Power plant system equipment fault diagnosis method and system based on artificial intelligence

Abstract

The invention provides a power plant system equipment fault diagnosis method and system based on artificial intelligence, and the method comprises the steps: firstly obtaining a continuous operation state data flow of key equipment of a power plant, then carrying out the feature extraction of the continuous operation state data flow, and generating a fusion feature set reflecting the equipment operation state coupling relation; calling a pre-trained fault dynamic diagnosis model to perform state evolution analysis processing on the fusion feature set to generate a diagnosis intermediate result which comprises a fault potential node and a fault evolution path, and performing fault attribute analysis processing based on the diagnosis intermediate result to obtain a fault attribute analysis result; according to the method, fault diagnosis information including fault trigger conditions, fault development trends and fault influence ranges is generated, and finally, an equipment diagnosis report including early warning strategies and maintenance priorities is generated according to the fault diagnosis information and is transmitted to a power plant intelligent operation and maintenance terminal, so that the accuracy and timeliness of power plant equipment fault diagnosis can be improved, and the power plant equipment fault diagnosis efficiency is improved. Safe and stable operation of a power plant is ensured.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology. Specifically, it relates to a method and system for fault diagnosis of power plant system equipment based on artificial intelligence. Background Art

[0002] In the field of modern power production, the stable operation of power plant system equipment is crucial for ensuring the reliability and safety of power supply. With the expansion of power plant scale and the increase in equipment complexity, the probability of equipment failure and potential hazards also increase accordingly. Traditional power plant equipment fault diagnosis methods mainly rely on manual experience and simple monitoring systems based on fixed thresholds. Manual experience diagnosis has problems such as strong subjectivity, low efficiency, and limited ability to judge complex faults, and it is difficult to meet the requirements of modern power plants for rapid and accurate diagnosis of equipment faults. The monitoring system based on fixed thresholds can only judge whether the equipment operation parameters exceed the preset thresholds, and cannot comprehensively consider the coupling relationship between equipment operation states and the dynamic evolution process of faults, resulting in insufficient ability to identify some early and potential faults, easy to miss the best treatment opportunity, and then may cause more serious equipment damage and production accidents, bringing huge economic losses and safety risks to the power plant.

[0003] Therefore, there is an urgent need for a more advanced and efficient fault diagnosis method to improve the accuracy and timeliness of power plant equipment fault diagnosis. Summary of the Invention

[0004] In view of the problems mentioned above, in combination with the first aspect of the present invention, embodiments of the present invention provide a method for fault diagnosis of power plant system equipment based on artificial intelligence. The method includes: [[ID=I8]]Obtain the continuous operation state data stream of key equipment in the power plant. The continuous operation state data stream includes the equipment real-time sensing signal sequence and historical fault association records; Extract features from the continuous operation state data stream to generate a fusion feature set reflecting the coupling relationship of equipment operation states; Call the pre-trained fault dynamic diagnosis model to perform state evolution analysis processing on the fusion feature set to generate a diagnosis intermediate result including fault potential nodes and fault evolution paths; Perform fault attribute analysis processing based on the diagnosis intermediate result to generate fault diagnosis information including fault trigger conditions, fault development trends, and fault influence scopes; Generate an equipment diagnosis report including a warning strategy and a maintenance priority based on the fault diagnosis information, and transmit the equipment diagnosis report to the power plant intelligent operation and maintenance terminal.

[0005] In another aspect, an embodiment of the present invention further provides an artificial intelligence-based power plant system equipment fault diagnosis system, including a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor. The machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0006] Based on the above aspects, in the embodiment of the present invention, by comprehensively acquiring the continuous operation status data stream of key equipment in the power plant, the continuous operation status data stream not only includes the equipment real-time sensing signal sequence, but also covers the historical fault association records. On this basis, feature extraction is performed on the continuous operation status data stream to generate a fusion feature set, which can accurately reflect the coupling relationship between equipment operation states, overcoming the defect that traditional methods lack consideration of equipment state relationships. The pre-trained fault dynamic diagnosis model is called to perform state evolution analysis and processing on the fusion feature set, generating a diagnostic intermediate result including fault potential nodes and fault evolution paths, realizing the accurate capture of the fault dynamic process. Based on the diagnostic intermediate result, fault attribute parsing and processing are performed, and the generated fault diagnosis information details the fault triggering conditions, fault development trends and fault impact ranges. Finally, an equipment diagnostic report including a warning strategy and a maintenance priority is generated according to the fault diagnosis information and transmitted to the intelligent operation and maintenance terminal of the power plant, enabling the operation and maintenance personnel to take timely and targeted measures, effectively improving the accuracy, timeliness and comprehensiveness of power plant equipment fault diagnosis, significantly reducing the impact of equipment faults on power plant production, and improving the safety and reliability of power plant operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 is a schematic execution flowchart of an artificial intelligence-based power plant system equipment fault diagnosis method provided by an embodiment of the present invention.

[0008] Figure 2 is a schematic diagram of exemplary hardware and software components of an artificial intelligence-based power plant system equipment fault diagnosis system provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0009] The present invention will be specifically described below with reference to the accompanying drawings of the specification. Figure 1 is a schematic flowchart of an artificial intelligence-based power plant system equipment fault diagnosis method provided by an embodiment of the present invention. The artificial intelligence-based power plant system equipment fault diagnosis method will be introduced in detail below.

[0010] Step S110: Acquire a continuous operation status data stream of key equipment in the power plant, where the continuous operation status data stream includes an equipment real-time sensing signal sequence and historical fault association records.

[0011] In the actual operation scenario of a power plant, it is necessary to obtain the continuous operation status data stream of key equipment. Key equipment such as boilers, steam turbines, generators, etc. The changes in their operation status will generate corresponding signals through various sensors. For specific details, please refer to the following sub-steps.

[0012] Step S111: Collect a real-time sensing signal sequence from the power plant distributed control system. The real-time sensing signal sequence contains sensing signal measurement values at consecutive time points.

[0013] In this embodiment, the power plant distributed control system is connected to numerous sensors distributed on various key equipment in the power plant, such as temperature sensors, pressure sensors, flow sensors, vibration sensors, etc., and can monitor the sensing signal measurement values of the equipment at consecutive time points in real time.

[0014] Step S112: Extract historical fault association records aligned with the time stamps of the real-time sensing signal sequence from the power plant equipment management system. The historical fault association records contain the equipment operation parameters, abnormal phenomenon descriptions, and corresponding maintenance measures at the time of the fault.

[0015] In this embodiment, the power plant equipment management system contains a large number of historical fault association records. These historical fault association records are recorded by staff or automated systems when equipment failures occur. When extracting historical fault association records, it is necessary to ensure that their time stamps are aligned with the time stamps of the real-time sensing signal sequence, so that candidates can accurately associate historical fault information with the current real-time operation status for analysis.

[0016] Among them, the equipment operation parameters in the historical fault association records include temperature, pressure, flow rate, rotational speed, etc. The above parameters are recorded at the time of the fault and can reflect the operation status of the equipment at the moment of the fault. The abnormal phenomenon description is a text description of the abnormal conditions that occurred to the equipment at the time of the fault, such as smoking, abnormal noise, increased vibration, etc. The corresponding maintenance measures are the repair and maintenance methods taken for this fault, including replacing parts, adjusting equipment parameters, and performing maintenance.

[0017] Step S113: Perform time dimension alignment processing on the real-time sensing signal sequence and the historical fault association records to make the real-time signals and historical records at the same time node have consistent time identifiers.

[0018] In this embodiment, in order to ensure that the real-time signals and historical records at the same time node have exactly the same time identifiers, time dimension alignment processing is also required. For example, linear interpolation method or time synchronization protocol, such as Network Time Protocol (NTP), can be used to ensure the consistency of time records from different data sources.

[0019] Step S114: Group the aligned real-time sensing signal sequences and historical fault association records by device functional modules to generate a continuous operating status data stream including a combustion system data stream, a steam turbine system data stream, and an electrical system data stream.

[0020] In this embodiment, the key equipment of the power plant can be divided into different modules according to their functions, mainly including a combustion system, a steam turbine system, and an electrical system. Grouping the aligned real-time sensing signal sequences and historical fault association records according to these device functional modules can better analyze and manage the operating status of devices in different modules.

[0021] For the combustion system, its main function is to convert fuel combustion into heat energy. Therefore, the real-time sensing signal sequences related to the combustion system include fuel flow rate, air flow rate, combustion temperature, combustion pressure, etc., and the historical fault association records include fault information such as unstable combustion, flameout, etc. and corresponding maintenance measures. After grouping these data, a combustion system data stream is formed.

[0022] The main function of the steam turbine system is to convert heat energy into mechanical energy. The real-time sensing signal sequences related to the steam turbine system include steam temperature, steam pressure, steam turbine speed, vibration, etc., and the historical fault association records include fault information such as steam turbine blade damage, bearing failure, etc. and corresponding maintenance measures. After grouping, a steam turbine system data stream is formed.

[0023] The main function of the electrical system is to convert mechanical energy into electrical energy. The real-time sensing signal sequences related to the electrical system include voltage, current, power, frequency, etc., and the historical fault association records include fault information such as short circuit, overload, etc. and corresponding maintenance measures. After grouping, an electrical system data stream is formed.

[0024] Step S115: Perform an integrity verification process on the continuous operating status data stream, eliminate record units with time breaks or data missing, and generate a continuous operating status data stream that meets the time continuity requirement.

[0025] During the data acquisition and transmission process, due to reasons such as sensor failures and communication interruptions, there may be time breaks or data missing in the continuous operating status data stream, which may affect the subsequent fault diagnosis and analysis results. Therefore, an integrity verification process is required. For data with time breaks, interpolation or extrapolation can be performed based on the data before and after it to fill in the time breaks. Also, for each data point, if the measured value exceeds a reasonable range or is an invalid value (such as a null value, error code, etc.), it is considered that the data point has data missing. It is necessary to consider eliminating this record unit. After the integrity verification process, the obtained continuous operating status data stream meets the time continuity requirement.

[0026] Step S120: Extract features from the continuous operating state data stream to generate a fusion feature set reflecting the coupling relationship of the equipment operating state.

[0027] To deeply understand the operating state of key equipment in a power plant, it is necessary to extract features from the continuous operating state data stream that can reflect the coupling relationship of the equipment operating state. The operating state of equipment is often not isolated, and there are complex interactions and coupling relationships between different equipment and between different parameters of the same equipment. Through feature extraction, these coupling relationships can be represented in the form of features.

[0028] Step S121: Input the continuous operating state data stream into the time series feature extraction module, and perform time-dependent relationship modeling on the single-device real-time sensing signal sequence through a long short-term memory network to generate a time series feature vector containing historical state memory.

[0029] The long short-term memory network (LSTM) is a recurrent neural network used to process sequence data, which can effectively capture the time-dependent relationships in sequence data. In this step, the single-device real-time sensing signal sequence in the continuous operating state data stream is input into the LSTM network for time-dependent relationship modeling.

[0030] The LSTM network consists of an input gate, a forget gate, an output gate, and a cell state. The input gate is responsible for determining how much of the current input information can enter the cell state; the forget gate is responsible for determining which information in the cell state needs to be forgotten; the output gate is responsible for determining how much information in the cell state can be output to the next time step. The cell state is used to store and transmit historical information.

[0031] At each time step, the LSTM network receives the input information at the current time point, combines the hidden state and cell state at the previous time step, and updates the cell state and hidden state through the calculations of the input gate, forget gate, and output gate. Specifically, the input gate calculates an input gate vector based on the current input information and the hidden state at the previous time step, and this input gate vector represents which parts of the current input information can enter the cell state. The forget gate calculates a forget gate vector based on the current input information and the hidden state at the previous time step, and this forget gate vector represents which information in the cell state needs to be forgotten. The output gate calculates an output gate vector based on the current input information and the hidden state at the previous time step, and this vector represents how much information in the cell state can be output to the next time step.

[0032] By continuously updating the cell state and hidden state, the LSTM network can record the time-dependent information in the sequence data. Ultimately, the hidden state sequence output by the LSTM network can be used as a time series feature vector containing historical state memory, which contains the operating state information of the device at different time points and the time dependencies between these states.

[0033] Training an LSTM network requires a large amount of historical, real-time sensor signal sequence data from a single device. During training, the LSTM network parameters are adjusted so that the network output accurately reflects the temporal dependencies in the sequence data. The training goal is to minimize the error between the network output and the true value. Backpropagation can be used to update the network parameters, gradually reducing the error through continuous iteration.

[0034] Step S122: Input the continuous operation status data stream into the spatial association analysis module, and use the graph neural network to perform collaborative operation status modeling on the physical connection relationship between devices to generate a spatial association feature vector reflecting the device linkage status. The physical connection relationship is determined by the device association relationship matrix.

[0035] The complex physical connections between devices in a power plant determine their coordinated operating states. Graph neural networks (GNNs) are neural networks specifically designed to process graph-structured data, effectively capturing the relationships between nodes within the graph. In this step, the continuous operating state data stream is fed into the spatial correlation analysis module, where GNNs are used to model the coordinated operating states of the physical connections between devices.

[0036] First, we need to construct a device association matrix to represent the physical connections between devices. This is a two-dimensional matrix, where the rows and columns represent different devices, and the matrix elements represent the connections between the devices. If a physical connection exists between two devices, the matrix element is 1; otherwise, it is 0. For example, if a steam pipe connects a boiler to a steam turbine, the matrix elements corresponding to the boiler and the steam turbine in the device association matrix are 1.

[0037] A GNN network consists of node features and edge features. Node features represent the operating status of each device, such as temperature, pressure, and flow rate; edge features represent the connections between devices. At each time step, the GNN network receives both node and edge features from each device and updates the node features through a message passing mechanism. This message passing mechanism is the core of GNNs, transmitting information through edges between nodes, enabling each node to obtain information about its neighbors.

[0038] Specifically, during the message passing process, each node calculates a message vector based on the features of its neighbor nodes and edge features. Then, these message vectors are aggregated to obtain the new features of the node. By continuously performing message passing and feature updating, the GNN network can capture the collaborative operation state between devices.

[0039] Finally, the node feature vector output by the GNN network can be used as a spatial correlation feature vector reflecting the device linkage state, which contains the spatial correlation information between devices.

[0040] When training the GNN network, a large amount of historical continuous operation state data streams are required for training. During the training process, by adjusting the parameters of the GNN network, the output of the network can reflect the collaborative operation state between devices as accurately as possible. The training objective is to minimize the error between the network output and the true value. The backpropagation algorithm can be used to update the parameters of the network, and through continuous iteration, the error is gradually reduced.

[0041] Step S123: Input the continuous operation state data stream into the log semantic parsing module, and perform text feature extraction processing on the historical fault association records through a pre-trained natural language processing model to generate a log feature vector containing fault semantic information.

[0042] The historical fault association records contain a large amount of text information, such as abnormal phenomenon descriptions, maintenance measures, etc. The above text information contains rich fault semantic information. By performing text feature extraction processing on it, these semantic information can be represented in the form of vectors.

[0043] In this step, the historical fault association records in the continuous operation state data stream are input into the log semantic parsing module, and text feature extraction processing is performed through a pre-trained natural language processing model. The pre-trained natural language processing model is a model pre-trained on a large amount of text data, which can learn the semantic and syntactic information of the language. Commonly used pre-trained natural language processing models include BERT, GPT, etc.

[0044] First, it is necessary to preprocess the text in the historical fault association records. The preprocessing includes operations such as word segmentation, stop word removal, and stemming. Word segmentation is to split the text into individual words or phrases; stop word removal is to remove some words with no actual meaning in the text, such as "of", "is", "in", etc.; stemming is to restore the word to its stem form, such as restoring "running" to "run".

[0045] After preprocessing, the text is input into a pre-trained natural language processing model, which encodes the text and converts it into a vector representation. Specifically, each word or term in the text is mapped into a high-dimensional vector space, and then by aggregating these word vectors, a vector representation of the entire text is obtained.

[0046] Finally, the obtained text vector can be used as a log feature vector containing fault semantic information. This log feature vector contains the fault semantic information in the historical fault association records.

[0047] When using a pre-trained natural language processing model, the model needs to be fine-tuned according to the specific task. Fine-tuning is based on the pre-trained model and further training is carried out using the historical fault association record data, so that the model can better adapt to the task of extracting fault semantic information. During the fine-tuning process, by adjusting the parameters of the model, the output of the model can reflect the fault semantic information in the text as accurately as possible.

[0048] Step S124: Input the time series feature vector, the spatial association feature vector, and the log feature vector into the feature fusion module, and perform importance weighting processing on the feature vectors of each dimension through the attention mechanism, and calculate the weight coefficient of each dimension feature vector.

[0049] In this embodiment, the time series feature vector, the spatial association feature vector, and the log feature vector reflect the operating state of the device from different perspectives, but their respective importance may be different. In order to better comprehensively utilize these feature vectors, it is necessary to perform importance weighting processing on the feature vectors of each dimension through the attention mechanism.

[0050] The attention mechanism is a mechanism that can automatically learn the importance of features. In this step, the time series feature vector, the spatial association feature vector, and the log feature vector are input into the feature fusion module, and the weight coefficient of each dimension feature vector is calculated through the attention mechanism.

[0051] First, the time series feature vector, the spatial association feature vector, and the log feature vector are concatenated to obtain a comprehensive feature vector. Then, the comprehensive feature vector is input into a fully connected layer, and an attention score vector is calculated through an activation function. Each element of the attention score vector represents the importance score of the corresponding dimension feature vector.

[0052] In order to convert the attention score vector into a weight coefficient, it needs to be normalized. The softmax function can be used to normalize the attention score vector, so that the value of each element is between 0 and 1, and the sum of all elements is 1. Finally, the obtained normalized attention score vector can be used as the weight coefficient of each dimension feature vector.

[0053] The weight coefficients calculated through the attention mechanism can automatically reflect the importance of feature vectors in each dimension. The larger the weight coefficient, the higher the importance of the feature vector in that dimension in the comprehensive analysis.

[0054] Step S125: Generate a dynamic weight matrix according to the weight coefficients. After performing weighted processing on the time-series feature vector, the spatial correlation feature vector, and the log feature vector, splice the weighted feature vectors along the feature dimension to generate an intermediate fusion feature vector with dynamic weight allocation.

[0055] In step S124, the weight coefficients of the feature vectors in each dimension have been calculated. Based on these weight coefficients, a dynamic weight matrix can be generated. The dynamic weight matrix is a diagonal matrix, and the elements on its diagonal are the weight coefficients of the time-series feature vector, the spatial correlation feature vector, and the log feature vector respectively.

[0056] Next, multiply the dynamic weight matrix by the time-series feature vector, the spatial correlation feature vector, and the log feature vector respectively to obtain the weighted feature vectors. The purpose of the weighted processing is to perform different degrees of weighting on the feature vectors according to their importance. The larger the weight coefficient, the greater the proportion of the corresponding feature vector in the weighted result.

[0057] Finally, splice the weighted time-series feature vector, the spatial correlation feature vector, and the log feature vector along the feature dimension. The splicing process is to arrange the three vectors in sequence according to the feature dimension, so as to integrate the feature information in different dimensions and form an intermediate fusion feature vector with dynamic weight allocation. It combines the time-series features, spatial correlation features, and log semantic features, and performs dynamic weighting according to the importance of each feature, which can more comprehensively and accurately reflect the operating state of the device.

[0058] Step S126: Perform dimension unification processing on the intermediate fusion feature vector, and map the features of different modalities to the feature space of the same dimension through linear transformation to generate a fusion feature set.

[0059] Since the time-series feature vector, the spatial correlation feature vector, and the log feature vector may come from different data sources and processing methods, their dimensions may be different. In order to effectively analyze and process these features, it is necessary to perform dimension unification processing on the intermediate fusion feature vector.

[0060] Linear transformation is a commonly used method for dimension unification. Through linear transformation, features of different modalities can be mapped to a feature space of the same dimension. Specifically, a linear transformation matrix can be constructed. The number of rows of this matrix is equal to the dimension of the target feature space, and the number of columns is equal to the dimension of the intermediate fusion feature vector. Multiply the intermediate fusion feature vector by the linear transformation matrix to obtain the feature vector after dimension unification.

[0061] When constructing the linear transformation matrix, it is necessary to design according to the dimension of the target feature space and the characteristics of the intermediate fusion feature vector. Methods such as principal component analysis (PCA) can be used to determine the parameters of the linear transformation matrix. Principal component analysis is an unsupervised dimensionality reduction method, which can find the most important feature components in the data and project the data into the space composed of these feature components. Through principal component analysis, a projection matrix can be obtained. Using this projection matrix as the linear transformation matrix can effectively map the intermediate fusion feature vector to a feature space of the same dimension.

[0062] After the dimension unification process, the obtained feature vectors form a fusion feature set. All the feature vectors in this fusion feature set have the same dimension, integrating temporal features, spatial correlation features, and log semantic features, and can more comprehensively reflect the coupling relationship of the device's operating state.

[0063] Furthermore, before step S125, it may further include: Step S1241: Obtain the operation stage identifier of the current device, where the operation stage identifier is determined by the change trend in the temporal feature vector of the current device, and the change trend includes an upward trend, a stable trend, and a downward trend.

[0064] The operating state of the device changes over time, and different operating stages have different characteristics. To analyze the operating state of the device more accurately, it is necessary to obtain the operation stage identifier of the current device.

[0065] The operation stage identifier can be determined by analyzing the change trend in the temporal feature vector of the current device. The temporal feature vector contains the operating state information of the device at different time points. By observing the change trend of this information, it can be judged whether the device is in an upward trend, a stable trend, or a downward trend.

[0066] Specifically, each element in the temporal feature vector can be analyzed, and the difference between adjacent elements can be calculated. If the difference is positive and the absolute value of the difference gradually increases, it is considered that the device is in an upward trend; if the difference is close to zero, it is considered that the device is in a stable trend; if the difference is negative and the absolute value of the difference gradually increases, it is considered that the device is in a downward trend.

[0067] For example, for the temperature time-series feature vector, if the temperature values at adjacent time points gradually increase and the increasing amplitude becomes larger and larger, it can be determined that the device is in an upward trend; if the temperature values at adjacent time points basically remain unchanged, it can be determined that the device is in a stable trend; if the temperature values at adjacent time points gradually decrease and the decreasing amplitude becomes larger and larger, it can be determined that the device is in a downward trend.

[0068] According to different change trends, corresponding operation stage identifiers can be assigned to the device. For example, the upward trend can correspond to the startup stage, the stable trend can correspond to the normal operation stage, and the downward trend can correspond to the shutdown stage.

[0069] Step S1242: Extract the stage feature verification records matching the operation stage identifier from the historical feature verification database, where the stage feature verification records include the verification validity information of the time-series feature vector, the spatial association feature vector, and the log feature vector under each operation stage.

[0070] A large number of historical feature verification records are stored in the historical feature verification database. The above records contain the feature information of the device under different operation stages and the verification validity information of these feature information. To evaluate the validity of each feature vector of the current device under different operation stages, it is necessary to extract the stage feature verification records matching the current operation stage identifier from the historical feature verification database.

[0071] The operation stage identifier can be used as an index to query in the historical feature verification database. When a record matching the current operation stage identifier is found, the stage feature verification records therein are extracted. These records contain the verification validity information of the time-series feature vector, the spatial association feature vector, and the log feature vector under this operation stage.

[0072] The verification validity information can be obtained through the analysis and verification of historical data. For example, for the time-series feature vector, its validity can be evaluated by comparing the error between its predicted value and the actual value; for the spatial association feature vector, its validity can be evaluated by analyzing its correlation with the actual operation state of the device; for the log feature vector, the accuracy and integrity of its semantic information can be evaluated through manual review or machine learning algorithms.

[0073] By extracting the stage feature verification records matching the current operation stage identifier, the validity of each feature vector under the current operation stage can be understood.

[0074] Step S1243: Calculate the feature validity score of each feature vector in the current operation stage according to the stage feature verification records, where the feature validity score is determined by the correct recognition rate of the feature vector in the historical fault samples corresponding to the stage.

[0075] To more quantitatively evaluate the effectiveness of each eigenvector in the current operating stage, it is necessary to calculate their feature effectiveness scores. The feature effectiveness score can be determined by the correct recognition rate of the eigenvector in the historical fault samples of the corresponding stage.

[0076] For each eigenvector, extract its recognition results in the historical fault samples of the corresponding stage from the stage feature verification record. The recognition results can be divided into two cases: correct recognition and incorrect recognition. Correct recognition means that the eigenvector can accurately identify the faults in the historical fault samples; incorrect recognition means that the eigenvector fails to accurately identify the faults in the historical fault samples.

[0077] Statistically count the number of correct recognitions and the total number of recognitions of each eigenvector in the historical fault samples of the corresponding stage, and then calculate the correct recognition rate. The correct recognition rate is equal to the number of correct recognitions divided by the total number of recognitions. Take the correct recognition rate as the feature effectiveness score of the eigenvector in the current operating stage.

[0078] For example, for the time series eigenvector, in the historical fault samples of the corresponding stage, there are a certain number of total recognition times, and the number of correct recognitions is a certain number. Then the feature effectiveness score of this time series eigenvector is equal to the number of correct recognitions divided by the total number of recognitions. By calculating the feature effectiveness score, the effectiveness of each eigenvector in the current operating stage can be intuitively understood.

[0079] Step S1244: Adjust the weight coefficients in the dynamic weight matrix based on the feature effectiveness scores to generate a corrected weight matrix including stage effectiveness constraints.

[0080] In step S124, the weight coefficients in the dynamic weight matrix are calculated according to the attention mechanism. However, the above weight coefficients do not consider the effectiveness of each eigenvector in different operating stages of the device. To more accurately reflect the importance of each eigenvector in the current operating stage, it is necessary to adjust the weight coefficients in the dynamic weight matrix based on the feature effectiveness scores.

[0081] Specifically, multiply the feature effectiveness score of each eigenvector by the weight coefficient of the corresponding eigenvector in the dynamic weight matrix, and use the multiplication result as the adjusted weight coefficient. In this way, the weight coefficients of the eigenvectors with high feature effectiveness scores will increase accordingly, and the weight coefficients of the eigenvectors with low feature effectiveness scores will decrease accordingly.

[0082] For example, for a time-series feature vector, if its feature validity score is high, then after multiplying its feature validity score by the weight coefficient of the time-series feature vector in the dynamic weight matrix, the adjusted weight coefficient obtained will be larger than the original weight coefficient; conversely, if its feature validity score is low, the adjusted weight coefficient will be smaller than the original weight coefficient.

[0083] The adjusted weight coefficients are used to form a corrected weight matrix that includes stage validity constraints. This matrix takes into account the validity of each feature vector in the current operating stage of the device and can more accurately reflect the importance of each feature vector.

[0084] Step S1245: Use the corrected weight matrix to perform a secondary weighting process on the time-series feature vector, the spatial correlation feature vector, and the log feature vector.

[0085] After obtaining the corrected weight matrix, use this matrix to perform a secondary weighting process on the time-series feature vector, the spatial correlation feature vector, and the log feature vector. The purpose of the secondary weighting process is to perform more reasonable weighting based on the validity of each feature vector in the current operating stage.

[0086] Multiply the corrected weight matrix by the time-series feature vector, the spatial correlation feature vector, and the log feature vector respectively to obtain the feature vectors after secondary weighting. The secondary weighting process makes the feature vectors with high feature validity scores account for a larger proportion in the final result, and the feature vectors with low feature validity scores account for a smaller proportion in the final result.

[0087] For example, for a time-series feature vector, if its feature validity score is high, its influence in the final result will be greater after the secondary weighting process; conversely, if its feature validity score is low, its influence in the final result will be smaller. Through the secondary weighting process, the importance of each feature vector in the current operating stage of the device can be more accurately reflected.

[0088] Step S1246: Re-piece together the feature vectors after secondary weighting along the feature dimension to generate a stage-adaptive fusion feature vector that reflects the feature validity in the current operating stage as the updated intermediate fusion feature vector.

[0089] After completing the secondary weighting process, re-piece together the time-series feature vector, the spatial correlation feature vector, and the log feature vector after secondary weighting along the feature dimension. The way of re-piece together is similar to the piece-together process in step S125, that is, arranging the three vectors in sequence according to the feature dimension to form a longer vector.

[0090] Through the re - splicing process, the feature information of different dimensions after secondary weighting is integrated together to form a stage - adapted fusion feature vector that reflects the effectiveness of the features in the current operation stage. This stage - adapted fusion feature vector takes into account the effectiveness of each feature vector of the device in the current operation stage and can more accurately reflect the operation state of the device in the current operation stage. Take this stage - adapted fusion feature vector as the updated intermediate fusion feature vector for subsequent analysis and processing.

[0091] Step S130: Invoke the pre - trained fault dynamic diagnosis model to perform state evolution analysis on the fusion feature set, and generate a diagnostic intermediate result containing potential fault nodes and fault evolution paths.

[0092] After obtaining the fusion feature set, it is necessary to invoke the pre - trained fault dynamic diagnosis model to perform state evolution analysis on it. The fault dynamic diagnosis model is a model trained with a large amount of data. It can analyze the operation state evolution process of the device according to the input feature information, and find potential fault nodes and fault evolution paths that may exist.

[0093] Step S131: Obtain the operation mode identifier of the current device from the power plant operation management system, and the operation mode identifier includes the base - load operation mode, the peak - shaving operation mode, and the start - stop transition mode.

[0094] The power plant operation management system records the operation mode information of the device. Different operation modes have different impacts on the operation state of the device and the probability of faults occurring. To perform more accurate fault diagnosis, it is necessary to obtain the operation mode identifier of the current device from the power plant operation management system.

[0095] The operation mode identifier mainly includes the base - load operation mode, the peak - shaving operation mode, and the start - stop transition mode. The base - load operation mode means that the device operates at a stable load, usually for a long time; the peak - shaving operation mode means that the device adjusts according to the change of the power grid load, and the operation load fluctuates greatly; the start - stop transition mode means the operation state of the device during the start - up or shutdown process, and the operation state of the device changes greatly during this stage.

[0096] Step S132: Extract the time - series feature vector from the fusion feature set, analyze the fluctuation frequency and amplitude change characteristics of the device real - time sensing signal sequence therein, and generate an operation mode feature descriptor.

[0097] The fusion feature set contains a time - series feature vector, and this time - series feature vector contains the information of the device real - time sensing signal sequence. By analyzing the fluctuation frequency and amplitude change characteristics of the device real - time sensing signal sequence in the time - series feature vector, an operation mode feature descriptor can be generated.

[0098] The fluctuation frequency refers to the number of fluctuations of a signal within a unit of time, and the amplitude change characteristic refers to the change of the signal amplitude at different time points. Under different operating modes, the fluctuation frequency and amplitude change characteristics of the real-time sensing signal sequence of the device will be different.

[0099] For example, in the base load operating mode, the operating state of the device is relatively stable, the fluctuation frequency of the real-time sensing signal sequence is low, and the amplitude change is also small; in the peak shaving operating mode, the operating load of the device fluctuates greatly, the fluctuation frequency of the real-time sensing signal sequence is high, and the amplitude change is also large; in the start-stop transition mode, the operating state of the device changes violently, and both the fluctuation frequency and amplitude change of the real-time sensing signal sequence are very obvious.

[0100] By analyzing and extracting the fluctuation frequency and amplitude change characteristics of the real-time sensing signal sequence of the device in the time series feature vector, these characteristics are combined into a vector as the operating mode feature descriptor. This operating mode feature descriptor can reflect the current operating mode characteristics of the device.

[0101] Step S133: Match the operating mode feature descriptor with a preset mode feature template library to determine the actual operating mode of the current device.

[0102] The preset mode feature template library stores feature templates for different operating modes. These templates are obtained through the analysis and summary of a large amount of historical data and can represent the typical characteristics of different operating modes.

[0103] Match the operating mode feature descriptor with each template in the preset mode feature template library. The matching process can adopt methods such as similarity calculation, for example, calculating the Euclidean distance, cosine similarity, etc. between the operating mode feature descriptor and each template. The higher the similarity, the more the operating mode feature descriptor matches the operating mode represented by the template.

[0104] Select the operating mode corresponding to the template with the highest similarity as the actual operating mode of the current device. In this way, the current operating mode of the device can be accurately determined.

[0105] Step S134: Obtain the corresponding set of evolution parameters from the model parameter configuration library according to the actual operating mode. The set of evolution parameters includes the hidden state update rate parameter of the recurrent neural network and the anomaly detection threshold of the state transition vector.

[0106] The model parameter configuration library stores the corresponding set of evolution parameters for different operating modes. These parameters are set according to the operating characteristics of the device and the possibility of faults occurring under different operating modes.

[0107] According to the determined actual operation mode, find the corresponding set of evolution parameters from the model parameter configuration library. The set of evolution parameters includes the hidden state update rate parameter of the recurrent neural network and the anomaly detection threshold of the state transition vector.

[0108] The hidden state update rate parameter determines the speed at which the recurrent neural network updates the hidden state at each time step. Under different operation modes, the change speed of the device's operation state is different, so different hidden state update rate parameters are required. For example, in the start-stop transition mode, the device's operation state changes violently, and a higher hidden state update rate parameter is required; in the base load operation mode, the device's operation state is relatively stable, and a lower hidden state update rate parameter is required.

[0109] The anomaly detection threshold of the state transition vector is used to judge whether the state transition vector is abnormal. Under different operation modes, the normal operation state range of the device is different, so different anomaly detection thresholds are required. For example, in the peak shaving operation mode, the operation load of the device fluctuates greatly, and the normal operation state range is wider, so a higher anomaly detection threshold is required; in the base load operation mode, the device's operation state is relatively stable, and the normal operation state range is narrower, so a lower anomaly detection threshold is required.

[0110] Step S135: Input the set of evolution parameters into the time step deduction layer of the fault dynamic diagnosis model, and adjust the activation function parameters of the input gate, forget gate, and output gate of the recurrent neural network.

[0111] After obtaining the set of evolution parameters, input it into the time step deduction layer of the fault dynamic diagnosis model. The recurrent neural network in the time step deduction layer is the core part of the fault dynamic diagnosis model, and it controls the flow and update of information through the input gate, forget gate, and output gate.

[0112] The hidden state update rate parameter and the anomaly detection threshold of the state transition vector in the set of evolution parameters will affect the activation function parameters of the input gate, forget gate, and output gate of the recurrent neural network. By adjusting these activation function parameters, the recurrent neural network can better adapt to the operation state of the device under the current operation mode.

[0113] Specifically, according to the hidden state update rate parameter, adjust the weight coefficients of the input gate, forget gate, and output gate, so that the recurrent neural network can update the hidden state more accurately at each time step. According to the anomaly detection threshold of the state transition vector, adjust the threshold of the activation function, so that the recurrent neural network can more accurately judge whether the state transition vector is abnormal.

[0114] Step S136: Input the fusion feature set into the initial state encoding layer of the fault dynamic diagnosis model, and map it into a device initial operating state vector through a fully connected network.

[0115] The initial state encoding layer of the fault dynamic diagnosis model is the first layer of the model, and its function is to map the input fusion feature set into a device initial operating state vector. A fully connected network is a commonly used neural network structure, and each neuron in it is connected to all neurons in the previous layer.

[0116] Input the fusion feature set into the fully connected network of the initial state encoding layer. The fully connected network will perform weighted summation on each element in the fusion feature set according to the preset weight coefficients, and perform a non-linear transformation through an activation function, and finally output a device initial operating state vector. This device initial operating state vector is an abstract representation of the current operating state of the device, and it contains the key information in the fusion feature set.

[0117] Step S137: Input the initial operating state vector into the time step deduction layer of the fault dynamic diagnosis model, and simulate the evolution process of the device state within continuous time steps through a recurrent neural network.

[0118] The time step deduction layer is the core part of the fault dynamic diagnosis model, and it simulates the evolution process of the device state within continuous time steps through a recurrent neural network. The recurrent neural network has a memory function, and it can update the hidden state of the current time step according to the input of the current time step and the hidden state of the previous time step, so as to realize the simulation of the device state evolution process.

[0119] Step S1371: Extract the key operating parameters in the device real-time sensing signal sequence, and the key operating parameters include temperature parameters, pressure parameters and flow parameters.

[0120] The device real-time sensing signal sequence contains multiple operating parameters, among which temperature parameters, pressure parameters and flow parameters are the key parameters reflecting the device operating state. The changes of these parameters can directly affect the performance and safety of the device.

[0121] Screen out the temperature parameters, pressure parameters and flow parameters from the device real-time sensing signal sequence. These parameters can be measured in real time by sensors and play an important role in the fault dynamic diagnosis model.

[0122] Step S1372: After performing normalization processing on the key operating parameters, obtain the parameter correlation relationship corresponding to the key operating parameters from the device operating characteristic database, and the parameter correlation relationship includes the coupling change law between temperature and pressure, and the co-variation mode between pressure and flow.

[0123] In order to eliminate the dimensional differences and numerical range differences between key operating parameters, it is necessary to standardize them. Common methods can be used for standardization, such as subtracting the mean of each parameter and then dividing by its standard deviation.

[0124] The device operation characteristic database stores the correlation relationships between key operating parameters. These correlation relationships are obtained through a large number of experiments and data analyses, reflecting the coupling change laws and co-variation patterns between key operating parameters under normal operating conditions of the device.

[0125] Search for the parameter correlation relationships corresponding to the standardized key operating parameters in the device operation characteristic database. For example, search for the coupling change law between temperature and pressure to understand the change trend of pressure at different temperatures; search for the co-variation pattern between pressure and flow rate to understand the corresponding change of flow rate when pressure changes.

[0126] Step S1373: Construct a parameter co-variation index according to the parameter correlation relationship, and the parameter co-variation index is determined by the synchronous change degree of key operating parameters.

[0127] Construct a parameter co-variation index according to the obtained parameter correlation relationship. The parameter co-variation index is used to measure the synchronous change degree between key operating parameters.

[0128] The parameter co-variation index can be constructed by calculating statistical quantities such as the correlation coefficient and covariance between key operating parameters. The larger the correlation coefficient, the higher the synchronous change degree between key operating parameters; the larger the covariance, the more obvious the co-variation between key operating parameters. By comprehensively considering these statistical quantities, a comprehensive parameter co-variation index can be constructed. For example, the correlation coefficient and covariance can be weighted and combined to obtain a more comprehensive parameter co-variation index. The specific weighting coefficients can be determined according to the importance of different parameters in device operation. For parameters that have a greater impact on device performance, higher weighting coefficients are assigned.

[0129] Step S1374: Input the parameter co-variation index into the hidden state update function of the recurrent neural network to generate a state update coefficient containing parameter co-variation constraints.

[0130] The hidden state update function of the recurrent neural network is used to update the hidden state of the current time step according to the current input and the hidden state of the previous time step. Input the parameter co-variation index into this hidden state update function, and it generates a state update coefficient containing parameter co-variation constraints according to the magnitude and change of the parameter co-variation index.

[0131] In the hidden state update function, the parameter co-variation index affects the calculation process inside the function. When the parameter co-variation index is large, it indicates that the co-variation among the key operating parameters is relatively obvious. At this time, the hidden state update function will generate a large state update coefficient to emphasize the impact of parameter co-variation on the hidden state update; conversely, when the parameter co-variation index is small, the state update coefficient will also become smaller accordingly. In this way, the state update coefficient contains the constraint information of parameter co-variation, enabling the recurrent neural network to better consider the co-variation relationship among the key operating parameters when simulating the evolution process of the device state.

[0132] Step S1375: In the calculation process of the input gate of the recurrent neural network, multiply the state update coefficient by the activation value of the input gate to generate a co-variation constraint input gate value.

[0133] The input gate of the recurrent neural network is used to control the degree to which the current input information enters the hidden state. When calculating the activation value of the input gate, it is usually calculated based on the current input and the hidden state of the previous time step. In this step, the generated state update coefficient containing parameter co-variation constraints is multiplied by the activation value of the input gate.

[0134] Through this multiplication process, the activation value of the input gate is affected by the parameter co-variation. If the state update coefficient is large, it indicates that the parameter co-variation is obvious. At this time, the activation value of the input gate will increase accordingly, meaning that more current input information can enter the hidden state to reflect the impact of parameter co-variation on the device state; conversely, if the state update coefficient is small, the activation value of the input gate will decrease, and the current input information entering the hidden state will also decrease accordingly. In this way, the co-variation constraint input gate value is generated, enabling the calculation process of the input gate to consider the factor of parameter co-variation.

[0135] Step S1376: In the calculation process of the forget gate of the recurrent neural network, multiply the state update coefficient by the activation value of the forget gate to generate a co-variation constraint forget gate value.

[0136] The role of the forget gate is to control which information in the hidden state of the previous time step needs to be forgotten. When calculating the activation value of the forget gate, it is also based on the current input and the hidden state of the previous time step. Multiply the state update coefficient by the activation value of the forget gate to introduce the constraint of parameter co-variation.

[0137] When the state update coefficient is relatively large, the activation value of the forget gate increases, meaning that more information in the hidden state of the previous time step will be retained. Because when the parameters co-vary significantly, historical information is more important for understanding the evolution of the current device state; when the state update coefficient is relatively small, the activation value of the forget gate decreases, and more information in the hidden state of the previous time step will be forgotten to adapt to the rapid changes in the device state when the parameter co-variation is not obvious. In this way, the co-constrained forget gate value is generated, making the calculation process of the forget gate also consider the influence of parameter co-variation.

[0138] Step S1377: In the calculation process of the output gate of the recurrent neural network, multiply the state update coefficient by the activation value of the output gate to generate a co-constrained output gate value.

[0139] The output gate is used to control how much information in the hidden state of the current time step will be output to the next time step or as the final output. When calculating the activation value of the output gate, the current input and the hidden state of the previous time step are combined. Multiply the state update coefficient by the activation value of the output gate to reflect the influence of parameter co-variation on the output information.

[0140] When the state update coefficient is relatively large, the activation value of the output gate increases, and more hidden state information will be output. Because when the parameters co-vary significantly, the hidden state contains more important information related to the evolution of the device state; when the state update coefficient is relatively small, the activation value of the output gate decreases, and the output hidden state information will also decrease accordingly. Thus, the co-constrained output gate value is obtained, making the calculation process of the output gate consider the factor of parameter co-variation.

[0141] Step S1378: Use the co-constrained input gate value, co-constrained forget gate value, and co-constrained output gate value to update the hidden state of the recurrent neural network, generating a state transition vector reflecting the influence of parameter co-variation.

[0142] After obtaining the co-constrained input gate value, co-constrained forget gate value, and co-constrained output gate value, they are used to update the hidden state of the recurrent neural network. Specifically, determine the amount of current input information entering the hidden state according to the co-constrained input gate value, determine the information to be retained in the hidden state of the previous time step according to the co-constrained forget gate value, and determine the information to be output in the current hidden state according to the co-constrained output gate value.

[0143] By comprehensively considering these three co-constrained gate values, perform an update operation on the hidden state. The updated hidden state contains the influence information of parameter co-variation, and it is used as the state transition vector of the current time step. This state transition vector reflects the transfer of the device state from one time step to the next considering the co-variation of key operating parameters.

[0144] Step S1379: Record the state transition vectors at each time step to form a state transition sequence containing parameter co-variation information, where the state transition sequence is used for subsequent abnormal state node identification and fault evolution path tracing.

[0145] During the process of simulating the evolution of the device state at consecutive time steps by the recurrent neural network, a state transition vector is generated at each time step. Recording these state transition vectors in chronological order forms a state transition sequence containing parameter co-variation information.

[0146] This state transition sequence is a detailed record of the device state evolution process, which not only includes the transfer of the device state at each time step but also incorporates the information of the co-variation of key operating parameters. During the subsequent abnormal state node identification process, by analyzing the characteristics and change trends of each state transition vector in the state transition sequence, vectors significantly different from the normal state can be found, and the corresponding time step nodes are marked as abnormal state nodes. During the fault evolution path tracing process, based on the sequence and change relationship between the state transition vectors in the state transition sequence, the key fault evolution path from the initial operating state to the abnormal state node can be determined.

[0147] Step S138: Input the state transition vector into the abnormal trigger detection layer of the fault dynamic diagnosis model to identify abnormal state nodes during the state transition process.

[0148] The abnormal trigger detection layer is an important module in the fault dynamic diagnosis model, and its main function is to analyze the input state transition vector to identify possible abnormal state nodes. This layer will check each dimension feature of the state transition vector one by one according to the preset abnormal detection rules and thresholds.

[0149] For each dimension feature in the state transition vector, the abnormal trigger detection layer calculates the deviation degree from the mean and standard deviation of this dimension feature in the normal state. If the deviation degree of a certain dimension feature exceeds the preset threshold, then it is considered that there is an abnormality in this dimension of the state transition vector. When there are multiple dimension features in the state transition vector that exceed the threshold, or the deviation degree of a certain key dimension feature is very large, the time step node corresponding to this state transition vector is marked as an abnormal state node.

[0150] The anomaly trigger detection layer can adopt various methods for anomaly detection, such as methods based on statistical distribution, methods based on machine learning, etc. The methods based on statistical distribution assume that the feature of each dimension of the state transition vector follows a certain statistical distribution under normal conditions, such as normal distribution, and then determine the anomaly detection threshold according to the parameters of this distribution. The methods based on machine learning will use trained classification models, such as support vector machines, decision trees, etc., take the state transition vector as the input, and judge whether it is an abnormal state.

[0151] Step S139: Input the abnormal state node into the path tracing layer of the fault dynamic diagnosis model, and determine the key fault evolution path from the initial operation state vector to the abnormal state node through the backtracking algorithm.

[0152] After receiving the abnormal state node, the path tracing layer will use the backtracking algorithm to determine the key fault evolution path from the initial operation state vector to this abnormal state node. The backtracking algorithm is a way of reverse search. Starting from the abnormal state node, it gradually backtracks to the initial operation state vector, finds all the state nodes passed in the middle, and determines the key fault evolution path according to the transfer relationship between these nodes.

[0153] During the backtracking process, the path tracing layer will refer to the sequence and change relationship between the state transition vectors in the state transition sequence. For each abnormal state node, it will find the state transition vector at the previous time step and judge whether there is an obvious association between this vector and the abnormal state node. If the degree of association is high, then the time step node corresponding to this vector will be added to the fault evolution path. Then continue to backtrack forward from this node and repeat the above process until backtracking to the initial operation state vector.

[0154] When determining the key fault evolution path, the path tracing layer will screen and sort the nodes in the path according to the change degree and importance between the state transition vectors. For those nodes that play a key role in the fault evolution, higher priorities will be given; for those nodes that have less influence on the fault evolution, they may be excluded from the key fault evolution path. The finally obtained key fault evolution path is a directed path from the initial operation state vector to the abnormal state node, which clearly shows the evolution process of the device state from normal to abnormal.

[0155] Step S1310: Input the key fault evolution path into the fault impact range evaluation layer of the fault dynamic diagnosis model, and calculate the fault impact range information of the abnormal state node on the associated devices through the device association relationship matrix.

[0156] The equipment association relationship matrix is a matrix that represents the physical connections and operational association relationships among various equipment in a power plant. The elements in the equipment association relationship matrix represent the association strength between two pieces of equipment, and the association strength can be determined based on factors such as the physical connection method, energy transfer relationship, and control relationship between the equipment.

[0157] After the fault impact scope evaluation layer receives the critical fault evolution path, it calculates the impact degree of the abnormal state node on the associated equipment according to the abnormal state nodes in the critical fault evolution path and the equipment association relationship matrix. The specific calculation process is as follows: First, for the equipment corresponding to the abnormal state node, find all other equipment associated with this equipment from the equipment association relationship matrix. Then, according to the association strength and the severity of the abnormal state node, calculate the impact score of this abnormal state node on each associated equipment.

[0158] The calculation of the impact score can consider multiple factors, such as the association strength, the degree of abnormality of the abnormal state node, the vulnerability of the associated equipment, etc. For equipment with a higher association strength, the impact score of the abnormal state node on it will be relatively higher; for an abnormal state node with a greater degree of abnormality, its impact score on the associated equipment will also increase accordingly; for an associated equipment with a higher vulnerability, that is, an equipment more likely to be affected by faults, the impact score will also increase. Summarizing and sorting the impact scores of each associated equipment can obtain the evaluation result of the impact degree of the abnormal state node on the associated equipment, including information such as which equipment is more affected and which equipment is less affected.

[0159] Step S1311: Integrate the abnormal state node, the critical fault evolution path, and the fault impact scope information to generate a diagnostic intermediate result containing potential fault nodes and fault evolution paths.

[0160] After completing the identification of abnormal state nodes, the determination of critical fault evolution paths, and the evaluation of fault impact scopes, it is necessary to integrate this information to generate a diagnostic intermediate result. The diagnostic intermediate result is a comprehensive information set that contains important information such as potential fault nodes and fault evolution paths.

[0161] Organize and sort the identification information of the abnormal state node, the detailed description of the critical fault evolution path, and the fault impact scope information in a set format. It can be stored together in a data structure, such as a list, dictionary, etc. For example, a dictionary can be created, which contains keys such as "abnormal state node", "critical fault evolution path", "fault impact scope", etc., corresponding to the respective information values.

[0162] By integrating this information, the diagnostic intermediate results can comprehensively reflect the potential situation and evolution process of equipment failures. The potential failure nodes indicate the locations and times where equipment failures may occur, the failure evolution paths show the development process of failures from the initial state to the abnormal state, and the failure impact range information illustrates the associated equipment and the degree of impact that the failures may affect.

[0163] Step S140: Perform failure attribute analysis processing based on the diagnostic intermediate results to generate failure diagnosis information including failure trigger conditions, failure development trends, and failure impact ranges.

[0164] Step S141: Extract the abnormal state nodes from the diagnostic intermediate results and query historical failure cases that match the abnormal state nodes from the historical failure database.

[0165] The historical failure database stores a large amount of information on past failure cases of power plant equipment, including the equipment operating status, relevant parameters, abnormal phenomena, etc. when the failures occurred. After extracting the abnormal state nodes from the diagnostic intermediate results, the characteristic information of these abnormal state nodes will be used as the query conditions to search in the historical failure database to find the matching historical failure cases.

[0166] The matching process will consider multiple aspects of the characteristics of the abnormal state nodes, such as the characteristics of each dimension of the state transition vector, the time point when the abnormality occurs, the operating status of the associated equipment, etc. For each abnormal state node, the similarity between it and the failure nodes in the historical failure cases will be calculated. Various methods can be used to calculate the similarity, such as calculating the Euclidean distance, cosine similarity, etc. between the state transition vectors. Select the historical failure case with the highest similarity as the case that matches the abnormal state node.

[0167] Step S142: Analyze the equipment operating parameter records of the historical failure case and extract the equipment operating parameter sequence before the occurrence of the abnormal state node.

[0168] After finding the historical failure case that matches the abnormal state node, it is necessary to analyze the equipment operating parameter records of this case in detail. The equipment operating parameter records contain various operating parameter values of the equipment before and after the failure, such as temperature, pressure, flow rate, rotational speed, etc.

[0169] Extract the equipment operating parameter sequence within a period of time before the occurrence of the abnormal state node from the equipment operating parameter records. The selection of this period can be determined according to the characteristics of the equipment and the type of failure. Generally, a period that can reflect the trend of the equipment state change before the failure will be selected. For example, for some slowly developing failures, a longer parameter sequence before the occurrence of the abnormal state node can be selected; for some sudden failures, a shorter parameter sequence within a short time can be selected.

[0170] The extracted sequence of device operation parameters is a multi-dimensional time series data. Each dimension corresponds to a device operation parameter, and each time point corresponds to a parameter value. This sequence of device operation parameters can show the changes in the operation state of the device before a fault occurs.

[0171] Step S143: Calculate the mean, maximum value, and minimum value of the sequence of device operation parameters to generate statistical features of parameter distribution.

[0172] To better understand the characteristics and variation laws of the sequence of device operation parameters, it is necessary to calculate statistical features such as the mean, maximum value, and minimum value of this sequence. The mean reflects the average level of the device operation parameters over a period of time, the maximum value represents the highest value reached by the parameter during this period, and the minimum value represents the lowest value reached by the parameter during this period.

[0173] For each dimension (i.e., each device operation parameter) in the sequence of device operation parameters, calculate its mean, maximum value, and minimum value respectively. When calculating the mean, add up the parameter values at all time points on this dimension and then divide by the number of time points; when calculating the maximum value, find the maximum value among all parameter values on this dimension; when calculating the minimum value, find the minimum value among all parameter values on this dimension.

[0174] Combine the calculated mean, maximum value, and minimum value together to form the statistical features of parameter distribution for this dimension. For multiple dimensions of the sequence of device operation parameters, multiple statistical features of parameter distribution will be obtained. These statistical features can intuitively show the distribution and variation range of the device operation parameters before the occurrence of abnormal state nodes.

[0175] Step S144: Calculate the mutual information value between each operation parameter and the abnormal state node according to the statistical features of parameter distribution, and filter out the operation parameters with mutual information values greater than the preset threshold as candidate key parameters.

[0176] Mutual information is an index to measure the correlation between two random variables. In this step, it is used to measure the correlation between each operation parameter and the abnormal state node. According to the statistical features of parameter distribution of the sequence of device operation parameters and the feature information of the abnormal state node, calculate the mutual information value between each operation parameter and the abnormal state node.

[0177] The process of calculating the mutual information value will consider the value distribution of the operation parameter and the occurrence probability of the abnormal state node. For each operation parameter, divide its value range into multiple intervals according to its statistical features of parameter distribution, and then count the frequency of occurrence of the abnormal state node in each interval. Through this frequency information, the mutual information value between this operation parameter and the abnormal state node can be calculated.

[0178] The preset threshold is a pre-set critical value used to screen out the operating parameters that are strongly correlated with the abnormal state nodes. Compare the mutual information value of each operating parameter with the preset threshold. If the mutual information value of a certain operating parameter is greater than the preset threshold, then this operating parameter is regarded as a candidate key parameter. Candidate key parameters are those operating parameters that have a high correlation with the abnormal state nodes and may have played an important role during the fault occurrence process.

[0179] Step S145: Calculate the feature importance scores of the candidate key parameters through the random forest algorithm, and select the top N operating parameters with the highest feature importance scores as the key operating parameters strongly correlated with the abnormal state nodes.

[0180] The random forest algorithm is an ensemble learning algorithm composed of multiple decision trees, which performs classification and regression analysis by synthesizing the results of multiple decision trees. In this step, the random forest algorithm is used to calculate the feature importance scores of the candidate key parameters.

[0181] Take the candidate key parameters as input features and the identifier of the abnormal state node as the output label to construct a random forest model. During the process of training the random forest model, each decision tree will make decisions and classifications according to different combinations of input features. The random forest algorithm will calculate the feature importance scores of each candidate key parameter based on the usage frequency and effect of each feature when dividing nodes by each decision tree.

[0182] The higher the feature importance score, the greater the role that the candidate key parameter plays in distinguishing between abnormal state nodes and normal state nodes. Sort the feature importance scores of all candidate key parameters, and select the top N operating parameters with the highest scores as the key operating parameters strongly correlated with the abnormal state nodes. Here, N can be adjusted according to the actual situation, and generally it is determined according to the complexity of the device and the characteristics of the fault to ensure that the selected key operating parameters can accurately reflect the fault triggering conditions.

[0183] Step S146: Extract the value ranges of the key operating parameters when the abnormal state nodes occur, and generate the fault triggering conditions including the parameter names and value ranges.

[0184] After obtaining the key operating parameters strongly correlated with the abnormal state nodes, further extract the value ranges of these key operating parameters when the abnormal state nodes occur. Through a detailed analysis of the real-time sensing signal sequence of the device and historical fault data, determine the specific value conditions of each key operating parameter before and after the occurrence of the abnormal state node.

[0185] For each key operating parameter, collect all the measurement values within a small time period near the occurrence time of the abnormal state node, and then find the minimum and maximum values among these measurement values to determine the value range of the parameter in the abnormal state. For example, for the key operating parameter of temperature, within a few minutes before and after the occurrence time of the abnormal state node, the collected temperature measurement values fluctuate within a set interval, and the lower and upper limits of this interval constitute the value range of the temperature parameter in the abnormal state.

[0186] Combine the name of the key operating parameter and the corresponding value range to form a fault trigger condition that includes the parameter name and the value range. The above fault trigger condition clarifies under what parameter value conditions the device may enter an abnormal state.

[0187] Step S147: Extract the key fault evolution path from the diagnostic intermediate results, analyze the continuity of state transitions in the path and the degree of abnormal accumulation at each node, and generate a description of the fault development trend that reflects the fault development rate.

[0188] Extract the key fault evolution path from the diagnostic intermediate results. This key fault evolution path records the state transition process of the device from the initial operating state to the abnormal state node. Conduct an in-depth analysis of the key fault evolution path. First, focus on the continuity of state transitions in the path. The continuity of state transitions reflects whether the fault development process is smooth and whether there are sudden jumps or interruptions.

[0189] If the state transitions are relatively continuous, it indicates that the fault develops gradually, which may be caused by the gradual wear or performance degradation of a certain component of the device; if there are obvious jumps or interruptions in the state transitions, it indicates that the fault may be caused by sudden factors, such as external impacts, sudden component failures, etc.

[0190] At the same time, analyze the degree of abnormal accumulation at each node in the path. The degree of abnormal accumulation at each node can be measured by the deviation degree of the state transition vector corresponding to this node from the normal state vector. The greater the deviation degree, the more serious the abnormal accumulation at this node.

[0191] Comprehensively consider the continuity of state transitions and the degree of abnormal accumulation at each node to generate a description of the fault development trend that reflects the fault development rate. If the state transitions are continuous and the degree of abnormal accumulation gradually increases, then the fault development trend may be a slow upward trend; if there are jumps in the state transitions and the degree of abnormal accumulation suddenly increases, then the fault development trend may be a rapid upward trend. The above description of the fault development trend helps to predict the development of the fault.

[0192] Step S148: Extract the fault impact scope information from the diagnostic intermediate results, analyze the importance level of associated devices and the possibility of fault propagation, and generate a fault impact scope description including a list of associated devices and the impact degree.

[0193] In this embodiment, the fault impact scope information includes the evaluation results of the impact degree of abnormal state nodes on associated devices. Through further analysis, the importance level of associated devices is first determined. The importance level of associated devices can be determined according to factors such as their roles in the power plant production process and their impacts on the stability of the entire system. For example, some key devices directly involved in the power generation process, such as generators and steam turbines, have a higher importance level; while some auxiliary devices, such as cooling water pumps and lubrication systems, have a relatively lower importance level.

[0194] Then, analyze the possibility of fault propagation. The possibility of fault propagation is related to the physical connection relationship, operation association relationship between devices, and the characteristics of the devices themselves. If there is a close physical connection and energy transfer relationship between devices, the possibility of fault propagation is greater; if the devices have good isolation measures and self - protection mechanisms, the possibility of fault propagation is smaller.

[0195] Taking into account the importance level of associated devices and the possibility of fault propagation, generate a fault impact scope description including a list of associated devices and the impact degree, which details the associated devices affected by the fault and evaluates the impact degree of each associated device, such as "high impact", "medium impact", "low impact", etc. This helps the operation and maintenance personnel quickly understand the fault impact scope and reasonably arrange resources for repair and prevention.

[0196] Step S149: Structurally integrate the fault trigger conditions, fault development trend descriptions, and fault impact scope descriptions to generate fault diagnosis information including text descriptions and data identifiers.

[0197] Structurally integrate the fault trigger conditions, fault development trend descriptions, and fault impact scope descriptions generated in the previous steps. First, determine a suitable structure to organize this information. For example, a hierarchical structure can be adopted, with the fault trigger conditions as the first - layer information, the fault development trend descriptions as the second - layer information, and the fault impact scope descriptions as the third - layer information.

[0198] For the fault trigger conditions, present them in a clear text form, and add data identifiers to the value ranges of each key operating parameter for subsequent query and analysis. For example, "Temperature: value range [X1, X2]", where "Temperature" is the parameter name, "[X1, X2]" is the value range, and assign a unique data identifier to this value range.

[0199] For the description of the fault development trend, concisely describe the fault development rate and trend characteristics in text. Similarly, data identifiers can be added to the key information in the description, such as "Fault development trend: slowly rising (identifier: T1)".

[0200] For the description of the fault impact scope, present the list of associated devices and the impact degree in a list form, and add data identifiers to each associated device and its impact degree, such as "Associated device: generator (identifier: E1), Impact degree: high (identifier: I1)".

[0201] Combine this processed information according to a determined structure to generate fault diagnosis information containing text descriptions and data identifiers. The above-structured fault diagnosis information is convenient for storage, query, and sharing.

[0202] Step S150: Generate a device diagnosis report containing an early warning strategy and a maintenance priority based on the fault diagnosis information, and transmit the device diagnosis report to the intelligent operation and maintenance terminal of the power plant.

[0203] Based on the previously generated fault diagnosis information, further generate a device diagnosis report containing an early warning strategy and a maintenance priority, which will provide clear decision-making basis for the operation and maintenance personnel of the power plant, so that they can take timely measures to deal with equipment failures.

[0204] Step S151: According to the fault trigger conditions in the fault diagnosis information, match the corresponding early warning trigger timing and early warning level from the pre-stored early warning rule library.

[0205] The pre-stored early warning rule library stores the mapping relationships between various fault trigger conditions and the corresponding early warning trigger timing and early warning levels. These rules are formulated based on the historical fault data and operation and maintenance experience of the power plant.

[0206] According to the fault trigger conditions in the fault diagnosis information, perform a matching search in the early warning rule library. For the value range of each key operating parameter, find the rule in the early warning rule library that matches it. For example, if the value range of the temperature in the fault trigger condition is [X1, X2], find the rule in the early warning rule library that contains this value range, and find the corresponding early warning trigger timing and early warning level.

[0207] The early warning trigger timing can be divided into real-time early warning, short-term early warning, long-term early warning, etc. Real-time early warning means that when the key operating parameter reaches the fault trigger condition, an early warning is immediately issued; short-term early warning means that within a set time (such as several hours or days), if the parameter continuously approaches the fault trigger condition, an early warning is issued; long-term early warning means that within a relatively long time (such as several weeks or months), the change trend of the parameter shows that it may reach the fault trigger condition, and an early warning is issued.

[0208] The warning levels can be divided into Level 1 warning, Level 2 warning, Level 3 warning, etc., indicating different severities of the faults. Level 1 warning means that the fault may cause the equipment to stop immediately or seriously affect the production safety of the power plant, and immediate measures need to be taken; Level 2 warning means that the fault may have a certain impact on the equipment performance and requires timely attention and handling; Level 3 warning means that the impact of the fault is relatively small, but still requires monitoring and prevention.

[0209] Step S152: According to the description of the fault development trend in the fault diagnosis information, determine the urgency of the maintenance operation based on the fault development rate.

[0210] The description of the fault development trend in the fault diagnosis information provides information on the fault development rate. Based on the fault development rate, the urgency of the maintenance operation can be determined.

[0211] If the description of the fault development trend shows that the fault development rate is fast, that is, the fault may deteriorate in a short time, then the urgency of the maintenance operation is high and maintenance personnel need to be arranged immediately for handling; if the description of the fault development trend shows that the fault development rate is slow, that is, the fault will reach a severe level only after a long time, then the urgency of the maintenance operation is low and the time for maintenance can be arranged appropriately.

[0212] Different urgency levels can be divided according to different intervals of the fault development rate, such as "urgent", "relatively urgent", "general", "not urgent", etc. For example, if the fault development rate exceeds a preset fast threshold, then the urgency of the maintenance operation is "urgent"; if the fault development rate is between the fast threshold and the slow threshold, then the urgency of the maintenance operation is "relatively urgent"; if the fault development rate is lower than the slow threshold, then the urgency of the maintenance operation is "general".

[0213] Step S153: Query the maintenance resource requirement information of the associated equipment from the equipment ledger database and generate a maintenance resource requirement list.

[0214] The equipment ledger database records the detailed information of all the equipment in the power plant, including the equipment model, specifications, maintenance history, maintenance resource requirements, etc. Query the maintenance resource requirement information of the equipment associated with the fault from the equipment ledger database.

[0215] The maintenance resource requirement information includes the spare parts, tools, manpower, etc. required for maintenance. For each associated equipment, according to its equipment model and fault situation, find the corresponding maintenance resource requirement record in the equipment ledger database. For example, for a generator, specific types of bearings, insulating materials and other spare parts, as well as professional maintenance tools and maintenance personnel with relevant skills may be required.

[0216] Sort out the queried maintenance resource requirement information to generate a maintenance resource requirement list, which details the maintenance resources required for each associated device, including information such as the name, quantity, and specifications of the resources, providing clear guidance for the preparation of maintenance work.

[0217] Step S154: Integrate and process the warning trigger time, warning level, maintenance urgency, and maintenance resource requirement list in a formatted manner to generate a structured diagnostic report containing a title, fault overview, warning strategy, and maintenance guidelines, and add timestamp information and diagnostic system identification to the structured diagnostic report to generate a traceable device diagnostic report.

[0218] Integrate and process the warning trigger time, warning level, maintenance urgency, and maintenance resource requirement list according to the set format. First, determine the structure of the diagnostic report, which generally includes parts such as a title, fault overview, warning strategy, and maintenance guidelines.

[0219] The title should concisely summarize the theme of the report, such as "Fault Diagnostic Report of XX Equipment". The fault overview part briefly introduces the basic situation of the fault, including information such as the device where the fault occurred and the fault trigger conditions. The warning strategy part details the warning trigger time and warning level, providing timely warning information for operation and maintenance personnel. The maintenance guidelines part provides specific maintenance suggestions and operation steps based on the maintenance urgency and maintenance resource requirement list.

[0220] When integrating the above information, pay attention to the logic and readability of the information. For example, sort the warning strategy and maintenance guidelines according to the urgency level, and list the information with a higher urgency level first.

[0221] Add timestamp information and diagnostic system identification to the structured diagnostic report. The timestamp information records the specific time when the report was generated, which helps to track the chronological order of fault diagnosis; the diagnostic system identification clarifies the source of the device diagnostic report, facilitating the traceability and verification of the diagnostic results. By adding these information, a traceable device diagnostic report is generated to ensure the accuracy and reliability of the device diagnostic report.

[0222] Step S155: Transmit the device diagnostic report to the message queue of the intelligent operation and maintenance terminal through the internal data interface of the power plant, trigger the message push service, and send the device diagnostic report to the preset mobile terminal of the operation and maintenance personnel.

[0223] The internal data interface of the power plant is a channel for data transmission between different systems within the power plant. Through this interface, the generated device diagnostic report is transmitted to the message queue of the intelligent operation and maintenance terminal. The message queue is a middleware for storing and managing messages, which can ensure the orderly transmission and reliable reception of messages.

[0224] When the device diagnostic report enters the message queue, the message push service is triggered. The message push service is a mechanism that sends messages to target users in a timely manner. The preset mobile terminals of operation and maintenance personnel are the terminal devices that receive the device diagnostic report, such as smartphones, tablets, etc.

[0225] The message push service will send the device diagnostic report to the mobile terminal of the operation and maintenance personnel according to the preset rules. Multiple push methods can be adopted, such as text messages, in-app notifications, etc. After receiving the device diagnostic report, the operation and maintenance personnel can timely understand the fault situation of the device and handle it according to the warning strategy and maintenance guidelines in the report, thereby improving the operation and maintenance efficiency and reliability of power plant equipment.

[0226] Figure 2 FIG. shows a schematic diagram of exemplary hardware and software components of an artificial intelligence-based power plant system equipment fault diagnosis system 100 that can implement the ideas of the present application. For example, the processor 120 can be used on the artificial intelligence-based power plant system equipment fault diagnosis system 100 and is used to execute the functions in the present application.

[0227] The artificial intelligence-based power plant system equipment fault diagnosis system 100 can be a general-purpose server or a special-purpose server, both of which can be used to implement the artificial intelligence-based power plant system equipment fault diagnosis method of the present application. Although only one server is shown in the present application, for convenience, the functions described in the present application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0228] For example, the artificial intelligence-based power plant system equipment fault diagnosis system 100 can include a network port 110 connected to the network, one or more processors 120 for executing program instructions, a communication bus 130, and different forms of storage media 140, such as disks, ROM, or RAM, or any combination thereof. Exemplarily, the artificial intelligence-based power plant system equipment fault diagnosis system 100 can also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The method of the present application can be implemented according to these program instructions. The artificial intelligence-based power plant system equipment fault diagnosis system 100 also includes an I / O interface 150 between the computer and other input / output devices.

[0229] In addition, an embodiment of the present invention also provides a readable storage medium, in which computer-executable instructions are preset. When the processor executes the computer-executable instructions, the above-mentioned artificial intelligence-based power plant system equipment fault diagnosis method is implemented.

[0230] It should be noted that, in order to simplify the description of the disclosure of the present invention and thus help the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, sometimes a plurality of features are combined into one embodiment, drawing or description thereof.

Claims

1. A method for diagnosing faults in power plant system equipment based on artificial intelligence, characterized in that, The method includes: Obtaining a continuous operation status data stream of key equipment in a power plant, where the continuous operation status data stream includes a sequence of device real-time sensing signals and historical fault association records; Performing feature extraction on the continuous operation status data stream to generate a fusion feature set reflecting the coupling relationship of the device operation status; Invoking a pre-trained fault dynamic diagnosis model to perform state evolution analysis processing on the fusion feature set, and generating a diagnostic intermediate result including potential fault nodes and fault evolution paths; Performing fault attribute parsing processing based on the diagnostic intermediate result to generate fault diagnosis information including fault triggering conditions, fault development trends, and fault impact scopes; Generating a device diagnostic report including a warning strategy and a maintenance priority according to the fault diagnosis information, and transmitting the device diagnostic report to the intelligent operation and maintenance terminal of the power plant.

2. The method for diagnosing faults in power plant system equipment based on artificial intelligence according to claim 1, wherein The obtaining of the continuous operation status data stream of key equipment in the power plant, where the continuous operation status data stream includes a sequence of device real-time sensing signals and historical fault association records, includes: Collecting a sequence of real-time sensing signals from the distributed control system of the power plant, where the sequence of real-time sensing signals includes sensing signal measurement values at continuous time points; Extracting historical fault association records aligned with the time stamps of the sequence of real-time sensing signals from the device management system of the power plant, where the historical fault association records include device operation parameters, abnormal phenomenon descriptions, and corresponding maintenance measures when the faults occur; Performing time dimension alignment processing on the sequence of real-time sensing signals and the historical fault association records so that the real-time signals and historical records at the same time node have consistent time identifiers; Grouping the aligned sequence of real-time sensing signals and historical fault association records according to device function modules to generate a continuous operation status data stream including a combustion system data stream, a steam turbine system data stream, and an electrical system data stream; Performing integrity verification processing on the continuous operation status data stream, and removing record units with time breaks or data missing to generate a continuous operation status data stream meeting the time continuity requirement.

3. The method for diagnosing faults of power plant system equipment based on artificial intelligence according to claim 1, wherein, The performing of feature extraction on the continuous operation status data stream to generate a fusion feature set reflecting the coupling relationship of the device operation status includes: Inputting the continuous operation status data stream into a time series feature extraction module, and performing time-dependent relationship modeling processing on the sequence of single-device real-time sensing signals through a long short-term memory network to generate a time series feature vector including historical state memories; Inputting the continuous operation status data stream into a spatial association analysis module, and performing collaborative operation status modeling processing on the physical connection relationship between devices through a graph neural network to generate a spatial association feature vector reflecting the linkage state of the devices, where the physical connection relationship is determined by a device association relationship matrix; Inputting the continuous operation status data stream into a log semantic parsing module, and performing text feature extraction processing on the historical fault association records through a pre-trained natural language processing model to generate a log feature vector including fault semantic information; Inputting the time series feature vector, the spatial association feature vector, and the log feature vector into a feature fusion module, and performing importance weighting processing on the feature vectors of each dimension through an attention mechanism to calculate the weight coefficients of the feature vectors of each dimension; Generate a dynamic weight matrix according to the weight coefficients. After performing weighted processing on the time series feature vector, the spatial correlation feature vector, and the log feature vector, splice the weighted feature vectors along the feature dimension to generate an intermediate fusion feature vector with dynamic weight allocation; Perform dimension unification processing on the intermediate fusion feature vector, and map different modality features to the feature space of the same dimension through linear transformation to generate a fusion feature set.

4. The method for diagnosing faults of power plant system equipment based on artificial intelligence according to claim 3, characterized in that, After splicing the weighted feature vectors along the feature dimension to generate an intermediate fusion feature vector with dynamic weight allocation, the method further includes: Obtain the operation stage identifier of the current device, where the operation stage identifier is determined by the change trend in the time series feature vector of the current device, and the change trend includes an upward trend, a stable trend, and a downward trend; Extract the stage feature verification record matching the operation stage identifier from the historical feature verification database, where the stage feature verification record includes the verification validity information of the time series feature vector, the spatial correlation feature vector, and the log feature vector in each operation stage; Calculate the feature validity score of each feature vector in the current operation stage according to the stage feature verification record, where the feature validity score is determined by the correct recognition rate of the feature vector in the historical fault samples in the corresponding stage; Adjust the weight coefficients in the dynamic weight matrix based on the feature validity score to generate a corrected weight matrix including stage validity constraints; Perform secondary weighted processing on the time series feature vector, the spatial correlation feature vector, and the log feature vector using the corrected weight matrix; Re-splice the secondary weighted feature vectors along the feature dimension to generate a stage-adapted fusion feature vector reflecting the feature validity of the current operation stage as the updated intermediate fusion feature vector.

5. The method for diagnosing faults in power plant system equipment based on artificial intelligence according to claim 1, wherein The step of calling the pre-trained fault dynamic diagnosis model to perform state evolution analysis processing on the fusion feature set to generate a diagnostic intermediate result including fault potential nodes and fault evolution paths includes: Input the fusion feature set into the initial state encoding layer of the fault dynamic diagnosis model, and map it to the device initial operation state vector through a fully connected network; Input the initial operation state vector into the time step deduction layer of the fault dynamic diagnosis model, and simulate the evolution process of the device state within continuous time steps through a recurrent neural network; In the recurrent neural network, update the hidden state through the input gate, forget gate, and output gate mechanisms to generate the state transition vector at each time step; Input the state transition vector into the abnormal trigger detection layer of the fault dynamic diagnosis model to identify the abnormal state nodes in the state transition process; Input the abnormal state nodes into the path tracing layer of the fault dynamic diagnosis model, and determine the key fault evolution path from the initial operation state vector to the abnormal state nodes through the backtracking algorithm; Input the key fault evolution path into the fault impact range evaluation layer of the fault dynamic diagnosis model, and calculate the fault impact range information of the abnormal state nodes on the associated devices through the device association relationship matrix; Integrate the abnormal state nodes, the critical fault evolution paths, and the fault impact range information to generate a diagnostic intermediate result including potential fault nodes and fault evolution paths.

6. The method for diagnosing faults of power plant system equipment based on artificial intelligence according to claim 5, characterized in that, The simulation of the device state evolution process within continuous time steps by the recurrent neural network includes: Extract the key operating parameters from the real-time sensing signal sequence of the device, where the key operating parameters include temperature parameters, pressure parameters, and flow parameters; After standardizing the key operating parameters, obtain the parameter correlation relationships corresponding to the key operating parameters from the device operation characteristic database, where the parameter correlation relationships include the coupled change law of temperature and pressure, and the co-variation mode of pressure and flow; Construct a parameter co-variation index according to the parameter correlation relationships, where the parameter co-variation index is determined by the synchronous change degree of the key operating parameters; Input the parameter co-variation index into the hidden state update function of the recurrent neural network to generate a state update coefficient including parameter co-variation constraints; During the calculation of the input gate of the recurrent neural network, multiply the state update coefficient by the activation value of the input gate to generate a co-variation constraint input gate value; During the calculation of the forget gate of the recurrent neural network, multiply the state update coefficient by the activation value of the forget gate to generate a co-variation constraint forget gate value; During the calculation of the output gate of the recurrent neural network, multiply the state update coefficient by the activation value of the output gate to generate a co-variation constraint output gate value; Use the co-variation constraint input gate value, the co-variation constraint forget gate value, and the co-variation constraint output gate value to update the hidden state of the recurrent neural network to generate a state transition vector reflecting the influence of parameter co-variation; Record the state transition vectors at each time step to form a state transition sequence containing parameter co-variation information, and the state transition sequence is used for subsequent abnormal state node identification and fault evolution path tracking.

7. The method for diagnosing faults of power plant system equipment based on artificial intelligence according to claim 1, wherein The fault attribute analysis process based on the diagnostic intermediate result to generate fault diagnosis information including fault trigger conditions, fault development trends, and fault impact ranges includes: Extract the abnormal state nodes from the diagnostic intermediate result and query the historical fault cases matching the abnormal state nodes from the historical fault database; Analyze the device operation parameter records of the historical fault cases and extract the device operation parameter sequence before the occurrence of the abnormal state nodes; Calculate the mean, maximum, and minimum values of the device operation parameter sequence to generate parameter distribution statistical features; Calculate the mutual information value between each operating parameter and the abnormal state node according to the parameter distribution statistical features, and filter out the operating parameters with mutual information values greater than the preset threshold as candidate key parameters; Calculate the feature importance scores of the candidate key parameters through the random forest algorithm, and select the top N operating parameters with the highest feature importance scores as the key operating parameters strongly related to the abnormal state nodes; Extract the value range of the key operating parameters when the abnormal state node occurs to generate a fault trigger condition including the parameter name and the value range. Extract the key fault evolution paths in the diagnostic intermediate results, analyze the continuity of state transitions and the degree of abnormal accumulation at each node in the key fault evolution paths, and generate a description of the fault development trend reflecting the fault development rate; Extract the fault impact scope information in the diagnostic intermediate results, analyze the importance levels of associated devices and the possibility of fault propagation, and generate a description of the fault impact scope including a list of associated devices and the degree of impact; Perform a structured integration process on the fault trigger conditions, the description of the fault development trend, and the description of the fault impact scope to generate fault diagnosis information including a text description and a data identifier.

8. The method for diagnosing faults of power plant system equipment based on artificial intelligence according to claim 1, characterized in that, Generating an equipment diagnosis report including a warning strategy and a maintenance priority according to the fault diagnosis information, and transmitting the equipment diagnosis report to the intelligent operation and maintenance terminal of the power plant, including: According to the fault trigger conditions in the fault diagnosis information, match the corresponding warning trigger timing and warning level from the pre-stored warning rule library; According to the description of the fault development trend in the fault diagnosis information, determine the urgency of maintenance operations according to the fault development rate; Query the maintenance resource requirement information of associated devices from the equipment inventory database to generate a list of maintenance resource requirements; Perform a format integration process on the warning trigger timing, warning level, maintenance urgency, and the list of maintenance resource requirements to generate a structured diagnosis report including a title, a fault overview, a warning strategy, and a maintenance guide, and add a timestamp information and a diagnosis system identifier to the structured diagnosis report to generate a traceable equipment diagnosis report; Transmit the equipment diagnosis report to the message queue of the intelligent operation and maintenance terminal through the internal data interface of the power plant, trigger the message push service, and send the equipment diagnosis report to the preset operation and maintenance personnel's mobile terminal; The determining the urgency of maintenance operations according to the fault development rate according to the description of the fault development trend in the fault diagnosis information includes: Parse the fault development rate information in the description of the fault development trend, where the fault development rate information is determined by the time interval of state transitions and the degree of abnormal accumulation in the key fault evolution paths; Match the fault development rate with the pre-stored mapping table, determine the urgency of maintenance operations according to the matching result, and perform an association process between the urgency and the importance level of the equipment to generate a final maintenance urgency identifier, where the pre-stored mapping table includes the description of the urgency corresponding to different rate intervals; Bind the maintenance urgency identifier to the list of maintenance resource requirements to generate maintenance guide information including the urgency and resource requirements.

9. The method for diagnosing faults of power plant system equipment based on artificial intelligence according to claim 5, wherein, Before mapping the fusion feature set to the initial operating state vector of the equipment through the initial state encoding layer of the fault dynamic diagnosis model, the method further includes: Obtain the operating mode identifier of the current equipment from the power plant operation management system, where the operating mode identifier includes a base load operating mode, a peak shaving operating mode, and a start-stop transition mode; Extract the time series feature vector from the fusion feature set, analyze the fluctuation frequency and amplitude change characteristics of the equipment real-time sensing signal sequence therein, and generate an operating mode feature descriptor; Match the operating mode feature descriptor with a preset mode feature template library to determine the actual operating mode of the current device; Obtain a corresponding set of evolution parameters from the model parameter configuration library according to the actual operating mode, where the set of evolution parameters includes the hidden state update rate parameter of the recurrent neural network and the anomaly detection threshold of the state transition vector; Input the set of evolution parameters into the time step deduction layer of the fault dynamic diagnosis model to adjust the activation function parameters of the input gate, forget gate, and output gate of the recurrent neural network.

10. An artificial intelligence-based power plant system equipment fault diagnosis system, characterized in that, It includes a processor and a memory. The memory is connected to the processor. The memory is used to store programs, instructions, or codes, and the processor is used to execute the programs, instructions, or codes in the memory to implement the method for fault diagnosis of power plant system equipment based on artificial intelligence according to any one of claims 1-9 above.

Images