Intelligent fault early warning, diagnosis and self-healing integrated system and method for thermal power generating unit

Abstract

The invention provides a thermal power generating unit intelligent fault early warning, diagnosis and self-healing integrated system and method, and relates to the technical field of thermal power generating unit fault processing, and the system comprises a data collection module which is used for collecting the operation data of a thermal power generating unit in real time; the intelligent early warning module is connected with the data acquisition module and is used for encoding and reconstructing the operation data and triggering fault diagnosis according to a comparison result of a reconstruction error and a dynamic threshold value; and the fault diagnosis module is connected with the intelligent early warning module and is used for performing feature extraction and probability fusion on the fault and outputting a fault diagnosis result when the fault diagnosis is triggered. An early warning, diagnosis and self-healing integrated system is constructed, a diagnosis result is directly mapped into a control strategy through dominant function cooperation reinforcement learning, the control strategy is automatically executed by a progressive control unit, and automation of the whole process from fault recognition to self-recovery is achieved.

Description

Technical Field

[0001] This invention relates to the field of thermal power unit fault handling technology, specifically to an integrated system and method for intelligent fault early warning, diagnosis and self-healing of thermal power units. Background Technology

[0002] Against the backdrop of dual-carbon goals and the construction of a new power system, my country is accelerating the construction of a new power system with new energy as the main body. Among them, the desert and Gobi large-scale base has become an important scenario for the large-scale development of new energy. Such bases mostly adopt a multi-energy complementary mode of wind, solar, thermal and energy storage, which puts forward higher requirements for the flexibility, reliability and intelligence level of the supporting coal-fired power units. The DCS control system of traditional thermal power units generally adopts foreign software and hardware platforms, which poses a risk of bottleneck in key technologies. At the same time, existing intelligent applications are mostly deployed outside the DCS, and the early warning and diagnosis results rely on manual intervention, which cannot form an integrated control of fault early warning, diagnosis and self-healing, making it difficult to implement the project.

[0003] While existing data-driven fault warning and diagnosis methods exist, they typically suffer from the following problems: 1. Warning models are mostly static thresholds or simple statistical models, which cannot adapt to the time-varying and nonlinear operating characteristics of the unit, easily leading to false alarms or missed alarms; 2. Fault diagnosis often relies on a single model or expert system, and diagnostic accuracy decreases when the model is mismatched or knowledge is incomplete; 3. Diagnostic results are isolated from the control system, unable to directly trigger control actions, and lack self-healing capabilities; 4. System parameters are fixed, lacking online learning and adaptive capabilities, making it difficult to cope with long-term challenges such as fuel changes and equipment degradation.

[0004] Therefore, in order to solve the above problems, the present invention provides an integrated system and method for intelligent fault early warning, diagnosis and self-healing of thermal power units. Summary of the Invention

[0005] The purpose of this invention is to provide an integrated system and method for intelligent fault early warning, diagnosis and self-healing of thermal power units, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an integrated intelligent fault early warning, diagnosis, and self-healing system for thermal power units, comprising:

[0007] The data acquisition module is used to collect real-time operating data of thermal power units;

[0008] The intelligent early warning module is connected to the data acquisition module and is used to encode and reconstruct the operating data, and trigger fault diagnosis based on the comparison result of the reconstruction error and the dynamic threshold.

[0009] The fault diagnosis module, connected to the intelligent early warning module, is used to extract features and fuse probabilities of the fault when fault diagnosis is triggered, and output the fault diagnosis result.

[0010] A fault self-healing module, connected to the fault diagnosis module, is used to generate and execute a self-healing control strategy based on the fault diagnosis results.

[0011] The parameter optimization module is connected to the intelligent early warning module, the fault diagnosis module, and the fault self-healing module, respectively, and is used to perform online learning and adaptive updates of the parameters in the system.

[0012] A preferred technical solution of this application: the intelligent early warning module includes:

[0013] A gated time-space encoder is used to process input operational data. Encode the features to obtain the encoded features. The calculation process is as follows:

[0014]

[0015] in, For Long Short-Term Memory (LSTM) network encoding functions, It is a one-dimensional convolutional coding function. The gating weight is calculated using the following formula:

[0016]

[0017] It is the Sigmoid activation function. For the gated weight matrix, This represents a vector concatenation operation;

[0018] The reconstruction error calculation unit is used to calculate the error based on the encoded features. Reconstruct the data and calculate the reconstruction error. The calculation process is as follows:

[0019]

[0020] in, For decoder functions;

[0021] The dynamic threshold calculation unit is used to calculate the adaptive warning threshold at the current moment. The calculation process is as follows:

[0022]

[0023] in, and These are the mean and standard deviation of the historical reconstruction error, respectively. , is the confidence coefficient For adaptive adjustment coefficient, The information entropy at the current moment, This represents the system's maximum theoretical entropy.

[0024] Early warning triggering unit, used to compare the reconstruction error With the dynamic threshold ,like This will trigger fault diagnosis.

[0025] A preferred technical solution of this application: the fault diagnosis module includes:

[0026] Causal feature map network units are used to extract fault features. The calculation process is as follows:

[0027]

[0028] in, It is a multilayer perceptron. It is a causal adjacency matrix;

[0029] A confidence-weighted probability fusion unit is used to predict probabilities based on the model. and expert rule probability Calculate the final failure probability The calculation process is as follows:

[0030]

[0031] in, and These represent the model credibility and expert system credibility at the current moment, respectively. This credibility is updated online using the following formula:

[0032]

[0033]

[0034] in, For credibility learning rate, and These are the accuracy evaluation values ​​of the model and the expert system at the current moment, respectively.

[0035] A preferred technical solution of this application: the fault self-healing module includes:

[0036] Advantage function collaboration Learning units, used to generate self-healing strategies, are jointly... value The calculation process is as follows:

[0037]

[0038] in, For average value, and Two intelligent agents respectively value, For the joint advantage function, For system status, and For the actions of two intelligent agents, The calculation process is as follows:

[0039]

[0040] in, and These are the advantage functions of the two agents, respectively. and For agent weights, and satisfying , To coordinate the penalty coefficient;

[0041] An adaptive progressive control unit is configured to output control commands according to the self-healing strategy. The calculation process is as follows:

[0042]

[0043] in, This is the proportionality coefficient. For systematic error, It is a saturation function. To control the time constant.

[0044] A preferred technical solution of this application: the parameter optimization module includes:

[0045] Model parameter update unit, used to update model parameters The update process is as follows:

[0046]

[0047] in, For adaptive decay learning rate, The gradient of the loss function. For memory coefficient, These are the historically optimal parameters;

[0048] The learning rate The calculation process is as follows:

[0049]

[0050] in, The initial learning rate, The attenuation coefficient is... For training steps;

[0051] Transfer learning units are used to transfer the source model Transferring to the target scenario to obtain the target model The calculation process is as follows:

[0052]

[0053] in, For the first The amount of change in the layer parameters, For the first Feature extraction function of layer The total number of network layers. For target scene data.

[0054] This application also provides an integrated method for intelligent fault early warning, diagnosis and self-healing of thermal power units, including the following steps:

[0055] S1. Real-time acquisition of operating data of thermal power units;

[0056] S2. Perform intelligent early warning processing on the running data, including encoding and reconstructing the data, calculating the reconstruction error, comparing the error with a dynamic threshold calculated based on historical data and information entropy, and triggering fault diagnosis if the threshold is exceeded.

[0057] S3. When fault diagnosis is triggered, fault diagnosis processing is performed, including extracting fault features, and weighted fusion of model prediction results and expert rule results based on online updated model credibility and expert system credibility to obtain the final fault diagnosis result.

[0058] S4. Based on the fault diagnosis results, perform fault self-healing processing, including generating a self-healing strategy by coordinating the reinforcement learning decisions of multiple agents, and executing self-healing control based on a control law containing a proportional term and an asymptotic adjustment term.

[0059] S5. Online optimization and adaptive updates of model parameters involved in the early warning, diagnosis and self-healing process, including updating parameters with adaptive decay learning rate, and adapting the model to different scenarios through hierarchical transfer learning.

[0060] A preferred technical solution of this application: In step S2, the calculation of the dynamic threshold in the intelligent early warning processing is specifically as follows: combining the mean and standard deviation of historical reconstruction errors, and introducing an adaptive adjustment term based on the information entropy of the current running data, to jointly determine the early warning threshold at the current moment.

[0061] A preferred technical solution of this application is as follows: In step S3, the method for updating the model credibility and expert system credibility in the fault diagnosis process is: dynamically adjust their credibility weights according to their respective accuracy performance in recent diagnoses with a certain learning rate.

[0062] A preferred technical solution of this application: In step S4, the method for generating a self-healing strategy in the fault self-healing process is as follows: a multi-agent cooperative reinforcement learning method based on the advantage function is adopted to integrate the value assessment and advantage function of each agent and introduce a coordination penalty mechanism to generate the optimal self-healing strategy. According to the self-healing strategy, a control method combining proportional control and saturation function compensation is adopted to generate control commands and execute them.

[0063] A preferred technical solution of this application is as follows: In step S5, in the online optimization and adaptive update of parameters, the model parameters are updated using a learning rate that decays with the time step, and fine-tuned by combining the memory of historical optimal parameters. The model transfer adopts a hierarchical parameter adjustment method to perform differentiated adaptation for different layers of the network.

[0064] Compared with the prior art, the technical solution provided by this invention has the following advantages:

[0065] 1. This invention constructs an integrated system for early warning, diagnosis, and self-healing. Through cooperative reinforcement learning of advantageous functions, the diagnostic results are directly mapped to control strategies, which are then automatically executed by the progressive control unit, realizing full automation from fault identification to autonomous recovery.

[0066] 2. This invention adopts a dynamic early warning model based on a gated spatiotemporal encoder, and combines historical error statistics with real-time information entropy to adaptively calculate the early warning threshold. It can accurately track the time-varying and nonlinear operating characteristics of the unit, significantly reduce the false alarm and missed alarm rates, and proposes a credibility-weighted probability fusion mechanism to integrate the diagnostic results of the data-driven model and the expert knowledge base online, and dynamically adjust the weights according to their real-time performance. Even when the model is mismatched or the knowledge is incomplete, it can still maintain high diagnostic accuracy.

[0067] 3. This invention has a complete online learning and transfer optimization mechanism. Through adaptive learning rate updates and hierarchical transfer strategies, the system can continuously adapt to long-term operational challenges such as fuel changes and equipment degradation, and maintain optimal performance. Attached Figure Description

[0068] Figure 1 This is a system flowchart of the present invention. Detailed Implementation

[0069] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. The present invention will be further described below with reference to the embodiments.

[0070] Please see Figure 1 The embodiments of this application provide an integrated intelligent fault early warning, diagnosis and self-healing system for thermal power units, which is deployed on an intelligent control system (ICS) based on independently controllable software and hardware, including:

[0071] The data acquisition module is used to collect real-time operating data of thermal power units;

[0072] The data acquisition module connects to the power plant's DCS / ICS, fieldbus network, and intelligent sensing devices to collect real-time operating data from the entire plant, including temperature, pressure, flow rate, current, and vibration. It also performs preprocessing, such as filtering and normalization. The data formats cover real-time data, historical data, and unstructured data, providing a high-quality data source for upper-level analysis.

[0073] The intelligent early warning module, connected to the data acquisition module, is used to encode and reconstruct the operating data, and trigger fault diagnosis based on the comparison result of the reconstruction error and the dynamic threshold.

[0074] The fault diagnosis module, connected to the intelligent early warning module, is responsible for accurately locating the fault after the early warning is triggered. It is used to extract features and perform probability fusion on the fault when fault diagnosis is triggered, and output the fault diagnosis results.

[0075] The fault self-healing module, connected to the fault diagnosis module, aims to achieve autonomous fault elimination. It is used to generate and execute self-healing control strategies based on fault diagnosis results.

[0076] The parameter optimization module is connected to the intelligent early warning module, the fault diagnosis module, and the fault self-healing module, respectively, and is used to learn and adaptively update the parameters in the system online.

[0077] In this embodiment, the intelligent early warning module includes:

[0078] A gated time-space encoder is used to process input operational data. Encoding is performed on the operational data, which includes multi-dimensional time-series data of the unit's real-time operation, such as temperature, pressure, and flow rate, to obtain coding features. The calculation process is as follows:

[0079]

[0080] in, For Long Short-Term Memory (LSTM) network encoding functions, This is the output of the LSTM encoder, used to capture long-term temporal dependencies and dynamic patterns in the data. It is a one-dimensional convolutional coding function. This is the output of a one-dimensional convolutional encoder, used to extract local spatial features and short-time patterns from the data. As gating weights, the contribution ratio of the LSTM and CNN encoding results is dynamically adjusted, ranging from... Calculated using the following formula:

[0081]

[0082] The Sigmoid activation function is used to compress the gate value to... Ranges enable soft selection. This is the gating weight matrix, used to learn how to combine the two encoded features to generate the optimal gating value. This indicates a vector concatenation operation that merges two types of encoded features to provide complete information for gating decisions.

[0083] Reconstruction error calculation unit, used to calculate errors based on coding features Reconstruct the data and calculate the reconstruction error. The calculation process is as follows:

[0084]

[0085] in, The decoder function is a mapping function that reconstructs high-dimensional data from a low-dimensional feature space, encoding features... Reverse mapping back to the original data space generates reconstructed data. is the Euclidean norm, used to calculate the Euclidean distance between the original data and the reconstructed data;

[0086] The dynamic threshold calculation unit is used to calculate the adaptive warning threshold at the current moment. , Indicates time The warning threshold is a boundary value used to dynamically determine whether a warning is triggered. Its calculation process is as follows:

[0087]

[0088] in, and These are the mean and standard deviation of the historical reconstruction error, respectively. This represents the historical reconstruction error mean, reflecting the average error level during normal system operation. The standard deviation of the historical reconstruction error is used to reflect the degree of error fluctuation and to adjust the threshold width. , where is the confidence coefficient, used to control how lenient the threshold is relative to the historical error distribution. This is an adaptive adjustment coefficient used to adjust the strength of the influence of information entropy on the threshold. The information entropy at the current moment is used to quantify the degree of disorder or uncertainty in the data distribution. This is the maximum theoretical entropy value of the system, used to normalize the information entropy so that its impact is within a reasonable range;

[0089] Early warning triggering unit, used to compare reconstruction errors With dynamic threshold ,like If this is triggered, fault diagnosis is established. This design solves the problem of poor adaptability of fixed thresholds by adaptively adjusting the threshold tightness through information entropy.

[0090] In this embodiment, the fault diagnosis module includes:

[0091] Causal feature map network units are used to extract fault features. It integrates data-driven characteristics and system causal relationships into a comprehensive representation, and its calculation process is as follows:

[0092]

[0093] in, It is a multilayer perceptron. The output of the multilayer perceptron learns and extracts deep nonlinear features from the data. This is a causal adjacency matrix used to encode physical or logical causal relationships between system devices / parameters;

[0094] A confidence-weighted probability fusion unit is used to predict probabilities based on the model. and expert rule probability , The probability of failure is calculated based on a data-driven model. The final failure probability is calculated based on the failure probability inferred from the expert knowledge base and rule engine. The confidence level of the diagnostic results, which integrates the model and expert opinions, achieves intelligent fusion of diagnostic results that dynamically adjusts with the performance of both. The calculation process is as follows:

[0095]

[0096] in, and These represent the model credibility and the expert system credibility at the current moment, respectively. This represents the model's credibility at the current moment, used to quantify the reliability of the data-driven model at that current moment. The current credibility of the expert system is used to quantify its reliability at the current moment. This credibility is updated online using the following formula:

[0097]

[0098]

[0099] in, The confidence learning rate is used to control the speed and stability of confidence updates. and These are the accuracy evaluation values ​​of the model and the expert system at the current moment, respectively. Real-time evaluation of model performance based on recent prediction performance. Real-time evaluation of the expert system based on recent rule matching performance.

[0100] In this embodiment, the fault self-healing module includes:

[0101] Advantage function collaboration Learning units, used to generate self-healing strategies, are jointly... value The calculation process is as follows:

[0102]

[0103] in, For the United Value, representing the state of the system The following joint actions were taken. Long-term expected return For average value, and Two intelligent agents respectively The values ​​represent each agent's estimate of the long-term value of its own actions. The joint advantage function quantifies the additional advantage of a joint action relative to its individual actions. For system status, and For the actions of two intelligent agents, The calculation process is as follows:

[0104]

[0105] in, and Let be the advantage functions of the two agents, representing the advantage of each agent's action relative to the average value of the state. and Assign weights to agents, allocating the importance of each agent's strengths in joint decision-making, and satisfying the following conditions: , To coordinate the penalty coefficient, penalties are imposed on two agents whose advantage differences are too large, in order to encourage coordinated and consistent decision-making.

[0106] An adaptive progressive control unit is used to output control commands based on a self-healing strategy. As the specific control signal sent to the actuator, such as a valve, pump, or motor, the calculation process is as follows:

[0107]

[0108] in, This is the proportional coefficient, used to adjust the intensity of the direct influence of the error signal on the control output. Systematic error represents the difference between the desired state and the actual state. This is a saturation function used to limit the amplitude of control commands, preventing actuator saturation or system overshoot. To control the time constant, the saturation function is used to adjust the smoothness of the error, affecting the control response speed and stability. This control law combines the speed of proportional control with the smoothing and limiting characteristics of the saturation function, avoiding large fluctuations in the actuator and achieving safe and gradual self-healing.

[0109] In this embodiment, the parameter optimization module includes:

[0110] Model parameter update unit, used to update model parameters , The update process for the model parameters at the current time, such as the weights and biases of the neural network or other models, is as follows:

[0111]

[0112] in, The adaptively decaying learning rate controls the step size of parameter updates and decreases over time. The gradient of the loss function indicates the direction in which the parameters should be adjusted to reduce model error. The memory factor is used to control the strength of the guidance of historically optimal parameters on the current update. These are the historically optimal parameters, representing the best-performing parameter combination on the validation set to date.

[0113] Learning rate The calculation process is as follows:

[0114]

[0115] in, The initial learning rate represents the size of the learning rate at the start of training. This is the decay factor, used to control the rate at which the learning rate decays with the number of training steps. The training steps are the number of iterations or time steps counted from the start of training.

[0116] Transfer learning units are used to transfer the source model Transferring to the target scene means obtaining the target model from the model already trained on the base scene. This represents the final model after adaptation to the new scenario. It involves fine-tuning the underlying general feature layer and making larger adjustments to the top-level task-specific layer, achieving fast and efficient cross-scenario adaptation. The calculation process is as follows:

[0117]

[0118] in, For the first The parameter changes of the layer, the source model The layer parameters need to be adjusted incrementally for the target scenario. For the first The feature extraction function of the layer, the source model's first layer The feature mapping or transformation function corresponding to the layer, The total number of layers in the network represents the number of layers in the source or target model. For target scenario data, it represents a sample of runtime data from a new scenario.

[0119] This embodiment provides an integrated method for intelligent fault early warning, diagnosis, and self-healing of thermal power units, including the following steps:

[0120] S1. Real-time acquisition of operating data of thermal power units;

[0121] S2. Perform intelligent early warning processing on the running data, including encoding and reconstructing the data, calculating the reconstruction error, comparing the error with a dynamic threshold calculated based on historical data and information entropy, and triggering fault diagnosis if the threshold is exceeded.

[0122] In step S2, the calculation of the dynamic threshold in the intelligent early warning processing is as follows: combining the mean and standard deviation of the historical reconstruction error, and introducing an adaptive adjustment term based on the information entropy of the current running data, to jointly determine the early warning threshold at the current moment.

[0123] S3. When fault diagnosis is triggered, fault diagnosis processing is performed, including extracting fault features, and weighted fusion of model prediction results and expert rule results based on online updated model credibility and expert system credibility to obtain the final fault diagnosis result.

[0124] In step S3, during fault diagnosis and processing, the update method for model credibility and expert system credibility is as follows: based on their respective accuracy performance in recent diagnoses, their credibility weights are dynamically adjusted with a certain learning rate.

[0125] S4. Based on the fault diagnosis results, perform fault self-healing processing, including generating a self-healing strategy by coordinating the reinforcement learning decisions of multiple agents, and executing self-healing control based on a control law containing a proportional term and an asymptotic adjustment term.

[0126] In step S4, the method for generating the self-healing strategy in the fault self-healing process is as follows: a multi-agent cooperative reinforcement learning method based on the advantage function is adopted to integrate the value evaluation and advantage function of each agent and introduce a coordination penalty mechanism to generate the optimal self-healing strategy. According to the self-healing strategy, a control method combining proportional control and saturation function compensation is adopted to generate control commands and execute them.

[0127] S5. Online optimization and adaptive updating of model parameters involved in the early warning, diagnosis and self-healing process, including updating parameters with adaptive decay learning rate, and adapting the model to different scenarios through hierarchical transfer.

[0128] In step S5, during online parameter optimization and adaptive update, the model parameters are updated using a learning rate that decays with the time step, combined with fine-tuning by remembering the historical optimal parameters, and the model transfer adopts a hierarchical parameter adjustment method to perform differentiated adaptation for different layers of the network.

[0129] The present invention and its embodiments have been described above illustratively. This description is not restrictive, and the figures shown are only one embodiment of the present invention; the actual structure is not limited thereto. Therefore, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present invention, such designs should fall within the protection scope of the present invention.

[0130] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only independent technical solutions. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. An integrated intelligent fault early warning, diagnosis, and self-healing system for thermal power units, characterized in that: include: The data acquisition module is used to collect real-time operating data of thermal power units; The intelligent early warning module is connected to the data acquisition module and is used to encode and reconstruct the operating data, and trigger fault diagnosis based on the comparison result of the reconstruction error and the dynamic threshold. The fault diagnosis module, connected to the intelligent early warning module, is used to extract features and fuse probabilities of the fault when fault diagnosis is triggered, and output the fault diagnosis result. A fault self-healing module, connected to the fault diagnosis module, is used to generate and execute a self-healing control strategy based on the fault diagnosis results. The parameter optimization module is connected to the intelligent early warning module, the fault diagnosis module, and the fault self-healing module, respectively, and is used to perform online learning and adaptive updates of the parameters in the system.

2. The integrated intelligent fault early warning, diagnosis, and self-healing system for thermal power units according to claim 1, characterized in that, The intelligent early warning module includes: A gated time-space encoder is used to process input operational data. Encode the features to obtain the encoded features. The calculation process is as follows: in, For Long Short-Term Memory (LSTM) network encoding functions, It is a one-dimensional convolutional coding function. The gating weight is calculated using the following formula: It is the Sigmoid activation function. For the gated weight matrix, This represents a vector concatenation operation; The reconstruction error calculation unit is used to calculate the error based on the encoded features. Reconstruct the data and calculate the reconstruction error. The calculation process is as follows: in, For decoder functions; The dynamic threshold calculation unit is used to calculate the adaptive warning threshold at the current moment. The calculation process is as follows: in, and These are the mean and standard deviation of the historical reconstruction error, respectively. Here is the confidence coefficient. For adaptive adjustment coefficients, The information entropy at the current moment, This represents the system's maximum theoretical entropy. Early warning triggering unit, used to compare the reconstruction error With the dynamic threshold ,like This will trigger fault diagnosis.

3. The integrated intelligent fault early warning, diagnosis, and self-healing system for thermal power units according to claim 1, characterized in that, The fault diagnosis module includes: Causal feature map network units are used to extract fault features. The calculation process is as follows: in, It is a multilayer perceptron. It is a causal adjacency matrix; A confidence-weighted probability fusion unit is used to predict probabilities based on the model. and expert rule probability Calculate the final failure probability The calculation process is as follows: in, and These represent the model credibility and expert system credibility at the current moment, respectively. This credibility is updated online using the following formula: in, For credibility learning rate, and These are the accuracy evaluation values ​​of the model and the expert system at the current moment, respectively.

4. The integrated intelligent fault early warning, diagnosis, and self-healing system for thermal power units according to claim 1, characterized in that, The fault self-healing module includes: Advantage function collaboration Learning units, used to generate self-healing strategies, are jointly... value The calculation process is as follows: in, For average value, and Two intelligent agents respectively value, For the joint advantage function, For system status, and For the actions of two intelligent agents, The calculation process is as follows: in, and These are the advantage functions of the two agents, respectively. and For agent weights, and satisfying , To coordinate the penalty coefficient; An adaptive progressive control unit is configured to output control commands according to the self-healing strategy. The calculation process is as follows: in, This is the proportionality coefficient. For systematic error, It is a saturation function. To control the time constant.

5. The integrated intelligent fault early warning, diagnosis, and self-healing system for thermal power units according to claim 1, characterized in that, The parameter optimization module includes: Model parameter update unit, used to update model parameters The update process is as follows: in, For adaptive decay learning rate, The gradient of the loss function. For memory coefficient, These are the historically optimal parameters; The learning rate The calculation process is as follows: in, The initial learning rate, The attenuation coefficient is... For training steps; Transfer learning units are used to transfer the source model Transferring to the target scenario to obtain the target model The calculation process is as follows: in, For the first The amount of change in the layer parameters, For the first Feature extraction function of layer The total number of network layers. For target scene data.

6. An integrated method for intelligent fault early warning, diagnosis, and self-healing of thermal power units, characterized in that: Includes the following steps: S1. Real-time acquisition of operating data of thermal power units; S2. Perform intelligent early warning processing on the running data, including encoding and reconstructing the data, calculating the reconstruction error, comparing the error with a dynamic threshold calculated based on historical data and information entropy, and triggering fault diagnosis if the threshold is exceeded. S3. When fault diagnosis is triggered, fault diagnosis processing is performed, including extracting fault features, and weighted fusion of model prediction results and expert rule results based on online updated model credibility and expert system credibility to obtain the final fault diagnosis result. S4. Based on the fault diagnosis results, perform fault self-healing processing, including generating a self-healing strategy by coordinating the reinforcement learning decisions of multiple agents, and executing self-healing control based on a control law containing a proportional term and an asymptotic adjustment term. S5. Online optimization and adaptive updates of model parameters involved in the early warning, diagnosis and self-healing process, including updating parameters with adaptive decay learning rate and adapting the model to different scenarios through hierarchical transfer learning.

7. The integrated method for intelligent fault early warning, diagnosis, and self-healing of thermal power units according to claim 6, characterized in that, In step S2, the calculation of the dynamic threshold in the intelligent early warning processing is specifically as follows: combining the mean and standard deviation of historical reconstruction errors, and introducing an adaptive adjustment term based on the information entropy of the current running data, to jointly determine the early warning threshold at the current moment.

8. The integrated method for intelligent fault early warning, diagnosis, and self-healing of thermal power units according to claim 6, characterized in that, In step S3, the method for updating the model credibility and expert system credibility in the fault diagnosis process is as follows: based on their respective accuracy performance in recent diagnoses, their credibility weights are dynamically adjusted with a certain learning rate.

9. The integrated method for intelligent fault early warning, diagnosis, and self-healing of thermal power units according to claim 6, characterized in that, In step S4, the method for generating the self-healing strategy in the fault self-healing process is as follows: a multi-agent cooperative reinforcement learning method based on the advantage function is adopted to integrate the value assessment and advantage function of each agent and introduce a coordination penalty mechanism to generate the optimal self-healing strategy. According to the self-healing strategy, a control method combining proportional control and saturation function compensation is adopted to generate control commands and execute them.

10. The integrated method for intelligent fault early warning, diagnosis, and self-healing of thermal power units according to claim 6, characterized in that, In step S5, the online optimization and adaptive update of parameters involves updating the model parameters using a learning rate that decays with the time step, combined with fine-tuning by remembering the historical optimal parameters, and using a hierarchical parameter adjustment method for model transfer to perform differentiated adaptation for different layers of the network.

Images