Water turbine unit anomaly detection method based on combination of space-time large model and GAN network

Through a method based on a space-time large model combined with GAN network, combining STM and GLLM to capture space-time correlation, and adopting an adversarial training strategy, the problems of space-time correlation identification and data distribution identification in the flow-solid coupling surface abnormal detection of the turbine overcurrent components are solved. Improves stress value estimation accuracy and detection performance.

CN118981656APending Publication Date: 2024-11-19HUANENG LANCANG RIVER HYDROPOWER CO LTD +1

Patent Information

Application Number
CN202410972644.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-19
Publication Date
2024-11-19

AI Technical Summary

Technical Problem

The prior art is difficult to effectively capture the spatial and temporal correlation in the detection of flow-solid coupling surface abnormalities of the overcurrent components of the turbine, and the supervision and training strategy cannot fully identify the data distribution, resulting in inaccurate estimates.

Method used

Using a method based on space-time large model combined with GAN network, the space-time correlation is captured through STM and GLLM, and combined with an adversarial training strategy of unsupervised and supervised learning, the generator and discriminator are built to generate and judge grid point stress values ​​and optimize the model. Training process.

Benefits of technology

The accuracy of grid point stress value estimation and comprehensive training of the model are improved, the performance of abnormal detection of turbine units is enhanced, and abnormal situations can be identified more accurately.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118981656A_ABST
    Figure CN118981656A_ABST
Patent Text Reader

Abstract

The invention discloses a water turbine unit anomaly detection method based on combination of a space-time large model and a GAN network. The method comprises the following steps: inputting clustered water turbine unit fluid-structure interaction surface grid point data under a current working condition; constructing an anomaly detection model; training an anomaly detection model by using the collected data; calculating the accuracy rate of the stress value of each grid point estimated by the model, and if the accuracy rate exceeds a preset threshold value, inputting the clustered grid point data of the fluid-structure interaction surface of the flow passage component of the water turbine into the trained anomaly detection model to obtain the stress estimation value of each grid point under the current working condition, and performing water turbine unit anomaly detection according to the stress estimation value. And if the accuracy is low, continuing to train the anomaly detection model. Compared with the prior art, the method has the advantages of good anomaly detection effect, good practicability and the like.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a method for detecting anomalies of a turbine unit based on a large spatiotemporal model combined with a GAN network, which is used for detecting anomalies of a fluid-solid coupling surface of a flow-through component of a turbine unit and belongs to the technical field of water conservancy detection. Background Art

[0002] Accurately detecting the abnormality of the fluid-structure coupling surface of the turbine flow passage is of great significance. However, since the fluid-structure coupling surface is regarded as a grid surface, the grid points in the grid surface influence and interact with each other. Therefore, the abnormality detection of the fluid-structure coupling surface of the turbine flow passage can be regarded as a spatiotemporal analysis task.

[0003] In the spatiotemporal analysis task, the existing methods are mainly divided into three categories: 1) Classical statistical methods: This is a low-cost and easy-to-apply technique for capturing mathematical relationships in time series data. Classical statistical methods include AR (Auto Regression, AR for short) and ARIMA (Autoregressive Integrated Moving Average Model, ARIMA for short). However, the nonlinear capabilities of these methods limit their performance in spatiotemporal analysis; 2) Shallow machine learning methods: In order to break through the defects of classical statistical methods, some studies have begun to use some shallow machine learning models such as SVR (Support Vector Regression, SVR for short) for spatiotemporal analysis. Compared with classical statistical methods, although these methods have improved in nonlinear capabilities to a certain extent, the generalization ability of these models is weak and they cannot extract complex and dynamic spatiotemporal correlations from spatiotemporal data. In addition, these methods have limited ability to process massive data; 3) Deep learning methods: With the rapid development of deep learning methods, the potential of these methods in modeling dynamic spatiotemporal correlations in spatiotemporal big data has attracted much attention from researchers. Some researchers have begun to use deep learning methods for modeling. Specifically: RNN (Recurrent Neural Network, RNN for short) was first used by researchers to extract temporal dependencies, but these methods have difficulty capturing spatial dependencies. To this end, some studies designed CNN-LSTM (Convolutional Neural Network-Long Short-Term Memory, CNN-LSTM for short), and convolutional neural networks (Convolutional Neural Network, CNN for short) were used to capture spatial dependencies. However, CNN models the grid surface as a standard format, such as a two-dimensional matrix, thereby losing some important spatial relationships, such as the dependencies between distant grid points. In view of this, researchers proposed GC-LSTM (Graph Convolutional Network-Long Short-Term Memory, GC-LSTM for short), in which GCN treats complex grid surfaces as graphs to fully maintain the original spatial relationships between grid points. With the emergence of Transformer, some studies have used self-attention (self-attention mechanism, referred to as self-attention) to extract spatiotemporal correlations to achieve effective spatiotemporal analysis.For example, ST-Transformer (Spatio-Temporal Transformer, ST-Transformer for short) and ASTemCN (spatial–temporal fusion attention model, ASTemCN for short) replace GCN and LSTM with spatiotemporal attention, thereby improving the ability to model spatiotemporal features. However, the Transformer has a complex structure and requires many parameters to be trained. This requires a large amount of training data and consumes a lot of computing resources. At present, we have witnessed the success of LLM (Large Language Models, LLM for short), which has promoted revolutionary progress in the field of natural language processing. LLM also performs well in spatiotemporal sequence analysis. For example, some studies have proposed a unified architecture, namely FPT (Frozen Pretrained Transformer, FPT for short), to analyze spatiotemporal sequences using pre-trained GPT-2 (chat GPT-2, chatGPT-2, GPT-2 for short). Although LLM can further improve the performance of spatiotemporal analysis, few studies have used LLM for anomaly detection of fluid-structure coupling surfaces of turbine flow passage components. If we directly apply the existing LLM to anomaly detection of fluid-structure coupling surfaces of turbine flow passage components, we will face two key challenges: 1) Previous studies have shown that extracting the spatiotemporal correlation between grid points can significantly improve the performance of fluid-structure coupling surface anomaly detection. However, LLM is a pure decoder structure, that is, it can only generate stress estimates for each grid point under the current working condition based on the three-dimensional coordinates of all fluid-structure coupling surface grid points, turbine head, and turbine opening under past working conditions, lacking consideration of spatiotemporal correlation. Therefore, it is crucial to capture spatiotemporal features in LLM; 2) In order to adapt to downstream tasks, LLM usually needs to be fine-tuned during the training process. At present, most LLM-based spatiotemporal sequence analysis models adopt a supervised training strategy, that is, minimizing the MSE (mean squared error, MSE for short) loss to make the estimated value close to the true value. However, supervised training is not comprehensive because it cannot help the model identify the true data distribution, resulting in inaccurate estimates. It is not easy to design an appropriate training strategy to fully train LLM. Summary of the invention

[0004] Purpose of the invention: In view of the problems and shortcomings of the prior art, the present invention provides a method for detecting anomalies of a turbine unit based on a large spatiotemporal model combined with a GAN network, which has good performance and strong practicability.

[0005] Technical solution: A method for detecting abnormalities of a turbine unit based on a spatiotemporal large model combined with a GAN network, comprising the following steps:

[0006] Step 1: Input the clustered fluid-structure interaction surface grid point data under each working condition;

[0007] Step 2: Construct a fluid-solid coupling surface anomaly detection model;

[0008] Step 3: Use the fluid-solid coupling surface grid point data to train the anomaly detection model constructed in step 2;

[0009] Step 4: Calculate the accuracy of the stress value predicted by the model. If the accuracy exceeds the preset threshold, execute step 5; otherwise, return to step 3.

[0010] Step 5: Input the generated fluid-solid coupling surface grid point data into the trained anomaly detection model to obtain the stress prediction value of each grid point under the current working condition, and then perform turbine unit anomaly detection based on the stress estimation value.

[0011] Preferably, the step 1 is specifically:

[0012] The grid point data of the fluid-solid coupling surface of the turbine flow passage components under different working conditions are collected. The data features contained in each grid point are three-dimensional coordinates, turbine head, turbine opening and stress. The missing values ​​of these grid point data are filled and outliers are deleted. Then, FCs (Fully Connected Neural Networks, FCs for short) are used to process the data features contained in each grid point. After that, the K-Means algorithm is applied to cluster all grid point data to generate multiple clusters of data. For each cluster of grid points, an undirected graph is generated.

[0013] Preferably, the step 2 is specifically:

[0014] Based on the principle of deep learning, a generator and a discriminator are constructed to establish a fluid-solid coupling surface anomaly detection model. The generator is used to generate the stress estimation value of the grid point under the current working condition, and the discriminator is used to determine whether the stress value under the current working condition is real or generated by the generator.

[0015] More preferably, the generator is composed of STLLM, which is specifically composed of two parts, namely STM and GLLM, STM is used to mine the spatiotemporal correlation between each cluster of grids, and GLLM is used to generate the stress estimation value of each cluster of grid points under the current working condition.

[0016] The input of STM is an undirected graph composed of each cluster of grid points. STM is composed of stacked STAtt blocks (Spatio-Temporal Attention Block, STAtt for short). Each STAtt block contains a spatial attention and a temporal attention. Spatial attention is used to mine the spatial dependencies between nodes in the graph, and temporal attention is used to capture the temporal dependencies of each graph node.

[0017] The input of GLLM is the output representation of STM. GLLM consists of tokenization, tag and position encoding, and GPT-2. Tokenization is used to make the output representation of STM suitable for GPT-2, tag encoding is used to align the dimension of the tokenized output with the embedding dimension of GPT-2, position encoding is used to learn the positional relationship of different working conditions, and GPT-2 uses its rich intrinsic knowledge to generate stress estimates under the current working condition.

[0018] More preferably, the discriminator is specifically composed of three parts, namely, a GCN layer, an LSTM layer, and a classification layer. The GCN layer is used to mine the spatial dependencies between the grid point stress output by the generator and the real grid point stress under the current working condition. The LSTM is then used to capture the temporal dependencies therein. Finally, the classification layer is used to determine whether the input of the discriminator comes from the grid point stress output by the generator or the real grid point stress.

[0019] Preferably, the step 3 is specifically:

[0020] The anomaly detection model is trained using adversarial training, which uses an overall objective function as the loss function.

[0021] More preferably, the overall objective function is specifically:

[0022]

[0023] Where Dis represents the output generated by the discriminator, Gen represents the output generated by the generator, E is the empirical estimate of the expected probability value, z~P(R E ) indicates that the training data z follows the true sample distribution P(R E ); is a given condition, representing the data characteristics of multiple clusters of grid points under different working conditions in the past; is the output of the generator under the given condition Down; Indicates the probability value of the training data z generated by the discriminator from the real sample under the given conditions The base of the log function is determined according to the specific task, and in the present invention, the base of the log function is set to 10; the discriminator aims to maximize Through unsupervised training, the generator aims to minimize Through supervised training.

[0024] Preferably, the step 5 is specifically:

[0025] After obtaining the stress estimation value of the grid point under the current working condition, the present invention determines whether the grid point is abnormal by whether the stress estimation value exceeds or is lower than the upper limit or lower limit of the preset normal stress value. If 10-20 abnormal grid points appear in the same small area, it means that the abnormal grid points are gathered in large numbers, indicating that the turbine unit is abnormal under the current working condition. If the abnormal grid points are dispersed, it indicates that the turbine unit is normal under the current working condition.

[0026] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for detecting anomalies of a turbine unit based on a large spatiotemporal model combined with a GAN network as described above is implemented.

[0027] A computer-readable storage medium having a computer program / instruction stored thereon, which, when executed by a processor, implements the steps of the method for detecting anomalies of a turbine unit based on a large spatiotemporal model combined with a GAN network as described above.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] 1. This invention proposes a method for detecting anomalies of a turbine unit based on a spatiotemporal large model combined with a GAN network. In order to extract spatiotemporal correlations and generate stress estimates under current working conditions, this invention designs an STLLM with STM and GLLM. The former is used to capture spatiotemporal correlations, and the latter uses GPT-2 as the backbone to infer the stress values ​​of all grid points under current working conditions.

[0030] Second, in order to optimize the training of STLLM, we developed a GAN (Generative Adversarial Network, GAN for short), which combines unsupervised learning and supervised learning to form an adversarial training strategy. In unsupervised training, STLLM can minimize the adversarial loss to learn the distribution of real data. In supervised training, the goal of STLLM is to minimize the MSE loss so that the estimated value is consistent with the true value. In this way, the adversarial loss and the MSE loss complement each other, promote more comprehensive training of the model, and further improve the estimation accuracy of grid point stress values. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1A flow chart of abnormality detection of a water turbine unit according to an embodiment of the present invention;

[0032] Figure 2 It is an overall structural diagram of a water turbine unit abnormality detection model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0033] The present invention is further explained below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0034] The present invention uses the learning ability of deep learning on big data to establish an abnormality detection model for the fluid-solid coupling surface of the flow-through components of a turbine unit. A method for detecting abnormalities of a turbine unit based on a spatiotemporal large model combined with a GAN network, such as Figure 1 As shown, the following steps are included:

[0035] Step 1: Input the clustered fluid-structure interaction surface grid point data under each working condition;

[0036] The grid point data of the fluid-solid coupling surface of the turbine flow components under different working conditions are collected. The data features contained in each grid point are three-dimensional coordinates, turbine head, turbine opening, and stress. The turbine head and opening correspond to one working condition. A total of 100,000 grid point data under 10,000 working conditions are collected. The missing values ​​of these grid point data are filled and outliers are deleted. Then, FCs (Fully Connected Neural Networks, FCs for short) are used to process the data features contained in each grid point. After that, the K-Means algorithm is applied to cluster all grid point data to generate multiple clusters of data. For each cluster of grid points, an undirected graph is generated.

[0037] Given the grid point data of the fluid-solid coupling surface of the flow-through component of the turbine, after filling the missing values ​​and deleting the outliers of the N grid point data, the K-Means algorithm is proposed to divide the N grid points into K clusters. For the i-th cluster, an undirected graph G is defined i =(V i ,E i ), where V i represents the set of grid points of the undirected graph formed by the i-th cluster, E i Represents the set of edges connecting grid points, an undirected graph G I The data characteristics at working condition t are expressed as in represents the grid point k iIn the data features of operating condition t, C represents the dimension of the data features (e.g., three-dimensional coordinates, turbine head, turbine opening, and stress). Specifically, assuming there are P historical operating conditions, the undirected graph sequence of the i-th cluster can be expressed as k i represents the total number of grid points in the i-th cluster. A two-layer FCs is introduced to convert X i The dimension is converted from C to D and can be expressed as: That is, the past P working conditions k i The input hidden state of each grid point.

[0038] Step 2: Construct a fluid-solid coupling surface anomaly detection model;

[0039] Based on the principle of deep learning, a generator and a discriminator are constructed to establish a fluid-solid coupling surface anomaly detection model. The generator is used to generate stress estimation values ​​of 100,000 grid points under the current working conditions, and the discriminator is used to determine whether the stress value under the current working conditions is real or generated by the generator.

[0040] The specific structure of the model is as follows Figure 2 As shown in Figure 2, the model mainly consists of two parts, namely the generator and the discriminator. The generator consists of STLLM. STLLM is mainly composed of STM and GLLM, and STM is composed of B stacked STAtt blocks. The Bth STAtt (Spatio-Temporal Attention, STAtt for short) block generates the final spatio-temporal representation, which can be expressed as Each STAtt block consists of spatial attention and temporal attention. The former is first used to capture spatial dependencies. Then the latter is used to mine temporal dependencies. In the rth STAtt block, the spatiotemporal representation of the past P working conditions generated by the r-1th STAtt block is used as the input of the spatial attention, where the tth k The working condition is expressed as The spatial representation produced by spatial attention is in Represents the k Representation of a working condition.

[0041] In detail, for the working condition t k , the spatial attention operation formula of the h-th head is

[0042]

[0043] in, Indicates that in working condition t k The spatial representation of the h-th head. Key vector Through linear mapping Get. Query vector Through linear mapping Get the value vector Through linear mapping get. as well as N represents the parameters of the learnable linear mapping. h is the total number of heads, and α represents the scaling factor. After that, the working condition t k The space is represented as:

[0044]

[0045] in is the weight parameter, and [*,…,*] is the concatenation operation.

[0046] The input of temporal attention is the spatial representation of the past P working conditions in is the grid point s i spatial representation. is the spatiotemporal representation produced by temporal attention, where is the grid point s i For the grid point s i , the temporal attention operation formula of the h-th head is

[0047]

[0048] in is the space-time representation of the hth head, and They are key vector, query vector and value vector respectively. and represents the trainable parameters in the linear mapping. β is the scaling factor.

[0049] Having obtained After that, the grid point s i The space-time representation of is:

[0050]

[0051] in is the weight parameter.

[0052] GLLM consists of three structures, namely tokenization, tagging and position encoding, and GPT-2. In order to adapt GPT-2 to spatiotemporal representation, tokenization uses channel independence and patches to mark the spatiotemporal representation to meet the input requirements of GPT-2. The present invention first uses a two-layer FCs to transform the spatiotemporal representation into The dimension is restored from D to F, expressed as Where F represents the dimension of the stress feature, that is, F = 1. Then, channel independence converts the spatiotemporal representation of multiple grid points into the spatiotemporal representation of multiple single grid points, thereby converting the dimension of the representation into Finally, for the spatiotemporal representation of each site, a patch is designed to merge adjacent conditions into one patch, reducing the dimensionality from P to P v , where P v represents the total number of patches. Subsequently, the dimension of the spatiotemporal representation of each grid point is expanded from 1×F to L d × F, where L d is the length of each patch. Therefore, the tokenization can be expressed as:

[0053]

[0054] Where CI stands for channel independence, pa stands for patch, is the grid point S i The spatiotemporal representation of patches. The tag encoding is developed to align the dimensions of the patch with the embedding dimensions of GPT-2. The present invention uses 1D-CNN (One Dimensional-Convolutional Neural Network, 1D-CNN for short) for tag encoding. Given a grid point S i Patch spatiotemporal representation of The tag encoding is formalized as:

[0055]

[0056] Where 1D-Conv means 1D-CNN, is the patch embedding produced by the marker encoding. T d is the patch embedding dimension, which is consistent with the embedding dimension of GPT-2. In order to learn the positional relationship between various working conditions, the present invention introduces position encoding. Assume that the grid point S i Patch location index A learnable lookup table P posi For mapping patch locations:

[0057]

[0058] in is the position embedding generated by the position encoding. and Then, the present invention adds them together to obtain the aggregate embedding

[0059]

[0060] after, is input into GPT-2. Afterwards, the present invention uses the pre-trained GPT-2 to generate output embedding, which can be expressed as:

[0061]

[0062] Where PTG represents pre-trained GPT-2. Represents the grid point S i Finally, the present invention uses the flattening (Fla) and rearrangement (Rea) operations to infer the grid point S under the current working condition. i Stress

[0063]

[0064] in represents the weight parameter. For the k i The stress value of the grid point under the current working condition is inferred by the above-mentioned GLLM, which is expressed as

[0065] The present invention uses the real sample set R E And the fake sample set F A As input, a discriminator is designed to determine which samples are real and which are fake. Given the stress values ​​of all N grid points in K clusters for the past P working conditions And the actual stress values ​​of N grid points in the current working condition The present invention reconstructs R E By splicing Y and Expressed as F A Created by splicing F A and R E is input to the discriminator. The discriminator consists of a GCN layer, an LSTM layer and a classification layer. A First, use 2 layers of FCs to A The dimension of is converted from F to D, expressed as: In the working condition t q The embedding is The spatial representation generated by a single-layer GCN is In the working condition t q The expression is Specifically, under the working condition t q , the spectral GCN is expressed as:

[0066]

[0067] where ⊙ represents the Hadamard product. is the graph Laplacian matrix, where represents the degree matrix, represents the identity matrix, is the adjacency matrix. Represents a learnable weight parameter. Relu is regarded as an activation function. After processing the embedding of P+1 working conditions, the GCN layer outputs the spatial representation of the fake sample set after, is input to the LSTM layer. q , the LSTM unit can be expressed as:

[0068]

[0069]

[0070] in and They represent output gate, forget gate and input gate respectively. is the storage unit state. F , W I , W O , W C , U F , U I , U O and U C is a learnable weight parameter, b F , b I , b O and b C is the bias. g is the activation function sigmoid. After obtaining the output of each LSTM unit, the present invention aggregates them to obtain the spatiotemporal representation Finally, the classification layer is used to generate the probability Dis(F A ), which can be expressed as:

[0071]

[0072] in is a learnable weight parameter, Represents bias. Apply the same structure of the discriminator to process R E .

[0073] Step 3: Train the anomaly detection model built in step 2;

[0074] For the constructed model, a training set, a validation set and a test set are constructed, the model is initialized and trained using the training set data, the parameters of the model are adjusted using the validation set and evaluation indicators, and the model effect is tested using the test set; in this embodiment, the proportions of the training data set, the validation data set and the test data set are 70%, 10% and 20% respectively. The overall objective function is used as the loss function for training, and the model is fine-tuned to complete the model training, and finally the stress estimation values ​​of all grid points on the fluid-solid coupling surface of the turbine flow-through component are output.

[0075] The overall objective function is specifically:

[0076]

[0077] Where Dis represents the output generated by the discriminator, Gen represents the output generated by the generator, E is the empirical estimate of the expected probability value, z~P(R E ) indicates that the training data z follows the true sample distribution P(R E ); is a given condition, representing the data characteristics of multiple clusters of grid points under different working conditions in the past; is the output of the generator under the given condition Down; represents the probability value of the training data z generated by the discriminator coming from the real sample under the given condition x; the base of the log function is determined according to the specific task, and in the present invention, the base of the log function is set to 10; the discriminator aims to maximize Through unsupervised training, the generator aims to minimize Through supervised training.

[0078] Step 4: Calculate the accuracy of the model stress estimation value. If the accuracy exceeds the preset threshold, execute step 5; otherwise, return to step 3.

[0079] Step 5: After obtaining the stress estimation values ​​of 100,000 grid points under the current working condition, the present invention determines whether the grid points are abnormal by whether the stress estimation values ​​exceed or are lower than the upper or lower limit of the preset normal stress value. If a large number of abnormal grid points are gathered, it indicates that the turbine unit is abnormal under the current working condition. If the abnormal grid points are dispersed, it indicates that the turbine unit is normal under the current working condition.

[0080] The data of all grid points on the fluid-solid coupling surface of the turbine flow-through component under the past working conditions are input into the trained anomaly detection model to obtain the stress estimation values ​​of all grid points under the current working conditions. Then, the present invention determines whether the stress estimation value of each grid point exceeds or is lower than the upper or lower limit of the preset normal stress value, and records the three-dimensional coordinates of these abnormal grid points. Finally, the present invention uses AutoCAD software to generate the spatial position distribution of these abnormal grid points in a three-dimensional coordinate system. If a large number of abnormal grid points are gathered, it indicates that the turbine unit is abnormal under the current working conditions. If the abnormal grid points are dispersed, it indicates that the turbine unit is normal under the current working conditions.

[0081] This embodiment also relates to a storage medium, in which any of the above-mentioned abnormality detection methods is stored.

[0082] Obviously, those skilled in the art should understand that the various steps of the above-mentioned method for detecting abnormalities of a turbine unit based on a spatiotemporal large model combined with a GAN network in the embodiment of the present invention can be implemented by a general computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, optionally, they can be implemented with executable program codes of the computing device, so that they can be stored in a storage device and executed by the computing device, and in some cases, the steps shown or described can be executed in a different order than here, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. In this way, the embodiments of the present invention are not limited to any specific combination of hardware and software.

Claims

1. A method for detecting abnormality of a turbine unit based on a spatiotemporal large model combined with a GAN network, characterized in that: The steps include: Step 1: Input the clustered fluid-structure interaction surface grid point data under each working condition; Step 2: Construct a fluid-solid coupling surface anomaly detection model; Step 3: Use the fluid-solid coupling surface grid point data to train the anomaly detection model constructed in step 2; Step 4: Calculate the accuracy of the stress value predicted by the model. If the accuracy exceeds the preset threshold, execute step 5; otherwise, return to step 3; Step 5: Input the generated fluid-solid coupling surface grid point data into the trained anomaly detection model to obtain the stress prediction value of each grid point under the current working condition, and then perform turbine unit anomaly detection based on the stress estimation value.

2. The method for detecting anomalies of a turbine unit based on a spatiotemporal large model combined with a GAN network according to claim 1 is characterized in that: In step 1: The grid point data of the fluid-solid coupling surface of the turbine flow component under different working conditions are collected, and the data features contained in each grid point are three-dimensional coordinates, turbine head, turbine opening and stress; the missing values ​​of the grid point data are filled and the outliers are deleted. Then, the data features contained in each grid point are processed by a fully connected network, and all the grid point data are clustered by the K-Means algorithm to generate multiple clusters of data. For each cluster of grid points, an undirected graph is generated.

3. The method for detecting anomalies of a turbine unit based on a spatiotemporal large model combined with a GAN network according to claim 1 is characterized in that: In step 2: Based on the principle of deep learning, a generator and a discriminator are constructed, and a fluid-solid coupling surface anomaly detection model is established. The generator is used to generate the stress estimation value of the grid point under the current working condition, and the discriminator is used to determine whether the stress value under the current working condition is real or generated by the generator.

4. The method for detecting anomalies of a turbine unit based on a spatiotemporal large model combined with a GAN network according to claim 3 is characterized in that: The generator execution method is: The generator is composed of STLLM, which includes STM and GLLM. STM is used to mine the spatiotemporal correlation between each cluster of grids; GLLM is used to generate stress estimation values ​​of each cluster of grid points under current working conditions.

5. The method for detecting abnormality of a turbine unit based on a spatiotemporal large model combined with a GAN network according to claim 3 is characterized in that: The discriminator includes a GCN layer, an LSTM layer and a classification layer; the GCN layer is used to mine the spatial dependency in the grid point stress output by the generator and the real grid point stress under the current working condition; LSTM is then used to capture the temporal dependency in the grid point stress output by the generator and the true grid point stress under the current working condition; Finally, the classification layer is used to determine whether the input of the discriminator comes from the grid point stress output by the generator or the true grid point stress.

6. The method for detecting anomalies of a turbine unit based on a spatiotemporal large model combined with a GAN network according to claim 1 is characterized in that: In step 3: The anomaly detection model is trained using adversarial training, which uses an overall objective function as the loss function.

7. The method for detecting abnormality of a turbine unit based on a spatiotemporal large model combined with a GAN network according to claim 6 is characterized in that: The overall objective function is specifically: Where Dis represents the output generated by the discriminator, Gen represents the output generated by the generator, E is the empirical estimate of the expected probability value, z~P(R E ) indicates that the training data z follows the true sample distribution P(R E ); is a given condition, representing the data characteristics of multiple clusters of grid points under different working conditions in the past; is the output of the generator under the given condition Down; Indicates the probability value of the training data z generated by the discriminator from the real sample under the given conditions The base of the log function is determined according to the specific task, and in the present invention, the base of the log function is set to 10; the discriminator aims to maximize Through unsupervised training, the generator aims to minimize Through supervised training.

8. The method for detecting abnormality of a turbine unit based on a spatiotemporal large model combined with a GAN network according to claim 1 is characterized in that: In step 5: After obtaining the stress estimation value of the grid point under the current working condition, the grid point abnormality is judged by whether the stress estimation value exceeds or falls below the upper limit or lower limit of the preset normal stress value; If 10-20 abnormal grid points appear in the same small area, it means that the abnormal grid points are gathered in large numbers, indicating that the turbine unit is abnormal under the current operating conditions; If the abnormal grid points are dispersed, it indicates that the turbine unit is normal under the current operating conditions.

9. A computer device, characterized in that: The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for detecting anomalies of a turbine unit based on a large spatiotemporal model and a GAN network is implemented as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instruction is executed by the processor, the steps of the method for detecting anomalies of a turbine unit based on a large spatiotemporal model combined with a GAN network are implemented as described in any one of claims 1 to 7.

Citation Information

Patent Citations

  • Plant station site violation operation behavior processing method and device and storage medium

    CN114842408A

  • Method and terminal for detecting time series data exception by combining attention mechanism and LSTM (Long Short Term Memory)

    CN115983087A

Cited By

  • Traffic prediction method combining dynamic GCN and fine-tuning GPT2

    CN119763327A