graph of the first sample node iThe sum of the connection values ​​of at least one connecting edge to the j-th oil-using node.

[0108] Optionally, the determining of the total loss value based on the running state loss value and the node graph loss value performed by the processor 401 includes:

[0109] The total loss value between the running state loss value and the node graph loss value is calculated by a total loss function, and the total loss function is:

[0110] L t =(1-α)·L 1 -α·L 2

[0111] Here, α ranges from 0 to 1 and increases with the number of iterations.

[0112] Optionally, before inputting the intermediate node graph and the rated oil product parameters into the trained second spatiotemporal graph convolutional network for prediction processing to obtain the prediction result of the lubricating oil life of the gas turbine, the method executed by the processor 401 includes:

[0113] Obtaining a second graph convolutional network to be trained and a second training data set, wherein the second graph convolutional network to be trained includes a third output, the third output is a predicted value of lubricant life, the second training data set includes a second sample node graph, sample rated oil parameters, and a true value of lubricant life of the second sample node graph, and the second sample node graph is obtained by pruning the trained first spatiotemporal graph convolutional network;

[0114] Inputting the second sample node graph and the sample rated oil parameters into the second to-be-trained graph convolutional network for processing to obtain a predicted value of lubricant life of the second sample node graph;

[0115] Calculate the life loss value between the lubricant life prediction value of the second sample node graph and the lubricant life true value of the second sample node graph by using a preset third loss function;

[0116] Based on the life loss value, the model parameters of the second graph convolutional network to be trained are adjusted by a back propagation algorithm, and the parameter adjustment process is iterated until the life loss value reaches a preset condition, and the training is stopped to obtain a trained second spatiotemporal graph convolutional network.

[0117] Optionally, the second graph convolution network to be trained includes a main input layer to be trained, an auxiliary input layer to be trained, a second spatial convolution layer to be trained, a second time convolution layer to be trained, a second feature fusion layer, and a third output layer to be trained. The outputs of the main input layer to be trained and the auxiliary input layer to be trained are simultaneously connected to the input of the second feature fusion layer, the output of the second feature fusion layer is connected to the input of the second spatial convolution layer to be trained, the output of the second spatial convolution layer to be trained is connected to the input of the second time convolution layer to be trained, the output of the second time convolution layer to be trained is connected to the input of the third output layer to be trained, and the input of the third output layer to be trained is used to output a predicted value of lubricant life. The processor 401 performs the step of inputting the second sample node graph and the sample rated oil parameters into the second graph convolution network to be trained for processing to obtain a predicted value of lubricant life of the second sample node graph, including:

[0118] Inputting the second sample node graph into the main input layer to be trained for processing to obtain main input features, and inputting the sample rated oil parameters into the auxiliary input layer to be trained for processing to obtain auxiliary input features;

[0119] The main input feature and the auxiliary input feature are fused through the second feature fusion layer to obtain a fused input feature;

[0120] Inputting the fused input features into the second graph convolutional network to be trained, performing spatial feature extraction on the fused input features through the second spatial convolutional layer to be trained, and obtaining fused spatial features of the second sample node graph;

[0121] Performing time feature extraction on the fused spatial feature through the second time convolution layer to be trained to obtain a second spatiotemporal feature of the second sample node graph;

[0122] The second spatiotemporal feature is fully connected through the third output layer to obtain a predicted value of the lubricant life of the second sample node graph.

[0123] An embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the various processes of the online prediction method for the lubricating oil life of a gas turbine provided by the embodiment of the present invention are implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0124] A person skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The computer-readable storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM).

[0125] The above disclosure is only the preferred embodiment of the present invention, which certainly cannot be used to limit the scope of the present invention. Therefore, equivalent changes made according to the claims of the present invention are still within the scope of the present invention.

Claims

1. A method for online prediction of lubricating oil life of a gas turbine, characterized in that: include: Obtaining an oil consumption node graph of a target gas turbine and rated oil parameters of lubricating oil used by the target gas turbine, wherein the oil consumption node graph includes oil consumption nodes and connecting edges between the oil consumption nodes, wherein the connecting edges have connection values ​​determined according to oil line transmission volume, and the oil consumption nodes include corresponding historical operating condition data, historical temperature data, historical cleaning data, and historical external environment data; The oil consumption node graph is pruned at the node level and the data level by using a trained first spatiotemporal graph convolutional network to obtain an intermediate node graph, wherein the pruning process includes node pruning and data pruning, and the first spatiotemporal graph convolutional network is trained by performing operation status recognition of oil consumption nodes using a first training data set; Inputting the intermediate node graph and the rated oil parameters into a trained second spatiotemporal graph convolutional network for prediction processing to obtain a prediction result of the lubricant oil life of the gas turbine, wherein the second spatiotemporal graph convolutional network is trained for lubricant oil life prediction using a second training data set; The step of obtaining the oil consumption node diagram of the target gas turbine includes: Determine a first oil consumption node graph according to the main oil circuit structure of the target gas turbine, wherein the first oil consumption node graph includes first oil consumption nodes and first connection edges between the first oil consumption nodes, and the first connection edges have first connection values ​​determined according to the oil circuit transmission volume; determining a second oil consumption node graph according to the auxiliary oil circuit structure of the target gas turbine, wherein the second oil consumption node graph includes second oil consumption nodes and second connection edges between the second oil consumption nodes, and the second connection edges have second connection values ​​determined according to the oil circuit transmission volume; Determine a third oil consumption node diagram according to the emergency oil circuit structure of the target gas turbine, wherein the third oil consumption node diagram includes third oil consumption nodes and third connection edges between the third oil consumption nodes, and the third connection edges have third connection values ​​determined according to the oil circuit transmission volume; The first oil consumption node graph, the second oil consumption node graph and the third oil consumption node graph are merged to obtain an initial oil consumption node graph of the target gas turbine; The historical operating condition data, historical temperature data, historical cleaning data and historical external environment data corresponding to the oil consumption nodes are added to the initial oil consumption node diagram to obtain the oil consumption node diagram of the target gas turbine.

2. The method according to claim 1, characterized in that The step of fusing the first oil consumption node graph, the second oil consumption node graph, and the third oil consumption node graph to obtain the initial oil consumption node graph of the target gas turbine includes: Merging nodes belonging to the same oil consumption point among the first oil consumption node, the second oil consumption node and the third oil consumption node to obtain a merged oil consumption node; Merging the edges having the same oil consumption point among the first connection edge, the second connection edge, and the third connection edge to obtain a merged connection edge, and performing weighted calculation on the first connection value, the second connection value, and the third connection value according to a preset weight value to obtain a connection value of the merged connection edge; Based on the merged oil consumption nodes, the merged connection edges and the connection values ​​of the merged connection edges, an initial oil consumption node graph of the target gas turbine is constructed.

3. The method according to claim 1, characterized in that Before pruning the oil consumption node graph at the node level and the data level by using the trained first spatiotemporal graph convolutional network to obtain the intermediate node graph, the method further includes: Obtaining a first graph convolutional network to be trained and a first training data set, wherein the first graph convolutional network to be trained includes a first output and a second output, wherein the first output is a predicted value of an operating state of an oil-consuming node, and the second output is a pruned node graph, and the first training data set includes a first sample node graph and true values ​​of operating states of oil-consuming nodes corresponding to the first sample node graph; Inputting the first sample node graph into the first to-be-trained graph convolutional network for processing to obtain a running state prediction value of the first sample node graph and a pruned node graph of the first sample node graph; Calculate the operating state loss value between the operating state prediction value of the oil consumption node of the first sample node graph and the operating state true value of the oil consumption node of the first sample node graph by using a preset first loss function, and calculate the node graph loss value between the first sample node graph and the pruned node graph of the first sample node graph by using a preset second loss function, and determine the total loss value based on the operating state loss value and the node graph loss value; Based on the total loss value, the model parameters of the first graph convolutional network to be trained are adjusted by a back-propagation algorithm, and the parameter adjustment process is iterated until the total loss value reaches a preset condition, and the training is stopped to obtain a trained first spatiotemporal graph convolutional network, and the trained first spatiotemporal graph convolutional network only retains the second output.

4. The method according to claim 3, characterized in that The first graph convolution network to be trained includes a first spatial convolution layer to be trained, a first temporal convolution layer to be trained, a spatial mask layer to be trained, a temporal mask layer to be trained, a first feature fusion layer, a first output layer to be trained and a second output layer to be trained. The output of the first spatial convolution layer to be trained is respectively connected to the input of the spatial mask layer to be trained and the first temporal convolution layer to be trained, the output of the spatial mask layer to be trained is respectively connected to the input of the first temporal convolution layer to be trained and the first feature fusion layer, the output of the temporal convolution layer to be trained is respectively connected to the input of the first feature fusion layer and the temporal mask layer to be trained, the output of the temporal mask layer to be trained is connected to the input of the first feature fusion layer, the output of the first feature fusion layer is respectively connected to the input of the first output layer to be trained and the input of the second output layer to be trained, the first output layer to be trained is used to output the running state prediction value of the oil consumption node, the second output layer to be trained is used to output the pruned node graph, the first sample node graph is input into the first graph convolution network to be trained for processing, and the running state prediction value of the first sample node graph and the pruned node graph of the first sample node graph are obtained, including: Inputting the first sample node graph into the first graph convolution network to be trained, performing spatial feature extraction on the first sample node graph through the first spatial convolution layer to be trained, and obtaining the spatial features of the first sample node graph; Masking the spatial features of the first sample node graph in the spatial dimension by using the spatial mask layer to be trained to obtain spatial mask features; Performing temporal feature extraction on the spatial features and the spatial mask features through the first to-be-trained temporal convolutional layer to obtain spatiotemporal features of the first sample node graph; Masking the spatiotemporal features of the first sample node graph in the time dimension through the time mask layer to be trained to obtain a time mask feature; Fusing the spatial mask feature, the temporal mask feature, and the spatiotemporal feature to obtain a first fused feature; The first fused feature is fully connected through the first output layer to be trained to obtain a running state prediction value of the first sample node graph, and the first fused feature is deconvolved through the second output layer to be trained to obtain a pruned node graph of the first sample node graph.

5. The method according to claim 3, characterized in that The step of calculating the operating state loss value between the operating state prediction value of the oil consumption node of the first sample node graph and the operating state true value of the oil consumption node of the first sample node graph by using a preset first loss function, and calculating the node graph loss value between the first sample node graph and the pruned node graph of the first sample node graph by using a preset second loss function, comprises: The operating state loss value between the operating state prediction value of the oil consumption node of the first sample node graph and the operating state true value of the oil consumption node of the first sample node graph is calculated by a preset first loss function, and the first loss function is: Where N is the number of oil-using nodes in the first sample node graph, f(x i ) is the i-th oil-using node x in the first sample node graph i The predicted value of the running status, y i is the i-th oil-using node x in the first sample node graph i The true value of the running state, y j is the true value of the operating status of the jth oil-consuming node in the first sample node graph, d i，j is the sum of the connection values ​​of at least one connecting edge between the i-th oil consumption node and the j-th oil consumption node; The node graph loss value between the first sample node graph and the pruned node graph of the first sample node graph is calculated by a preset second loss function, where the second loss function is: Wherein, M is the total number of data dimensions corresponding to the historical operating condition data, historical temperature data, historical cleaning data and historical external environment data in the oil consumption node of the first sample node, and x i，k is the i-th oil-using node x in the first sample node graph i The k-th dimension data in G(x i，k ) is the i-th oil-using node x in the pruned node graph of the first sample node graph i The pruned data of the k-th dimension in N G is the first sample node graph |G(x i，k )-x i，k |≠0, the number of oil-consuming nodes, G(d i，j ) is the i-th oil-using node x in the pruning node graph of the first sample node i The sum of the connection values ​​of at least one connecting edge to the j-th oil-using node.

6. The method according to claim 5, characterized in that The determining of the total loss value based on the running state loss value and the node graph loss value comprises: The total loss value between the running state loss value and the node graph loss value is calculated by a total loss function, and the total loss function is: L t =(1-a)·L1-a·L2 Here, α ranges from 0 to 1 and increases with the number of iterations.

7. The method according to any one of claims 1 to 6, characterized in that: Before inputting the intermediate node graph and the rated oil product parameters into the trained second spatiotemporal graph convolutional network for prediction processing to obtain the prediction result of the lubricating oil life of the gas turbine, the method includes: Obtaining a second graph convolutional network to be trained and a second training data set, wherein the second graph convolutional network to be trained includes a third output, the third output is a predicted value of lubricant life, the second training data set includes a second sample node graph, sample rated oil parameters, and a true value of lubricant life of the second sample node graph, and the second sample node graph is obtained by pruning the trained first spatiotemporal graph convolutional network; Inputting the second sample node graph and the sample rated oil parameters into the second to-be-trained graph convolutional network for processing to obtain a predicted value of lubricant life of the second sample node graph; Calculate the life loss value between the lubricant life prediction value of the second sample node graph and the lubricant life true value of the second sample node graph by using a preset third loss function; Based on the life loss value, the model parameters of the second graph convolutional network to be trained are adjusted by a back propagation algorithm, and the parameter adjustment process is iterated until the life loss value reaches a preset condition, and the training is stopped to obtain a trained second spatiotemporal graph convolutional network.

8. The method according to claim 7, characterized in that The second graph convolution network to be trained includes a main input layer to be trained, an auxiliary input layer to be trained, a second spatial convolution layer to be trained, a second time convolution layer to be trained, a second feature fusion layer, and a third output layer to be trained. The outputs of the main input layer to be trained and the auxiliary input layer to be trained are connected to the input of the second feature fusion layer at the same time. The output of the second feature fusion layer is connected to the input of the second spatial convolution layer to be trained. The output of the second spatial convolution layer to be trained is connected to the input of the second time convolution layer to be trained. The output of the second time convolution layer to be trained is connected to the input of the third output layer to be trained. The input of the third output layer to be trained is used to output a predicted value of lubricant life. The second sample node graph and the sample rated oil parameters are input into the second graph convolution network to be trained for processing to obtain a predicted value of lubricant life of the second sample node graph, including: Inputting the second sample node graph into the main input layer to be trained for processing to obtain main input features, and inputting the sample rated oil parameters into the auxiliary input layer to be trained for processing to obtain auxiliary input features; The main input feature and the auxiliary input feature are fused through the second feature fusion layer to obtain a fused input feature; Inputting the fused input features into the second graph convolutional network to be trained, performing spatial feature extraction on the fused input features through the second spatial convolutional layer to be trained, and obtaining fused spatial features of the second sample node graph; Performing time feature extraction on the fused spatial feature through the second time convolution layer to be trained to obtain a second spatiotemporal feature of the second sample node graph; The second spatiotemporal feature is fully connected through the third output layer to obtain the predicted value of the lubricant life of the second sample node graph.

9. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the steps in the method for online prediction of lubricating oil life of a gas turbine as claimed in any one of claims 1 to 8 are implemented.

Images