Fault prediction method and system for combined cycle power generation system

Abstract

The invention discloses a combined cycle power generation system-oriented fault prediction method and system, belongs to the technical field of power generation system fault prediction, and solves the problem that an early warning level cannot comprehensively describe the operation state of equipment due to relatively weak comprehensive processing capability of an existing system early warning module for multi-dimensional features. The method comprises the steps of performing normalization enhancement processing on real-time state parameters, performing multi-dimensional splicing on node equipment labels and state feature sets by adopting a fusion convolutional layer, performing multi-dimensional evaluation on a multi-dimensional feature graph based on a fault prediction model, and outputting a node equipment evaluation result. According to the method, the node equipment is assessed by adopting the fault prediction model based on the Transform model, complex multi-dimensional data in the combined cycle power generation system can be better processed, multiple factors such as fault probability distribution and comprehensive severity are comprehensively considered, the fault risk of the equipment can be assessed more accurately, and the reliability of the equipment is improved. Therefore, the running state of the equipment is comprehensively described.

Description

Technical Field

[0001] This invention belongs to the field of power generation system fault prediction technology, specifically relating to fault prediction methods and systems for combined cycle power generation systems. Background Technology

[0002] Combined cycle power generation systems are a highly efficient and clean method that organically combines gas turbine and steam turbine power generation systems. In this system, the gas turbine burns fuel to produce high-temperature, high-pressure gas, which drives the gas turbine to generate electricity. Simultaneously, the high-temperature exhaust gas from the gas turbine enters a waste heat boiler to produce steam, which then drives a steam turbine to generate electricity. This combined cycle method can fully utilize fuel energy, improve power generation efficiency, reduce energy consumption and environmental pollution, and is widely used in the power generation field, forming an important component of the modern energy system.

[0003] However, the complexity of combined cycle systems places extremely high demands on their reliability. Problems such as thermal fatigue of gas turbine blades, vibration of steam turbine shaft systems, corrosion of waste heat boiler tube walls, and efficiency decline of auxiliary systems can all lead to unplanned shutdowns, reduced energy efficiency, or even safety accidents. Traditional operation and maintenance models rely on periodic inspections and manual experience-based judgment, which are insufficient to cope with gradual equipment failures and sudden anomalies. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by providing a fault prediction method and system for combined cycle power generation systems, which solves the problem that existing system early warning modules are unable to cope with fault prediction in complex combined cycle systems.

[0005] This invention is implemented as follows: a fault prediction method for combined cycle power generation systems, the method comprising: The system acquires real-time status parameters of monitoring nodes in the cyclic electronic power generation system, performs normalization enhancement processing on the real-time status parameters to obtain status enhancement information, and uploads the status enhancement information to the strategy database. Specifically, monitoring nodes are set up for key locations of each module and / or device in the cyclic electronic power generation system, and data is collected. Monitoring nodes are also set up for key locations and key devices of each subsystem in the cyclic electronic power generation system, and data is collected. The subsystems are the modules and / or devices in the cyclic electronic power generation system. State feature extraction is performed on the state enhancement information based on convolutional neural network to obtain state feature set. The potential relationship between state feature set and corresponding node device label is learned by autoencoder and node device label is generated. The node device label and state feature set are multi-dimensionally concatenated by fusion convolutional layer to output multi-dimensional feature map. A fault prediction model based on the Transformer model is pre-built, modeling samples are exported from the policy database, and the fault prediction model is iteratively trained using the modeling samples until convergence. The system loads multi-dimensional feature maps in real time, performs multi-dimensional evaluation of the feature maps based on the fault prediction model, and outputs the evaluation results of the node devices. Obtain the evaluation results of the node devices, traverse the policy database based on the evaluation results, match the evaluation results of the node devices with the policy database, and output the policy matching set.

[0006] On the other hand, the present invention also provides a fault prediction system for combined cycle power generation systems, the system comprising: The parameter enhancement module is used to acquire the real-time status parameters of the monitoring nodes in the cyclic electronic system, perform normalization enhancement processing on the real-time status parameters to obtain status enhancement information, and upload the status enhancement information to the strategy database; wherein, monitoring nodes are set up for key locations of each module and / or device in the cyclic electronic system and data is collected. The feature fusion module extracts state features from the state enhancement information based on the convolutional neural network to obtain a state feature set. It then uses an autoencoder to identify and learn the potential relationship between the state feature set and the corresponding node device label and assists in generating node device labels. Finally, it uses a fusion convolutional layer to perform multi-dimensional splicing of the node device label and the state feature set to output a multi-dimensional feature map. The equipment evaluation module is used to pre-build a fault prediction model based on the Transformer model, export modeling samples from the policy database, iteratively train the fault prediction model to convergence using the modeling samples, load multi-dimensional feature maps in real time, perform multi-dimensional evaluation on the multi-dimensional feature maps based on the fault prediction model, and output the node equipment evaluation results. The maintenance strategy generation module is used to obtain the evaluation results of node devices, traverse the strategy database based on the evaluation results of node devices, match the evaluation results of node devices with the strategy database, and output the strategy matching set.

[0007] Compared with the prior art, the embodiments of this application have the following main advantages: In this embodiment of the invention, a fault prediction model based on the Transformer model is used to evaluate the node equipment. The Transformer has powerful parallel computing capabilities and the ability to model long sequences and multi-dimensional features. It can better handle complex multi-dimensional data in combined cycle power generation systems and comprehensively considers multiple factors such as fault probability distribution and overall severity. It can more accurately assess the fault risk of the equipment, thereby achieving a comprehensive description of the equipment's operating status.

[0008] In this embodiment of the invention, when normalizing and enhancing real-time state parameters, outlier cleaning is performed based on the K-nearest neighbor method combined with local anomaly factors. This effectively identifies and removes these outliers from the dataset, thereby improving data purity and reliability and avoiding the adverse effects of outliers on subsequent processing. Furthermore, SG filtering of the adaptive subsequence effectively smooths the data, removes high-frequency noise, and retains the main features and trends. Finally, Chebyshev chaotic mapping is used to calculate the mutual information between the adaptive filter set and the static data of node devices in the policy database. This process filters the static data of node devices, identifying the most relevant static data to the real-time state parameters and eliminating data that is irrelevant or weakly correlated with the current state parameters. This reduces redundant information in the data and improves the feature relevance of the data.

[0009] In this embodiment of the invention, when cleaning outliers in real-time state parameters based on the K-nearest neighbor method combined with local anomaly factors, a two-stage cleaning process is adopted, which uses KNN distance threshold for initial screening and LOF coefficient of variation for in-depth verification. This process balances the sensitivity and reliability of anomaly detection, while avoiding misjudging normal fluctuations under complex operating conditions as anomalies. This achieves accurate identification and rational processing of outliers in the real-time state parameters of the combined cycle power generation system.

[0010] In this embodiment of the invention, state enhancement information is output by combining the results of static data filtering from node devices with an adaptive filter set. The adaptive filter set, after parameter weight optimization, highlights key features, while the filtered static data from node devices has a strong correlation with real-time state parameters. Combining the two allows for the effective fusion of the dynamic characteristics of real-time state parameters with the historical characteristics of the static data from node devices, generating more comprehensive and representative state enhancement information.

[0011] In this embodiment of the invention, a Transformer model is used to linearly transform the channel feature representation, and the failure probability distribution of the node device is calculated by combining the dynamic weights of the channel features. This failure probability distribution calculation method based on multi-dimensional features and dynamic weight allocation can more accurately assess the probability of different failure types occurring in the device. Furthermore, the AdaBoost model is used to weight and fuse the overall severity and failure probability distribution, and a strong classifier obtained through iterative combination is used to determine the typical failure types of the node. This allows for more precise identification of the main failure types of the device, providing clearer guidance for subsequent maintenance strategies. Attached Figure Description

[0012] Figure 1 This is a schematic diagram illustrating the implementation process of the fault prediction method for combined cycle power generation systems provided by the present invention.

[0013] Figure 2 A schematic diagram of the implementation process of the normalization enhancement method for real-time status parameters is shown.

[0014] Figure 3 A schematic diagram of the implementation process of an outlier cleaning method for real-time state parameters based on the K-nearest neighbor method combined with local anomaly factors is shown.

[0015] Figure 4 A schematic diagram of the implementation process of the method combining static data filtering results from node devices with adaptive filter sets is shown.

[0016] Figure 5 A schematic diagram of the implementation process of a multi-dimensional evaluation method based on a fault prediction model for multi-dimensional feature maps is shown.

[0017] Figure 6 A schematic diagram of the process for matching node device evaluation results with the policy database is shown.

[0018] Figure 7 A schematic diagram of the fault prediction system for combined cycle power generation systems is shown. Detailed Implementation

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings of this application are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings of this application are used to distinguish different objects, not to describe a particular order.

[0020] This disclosure proposes a fault prediction method and system for combined cycle power generation systems. The fault prediction model based on the Transformer model is used to evaluate node equipment. The Transformer has powerful parallel computing capabilities and the ability to model long sequences and multi-dimensional features, which can better handle complex multi-dimensional data in combined cycle power generation systems. It also comprehensively considers multiple factors such as fault probability distribution and overall severity, and can more accurately assess the fault risk of equipment, thereby achieving a comprehensive description of the operating status of the equipment.

[0021] This invention provides a fault prediction method for combined cycle power generation systems. Figure 1 This diagram illustrates the implementation flow of a fault prediction method for combined cycle power generation systems. The fault prediction method for combined cycle power generation systems specifically includes: S10: Obtain the real-time status parameters of the monitoring nodes in the combined cycle power generation system, perform normalization enhancement processing on the real-time status parameters to obtain status enhancement information, and upload the status enhancement information to the strategy database; wherein, monitoring nodes are set up for key locations of each module and / or device in the combined cycle power generation system and data is collected; In this embodiment of the invention, the real-time status parameters are first normalized and enhanced to solve the problems of inconsistent dimensions, noise interference, and outlier effects in multi-source sensor data in the combined cycle power generation system, thereby ensuring the reliability of the input data. Furthermore, the use of adaptive window filtering and mutual information screening enhances the ability to extract key features, laying a high-quality data foundation for subsequent feature analysis.

[0022] S20: Based on a convolutional neural network, state features are extracted from the state enhancement information to obtain a state feature set. An autoencoder learns the potential association between the state feature set and the corresponding node device labels to assist in label generation. A fusion convolutional layer is used to concatenate the node device labels and the state feature set in multiple dimensions, outputting a multi-dimensional feature map. Specifically, the state feature set is extracted using a convolutional neural network (CNN), combined with the node device labels generated with the assistance of the autoencoder, and then the fusion convolutional layer concatenates the device labels and state features in multiple dimensions to finally generate a multi-dimensional feature map. This fusion method not only preserves the detailed information of the original data but also enhances the expressive power of the features through cross-dimensional associations (such as the coupling between device type and real-time status), enabling the model to more comprehensively understand the device's operating status. S30: Pre-build a fault prediction model based on the Transformer model, export modeling samples from the policy database, and use the modeling samples to iteratively train the fault prediction model until convergence; S40 loads multi-dimensional feature maps in real time, performs multi-dimensional evaluation of the multi-dimensional feature maps based on the fault prediction model, and outputs the evaluation results of node devices. S50: Obtain the node device evaluation results, traverse the policy database based on the node device evaluation results, perform matching processing between the node device evaluation results and the policy database, and output the policy matching set.

[0023] In the embodiments of the present disclosure, key positions and / or key devices can be determined for each sub-module / device in a power generation system, including but not limited to gas turbines, steam turbines, intake air circulation systems, fuel supply systems, cooling systems, lubrication systems, etc. Data acquisition points are set at the key positions and key devices (i.e., monitoring nodes) in the cyclic power generation subsystem for real-time monitoring of the operating status and performance parameters of the cyclic power generation subsystem. Generally speaking, the monitoring nodes can include key components such as compressors, combustion chambers, turbines, etc., and can also include intake ports, exhaust ports, etc. Sensors matching the nodes can be set at the corresponding monitoring nodes, so as to monitor the real-time status parameters of the monitoring points.

[0024] Taking the compressor as an example of the monitoring node, the monitoring nodes corresponding to the compressor can include: the compressor intake port, each stage of the compressor impeller, and the compressor exhaust port. Then the real-time status parameters collected by the monitoring nodes corresponding to the compressor can include: Temperature (temperature sensor): The intake port temperature reflects the initial temperature of the air entering the compressor, the temperature of each stage of the impeller reflects the temperature increase of the air due to the compression work during the compression process, and the exhaust port temperature reflects the final temperature of the compressed air; Pressure (pressure sensor): The intake port pressure is the initial pressure of the outside air entering the compressor, the pressure change between each stage of the impeller reflects the compression ratio, and the exhaust port pressure affects the intake conditions of the subsequent combustion chamber. Abnormal pressure may cause problems such as compressor surge; Flow rate (flow sensor): The intake flow rate determines the operating load of the compressor, and the change in the flow rate will affect the output power and efficiency of the compressor.

[0025] In the embodiments of the present disclosure, the policy database stores multi-dimensional information related to equipment fault maintenance, aiming to provide comprehensive and accurate data support for formulating scientific and reasonable maintenance strategies when equipment abnormalities occur. Including but not limited to: typical fault type information, fault maintenance strategy information, maintenance effect information, strategy-fault mapping relationship, system configuration and parameters, user permission information.

[0026] Furthermore, in step S20, state feature extraction can be performed on the state enhancement information using a convolutional neural network (e.g., a 1D convolutional layer) to obtain local salient features (e.g., signal peaks, frequency abrupt changes, etc.). The autoencoder can be a stacked autoencoder, composed of multiple stacked autoencoders, capable of extracting deeper features from the data layer by layer. Training can be conducted using an unsupervised training method. In combined cycle power generation systems, obtaining a large amount of labeled data with accurate node device labels often requires significant manpower, resources, and time. An autoencoder is an unsupervised learning model that does not require data labeling; it only needs to be trained using an unlabeled set of state features. Through the autoencoder's self-learning and feature extraction of unlabeled data, the potential structure and patterns in the data can be discovered. In this invention, the autoencoder, through the nonlinear transformation of a multi-layer neural network, can deeply mine the intrinsic structure and patterns of the data, extracting features closely related to the node device labels, thereby more accurately identifying the state features of different devices and outputting the encoded deep features. A classifier can be trained using a small amount of labeled data with real labels. For example, a Softmax layer can be added after the encoded features of an autoencoder to map deep features to node device labels (e.g., normal, minor fault, serious fault, or specific fault types such as bearing wear, combustion chamber carbon buildup, etc.). Because the labels are based on higher quality features, the accuracy is higher than that of a classifier trained directly with CNN features.

[0027] In the step of using a fusion convolutional layer to concatenate device labels and state features in multiple dimensions, discrete node device labels (e.g., normal, fault A, fault B) can be transformed into numerical vectors (e.g., one-hot encoding: normal → [1,0,0], fault A → [0,1,0]), enabling mathematical fusion with the state feature set. The fusion convolutional layer is input with the state feature set output by the CNN (e.g., a 256-dimensional vector) and the encoded label vector (e.g., 3-dimensional). A 1D convolutional layer concatenates the two types of vectors in the channel dimension (e.g., 256-dimensional + 3-dimensional → 259-dimensional). Then, feature interaction is performed through convolutional kernels (e.g., 1×1 convolution) to generate cross-dimensional related features (e.g., a coupling mode of "fault A label + vibration feature"), outputting a multi-dimensional feature map (e.g., a feature vector of shape (1,128) or a feature matrix of shape (16,16), which contains both the detailed features of the original data and the label information of the device state.

[0028] In this embodiment of the invention, a fault prediction model based on the Transformer model is used to evaluate node devices. The Transformer model possesses powerful parallel computing capabilities and the ability to model long sequences and multi-dimensional features, enabling it to better handle complex multi-dimensional data in combined cycle power generation systems. Furthermore, it comprehensively considers multiple factors such as fault probability distribution and overall severity, thus more accurately assessing the fault risk of the equipment and achieving a comprehensive description of the equipment's operating status. Here, the modeling samples in the Transformer-based fault prediction system refer to the dataset exported from the policy database used to train the fault prediction model. These samples typically contain equipment operating status data, fault labels, and related feature information.

[0029] This invention provides a method for normalizing and enhancing real-time state parameters. Figure 2 This diagram illustrates the implementation flow of a method for normalizing and enhancing real-time state parameters. The method specifically includes: S101, load real-time state parameters, normalize the real-time state parameters using the maximum-minimum method, clean outliers of the real-time state parameters based on the K-nearest neighbor method combined with local anomaly factors, and output the real-time state parameters after outlier cleaning. In this embodiment of the invention, real-time status parameters include, but are not limited to, sensor data of different types such as temperature, pressure, vibration, and flow rate. These data vary greatly in dimensions and value ranges. This embodiment employs a minimization method to normalize the real-time status parameters, mapping all parameters to a specific interval and eliminating differences in dimensions and value ranges between different parameters. The K-nearest neighbor method determines whether a data point is an outlier by calculating the distance between each data point and its K nearest neighbors, while the local anomaly factor further measures the degree of anomaly of a data point within its local neighborhood. This accurately identifies outlier data points that significantly deviate from normal data patterns and removes them from the dataset, thereby improving data purity and reliability and avoiding the adverse effects of outliers on subsequent processing.

[0030] S102, obtain the real-time status parameters after outlier cleaning, extract an adaptive length window from the real-time status parameters, generate an adaptive subsequence based on the preset step size sliding window, perform SG filtering (Savitzky-Golay) on the adaptive subsequence, and output the adaptive filter set. In this step, the window length and step size can be set according to the data volatility. For example, when performing spectrum analysis, if the data signal fluctuates significantly, the window length can be set shorter to reduce smoothing and retain more details; if the data fluctuation is small, the window length can be set longer. In practice, if the fluctuation is large, the window length can be set to 3 seconds; if the fluctuation is small, it can be 10-15 seconds.

[0031] S103: Load the adaptive filter set, traverse the policy database, and use Chebyshev chaotic mapping to calculate the mutual information between the adaptive filter set and the static data of node devices in the policy database. Filter the static data of node devices, and combine the filtered results with the adaptive filter set to output state enhancement information. This state enhancement information not only includes the current operating state parameters of the device but also incorporates static features related to the device, providing a more comprehensive description of the device's operating status and characteristics. In this way, richer and more valuable data input is provided for subsequent fault prediction models, helping the models to learn and understand the device's operating modes more accurately, thus improving the accuracy and reliability of fault prediction.

[0032] In this step, the static data of the node equipment is related to the inherent characteristics of the equipment itself and does not change in real time with the equipment's operating time or operating status. It reflects various information determined during the design, manufacturing, installation, and initial configuration phases of the equipment and is a basic characteristic description of the equipment, which may include: equipment model, specifications, serial number, equipment dimensions, material properties, optimized maintenance strategies, etc.

[0033] In this embodiment of the invention, when normalizing and enhancing real-time state parameters, outlier cleaning is performed based on the K-nearest neighbor method combined with local anomaly factors. This effectively identifies and removes these outliers from the dataset, thereby improving data purity and reliability and avoiding the adverse effects of outliers on subsequent processing. Furthermore, SG filtering of the adaptive subsequence effectively smooths the data, removes high-frequency noise, and retains the main features and trends. Finally, Chebyshev chaotic mapping is used to calculate the mutual information between the adaptive filter set and the static data of node devices in the policy database. This process filters the static data of node devices, identifying the most relevant static data to the real-time state parameters and eliminating data that is irrelevant or weakly correlated with the current state parameters. This reduces redundant information in the data and improves the feature relevance of the data.

[0034] This invention provides a method for outlier cleaning of real-time state parameters based on the K-nearest neighbor method combined with local anomaly factors. Figure 3This diagram illustrates the implementation flow of a method for outlier cleaning of real-time state parameters based on the K-nearest neighbor method combined with local anomaly factors. The method specifically includes: S1011: Load real-time state parameters, calculate the average distance between the real-time state parameter data point and the K nearest neighbors in the historical state parameters in the strategy database, and determine the distance threshold based on the parameter weights of the real-time state parameters. It should be noted that the K-Nearest Neighbors (KNN) method calculates the average distance between the real-time state parameter data point and the K nearest neighbors in the historical state parameters in the strategy database, using the local proximity relationship of the data point in the feature space to determine its degree of anomaly. In combined cycle power generation systems, the state parameters of normally operating equipment usually exhibit clustering, meaning that parameter values ​​under similar operating conditions are close to each other; while outliers are often far from normal data clusters. By calculating the average distance between real-time data points and historical normal data, data points deviating from the normal pattern can be effectively captured—if the average distance of a data point is significantly greater than the proximity distance of normal data, it is likely an outlier. This distance-based judgment method is intuitive and conforms to the actual distribution characteristics of industrial data, and is particularly suitable for anomaly detection of multi-dimensional parameters (such as temperature, pressure, vibration, etc.).

[0035] The distance threshold calculation formula is expressed as follows:

[0036]

[0037]

[0038] in, Indicates the distance threshold. For weighted average distance, Indicates the current data point With the K nearest neighbors The weighted distance, B is the index of the parameter type corresponding to the data point, and B is the number of parameter types. Indicates parameter weights; S1012, Determine whether the average distance between real-time status parameter data points is less than a preset distance threshold; S1013, If the average distance of the real-time status parameter data points is less than the preset distance threshold, retain the real-time status parameter data points; S1014, If the average distance between real-time status parameter data points is not less than a preset distance threshold, the corresponding data point is determined to be abnormal, and abnormal data points in the real-time status parameters are identified. S1015, Obtain abnormal data points in the real-time status parameters, construct an anomaly evaluation matrix based on the historical status parameters corresponding to the abnormal data points, and determine the data variation coefficient of the abnormal data points based on the anomaly evaluation matrix. In this step, the anomaly assessment matrix is ​​used to evaluate the characteristics of anomalous data points and supports the calculation of Local Outlier Factor (LOF), thereby more accurately identifying local outlier factors. The elements in the matrix may include: historical state parameters of anomalous data points, local density information of historical data points, distance information between each historical data point and the current anomalous data point, label information of historical data points, and the data structure of data variation coefficient and label information.

[0039] S1016, Determine whether the coefficient of variation of abnormal data points exceeds the preset variation threshold; S1017, if the coefficient of variation of an abnormal data point exceeds a preset variation threshold, the abnormal data point is determined to be a local outlier. It should be noted that this embodiment introduces the Local Outlier Factor (LOF) to calculate its coefficient of variation. LOF quantifies the degree of anomalousness of a target data point relative to its local environment by comparing the local density of the target data point with that of its neighboring regions. Even if its global distance is not extremely far off, it may still be identified as a local outlier. This local density analysis overcomes the limitation of KNN relying solely on global distance, and can more accurately capture hidden outliers.

[0040] S1018, If the coefficient of variation of the abnormal data point does not exceed the preset variation threshold, retain the current abnormal data point and merge the current abnormal data point with the retained real-time status parameter data point; S1019 outputs the real-time status parameters after outlier cleaning.

[0041] In this embodiment of the invention, when cleaning outliers in real-time state parameters based on the K-nearest neighbor method combined with local anomaly factors, a two-stage cleaning process is adopted, which uses KNN distance threshold for initial screening and LOF coefficient of variation for in-depth verification. This process balances the sensitivity and reliability of anomaly detection, while avoiding misjudging normal fluctuations under complex operating conditions as anomalies. This achieves accurate identification and rational processing of outliers in the real-time state parameters of the combined cycle power generation system.

[0042] This invention provides a method for combining static data filtering results from node devices with an adaptive filter set. Figure 4 This diagram illustrates the implementation process of a method combining static data filtering results from node devices with an adaptive filter set. Specifically, this method includes: S1031, based on a Chebyshev chaotic mapping network, determines the parameter weight update probability of real-time state parameters in the adaptive filtering set. Combining the weight update probability, the parameter weights of the real-time state parameters are optimized, and the optimized weights are output. The Chebyshev chaotic mapping network possesses high randomness and ergodicity, capable of generating uniformly distributed chaotic sequences. Determining the parameter weight update probability of real-time state parameters in the adaptive filtering set using this network can fully consider the complex nonlinear relationships and dynamic changes between parameters. Compared with traditional fixed-weight or simple random assignment methods, it can more accurately reflect the degree of influence of each real-time state parameter on the equipment's operating state under the current conditions. In this step, when the dimensionality of the real-time state parameters is not very high and the nonlinear relationships between the data are not complex, the initial order of the Chebyshev chaotic mapping network can be set to 2 or 4, and the initial value can be a random value within an interval. Here, "network" can be understood as a computational module built based on the Chebyshev chaotic mapping, which can take the real-time state parameters of the combined cycle power generation system (such as real-time sampled values ​​of temperature, pressure, and vibration) and the current adaptive filtering parameter weights (such as the noise reduction weights of each parameter) as input. The Chebyshev chaotic mapping module can iteratively generate a chaotic sequence (typically in the range of [-1,1] or [0,1]) based on the fluctuation of the real-time parameters (e.g., a sudden fluctuation of a parameter beyond the norm may be noise or a change in the actual state). The chaotic sequence is combined with the degree of fluctuation of the real-time parameters (e.g., the greater the fluctuation, the more the value of the corresponding chaotic sequence is biased towards the high update probability interval), and the update probability of each parameter weight is calculated (the higher the probability value, the less suitable the current weight of the parameter is for the current data, and the more it needs adjustment). The update probability of the weight corresponding to each real-time state parameter is output, and the weights are then dynamically adjusted according to the probability. For example, for parameters with a high update probability, the adjustment range can be increased to allow the weights to quickly adapt to data changes; for parameters with a low update probability, the adjustment can be small or not adjusted at all to avoid frequent changes that could lead to filter instability. Ultimately, optimized weights that fit the current real-time state parameters are obtained.

[0043] S1032, Obtain the real-time state parameters after parameter weight optimization. Based on the optimized parameter weights, determine the search space, upper boundary, and lower boundary of the node device static data in the policy database. The optimized parameter weights reflect the importance and characteristic information of the real-time state parameters. According to these weights, the range of node device static data related to the current real-time state parameters can be more accurately defined. This search space determined by weights can focus on the static data most relevant to the real-time state parameters, avoiding a non-targeted full search in the entire policy database and greatly narrowing the search range. In this step, the search space is jointly determined by the parameter optimization weights and the real-time state parameters, and is used to filter the static data of node devices related to the real-time state parameters. The search space is first determined by weighting the real-time state parameters according to the parameter optimization weights to obtain a comprehensive real-time state feature value. Then, based on this, the search space is determined according to the relationship between the static data of node devices in the policy database and the weighted real-time state feature value. For example, samples with weighted feature values ​​of the static data of node devices in the policy database within a certain range can be selected as the search space. The lower boundary of the search space is the minimum value of the search space range. In one example, if the lower boundary is 100.32, this means that samples with weighted feature values ​​of the static data of node devices in the policy database less than 100.32 will be excluded from the search space. The upper boundary of the search space is the maximum value of the search space range. In the above example, the upper boundary is 150.32. Samples with weighted feature values ​​of the static data of node devices in the policy database greater than 150.32 will also be excluded from the search space.

[0044] S1033, calculate the mutual information between real-time status parameters and static data of node devices in the search space, and filter the static data of node devices based on the correlation threshold. In this embodiment of the invention, filtering the static data of node devices based on the correlation threshold can remove static data with weak correlation to real-time status parameters and retain only data with strong correlation to real-time status parameters. This ensures that the filtered static data of node devices is closely related to real-time status parameters, and can provide more valuable information for fault prediction. Through this precise filtering, the correlation of data is enhanced, so that the subsequently generated status enhancement information can more accurately reflect the actual operating status of the equipment. S1034 outputs state enhancement information by combining the static data filtering results of node devices with the adaptive filter set.

[0045] In this embodiment of the invention, state enhancement information is output by combining the results of static data filtering from node devices with an adaptive filter set. The adaptive filter set, after parameter weight optimization, highlights key features, and the filtered static data from node devices has a strong correlation with real-time state parameters. Combining the two effectively integrates the dynamic characteristics of real-time state parameters with the historical characteristics of static data from node devices, generating more comprehensive and representative state enhancement information. The "real-time dynamic key features" of the adaptive filter set compensate for the "insufficient timeliness" of static data, while the "historical correlation features" of static data strengthen the "physical meaning anchoring" of dynamic data, ultimately generating state enhancement information that balances "real-time" and "comprehensiveness."

[0046] During implementation, data association and alignment can be performed first to ensure that the information being fused is from the same object and within the same time period. Spatially, matching can be done by monitoring node ID: the static data filtering results of a certain node (such as a gas turbine) are bound to the adaptive filter set of that node to avoid data overlap between different nodes (e.g., static data from a gas turbine cannot be used to fuse real-time filtered data from a steam turbine). Temporally, association can be done along the time dimension: although static data is not real-time, relevant historical features need to be extracted by combining it with the time window of real-time parameters. Fusion can be performed by combining dynamic features enhanced by static data with associated static features along dimensions to form a complete feature set containing both "dynamic" and "static" information, i.e., state enhancement information.

[0047] In this embodiment of the invention, the fault prediction model is based on the Transformer model. The Transformer model possesses powerful parallel computing capabilities and the ability to model long sequences and multi-dimensional features, enabling it to better handle complex multi-dimensional data in combined cycle power generation systems. Compared to the LSTM model, the Transformer can more effectively capture long-range dependencies and complex nonlinear relationships in the data, improving the accuracy of fault prediction. The fault prediction model also includes an input layer and an output layer. A multi-scale fusion detection network is introduced between the input layer and the Transformer model. This network aggregates the feature dimensions of the multi-dimensional feature map into channel descriptors based on a compression excitation algorithm, and constructs a node dynamic relationship matrix based on a weak learner. The dynamic relationship matrix and channel descriptors are fused through an average pooling layer to generate a channel feature representation of the fused dynamic relationship. This multi-scale fusion detection network can further mine feature information in the data and improve the model's ability to perceive features at different scales.

[0048] The multi-scale fusion detection network includes a fully connected layer, a global pooling layer, an average pooling layer, and a weak learner. Based on a compression activation algorithm, the network aggregates the feature dimensions of multi-dimensional feature maps into channel descriptors. It then constructs a dynamic node relationship matrix using the weak learner. The average pooling layer fuses the dynamic relationship matrix and channel descriptors to generate a channel feature representation with fused dynamic relationships. A learnable projection matrix is ​​constructed using the weak learner, and this learnable projection matrix is ​​mapped to the channel feature representation, achieving dynamic weight allocation for channel features. The multi-scale fusion detection network incorporates a compression activation algorithm, and a device evaluation layer is placed between the Transformer model and the output layer. The AdaBoost model incorporates a generative adversarial network (GAN) in its device evaluation layer. The Transformer model performs a linear transformation on the channel feature representations and calculates the fault probability distribution of node devices by combining the dynamic weights of the channel features. The device evaluation layer determines the overall severity based on the fault probability distribution of node devices and converts the overall severity and fault probability distribution into feature vectors that can be processed by the AdaBoost model. The AdaBoost model weights and fuses the overall severity and fault probability distribution, and determines the typical fault types of nodes through a strong classifier obtained by iterative combination. The overall severity, fault probability distribution, and typical fault types of nodes are integrated into the node device evaluation result.

[0049] In one possible implementation, for the input layer and multi-scale fusion detection network, the core objective is to perform multi-scale feature aggregation, dynamic relationship modeling, and channel weight allocation on the input multi-dimensional feature map to generate a more discriminative channel feature representation. First, the processing of the input layer. Multi-dimensional feature maps can be input into the input layer.

[0050] Second, channel descriptors are generated based on a compression activation algorithm. Global average pooling compresses spatial dimensionality: Global average pooling is performed on each channel of the multi-dimensional feature map (e.g., temperature channel, pressure channel), compressing the temporal / spatial dimension (T) of each channel into a single scalar, resulting in a feature vector of shape [B, C] (each element corresponds to the global statistics of one channel). Fully connected layers learn channel dependencies: The [B, C] vector is input into two fully connected layers (with ReLU activation in between; the number of channels is first reduced to C / r, then increased to C, where r is the compression ratio, e.g., r=16), learning the non-linear dependencies between different channels (e.g., the correlation between the temperature channel and the pressure channel). Output: Channel descriptors of shape [B, C], where each element represents a preliminary importance score for the corresponding channel.

[0051] Third, construct a dynamic relationship matrix for nodes based on weak learners. Shallow decision trees (such as CART trees) or logistic regression can be used as weak learners because they are computationally efficient and suitable for capturing local feature relationships. Dynamic relationship modeling: Input the original channel features ([B,T,C]) of the multi-dimensional feature map, group them by node (such as gas turbine, steam turbine), and each group contains all monitoring channels of that node (such as the 5 key channels of a gas turbine). For the channel features within each group, use a weak learner to learn the correlation weights between channel i and channel j (such as the influence weight of temperature change on pressure change), generating a dynamic relationship matrix of shape [C, C] (matrix element (i,j) represents the real-time correlation strength between channel i and j). Compared with a fixed weight matrix, weak learners can adapt to changes in data distribution (such as changes in the relationship between channels when equipment load changes), which is more in line with the dynamic characteristics of industrial systems.

[0052] Fourth, the dynamic relationship matrix and channel descriptors are fused. Matrix and vector fusion: The dynamic relationship matrix [C,C] and the channel descriptors [B,C] are fused through an average pooling layer: First, the matrix is ​​averaged row-wise to obtain a vector of [C,1], then multiplied element-wise with the channel descriptors to obtain an intermediate feature [B,C] that integrates channel importance and inter-channel relationships. Dynamic weight allocation: A learnable projection matrix (obtained through iterative optimization by a weak learner, with a shape of [C,C]) is introduced to map the intermediate features to channel weight vectors of [B,C] (using Sigmoid activation to ensure weights are between 0 and 1). The weight vector is then multiplied element-wise with the channel features of the original multi-dimensional feature map to obtain the channel feature representation [B,T,C] that integrates dynamic relationships. At this point, the features of important channels are enhanced, while those of secondary channels are suppressed.

[0053] For the Transformer model, the core objective is to perform long-sequence modeling on the fused channel feature representation, capturing the long-range dependencies of multi-dimensional features (e.g., the correlation between temperature fluctuations 3 hours ago and the current fault), and outputting a fault probability distribution. The model takes the channel feature representation [B,T,C] of the fused dynamic relationship output by the multi-scale fusion detection network as input. Through a multi-head attention mechanism, the features are split according to the number of heads, and the attention weights of different subspaces are calculated to capture the correlation between different time steps and different channels. A feedforward network is used to perform non-linear transformations on the features at each time step, strengthening the local feature expression. The output feature sequence undergoes a linear transformation to map the number of channels to the number of fault types, resulting in a temporary output of shape [B,T,K] (where K is the number of fault types). Combining the channel feature dynamic weights [B, C] output by the multi-scale network, the temporary output is weighted and summed over time steps to obtain the final node device fault probability distribution [B, K], where each element represents the probability of that fault type (e.g., bearing wear probability 0.85, normal probability 0.15).

[0054] For the equipment evaluation layer, its core objective is to comprehensively assess the severity and typical types of failures based on the failure probability distribution, and output the final node equipment evaluation results. First, Generative Adversarial Networks (GANs) assist in severity assessment. Input: The failure probability distribution [B, K] output by the Transformer. Through adversarial training between the generator and discriminator in the GAN, the potential mapping relationship between the failure probability distribution and severity is learned (addressing the problem of insufficient severity labels in industrial scenarios). Output: The generator generates a comprehensive severity level (e.g., an integer from 0 to 10, where 0 represents normal and 10 represents fatal failure). The discriminator optimizes the generator using real severity labels (a small amount of labeled data) to ensure the reasonableness of the severity assessment.

[0055] Second, the AdamBoost model identifies typical fault types. The comprehensive severity (scalar) and fault probability distribution [B,K] are concatenated into a feature vector [B,K+1] that AdaBoost can process. Initial weak classifiers: Multiple decision trees are used, each tree outputting fault type predictions based on a different feature subset (e.g., one tree focuses on wear probability + severity, another on carbon buildup probability + severity). Weights are adjusted based on the error rate of the previous weak classifier (classifiers with lower error rates have higher weights), and through multiple iterations, the weak classifiers are combined into a strong classifier, outputting the typical fault type of the node (e.g., carbon buildup in the gas turbine combustion chamber). The comprehensive severity (e.g., 7), fault probability distribution (e.g., [0.15, 0.85, 0,...]), and typical fault type of the node (e.g., bearing wear) are integrated into a structured node equipment evaluation result, including textual descriptions and numerical indicators (for subsequent strategy matching).

[0056] In this embodiment of the invention, a Transformer model is used to linearly transform the channel feature representation, and the failure probability distribution of the node device is calculated by combining the dynamic weights of the channel features. This failure probability distribution calculation method based on multi-dimensional features and dynamic weight allocation can more accurately assess the probability of different failure types occurring in the device. Furthermore, the AdaBoost model is used to weight and fuse the overall severity and failure probability distribution, and a strong classifier obtained through iterative combination is used to determine the typical failure types of the node. This allows for more precise identification of the main failure types of the device, providing clearer guidance for subsequent maintenance strategies.

[0057] The training methods for the fault prediction model include: S201. Export modeling samples from the policy database, perform normalization enhancement processing on the modeling samples to obtain an enhanced sample set, and divide the enhanced sample set into a training set and a test set. Normalization enhancement processing can improve the quality of sample data, and transfer learning can utilize existing knowledge and experience to accelerate the training speed of the model and improve the generalization ability of the model. The ratio of the training set to the test set can be 5:1. S202, Load the pre-built fault prediction model, and preset the training rounds, hyperparameters, loss function and activation function of the fault prediction model; S203. Obtain the training set. The training set uses a transfer learning strategy to iteratively train the fault prediction model. During the model training process, the model hyperparameters are tracked in real time to evaluate the model training quality. The model hyperparameters are tuned in real time based on the preset hyperparameter optimization strategy. After training is completed, a converged fault prediction model is output. Transfer learning can utilize existing knowledge and experience to accelerate the training speed of the model and improve the generalization ability of the model. S204, Obtain the test set, use the test set as input, execute the fault prediction model, and output the test results; S205, determine whether the test difference between the test result and the actual result exceeds the preset test threshold; S206 If the difference between the test result and the actual result does not exceed the preset test threshold, output the converged fault prediction model.

[0058] If the difference between the test result and the actual result exceeds the preset test threshold, return to S203 and continue iterative training of the fault prediction model.

[0059] This invention provides a method for multi-dimensional evaluation of multi-dimensional feature maps based on a fault prediction model. Figure 5 This diagram illustrates the implementation flow of a method for multi-dimensional evaluation of multi-dimensional feature maps based on a fault prediction model. The method specifically includes: S301: Load multi-dimensional feature maps, aggregate the feature dimensions of the multi-dimensional feature maps into channel descriptors based on the compression activation algorithm, and construct a node dynamic relationship matrix based on a weak learner. The dynamic relationship matrix and channel descriptors are fused through an average pooling layer to generate a channel feature representation that fuses dynamic relationships. S302, combined with a weak learner, constructs a learnable projection matrix, maps the learnable projection matrix to the channel feature representation, and realizes dynamic weight allocation of channel features; S303 performs a linear transformation on the channel feature representation and calculates the failure probability distribution of the node device by combining the dynamic weights of the channel features. The linear transformation allows for further processing and transformation of the features to meet the needs of failure probability calculation. Combining the dynamic weights of the channel features enables a more reasonable calculation of the failure probability based on the importance of the features. S304, determine the overall severity based on the failure probability distribution of node devices, and convert the overall severity and failure probability distribution into feature vectors that can be processed by the AdaBoost model; The formula for calculating the failure probability distribution of node devices is as follows:

[0060]

[0061]

[0062] in, This represents the failure probability distribution of the node devices. Typical fault types The channel feature linear transformation score, The number of typical fault types, These are probability constraint coefficients. Indicates the channel feature index. The first Eigenvalues ​​and projection matrices of each channel; The S305 AdaBoost model weightedly fuses the overall severity and failure probability distribution, and uses an iterative combination to obtain a strong classifier to determine the typical failure types of nodes. The failure probability distribution clearly provides the probability of various typical failure types occurring on node equipment, quantifying the failure risk. This allows maintenance personnel to intuitively understand the current failure risk level of the equipment, providing a basis for developing targeted maintenance strategies. Accurate calculation of failure probabilities enables the early detection of potential failure hazards, allowing for corresponding preventative measures to reduce the probability and impact of equipment failures. Furthermore, the AdaBoost model has powerful classification capabilities, continuously optimizing the classifier through iterative training to improve classification accuracy. It comprehensively considers multiple feature factors to accurately identify the typical failure types of node equipment, providing clear targets for subsequent maintenance and repair. S306 integrates the overall severity, failure probability distribution, and typical failure types of nodes into the node equipment evaluation results.

[0063] In this embodiment, the overall severity, fault probability distribution, and typical fault types of nodes are integrated into the node equipment assessment result, providing a comprehensive evaluation of the equipment status. This comprehensive assessment result not only includes the type and probability of the fault but also considers the severity of the fault, providing maintenance personnel with more detailed and accurate information to help them formulate more reasonable maintenance plans and decisions, and ensure the stable operation of the combined cycle power generation system.

[0064] This invention provides a method for matching node device evaluation results with a policy database. Figure 6 This diagram illustrates the implementation flow of a method for matching node device evaluation results with a policy database. Specifically, this method includes: S401, Obtain the node equipment assessment results, identify the comprehensive severity, fault probability distribution, and typical fault types of the node equipment in the assessment results, and based on the node typical fault type index strategy database, filter strategy records that match the fault type tags, and export at least one set of strategy records. In this embodiment of the invention, considering that different types of faults in a combined cycle power generation system (such as gas turbine blade cracks, steam turbine shaft misalignment, waste heat boiler tube wall corrosion, etc.) require different maintenance strategies, by accurately indexing fault type tags, strategies related to the current equipment fault type can be quickly located, avoiding blindly searching among many strategies, greatly improving the efficiency of finding a suitable maintenance strategy, enabling maintenance work to be directly targeted at specific faults, and enhancing the pertinence of maintenance. S402, using overall severity as a constraint, indexes the severity levels associated with policy records, filters these policy records by severity level, and outputs the filtered policy records. The overall severity level comprehensively considers the potential impact of a fault on equipment operational safety, energy efficiency, and lifespan, reflecting the severity of the fault. Severity level filtering allows for the priority allocation of maintenance resources to high-severity faults. For high-severity policy records, the system focuses on and prioritizes their processing, ensuring timely resolution of faults that may severely impact the system. For low-severity faults, maintenance time and resources can be rationally allocated based on the overall system operation and maintenance plan. This avoids resource waste, concentrating limited maintenance manpower, material resources, and financial resources where they are most needed, achieving optimal allocation of maintenance resources. S403: Calculate the cosine similarity between the strategy record and the fault probability distribution, filter strategy records that exceed the preset similarity threshold based on the preset similarity threshold, and output the strategy records as a strategy matching set.

[0065] In this embodiment of the invention, cosine similarity filtering ensures that the selected strategy records match the current equipment's failure probability distribution, meaning the strategy aligns with the equipment's current actual failure status and risk level. This allows maintenance strategies to better suit the equipment's actual situation, improving the effectiveness of maintenance measures. If the strategy does not match the failure probability distribution, the adopted maintenance measures may fail to effectively solve the current problem, or even waste resources. This filtering method provides the most suitable maintenance strategy for the combined cycle power generation system's current failure condition, improving the success rate of maintenance work.

[0066] On the other hand, embodiments of the present invention also provide a fault prediction system for combined cycle power generation systems. Figure 7 A schematic diagram of a fault prediction system for combined cycle power generation systems is shown. The fault prediction system for combined cycle power generation systems specifically includes: The parameter enhancement module 100 is used to acquire the real-time status parameters of the monitoring nodes in the cyclic electronic system, perform normalization enhancement processing on the real-time status parameters to obtain status enhancement information, and upload the status enhancement information to the strategy database; wherein, monitoring nodes are set for key locations of each module and / or device in the cyclic electronic system and data is collected. The feature fusion module 200 extracts state features from the state enhancement information based on the convolutional neural network to obtain a state feature set. It learns the potential relationship between the state feature set and the corresponding node device label through an autoencoder and assists in generating node device labels. It uses a fusion convolutional layer to perform multi-dimensional splicing of node device labels and state feature set to output a multi-dimensional feature map. The equipment evaluation module 300 is used to pre-build a fault prediction model based on the Transformer model, export modeling samples from the policy database, use the modeling samples to iteratively train the fault prediction model until convergence, load multi-dimensional feature maps in real time, perform multi-dimensional evaluation on the multi-dimensional feature maps based on the fault prediction model, and output the node equipment evaluation results. The maintenance strategy generation module 400 is used to obtain the evaluation results of node devices, traverse the strategy database based on the evaluation results of node devices, match the evaluation results of node devices with the strategy database, and output the strategy matching set.

[0067] In this embodiment, the maintenance strategy generation module 400 includes: The first matching unit 410 is used to obtain the node device evaluation results, identify the comprehensive severity, fault probability distribution and typical fault types of the node device evaluation results, index the strategy database based on the typical fault types of the node, filter the strategy records that are consistent with the fault type labels, and export at least one set of strategy records. The second matching unit 420, with the overall severity as a constraint, indexes the severity level associated with the strategy record, performs severity level filtering on the strategy record, and outputs the strategy record after severity level filtering. The similarity filtering unit 430 is used to calculate the cosine similarity between the strategy record and the fault probability distribution, filter strategy records that exceed the preset similarity threshold based on the preset similarity threshold, and output the strategy records as a strategy matching set.

[0068] In summary, this invention provides a fault prediction method and system for combined cycle power generation systems. In the embodiments of this invention, a fault prediction model based on the Transformer model is used to evaluate node equipment. The Transformer has powerful parallel computing capabilities and the ability to model long sequences and multi-dimensional features, which can better handle complex multi-dimensional data in combined cycle power generation systems. Furthermore, it comprehensively considers multiple factors such as fault probability distribution and overall severity, and can more accurately assess the fault risk of equipment, thereby achieving a comprehensive description of the operating status of the equipment.

[0069] It should be noted that, for the sake of simplicity, the foregoing embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to the present invention. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0070] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on these embodiments, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art can still combine, add, delete, or otherwise adjust the features of the various embodiments of the present invention according to the circumstances without conflict or creative effort, thereby obtaining different technical solutions that do not fundamentally depart from the concept of the present invention. These technical solutions also fall within the scope of protection of the present invention.

Claims

1. A fault prediction method for combined cycle power generation systems, characterized in that, The method includes: The system acquires real-time status parameters of monitoring nodes in the cyclic electronic system, performs normalization enhancement processing on the real-time status parameters to obtain status enhancement information, and uploads the status enhancement information to the strategy database; wherein, monitoring nodes are set up for key locations of each module and / or device in the cyclic electronic system and data is collected. State feature extraction is performed on the state enhancement information based on convolutional neural network to obtain state feature set. The potential relationship between state feature set and corresponding node device label is learned by autoencoder and node device label is generated. The node device label and state feature set are multi-dimensionally concatenated by fusion convolutional layer to output multi-dimensional feature map. A fault prediction model based on the Transformer model is pre-built, modeling samples are exported from the policy database, and the fault prediction model is iteratively trained using the modeling samples until convergence. The system loads multi-dimensional feature maps in real time, performs multi-dimensional evaluation of the feature maps based on the fault prediction model, and outputs the evaluation results of the node devices. Obtain the evaluation results of the node devices, traverse the policy database based on the evaluation results, match the evaluation results of the node devices with the policy database, and output the policy matching set.

2. The fault prediction method for combined cycle power generation systems as described in claim 1, characterized in that, The method for normalizing and enhancing real-time state parameters includes: Load real-time state parameters, normalize the real-time state parameters using the maximum-minimum method, clean outliers of the real-time state parameters based on the K-nearest neighbor method combined with local anomaly factors, and output the real-time state parameters after outlier cleaning. Obtain the real-time status parameters after outlier cleaning, extract an adaptive length window from the real-time status parameters, generate an adaptive subsequence based on the preset step size sliding window, perform SG filtering on the adaptive subsequence, and output the adaptive filter set. Load the adaptive filter set, traverse the policy database, use Chebyshev chaotic mapping to calculate the mutual information between the adaptive filter set and the static data of node devices in the policy database, filter the static data of node devices, combine the filtered static data of node devices with the adaptive filter set, and output state enhancement information.

3. The fault prediction method for combined cycle power generation systems as described in claim 2, characterized in that, The outlier cleaning of real-time state parameters based on the K-nearest neighbor method combined with local anomaly factors includes: Load real-time status parameters, calculate the average distance between the real-time status parameter data points and the K nearest neighbors in the historical status parameters in the policy database, and determine the distance threshold based on the parameter weights of the real-time status parameters; Determine whether the average distance between real-time status parameter data points is less than a preset distance threshold; If the average distance between real-time status parameter data points is less than the preset distance threshold, the real-time status parameter data points are retained. If the average distance between real-time status parameter data points is not less than the preset distance threshold, the corresponding data point is determined to be abnormal, and abnormal data points in the real-time status parameters are identified. Obtain abnormal data points from real-time status parameters, construct an anomaly evaluation matrix based on the historical status parameters corresponding to the abnormal data points, and determine the data variation coefficient of the abnormal data points based on the anomaly evaluation matrix. Determine whether the coefficient of variation of the abnormal data points exceeds the preset variation threshold. If the coefficient of variation of the abnormal data points exceeds the preset variation threshold, the abnormal data points are determined to be local anomalies. If the coefficient of variation of the abnormal data point does not exceed the preset variation threshold, the current abnormal data point is retained, and the current abnormal data point is merged with the retained real-time status parameter data point to output the real-time status parameter after the abnormal value is cleaned.

4. The fault prediction method for combined cycle power generation systems as described in claim 2, characterized in that, The step of combining the static data filtering results of node devices with the adaptive filter set includes: The parameter weight update probability of the real-time state parameters in the adaptive filter set is determined based on the Chebyshev chaotic mapping network. The parameter weights of the real-time state parameters are optimized by combining the weight update probability, and the optimized parameter weights are output. Obtain the real-time status parameters after parameter weight optimization, and determine the node device static data search space, upper boundary, and lower boundary of the search space based on the parameter optimization weight; Calculate the mutual information between real-time status parameters and static data of node devices in the search space, and filter the static data of node devices based on the correlation threshold; The status enhancement information is output by combining the static data filtering results of node devices with the adaptive filter set.

5. The fault prediction method for combined cycle power generation systems as described in claim 1, characterized in that, The fault prediction model is based on the Transformer model and includes an input layer and an output layer. A multi-scale fusion detection network is introduced between the input layer and the Transformer model. A compression excitation algorithm is introduced in the multi-scale fusion detection network. A device evaluation layer is set between the Transformer model and the output layer. The device evaluation layer includes a generative adversarial network and an AdaBoost model.

6. The fault prediction method for combined cycle power generation systems as described in claim 5, characterized in that: Fault prediction model training methods include: Export modeling samples from the policy database, perform normalization enhancement processing on the modeling samples to obtain the enhanced sample set, and divide the enhanced sample set into training set and test set; Load a pre-built fault prediction model, and preset the training rounds, hyperparameters, loss function, and activation function of the fault prediction model; Obtain the training set, use the transfer learning strategy to iteratively train the fault prediction model, track the model hyperparameters in real time during the model training process, evaluate the model training quality, and fine-tune the model hyperparameters in real time based on the preset hyperparameter optimization strategy. After training is completed, output the converged fault prediction model. Obtain the test set, use the test set as input, execute the fault prediction model, and output the test results; Determine whether the difference between the test result and the actual result exceeds the preset test threshold. If the difference between the test result and the actual result does not exceed the preset test threshold, output the converged fault prediction model.

7. The fault prediction method for combined cycle power generation systems as described in claim 5, characterized in that: The multi-dimensional evaluation of the multi-dimensional feature map based on the fault prediction model includes: Load multi-dimensional feature maps, aggregate the feature dimensions of the multi-dimensional feature maps into channel descriptors based on the compression activation algorithm, construct a node dynamic relationship matrix based on a weak learner, and fuse the dynamic relationship matrix and channel descriptors through an average pooling layer to generate a channel feature representation that fuses dynamic relationships. By combining weak learners to construct a learnable projection matrix, the learnable projection matrix is ​​mapped to the channel feature representation, thereby realizing dynamic weight allocation of channel features. A linear transformation is performed on the channel feature representation, and the failure probability distribution of the node device is calculated by combining the dynamic weights of the channel features. The overall severity is determined based on the failure probability distribution of node devices, and the overall severity and failure probability distribution are converted into feature vectors that can be processed by the AdaBoost model. The AdaBoost model weights and fuses the overall severity and failure probability distribution, and determines the typical failure type of a node by obtaining a strong classifier through iterative combination. The overall severity, failure probability distribution, and typical failure types of nodes are integrated to form the node equipment evaluation results.

8. The fault prediction method for combined cycle power generation systems as described in claim 7, characterized in that, The matching process between the node device evaluation results and the policy database includes: Obtain the node device evaluation results, identify the comprehensive severity, failure probability distribution and typical node failure types in the node device evaluation results, index the strategy database based on the typical node failure types, filter the strategy records that are consistent with the failure type labels, and export at least one set of strategy records. Based on the overall severity level as a constraint, the severity level associated with the indexing strategy records is used to filter the strategy records by severity level, and the strategy records after severity level filtering are output. Calculate the cosine similarity between the strategy record and the failure probability distribution, filter strategy records that exceed the preset similarity threshold based on the preset similarity threshold, and output the strategy records as a strategy matching set.

9. A fault prediction system for combined cycle power generation systems, used to implement the fault prediction method for combined cycle power generation systems as described in any one of claims 1-8, characterized in that, The system includes: The parameter enhancement module is used to acquire the real-time status parameters of the monitoring nodes in the cyclic electronic system, perform normalization enhancement processing on the real-time status parameters to obtain status enhancement information, and upload the status enhancement information to the strategy database; wherein, monitoring nodes are set up for key locations of each module and / or device in the cyclic electronic system and data is collected. The feature fusion module extracts state features from the state enhancement information based on the convolutional neural network to obtain a state feature set. It learns the potential relationship between the state feature set and the corresponding node device label through an autoencoder and assists in generating node device labels. It uses a fusion convolutional layer to perform multi-dimensional splicing of node device labels and state feature set to output a multi-dimensional feature map. The equipment evaluation module is used to pre-build a fault prediction model based on the Transformer model, export modeling samples from the policy database, iteratively train the fault prediction model to convergence using the modeling samples, load multi-dimensional feature maps in real time, perform multi-dimensional evaluation on the multi-dimensional feature maps based on the fault prediction model, and output the node equipment evaluation results. The maintenance strategy generation module is used to obtain the evaluation results of node devices, traverse the strategy database based on the evaluation results of node devices, match the evaluation results of node devices with the strategy database, and output the strategy matching set.

10. The fault prediction system for combined cycle power generation systems as described in claim 9, characterized in that, The maintenance strategy generation module includes: The first matching unit is used to obtain the node device evaluation results, identify the comprehensive severity, fault probability distribution and typical fault types of the node device evaluation results, index the strategy database based on the typical fault types of the node, filter the strategy records that are consistent with the fault type labels, and export at least one set of strategy records. The second matching unit, constrained by the overall severity level, indexes the severity levels associated with the policy records, filters the policy records by severity level, and outputs the policy records after severity level filtering. The similarity filtering unit is used to calculate the cosine similarity between the strategy record and the fault probability distribution, filter strategy records that exceed the preset similarity threshold based on the preset similarity threshold, and output the strategy records as a strategy matching set.

Images