Complex equipment performance parameter prediction method based on space-time dynamic association diagram

Through the combination of space-time dynamic correlation diagram and Transformer-GNN, the correlation variables and correlation modes of complex equipment performance parameters are dynamically identified, and the multivariable and dynamic change problems of performance parameters prediction of complex equipment in the existing technology are solved, achieving high-precision short-term and Long-term forecast.

CN118917182APending Publication Date: 2024-11-08ZHEJIANG UNIV
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202410950021.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-16
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

The existing complex equipment performance parameter prediction methods have insufficient calculations, are susceptible to noise interference, rely on external knowledge, and are unable to fully express the dynamic coupling rules of complex systems when dealing with multivariables and dynamic changes, making it difficult to achieve high-precision long-term predictions.

Method used

Using a method based on space-time dynamic correlation graph, a spatiotemporal dynamic correlation graph is constructed by capturing sparse units through the autoencoder, graph initialization unit, channel sparse unit and local correlation mode, dynamically identify performance parameters related state variables, and combine Transformer and graph neural The network performs timing encoding and decoding to realize dynamic modeling of the correlation modes between complex equipment performance parameters and related variables.

Benefits of technology

It improves the accuracy and adaptability of the prediction of performance parameters of complex equipment, can achieve high-precision prediction of short-term and long-term prediction tasks under variable operating conditions, reduces dependence on external knowledge, and enhances the degree of automation of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118917182A_ABST
    Figure CN118917182A_ABST
Patent Text Reader

Abstract

The invention discloses a complex equipment performance parameter prediction method based on a space-time dynamic association graph. The method comprises the following steps: acquiring and constructing an equipment performance parameter prediction data set; training the complex equipment performance parameter prediction model based on the space-time dynamic association diagram by using the equipment performance parameter prediction data set to obtain a trained complex equipment performance parameter prediction model; the latest state monitoring data obtained in real time are preprocessed and then input into the prediction model, and a multi-step prediction result of the equipment performance parameters is obtained in real time. According to the method, a space-time dynamic association graph and a dynamic graph structure generation module which take a local window-level sub-sequence of a single variable as nodes are introduced into a prediction model, so that automatic screening of performance parameter related state variables and dynamic modeling of correlation mode local characteristics of each related variable are realized; therefore, the accuracy of performance parameter prediction is effectively improved, and the method is suitable for short-term prediction and long-term prediction tasks of complex equipment performance parameters in various variable working condition scenes.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a complex equipment performance parameter prediction method in the field of multivariate time series prediction and deep learning in machine learning, and in particular to a complex equipment performance parameter prediction method based on a spatiotemporal dynamic association graph. Background Art

[0002] Complex equipment often has the characteristics of complex structure and changeable service environment. By analyzing and predicting the operating status of the equipment, we can guide the optimization of the equipment dynamic control strategy and the formulation of predictive maintenance strategy, so as to effectively ensure the stability and efficiency of the operation of complex equipment. The operating status of complex equipment is often reflected in its key performance parameters, such as the exhaust temperature deviation (△EGT) of the aircraft engine and the output frequency of the voltage-controlled oscillator in the radar frequency receiving unit. Therefore, predicting the key performance parameters of complex equipment is crucial for analyzing the future trend of equipment operating status and improving the level of equipment control and health management.

[0003] At present, the performance parameter prediction methods of complex equipment can be mainly divided into traditional methods and data-driven methods. Among them, traditional methods include those based on expert experience and those based on mathematical models. The former relies on the summary of the empirical knowledge and subjective understanding of experts in the field, while the latter relies on the construction of mechanism models and numerical simulation analysis of multi-physical field coupling such as heat, flow, and electronics. Their commonality is that they need to introduce a large amount of prior knowledge in the field and human intervention processes. They have the disadvantages of limited ability to express complex systems, reliance on idealized assumptions, and insufficient generalization ability. They are difficult to adapt to the complex composition and diverse operating conditions of complex equipment systems. Data-driven methods can be mainly divided into those based on statistical analysis and those based on machine learning. Among them, statistical analysis methods are represented by autoregressive moving average models and autoregressive sliding models. They are simple to implement and have good versatility, but they have shortcomings such as making specific assumptions on the distribution of input data and being sensitive to noise outliers. It is difficult to obtain sufficiently high accuracy when predicting the performance parameters of complex equipment with highly dynamic and nonlinear characteristics. Machine learning methods are represented by deep learning methods based on RNN, TCN, Transformer and other architectures. With their powerful nonlinear modeling capabilities and good dynamic characteristics, they have gradually become the mainstream method for predicting performance parameters of complex equipment.

[0004] However, the existing performance parameter prediction methods of complex equipment based on deep learning still have the following problems: 1. The changes in key performance parameters of complex equipment are usually affected by multiple related state variables. Therefore, the cross-variable dependency relationship between performance parameter variables and related state variables needs to be modeled in the prediction model. Some existing methods consider the cross-variable dependency relationship between all state monitoring variables of complex equipment at the same time, which has the disadvantages of large computational complexity and susceptibility to noise correlation interference. Some existing models select fixed key variables based on prior knowledge or correlation analysis algorithms to construct the input of the model, which has the disadvantages of relying on external knowledge and being unable to fully express the dynamic coupling laws within the complex system. 2. When dealing with the long-term prediction problem of performance parameters, the existing methods usually model a single type of global cross-variable dependency relationship at all time steps of the entire input sequence, which makes it difficult to accurately characterize the local correlation characteristics and dynamic change laws of the cross-variable dependency pattern of complex equipment in long-term spans and multi-condition alternation scenarios. The above shortcomings make the current methods still have a lot of room for improvement in the performance of complex equipment performance parameter prediction tasks. Summary of the invention

[0005] In order to solve the problems and needs in the background technology, the present invention proposes a method for predicting complex equipment performance parameters based on a spatiotemporal dynamic association graph. The present invention introduces a spatiotemporal dynamic association graph to realize automatic identification of state variables related to equipment performance parameters and dynamic modeling of local association patterns between variables, thereby effectively improving the accuracy of complex equipment performance parameter prediction, and is suitable for short-term and long-term prediction tasks of complex equipment performance parameters under various variable working conditions.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] 1. A method for predicting performance parameters of complex equipment based on spatiotemporal dynamic correlation graph

[0008] S1: Continuously sample and record historical signals of multiple condition monitoring variables of complex equipment including performance parameters to obtain multi-channel condition monitoring data of complex equipment;

[0009] S2: After preprocessing the multi-channel condition monitoring data of complex equipment, a data set for predicting equipment performance parameters is constructed;

[0010] S3: After training the complex equipment performance parameter prediction model based on the spatiotemporal dynamic association graph using the equipment performance parameter prediction data set, a trained complex equipment performance parameter prediction model is obtained;

[0011] S4: Preprocess the new state monitoring variable signal obtained in real time and then input it into the trained complex equipment performance parameter prediction model to obtain the prediction result of the complex equipment performance parameter under the prediction step.

[0012] In S1, the state monitoring variables include temperature parameters, pressure parameters, and environmental parameters, and the performance parameters are power parameters or efficiency parameters.

[0013] In S2, the multi-channel condition monitoring data of complex equipment is preprocessed, specifically including missing value filling, normalization operation, and data division according to input time step and prediction step.

[0014] In S3, the complex equipment performance parameter prediction model based on the spatiotemporal dynamic association graph first calculates the input sequence X of any sample in the equipment performance parameter prediction data set. (i) Perform uniform and non-overlapping subsequence segmentation with a preset local window size in the time dimension to obtain local window-level subsequences satisfy Among them, C max is the total number of channels of the state monitoring variable, L represents the number of local windows, T L Indicates the sampling step size corresponding to each local window, The number of channels is C max ×L×T L The real number set of the local window level subsequence Input into the spatiotemporal dynamic association graph generation module, the spatiotemporal dynamic association graph generation module generates a spatiotemporal dynamic association graph satisfy where v (i) is a node set, and the characteristics of each node in the node set are satisfy The number of channels is D×d gh A real number set, D represents the number of nodes in each graph, satisfying D = C max ·L, d gh Represents the feature dimension of each node, ε (i) is an edge set, A (i) represents the adjacency matrix;

[0015] Then based on the spatiotemporal dynamic association graph And the preset order k solves the neighborhood subgraph centered on the node corresponding to the performance parameter variable Then according to the neighborhood subgraph The belonging relationship between the contained nodes and the corresponding state variables identifies the state variables related to each performance parameter, and then from the local window level subsequence Construct a subset of the input sequence containing performance parameters and related state variables satisfy in, The number of channels is (C relevant +1)×L×T L The set of real numbers, Crelevant is the number of relevant state variables; then, the input sequence subset and neighborhood subgraph Input the timing prediction module to obtain the prediction output of performance parameter variables satisfy The number of channels is T p The set of real numbers, T p is the prediction step length.

[0016] The spatiotemporal dynamic association graph generation module includes a self-encoding unit, a graph initialization unit, a channel sparsification unit and a local association pattern capture sparsification unit which are connected in sequence.

[0017] The autoencoder unit includes an autoencoder neural network, which is composed of a feature encoder and a feature decoder. The feature encoder encodes the local window-level subsequence in the time dimension. After encoding, the potential features of the low-dimensional embedding space are output And as the input of the feature decoder and the input of the graph initialization unit, satisfying in The number of channels is C max ×L×d enc The set of real numbers, d enc is the self-encoding feature dimension; the feature decoder embeds the potential features of the low-dimensional space Decode to the feature encoder input dimension to obtain the reconstructed subsequence satisfy According to the local window level subsequence and reconstruct subsequence The autoencoder loss of the spatiotemporal dynamic association graph generation module is calculated.

[0018] In the graph initialization unit, the latent features of the low-dimensional embedding space output by the self-encoding unit are Convert to initial node features of spatiotemporal dynamic association graph satisfy The number of channels is D×d enc A real number set; according to the initial node characteristics of the spatiotemporal dynamic association graph After calculating the feature similarity, we get the graph with edge weights to initialize the adjacency matrix satisfy is a real number set with D×D channels, and the adjacency matrix of the graph with edge weights is initialized Transformed into a set of variable pair subgraph adjacency matrices And as the output of the graph initialization unit, satisfying The number of channels is The set of real numbers.

[0019] In the channel sparsification unit, the subgraph adjacency matrix set is the variable pair output by the graph initialization unit As input, the feature similarity threshold is set as a learnable parameter, the adjacency matrix of the subgraph of variables whose average feature similarity is less than or equal to the feature similarity threshold is set to zero, and the nonlinear function is used to perform nonlinear transformation on the edge weights in the subgraph adjacency matrix of the variables that are not set to zero, and the channel sparse subgraph adjacency matrix set is obtained. And as the output of the channel sparsification unit, satisfying The number of channels is The set of real numbers;

[0020] The local correlation pattern captures the adjacency matrix of the channel-sparse subgraph in the sparse unit. As input, the convolutional neural network transforms the value of each edge weight in the subgraph adjacency matrix of non-zero variables, and uses the ReLU (·) activation function to set the negative edge weights to zero, and obtains the pattern capture sparse subgraph adjacency matrix set satisfy After channel transformation, the final spatiotemporal dynamic association graph adjacency matrix is ​​obtained satisfy Then construct the spatiotemporal dynamic association graph output by the spatiotemporal dynamic association graph generation module

[0021] The temporal prediction module comprises an embedding unit, a temporal encoding unit and a temporal decoding unit which are connected in sequence.

[0022] In the embedding unit, first, a subset of the input sequence is Perform Patch division to obtain block sequence subsets satisfy in The number of channels is (C relevant +1)×L×N p ×P real number set, N p represents the number of patches in each local window in the Patch partition, P represents the Patch length, and then the block sequence subset Use linear transformation to get the embedding features of each patch satisfy The number of channels is (C relevant +1)×L×N p ×d trans The set of real numbers, d trans is the hidden dimension of the model, and then the embedding features of each patch are extracted separately The average information of each patch corresponding to each local window in the , and use it as the initialization value to generate the head patch embedding satisfy The number of channels is (C relevant+ 1) × L × 1 × d trans A real number set; then the head patch corresponding to each local window is embedded into Embedded features with the original patches After splicing, the labeled Patch embedding features are obtained satisfy The number of channels is (C relevant +1)×L×(1+N p )×d trans A real number set; finally, the labeled patch is embedded into the feature Each patch under all local windows in the is concatenated along the time dimension and a learnable position encoding is performed to obtain the encoded patch embedding feature And as the output of the embedding unit, it satisfies The number of channels is (C relevant +1)×N fp ×d trans The real number set, N fp The total number of patches corresponding to each variable, satisfying N fp =L·(N P +1).

[0023] The temporal coding unit is formed by stacking J layers of temporal encoders, wherein the j-th layer of temporal encoders converts the j-1-th layer of temporal coding features into Encoded as the j-th layer temporal coding feature The number of channels is (C relevant +1)×N fp ×d trans The set of real numbers that satisfies Each layer of temporal encoder is composed of a Transformer encoding layer and a graph neural network (GNN) encoding layer in series, where the Transformer encoding layer converts the j-1th layer temporal encoding features into Each patch in is regarded as a word unit and is encoded with attention to obtain the intermediate encoding feature satisfy Then, we encode the features from the intermediate Separate the head patch features corresponding to each local window satisfy The number of channels is (C relevant +1)×L×1×d trans The real number set, the GNN encoding layer is based on the head Patch feature and neighborhood subgraph As input, the features of each head patch are further encoded based on the graph neural network algorithm to obtain the graph encoding feature Finally, the graph encoding features Re-encoding features The non-head Patch features in are concatenated to obtain the j-th layer temporal coding features. as output.

[0024] In the timing decoding unit, the output of the timing encoding unit is first removed. All the head patch features in and only retain the variable channels corresponding to the performance parameters to obtain the performance parameter encoding features satisfy The number of channels is 1×L·N p ×d trans Then use linear transformation to encode the performance parameters After decoding, the predicted output of performance parameter variables is obtained

[0025] In S3, the loss function in the training process of the complex equipment performance parameter prediction model based on the spatiotemporal dynamic association graph includes the autoencoding loss of the spatiotemporal dynamic association graph generation module.

[0026] In S3, the complex equipment performance parameter prediction model based on the spatiotemporal dynamic association graph is trained and verified by using an early stopping mechanism to obtain a trained complex equipment performance parameter prediction model.

[0027] 2. A storage medium

[0028] The medium stores a computer program, and when the computer program is executed by a processor, the method described above is implemented.

[0029] The computer program described herein is an instruction corresponding to the implementation of the method.

[0030] Without the need for external prior knowledge, the present invention realizes the automatic construction of a spatiotemporal dynamic association graph based on autoencoders, similarity algorithms, convolutional neural networks, etc. according to the input sequence characteristics of the model, and integrates the constructed spatiotemporal dynamic association graph into a timing prediction module based on Transformer and graph neural networks for timing encoding and on linear transformation for timing decoding. It effectively models and utilizes the dynamic and local characteristics of the correlation patterns between equipment performance parameters and their related variables, and realizes high-precision prediction of complex equipment performance parameters.

[0031] Compared with the existing technologies and methods, the present invention has the following beneficial effects:

[0032] The present invention effectively handles the problem of dynamic selection of variables related to performance parameters of complex equipment. In actual industrial scenarios, the performance parameters of complex equipment are often affected by multiple related state variables, and different state variables have different degrees of influence on performance parameters under different working conditions. The selection of related state variables by existing methods is often static and highly dependent on prior knowledge, and there are deficiencies such as inaccurate modeling of cross-variable dependencies and low degree of automation. The method proposed in the present invention can adaptively and dynamically identify state variables related to performance parameters based on the characteristics of the input sequence and the rules learned by the model from the data, fully considering the actual characteristics of variables related to the performance parameters of complex equipment, which is conducive to improving the performance of the prediction model.

[0033] The present invention realizes effective modeling of the dynamic local characteristics of the correlation pattern between related state variables. On a longer time scale, the state monitoring data of complex equipment often has obvious local random noise and distribution transformation, showing strong dynamic local characteristics, and the correlation patterns between different variables often have local differences. In long-term prediction tasks with a large number of time steps corresponding to the input sequence and the output sequence, modeling the dynamic local correlation pattern between variables is conducive to better learning the changing laws of performance parameters and improving the prediction effect. However, existing methods usually model a single type of global cross-variable dependency relationship at all time steps of the entire input sequence length, which is difficult to accurately characterize the above-mentioned dynamic local correlation pattern. The present invention uses a spatiotemporal dynamic correlation graph in the prediction model, and realizes effective modeling of the dynamic local correlation pattern between variables on the basis of introducing the concept of local window, so that the model can better cope with the long-term prediction task of complex equipment performance parameters.

[0034] Therefore, the present invention models the dynamic and local characteristics of the correlation patterns between equipment performance parameters and their related variables through a spatiotemporal dynamic correlation graph, and combines it with a time series prediction model to predict the equipment performance parameters. The present invention can improve the accuracy of predicting complex equipment performance parameters under variable operating conditions, and is also suitable for various short-term and long-term prediction tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] 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 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 labor.

[0036] Figure 1 It is a schematic flow chart of the specific steps of the present invention.

[0037] Figure 2It is a schematic diagram of the structure of the prediction model based on the spatiotemporal dynamic association graph in the present invention.

[0038] Figure 3 It is a visualization diagram of some experimental results in the examples of the present invention. DETAILED DESCRIPTION

[0039] 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.

[0040] The specific embodiments of the present invention are as follows:

[0041] As an advanced complete set of power machinery, gas turbines are composed of compressors, combustion chambers, gas turbines, control components and other parts. They play an important role in many fields such as electricity, ships, and transportation vehicles, and are typical complex equipment. For gas turbines in service, power is the most direct indicator to reflect their operating status. On the one hand, most component abnormalities or performance degradation phenomena will be directly reflected in power fluctuations. On the other hand, the optimal operating point of the gas turbine unit can be determined by analyzing the power. Therefore, predicting the power of gas turbines is of great significance to the health management of gas turbines and the optimization of real-time control schemes. The power prediction of gas turbines is a typical problem of predicting performance parameters of complex equipment. It faces difficulties such as a large number of related state variables, strong dynamic characteristics, and low long-term prediction accuracy. It has high research value.

[0042] In this example, by cooperating with relevant companies, historical status monitoring data of gas turbines in real service conditions are obtained, and a gas turbine power prediction model based on a spatiotemporal dynamic correlation graph is trained after preprocessing. The prediction performance under prediction steps covering multiple time spans is analyzed to verify the versatility and effectiveness of the prediction method proposed in the present invention.

[0043] Attached Figure 1 A flow chart of the equipment performance parameter prediction method proposed in the present invention is shown.

[0044] The specific implementation is:

[0045] S1. Gas turbine state monitoring data acquisition. Use industrial sensors to continuously sample and record historical signals of multiple state monitoring variables of complex equipment including performance parameters to obtain multi-channel state monitoring data of complex equipment; where state monitoring variables include temperature parameters, pressure parameters, and environmental parameters, and performance parameters are power parameters or efficiency parameters; the gas turbine state monitoring variables collected in this example include power output (unit MW), spindle speed (unit rpm), blade channel temperature (unit ℃), turbine exhaust temperature (unit ℃), gas flow (unit Nm3 / h), exhaust pressure (unit %), atmospheric pressure (unit kPa), etc. This example is aimed at the gas turbine power prediction problem, that is, power output is selected as the performance parameter.

[0046] The multi-channel condition monitoring data used in this example is obtained by continuously measuring multiple condition monitoring variables of the gas turbine in real service state with the same and fixed sampling frequency using multiple sensors, and the sampling time of each sensor is synchronized. Multi-channel condition monitoring data is essentially multivariate time series data, which can be expressed as: Among them C max is the total number of channels of the state monitoring variables (i.e. the number of state monitoring variables), T total is the total length of the sequence obtained from each sampling. In order to enhance the versatility of the subsequent experimental results of this example, the state monitoring data of two gas turbines of different models (GTC1 and GATOKUSUI) are sampled under different operating conditions and time interval lengths, and two sets of original data sets (abbreviated as GTC1 and GATOKUSUI) are obtained as candidate options for multi-channel state monitoring data. The basic information of the two original data sets is shown in Table 1 below. The operating data covers the start-up, shutdown and multiple variable operating conditions of the gas turbine, and has significant dynamic characteristics.

[0047] Table 1 shows the basic information of the two original data sets.

[0048]

[0049] S2. Data preprocessing and partitioning

[0050] After preprocessing the multi-channel condition monitoring data of complex equipment, a data set for predicting equipment performance parameters is constructed; the preprocessing of the multi-channel condition monitoring data of complex equipment includes missing value filling, normalization operation, and data division according to the input time step and prediction step. The data set obtained by data division consists of three parts: training set, validation set, and test set. Each sample in the data set includes the input sequence and the true value labels of the performance parameter output sequence Two parts, the two are connected in the time dimension, where x targetis the input sequence in the performance parameter variable, T s is the input time step, T p is the prediction step length, C max -1 Input sequence of non-performance parameter condition monitoring variables.

[0051] Specifically, for the multi-channel condition monitoring data collected by sensors First, the forward filling method is used to fill in the missing values ​​that may exist. That is, the missing values ​​are filled with the last observed value before the missing value; secondly, X is sorted according to the ratio of training set, validation set and test set 7:1:2. F Divide the data to obtain the training sequence Verify Sequence and test sequence Then calculate The mean of each channel data and standard deviation And Perform independent standardization operations on each channel in the time dimension with the same parameters, convert the condition monitoring data of each channel into a standard normal distribution with a mean of 0 and a standard deviation of 1, and obtain the preprocessed equipment condition monitoring data sequence division result As shown below:

[0052]

[0053] Wherein, c represents the channel number.

[0054] According to the division results of each equipment status monitoring data series, the time dimension is T s +T p is the window size, and a sliding window operation is performed with a step size of 1 to obtain the sample set corresponding to each data sequence, that is, the training set X train , validation set X val , and the test set X test Each sample in the above sample set is further divided into input sequences and the true value labels of the performance parameter output sequence The two are sequentially connected in the time dimension, that is, Where T s =336 is the input time step, T p is a prediction step size selected from [96, 192, 336].

[0055] S3: After training the complex equipment performance parameter prediction model based on the spatiotemporal dynamic association graph using the equipment performance parameter prediction data set, a trained complex equipment performance parameter prediction model is obtained, such as Figure 2 As shown, the complex equipment performance parameter prediction model realizes the time series prediction of complex equipment performance parameters on the basis of realizing the generation of spatiotemporal dynamic correlation graph;

[0056] In S3, the complex equipment performance parameter prediction model based on the spatiotemporal dynamic association graph includes a spatiotemporal dynamic association graph generation module and a time series prediction module; the complex equipment performance parameter prediction model based on the spatiotemporal dynamic association graph firstly performs a training set X of the equipment performance parameter prediction data set. train The input sequence X of any sample i in (i) Perform uniform and non-overlapping subsequence segmentation with a preset local window size in the time dimension to obtain local window-level subsequences satisfy Among them, C max is the total number of channels of the state monitoring variable, L represents the number of local windows, T L Represents the sampling step size corresponding to each local window, T L =42, L=8, The number of channels is C max ×L×T L The real number set of the local window level subsequence Input into the spatiotemporal dynamic association graph generation module, the spatiotemporal dynamic association graph generation module generates a spatiotemporal dynamic association graph satisfy where v (i) is a node set, and the characteristics of each node in the node set are satisfy The number of channels is D×d gh A real number set, D represents the number of nodes in each graph, satisfying D = C max ·L, d gh The feature dimension representing the features of the graph nodes, ε (i) is an edge set, the number of edges depends on the input sequence X (i) varies with the A (i) Represents the adjacency matrix, and the values ​​of the non-zero elements in the matrix correspond to the size of the weight of each edge; the constructed spatiotemporal dynamic association graph It is an undirected weighted graph, each node of which corresponds to each local window-level subsequence of each state monitoring variable, covering both time and space (variable) dimensions, and its edge set ε (i)The corresponding edge weight indicates the strength of the correlation between each subsequence. The stronger the correlation, the greater the edge weight. Since the spatiotemporal dynamic correlation graph is dynamically generated based on the input characteristics of each sample, there are dynamic differences in the spatiotemporal dynamic correlation graphs of different samples.

[0057] Then based on the spatiotemporal dynamic association graph And the preset order k solves the neighborhood subgraph centered on the node corresponding to the performance parameter variable Thus, the nodes (n≤k) and the corresponding associated edges that have n-hop adjacency relationships with the nodes corresponding to the subsequence in the performance parameter variable to be predicted are screened out, i.e., the neighborhood subgraph. It only contains the subsequence nodes corresponding to the performance parameter variables, the neighbor nodes and corresponding associated edges that have n-hop adjacency relationships with these nodes, and the nodes that belong to the same variables as the above nodes. Then, according to the neighborhood subgraph The belonging relationship between the contained nodes and the corresponding state variables identifies the state variables related to each performance parameter, and then from the local window level subsequence Construct a subset of the input sequence containing performance parameters and related state variables satisfy in, The number of channels is (C relevant +1)×L×T L The set of real numbers, C relevant is the number of relevant state variables; then, the input sequence subset and neighborhood subgraph Input the timing prediction module to obtain the prediction output of performance parameter variables satisfy The number of channels is T p The set of real numbers.

[0058] The spatiotemporal dynamic association graph generation module includes a self-encoding unit, a graph initialization unit, a channel sparsification unit and a local association pattern capture sparsification unit which are connected in sequence.

[0059] The autoencoder unit includes an autoencoder neural network, which consists of a feature encoder and a feature decoder. Both are implemented by a two-layer feedforward neural network. The intermediate hidden layer dimension is 16. The feature encoder encodes the local window-level subsequence in the time dimension. After encoding, the potential features of the low-dimensional embedding space are output And as the input of the feature decoder and the input of the graph initialization unit, satisfying in The number of channels is C max ×L×d enc The set of real numbers, d encThe dimension of the self-encoding feature is 3. The activation function of the last layer of the feature encoder uses the Tanh (·) function to limit the encoded feature value range to [-1, 1]. The feature decoder embeds the potential features of the low-dimensional space into Decode to the feature encoder input dimension to obtain the reconstructed subsequence satisfy According to the local window level subsequence and reconstruct subsequence Calculated by time series general evaluation index and X L The average error between the two is used to obtain the autoencoder loss of the spatiotemporal dynamic association graph generation module. Also consider C max The decoding effect of each channel is obtained by and X L The mean square error (MSE) between them is obtained and recorded as:

[0060]

[0061] In the graph initialization unit, the latent features of the low-dimensional embedding space output by the self-encoding unit are Convert to initial node features of spatiotemporal dynamic association graph satisfy The number of channels is D×d enc A real number set; according to the initial node characteristics of the spatiotemporal dynamic association graph After calculating the cosine similarity between each feature, the adjacency matrix of the graph with edge weights is initialized. satisfy is a real number set whose number of channels is D×D, where the cosine similarity calculation formula is:

[0062]

[0063] Where T represents transpose, Represents the initial adjacency matrix of a graph with edge weights The element in the mth row and nth column of Representing the initial node features of the spatiotemporal dynamic association graph The mth row vector in , |||| represents the modulus of the vector.

[0064] Initialize the adjacency matrix of the graph with edge weights Transformed into a set of variable pair subgraph adjacency matrices And as the output of the graph initialization unit, satisfying The number of channels is The real number set of the spatiotemporal dynamic association graph is regarded as The combination of variable-subgraph adjacency matrices. The adjacency matrix corresponding to a single variable pair subgraph consisting of nodes related to variable m as source nodes and nodes related to variable n as target nodes Satisfies the following relationship:

[0065]

[0066] in, Represents the initialization weight of the edge with node j as the source node and node k as the target node in the initialized spatiotemporal dynamic association graph.

[0067] In the channel sparsification unit, the subgraph adjacency matrix set of the variables output by the graph initialization unit As input, the overall characteristics of the connection weights corresponding to each variable pair subgraph are analyzed, the feature similarity threshold is set as a learnable parameter, the adjacency matrix of the variable pair subgraph whose average feature similarity is less than or equal to the feature similarity threshold is set to zero, and the nonlinear function is used to perform nonlinear transformation on the edge weights in the adjacency matrix of the variable pair subgraph that are not set to zero, so as to realize the identification of the related variables of each state variable and the enhanced expression of the difference in the degree of association between different variable pairs, and obtain the channel sparse subgraph adjacency matrix set And as the output of the channel sparsification unit, satisfying The number of channels is The set of real numbers; specifically, the set of adjacency matrices of the subgraphs of the variables Perform global average pooling in the last two dimensions, calculate the average edge weight of each variable on the subgraph, and obtain the global channel feature In order to enable the network to automatically control the sparsity degree according to the system characteristics of complex equipment, the global channel characteristics Plus a learnable bias factor b c Then use the ReLU(·) function to get the sparse global channel features Right now:

[0068]

[0069] For the sparse global channel features Perform nonlinear exponential transformation to achieve the goal of Nonlinear scaling of non-zero global channel eigenvalues ​​in , resulting in a channel sparsification multiplicative factor where k m It is a hyperparameter that controls the degree of scaling and its default value is 1.

[0070]

[0071] Sparse channels by multiplicative factor Multiply them by the edge weights in the corresponding subgraphs respectively to realize the identification of related variables of each state variable and the enhanced expression of the difference in the degree of correlation between different variable pairs, and obtain the channel sparse subgraph adjacency matrix set

[0072] Local correlation patterns are captured in the sparse unit, and the adjacency matrix of the channel-sparse subgraph is set The convolutional neural network is used as input to transform the values ​​of the edge weights in the subgraph adjacency matrix of non-zero variables, and the ReLU (·) activation function is used to set the negative edge weights to zero, so as to enhance the weights of the associated edges that can reflect specific local association patterns while suppressing the weights of other noise associated edges. The pattern-capturing sparse subgraph adjacency matrix set is obtained. satisfy After channel transformation, the final spatiotemporal dynamic association graph adjacency matrix is ​​obtained satisfy Then construct the spatiotemporal dynamic association graph output by the spatiotemporal dynamic association graph generation module Specifically: for the channel sparse subgraph adjacency matrix set A convolutional layer with learnable parameters is used to extract and enhance the local regularity features in the subgraph adjacency matrix, while suppressing the weights of other noise-related edges. Perform pattern capture convolution coding and Treated as multiple single-channel two-dimensional feature maps Symmetrically fill and expand its last two dimensions to obtain Where L p is the width and height of the padded feature map, K is the size of the convolution kernel captured in the two-dimensional mode, and the default value is 3, satisfying (K-1)+L=L p , then make The convolutional hidden layer features are obtained by sequentially passing through the pattern capture convolution layer and the ReLU (·) activation layer. As shown below:

[0073]

[0074] Among them, M is the number of pattern capture convolution kernels, The representative pattern captures the convolution bias value, It represents the symmetric mode capturing the convolution kernel weights, whose element values ​​are symmetric in the last two dimensions, b m and ω ms are all learnable parameters. Then the convolutional hidden layer features Perform mode fusion coding to make The pattern sparse hidden layer features are obtained by sequentially using the pattern fusion convolution layer with a 1×1 convolution kernel, the instance normalization layer, and the ReLU activation layer. As shown below:

[0075]

[0076] in, Represents the mode fusion convolution bias value, Represents the mode fusion convolution kernel weights, which are all learnable parameters. InstanceNorm2d represents instance normalization operations in the last two dimensions.

[0077] Sparse hidden layer features for patterns Perform dimension compression to obtain the pattern-capturing sparse subgraph adjacency matrix set Then transform to get the final spatiotemporal dynamic association graph adjacency matrix Construct the output of the entire module

[0078] The temporal prediction module includes an embedding unit, a temporal encoding unit and a temporal decoding unit which are connected in sequence.

[0079] In the embedding unit, the idea of ​​patch partitioning is first used to partition the input sequence subsets Perform Patch division to obtain block sequence subsets satisfy in The number of channels is (C relevant +1)×L×N p ×P real number set, N p Patch represents the number of patches in each local window in the patch partition, P represents the patch length, P = 12, step size S = 6, and each parameter satisfies the following formula:

[0080] T L =P+(N p -1)·S

[0081] T s =L·T L

[0082] Then, for the subset of block sequences Use linear transformation to get the embedding features of each patch satisfy The number of channels is (C relevant +1)×L×N p ×d trans The set of real numbers, d trans is the model hidden layer dimension, d trans =128, and then extract the embedding features of each patch separately The average information of each patch corresponding to each local window in the , and use it as the initialization value to generate the head patch embedding satisfy The number of channels is (C relevant +1)×L×1×d trans Then embed the head patch corresponding to each local window into Embedded features with the original patches After splicing, the labeled Patch embedding features are obtained satisfy The number of channels is (C relevant +1)×L×(1+N p )×d trans The set of real numbers; as shown below:

[0083]

[0084] Among them, Cat is the splicing operation, Represents the features corresponding to the kth Patch in each local window.

[0085] Finally, the labeled patch is embedded into the feature Each patch under all local windows in the is concatenated along the time dimension and a learnable position encoding is performed to obtain the encoded patch embedding feature And as the output of the embedding unit, it satisfies The number of channels is (C relevant +1)×N fp ×d trans The real number set, N fp The total number of patches corresponding to each variable, satisfying N fp =L·(N P +1). The learnable position encoding is specifically to construct a learnable absolute position encoding And use the uniform distribution initialization parameters of X~U(-0.2,0.2), After copying and concatenation along the channel dimension After adding and embedding the time and position information for each patch feature, the output is obtained.

[0086] The temporal coding unit is composed of J layers of temporal encoders stacked together, J = 2, where the j-th layer of temporal encoders converts the j-1-th layer of temporal coding features into Encoded as the j-th layer temporal coding feature The number of channels is (C relevant +1)×N fp ×d trans The set of real numbers that satisfies Each layer of temporal encoder is composed of a Transformer encoding layer and a graph neural network (GNN) encoding layer in series, where the Transformer encoding layer converts the j-1th layer temporal encoding features into Each patch in is regarded as a token and the classic Transformer encoding architecture is used for attention encoding. The encoding layer adopts a multi-head attention mechanism with 8 heads to obtain the intermediate encoding features. satisfy Then, we encode the features from the intermediate Separate the head patch features corresponding to each local window satisfy The number of channels is (C relevant +1)×L×1×d trans The real number set, the GNN encoding layer is based on the head Patch feature and neighborhood subgraph As input, the features of each head patch are further encoded based on the graph neural network algorithm. Specifically, the GraphSAGE neural network with edge weights is used for message transmission and aggregation to obtain the graph encoding feature Finally, the graph encoding features Re-encoding features The non-head Patch features in are concatenated to obtain the j-th layer temporal coding features. as output.

[0087] In the timing decoding unit, the output of the timing encoding unit is first Crop in the first two dimensions, remove all head patch features and retain only the variable channels corresponding to the performance parameters to obtain the performance parameter encoding features satisfy The number of channels is 1×L·N p ×d trans Then use linear transformation to encode the performance parameters After decoding, the predicted output of performance parameter variables is obtained

[0088] Loss function in the training process of complex equipment performance parameter prediction model based on spatiotemporal dynamic correlation graph Autoencoder loss including spatiotemporal dynamic association graph generation module And the forecast output calculated based on the general evaluation index of time series and the average error between the true value label Y (referred to as the prediction loss ), the specific formula is as follows:

[0089]

[0090] Among them, λ is the weighting factor, its value is 0.5, and the prediction loss Considering only the prediction effect of the performance parameter channel, we obtain The mean square error (MSE) between y and Y is obtained. During training, the learner uses Adams and the learning rate is set to 0.0001.

[0091] The early stopping mechanism is used to train and verify the complex equipment performance parameter prediction model based on the spatiotemporal dynamic association graph. Specifically, the training data is used to optimize the parameters of the model, the verification data is used to evaluate the overfitting degree of the model, and the training termination condition is determined to obtain a trained complex equipment performance parameter prediction model. In this embodiment, the model adopts a training mode of multiple rounds of training. After each round of training, the verification set X obtained in S2 is used. val Feed the model in the same way as the training set and calculate the prediction loss To evaluate the degree of overfitting of the model, the early stopping strategy with a Patience of 3 is used to determine the training termination condition. To prevent the early stopping condition from not being met, the maximum number of training rounds of the model is set to 100.

[0092] S4: Preprocess the new state monitoring variable signal obtained in real time and then input it into the trained complex equipment performance parameter prediction model to obtain the prediction result of the complex equipment performance parameter under the prediction step.

[0093] In the model testing phase, the test set X test The way to input the model is the same as the validation set X in S3 val The input method is the same as that of Where N test is the total number of samples in the test set. Two evaluation indicators are used for model evaluation, namely And the corresponding test set true value label Y test Table 2 shows the prediction step size T of the model when the two different gas turbine original data sets mentioned in S1 are used as multi-channel condition monitoring data. p ∈[96,192,336] experimental results. Figure 3 It shows the prediction results visualization corresponding to some test samples under different experimental conditions. Each figure contains the input sequence of performance parameters, predicted output and corresponding true value label. Figure 3 (a) and (b), Figure 3 (c) and (d), Figure 3 (e) and (f) correspond to the prediction step length T sPerformance parameter prediction results of randomly selected samples from the GTC1 and GATOKUSUI datasets when the values ​​are 96, 192, and 336.

[0094] Table 2 shows the experimental results using two different gas turbine raw data sets.

[0095]

[0096] From Table 2 and Figure 3 It can be seen that the MSE and MAE of the power prediction results of different gas turbines using the present invention are maintained at a low level, the trend of the prediction output is similar to the true value label, and the error is small. The experimental results show that the prediction method proposed in the present invention is suitable for gas turbine power prediction under different working conditions, and can achieve good performance under different prediction step sizes. It has the advantages of good versatility, high prediction accuracy, and good accuracy retention in long-term prediction tasks. In summary, the method proposed in the present invention is effective in the field of complex equipment performance parameter prediction.

[0097] When using the evaluated prediction model for real-time prediction, firstly, the C max The state monitoring variables are in the recent T s The state monitoring data within time steps are filled with missing values ​​and normalized in the same way as in S2 to construct the input Input to the prediction model to get output In the right After denormalization, the gas turbine power output can be obtained in the future T p The prediction results at each time step are used for multiple downstream tasks such as gas turbine state prediction, control optimization, and health management.

[0098] This example predicts the power output of a gas turbine based on the real state monitoring data of the gas turbine. First, the data is preprocessed to obtain an input sequence; secondly, a spatiotemporal dynamic association graph is constructed based on the input sequence, and the association graph is initialized through an autoencoder and similarity calculation, and the association graph is sparsely constructed through learnable parameters and convolutional neural networks. The resulting spatiotemporal dynamic association graph can automatically screen state variables related to power output and dynamically model the local characteristics of the association pattern of each related variable; finally, with the input sequence and the spatiotemporal dynamic association graph as input, a time series prediction module based on Transformer and graph neural network for time series encoding and linear layer for time series decoding is used to obtain multi-step prediction results of power output. Experimental results show that the prediction method proposed in the present invention has good accuracy and versatility.

[0099] The present invention proposes for the first time a method for predicting the performance parameters of complex equipment by introducing a spatiotemporal dynamic correlation diagram, which realizes the dynamic screening of performance parameter-related variables and the dynamic modeling of local characteristics of cross-variable correlation patterns. It effectively solves the problems of the existing performance parameter prediction methods, such as the difficulty in selecting related variables, poor dynamic modeling capabilities, and low long-term prediction accuracy. It can be used in the fields of health management and optimization control of complex equipment, and is innovative and practical.

[0100] The present invention uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only used to help understand the method and core ideas of the present invention. At the same time, for those skilled in the art, according to the ideas of the present invention, there will be changes in the specific implementation methods and application scopes. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for predicting performance parameters of complex equipment based on spatiotemporal dynamic correlation graphs. It is characterized in that The following steps are involved: S1: Continuously sample and record historical signals of multiple condition monitoring variables of complex equipment including performance parameters to obtain multi-channel condition monitoring data of complex equipment; S2: After preprocessing the multi-channel condition monitoring data of complex equipment, a data set for predicting equipment performance parameters is constructed; S3: After training the complex equipment performance parameter prediction model based on the spatiotemporal dynamic association graph using the equipment performance parameter prediction data set, a trained complex equipment performance parameter prediction model is obtained; S4: Preprocess the new state monitoring variable signal obtained in real time and then input it into the trained complex equipment performance parameter prediction model to obtain the prediction result of the complex equipment performance parameter under the prediction step.

2. According to claim 1, a method for predicting complex equipment performance parameters based on a spatiotemporal dynamic association graph, It is characterized in that In S3, the complex equipment performance parameter prediction model based on the spatiotemporal dynamic association graph first calculates the input sequence X of any sample in the equipment performance parameter prediction data set. (i) Perform uniform and non-overlapping subsequence segmentation with a preset local window size in the time dimension to obtain local window-level subsequences satisfy Among them, C max is the total number of channels of the state monitoring variable, L represents the number of local windows, T L Indicates the sampling step size corresponding to each local window, The number of channels is C max ×L×T L The real number set of the local window level subsequence Input into the spatiotemporal dynamic association graph generation module, the spatiotemporal dynamic association graph generation module generates a spatiotemporal dynamic association graph satisfy in is a node set, and the characteristics of each node in the node set are satisfy The number of channels is D×d gh A real number set, D represents the number of nodes in each graph, satisfying D = C max ·L, d gh Represents the feature dimension of each node, ε (i) is an edge set, A (i) Represents the adjacency matrix; then based on the spatiotemporal dynamic association graph And the preset order k solves the neighborhood subgraph centered on the node corresponding to the performance parameter variable Then according to the neighborhood subgraph The belonging relationship between the contained nodes and the corresponding state variables identifies the state variables related to each performance parameter, and then from the local window level subsequence Construct a subset of the input sequence containing performance parameters and related state variables satisfy in, The number of channels is (C relevant +1)×L×T L The set of real numbers, C relevant is the number of relevant state variables; then, the input sequence subset and neighborhood subgraph Input the timing prediction module to obtain the prediction output of performance parameter variables satisfy The number of channels is T p The set of real numbers, T p is the prediction step length.

3. According to claim 2, a method for predicting complex equipment performance parameters based on a spatiotemporal dynamic association graph, It is characterized in that The spatiotemporal dynamic association graph generation module includes a self-encoding unit, a graph initialization unit, a channel sparsification unit and a local association pattern capture sparsification unit which are connected in sequence.

4. According to claim 3, a method for predicting complex equipment performance parameters based on a spatiotemporal dynamic association graph, It is characterized in that The autoencoder unit includes an autoencoder neural network, which is composed of a feature encoder and a feature decoder. The feature encoder encodes the local window-level subsequence in the time dimension. After encoding, the potential features of the low-dimensional embedding space are output And as the input of the feature decoder and the input of the graph initialization unit, satisfying in The number of channels is C max ×L×d enc The set of real numbers, d enc is the self-encoding feature dimension; feature The decoder embeds the latent features into a low-dimensional space Decode to the feature encoder input dimension to obtain the reconstructed subsequence satisfy According to the local window level subsequence and reconstruct subsequence The autoencoder loss of the spatiotemporal dynamic association graph generation module is calculated.

5. According to claim 3, a method for predicting complex equipment performance parameters based on a spatiotemporal dynamic association graph, It is characterized in that In the graph initialization unit, the latent features of the low-dimensional embedding space output by the self-encoding unit are Convert to initial node features of spatiotemporal dynamic association graph satisfy The number of channels is D×d enc A real number set; according to the initial node characteristics of the spatiotemporal dynamic association graph After calculating the feature similarity, we get the graph with edge weights to initialize the adjacency matrix satisfy is a real number set with D×D channels, and the adjacency matrix of the graph with edge weights is initialized Transformed into a set of variable pair subgraph adjacency matrices And as the output of the graph initialization unit, satisfying The number of channels is The set of real numbers.

6. According to claim 3, a method for predicting complex equipment performance parameters based on a spatiotemporal dynamic association graph, It is characterized in that In the channel sparsification unit, the subgraph adjacency matrix set is the variable pair output by the graph initialization unit As input, the feature similarity threshold is set as a learnable parameter, the adjacency matrix of the subgraph of variables whose average feature similarity is less than or equal to the feature similarity threshold is set to zero, and the nonlinear function is used to perform nonlinear transformation on the edge weights in the subgraph adjacency matrix of the variables that are not set to zero, and the channel sparse subgraph adjacency matrix set is obtained. And as the output of the channel sparsification unit, satisfying The number of channels is The set of real numbers.

7. According to claim 3, a method for predicting complex equipment performance parameters based on a spatiotemporal dynamic association graph, It is characterized in that The local correlation pattern captures the adjacency matrix of the channel-sparse subgraph in the sparse unit. As input, the convolutional neural network transforms the value of each edge weight in the subgraph adjacency matrix of non-zero variables, and uses the ReLU (·) activation function to set the negative edge weights to zero, and obtains the pattern capture sparse subgraph adjacency matrix set satisfy After channel transformation, the final spatiotemporal dynamic association graph adjacency matrix is ​​obtained satisfy Then construct the spatiotemporal dynamic association graph output by the spatiotemporal dynamic association graph generation module 8. According to claim 2, a method for predicting complex equipment performance parameters based on a spatiotemporal dynamic association graph, It is characterized in that The temporal prediction module comprises an embedding unit, a temporal encoding unit and a temporal decoding unit which are connected in sequence.

9. A method for predicting complex equipment performance parameters based on a spatiotemporal dynamic association graph according to claim 8, It is characterized in that In the embedding unit, first, a subset of the input sequence is Perform Patch division to obtain block sequence subsets satisfy in The number of channels is (C relevant +1)×L×N p ×P real number set, N p represents the number of patches in each local window in the Patch partition, P represents the Patch length, and then the block sequence subset Use linear transformation to get the embedding features of each patch satisfy The number of channels is (C relevant +1)×L×N p ×d trans The set of real numbers, d trans is the hidden dimension of the model, and then the embedding features of each patch are extracted separately The average information of each patch corresponding to each local window in the , and use it as the initialization value to generate the head patch embedding satisfy The number of channels is (C relevant +1)×L×1×d trans A real number set; then the head patch corresponding to each local window is embedded into Embedded features with the original patches After splicing, the labeled Patch embedding features are obtained satisfy The number of channels is (C relevant +1)×L×(1+N p )×d trans A real number set; finally, the labeled patch is embedded into the feature Each patch under all local windows in the is concatenated along the time dimension and a learnable position encoding is performed to obtain the encoded patch embedding feature And as the output of the embedding unit, it satisfies The number of channels is (C relevant +1)×N fp ×d trans The real number set, N fp The total number of patches corresponding to each variable, satisfying N fp =L·(N P +1).

10. A method for predicting complex equipment performance parameters based on a spatiotemporal dynamic association graph according to claim 8, It is characterized in that The temporal coding unit is formed by stacking J layers of temporal encoders, wherein the j-th layer of temporal encoders converts the j-1-th layer of temporal coding features into Encoded as the j-th layer temporal coding feature The number of channels is (C relevant +1)×N fp ×d trans The set of real numbers that satisfies Each layer of temporal encoder is composed of a Transformer encoding layer and a graph neural network (GNN) encoding layer in series, where the Transformer encoding layer converts the j-1th layer temporal encoding features into Each patch in is regarded as a word unit and is encoded with attention to obtain the intermediate encoding feature satisfy Then, we encode the features from the middle Separate the head patch features corresponding to each local window satisfy The number of channels is (C relevant +1)×L×1×d trans The real number set, the GNN encoding layer is based on the head Patch feature and neighborhood subgraph As input, the features of each head patch are further encoded based on the graph neural network algorithm to obtain the graph encoding feature Finally, the graph encoding features Re-encoding features The non-head Patch features in are concatenated to obtain the j-th layer temporal coding features. as output.

Citation Information

Cited By

  • Flotation process working condition identification method and system based on dynamic time sequence diagram neural network

    CN120219843A