Fault diagnosis method and system for water turbine guide vanes

By combining the guide vane attention-enhanced cyclic feature identification algorithm and the physical prior parameter adaptive optimization model with the graph structure fault association inference function, the problem of insufficient robustness in turbine guide vane fault diagnosis is solved, and high-accuracy early fault warning and risk ranking are achieved.

CN123109933APending Publication Date: 2026-11-17HUADIAN SICHUAN POWER GENERATION CO LTD WAWUSHAN BRANCH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611278086.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-21
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies fail to effectively utilize the deep coupling relationship between multidimensional parameters and physical operating boundaries in the fault diagnosis of turbine guide vanes, resulting in insufficient diagnostic robustness under high head variable operating conditions, making it difficult to meet the needs of early warning and refined diagnosis.

Method used

An attention-enhanced cyclic feature identification algorithm for guide vanes is used to mine implicit dependencies between parameters. Combined with an adaptive optimization model of physical prior parameters and a graph-structured fault association inference function, a guide vane fault association graph is constructed, and joint criterion calculation is performed to output a comprehensive risk index.

Benefits of technology

It improves the accuracy and engineering practicality of guide vane fault diagnosis, avoids false alarms and missed alarms, and realizes the quantitative ranking of membership under multiple fault modes and the identification of early weak faults.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123109933A_ABST
    Figure CN123109933A_ABST
Patent Text Reader

Abstract

This invention discloses a fault diagnosis method and system for turbine guide vanes, comprising: collecting a multi-dimensional state parameter set consisting of guide vane bearing temperature, opening degree, water level difference, pressure pulsation, torque, vibration acceleration, sway displacement, and volute inlet velocity; extracting temporal latent feature sequences by calling a guide vane attention-enhanced cyclic feature identification algorithm; inputting physical prior parameters into an adaptive optimization model for physical consistency correction and outputting dynamic feature parameters; constructing a fault association graph reasoning state membership vector using a graph-structured fault association reasoning function; calculating a comprehensive fault risk index vector using joint criteria and pushing it to a turbine fault time-series data intelligent analysis platform for storage and display. This invention improves the diagnostic accuracy and robustness of turbine guide vanes under multiple fault modes through temporal attention fusion, physical constraint correction, and graph topology-structured reasoning.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology for turbine guide vanes, and in particular to a fault diagnosis method and system for turbine guide vanes. Background Technology

[0002] Hydropower generator units operate under complex conditions of high water pressure, strong turbulence, and variable loads for extended periods. As the core component for regulating the flow rate of the turbine, the guide vanes are frequently subjected to alternating hydraulic loads and mechanical wear. Fault characteristics such as abnormal bearing temperature, opening drift, increased pressure pulsation, and excessive vibration directly affect the unit's operational safety and power generation efficiency. Accurate and efficient fault diagnosis of the guide vanes is a key aspect of the condition-based maintenance strategy for hydropower stations.

[0003] In existing technologies, conventional guide vane fault diagnosis methods mostly rely on a single physical parameter threshold alarm mechanism, which compares the monitored values ​​of temperature, pressure, vibration, and sway collected by various sensors with preset fixed thresholds to trigger an over-limit alarm; or uses classical time-frequency analysis methods to extract spectral features from vibration acceleration and pressure pulsation signals, and completes fault type identification by comparing with a typical fault frequency feature database; some technical solutions introduce shallow neural networks to perform nonlinear mapping fitting of parameters such as guide vane opening and torque, and use historical fault samples to train a classifier to achieve fault discrimination; another diagnostic architecture based on graph theory abstracts each measuring point as a topological node, constructs an association graph by calculating the statistical correlation between measuring points, and combines the graph clustering results to help locate the source of the anomaly.

[0004] However, existing technical solutions of this kind still have two shortcomings. First, the signal processing and feature extraction of each measuring point are independent and isolated, failing to utilize the deep coupling relationship between the multidimensional parameters of the guide vane in the temporal evolution process, and lacking a constraint mechanism of physical operating boundary on the diagnostic process, resulting in severe feature drift and insufficient diagnostic robustness under high head variable operating conditions. Second, the graph topology construction method is limited to static correlation measurement and fails to establish a structured reasoning path between the prior knowledge of fault modes and the dynamic features of real-time measuring points, resulting in fuzzy membership determination and failure of risk ranking in multi-fault concurrent scenarios, making it difficult to meet the actual engineering needs of early warning and refined diagnosis of turbine guide vane faults. Summary of the Invention

[0005] In order to overcome the shortcomings and deficiencies of the existing technology, the present invention provides a fault diagnosis method and system for water turbine guide vanes.

[0006] The technical solution adopted in this invention is a fault diagnosis method for turbine guide vanes, comprising the following steps: S1, collecting turbine guide vane bearing temperature, guide vane opening, upstream and downstream water level difference, flow channel pressure pulsation, guide vane torque, vibration acceleration, sway displacement, and volute inlet velocity to form a multi-dimensional state parameter set; S2, based on the multi-dimensional state parameter set, calling the guide vane attention-enhanced cyclic feature identification algorithm to mine the implicit dependencies between parameters at the temporal evolution level, so as to form a guide vane operating condition latent feature sequence at each time point after attention reweighting; S3, inputting the guide vane operating condition latent feature sequence into the physical prior parameter adaptive optimization model, based on the turbine operating side... S4. Boundary constraints are used to adaptively optimize the adjustable prior coefficients within the model, outputting the guide vane dynamic characteristic parameters corrected for physical consistency; S5. Using a graph-structured fault association inference function, a fault association graph is constructed with each measuring point of the guide vane as a graph node and the mutual information between measuring points as edge weights. At the graph topology level, the state membership vector of the current guide vane operating state belonging to each preset fault mode is inferred; S6. Based on the state membership vector and the corrected guide vane dynamic characteristic parameters, a joint criterion calculation is performed to obtain the guide vane fault comprehensive risk index vector; S7. The guide vane fault comprehensive risk index vector is pushed to the turbine fault time series data intelligent analysis platform for storage and display.

[0007] Furthermore, the expression for the guide vane attention-enhanced cyclic feature identification algorithm is as follows: , ,and Attention weight ,in This indicates the cyclic hidden state at time 1. express The hidden state of the cycle at any moment, Represents a non-linear activation function. This represents the hidden state transition weight matrix. This represents the input mapping weight matrix. A multidimensional vector of state parameters input at any given time. Indicates the total length of the timing window. Let represent the original input vector at time i. Let represent the normalized attention score for the j-th input position at time t. This represents the unnormalized attention energy at time t for the j-th input position. This represents the activation function. This represents the hidden state mapping weight matrix in attention computation. This represents the input mapping weight matrix in attention calculation. This represents the input vector at time j. Let represent the final attention weight for the i-th input position at time t. Let represent the normalized attention score for the m-th input position at time t. This represents the product operator.

[0008] Furthermore, the expression for the adaptive optimization model of the physical prior parameters is: ,in ,in This represents the vector of highest priority test coefficients after adaptive optimization. This represents the vector of adjustable prior coefficients to be optimized. This represents a mapping function based on physical priors. This represents the input latent feature sequence of guide vane operating conditions. This represents the physical reference vector derived from the boundary constraints of the turbine operation. Represents the regularization coefficient. This represents the regularization constraint term on the prior coefficient vector. Describe the domain space of the prior coefficient vector. This represents the gradient operator for calculating the prior coefficient vector. Represents the physical consistency loss function. This represents the integration operator. The square operation represents the 2-norm of a vector. This represents the summation operation of the deviations of each component of the prior coefficients from their initial values, where... This represents the total dimension of the prior coefficient vector. This represents the k-th prior coefficient component. This represents the initial reference value of the k-th prior coefficient.

[0009] Furthermore, the expression for the graph structure fault association inference function is: ,in This indicates the membership degree of the current guide vane operating state to the r-th preset fault mode. Represents the normalized mapping function. This represents the total number of nodes in the graph. This represents the product operator. This represents the number of edges included in the path between two nodes in the graph topology. This represents the correlation strength coefficient on the nth edge between the r-th failure mode prior center node and the q-th measurement point node. Let represent the mutual information matching coefficient between the r-th fault mode and the q-th measurement point, where the measurement point node corresponds to the physical measurement point position of each guide vane, the fault mode prior center node corresponds to the cluster center of each preset fault type in the feature space, the association strength coefficient is calculated by the transfer entropy between nodes, and the mutual information matching coefficient is derived from the joint probability distribution between the current feature of the measurement point and the typical feature of the fault mode.

[0010] Furthermore, the multidimensional state parameter set includes a guide vane bearing temperature sampling frequency of 100Hz, a guide vane opening obtained by a displacement sensor at a sampling frequency of 200Hz, an upstream and downstream water level difference collected by a differential pressure transmitter at a sampling frequency of 50Hz, a flow channel pressure pulsation picked up by a high-frequency pressure sensor at a sampling frequency of 500Hz, a guide vane torque measured by a torque sensor at a sampling frequency of 150Hz, a vibration acceleration obtained by a piezoelectric accelerometer at a sampling frequency of 1kHz, a swing displacement collected by an eddy current displacement sensor at a sampling frequency of 200Hz, and a volute inlet flow velocity obtained by an ultrasonic flow meter at a sampling frequency of 60Hz. The turbine fault time-series data intelligent analysis platform has a built-in industrial server with a graphics processor acceleration unit. The industrial server is equipped with a real-time time-series database for storing intermediate features and final indicators generated in each step. The graphics processor acceleration unit is used to accelerate the attention weight calculation process in the guide vane attention enhancement cyclic feature identification algorithm in parallel. The real-time time-series database persistently stores the guide vane fault comprehensive risk index vector with timestamp as the primary key and supports historical trend backtracking query.

[0011] Further, S2 includes the following sub-steps: S21, arranging all parameters of each moment in the multidimensional state parameter set in chronological order to form the original input sequence; S22, for the input vector at each moment in the original input sequence, calculating the unnormalized attention energy between the hidden state at that moment and the input vectors at each historical moment; S23, performing a normalization operation on all unnormalized attention energies to obtain the attention weight distribution, and using the attention weights to weight and aggregate the input vectors at each historical moment to form the context feature vector; S24, inputting the context feature vector, the current moment's input vector, and the previous moment's hidden state into the cyclic update gating unit, and outputting the guide vane working condition hidden features at that moment after attention reweighting.

[0012] Further, step S3 includes the following sub-steps: S31, reading the upper and lower limits of guide vane opening, the safe threshold of bearing temperature, and the allowable range of flow channel pressure from the turbine operating boundary constraints as physical reference vectors; S32, substituting the hidden characteristic sequence of guide vane operating conditions into the mapping function of the physical prior parameter adaptive optimization model, and calculating the deviation between the model output under the current prior coefficients and the physical reference vector; S33, performing gradient descent iterative update on the prior coefficient vector based on the deviation, and using the sum of the squares of the deviations of each component of the prior coefficients from their initial reference values ​​as regularization constraints to participate in iterative optimization; S34, terminating the iteration when the deviation is less than a preset convergence threshold, and outputting the guide vane dynamic characteristic parameters corrected for physical consistency under the current prior coefficients.

[0013] Further, S4 includes the following sub-steps: S41, using the guide vane bearing temperature measurement point, guide vane opening measurement point, upstream and downstream water level difference measurement point, flow channel pressure pulsation measurement point, guide vane torque measurement point, vibration acceleration measurement point, sway displacement measurement point, and volute inlet velocity measurement point as graph nodes, initialize all graph nodes to a disconnected state; S42, calculate the mutual information of the state sequence between every two graph nodes, establish an edge connection between node pairs with mutual information greater than the association threshold, and use the mutual information value as the initial weight of the edge; S43, map the prior feature vectors of each preset fault mode to the graph node feature space to form the prior center node of each fault mode, and calculate the transfer entropy between each measurement point node and each prior center node as the association strength coefficient; S44, based on the association strength coefficient between the measurement point node and the prior center node and the mutual information matching coefficient between the current feature of the measurement point node and the typical feature of the fault mode, calculate the state membership vector of each fault mode through the graph structure fault association inference function.

[0014] Further, step S5 includes the following sub-steps: S51, comparing each membership value in the state membership vector with a preset membership alarm threshold to filter out a set of candidate fault modes whose membership exceeds the alarm threshold; S52, for each fault mode in the candidate fault mode set, extracting the feature components associated with the fault mode from the corrected guide vane dynamic feature parameters to form a fault feature subset; S53, weighting and fusing the amplitude deviation of the fault feature subset with the state membership value to obtain a single risk score for each candidate fault mode; S54, arranging the single risk scores of all candidate fault modes according to the fault mode number to form a guide vane fault comprehensive risk index vector, where the numerical value of each element in the vector represents the risk severity of the corresponding fault mode at the current moment.

[0015] A fault diagnosis system for turbine guide vanes includes: a guide vane multi-dimensional state parameter synchronous acquisition unit, used to synchronously sample and assemble guide vane bearing temperature, guide vane opening, upstream and downstream water level difference, flow channel pressure pulsation, guide vane torque, vibration acceleration, sway displacement, and volute inlet velocity into a multi-dimensional state parameter set; the output of this unit is connected to a guide vane temporal latent feature extraction unit; a guide vane temporal latent feature extraction unit, which incorporates a guide vane attention-enhanced cyclic feature identification algorithm, used to receive the multi-dimensional state parameter set and mine the implicit dependencies between temporal parameters to form a guide vane operating condition latent feature sequence; the output of this unit is connected to a physical consistency dynamic feature correction unit; and a physical consistency dynamic feature correction unit, which incorporates a physical prior parameter adaptive optimization model, used to receive the guide vane operating condition latent feature sequence and adjust the adjustable prior coefficients according to the operating boundary constraints. The unit performs adaptive optimization to output guide vane dynamic characteristic parameters corrected for physical consistency. Its output is connected to the fault graph reasoning membership calculation unit. This unit has a built-in graph-structured fault association reasoning function, used to construct a fault association graph with each measuring point of the guide vane as a graph node and infer the state membership vector of the current operating state belonging to each preset fault mode. Its output is connected to the joint criterion risk comprehensive assessment unit. This unit receives the state membership vector and the corrected guide vane dynamic characteristic parameters and performs joint criterion calculation to obtain the guide vane fault comprehensive risk index vector. Its output is connected to the turbine fault time-series data intelligent analysis platform interface unit. This interface unit receives, stores, and displays the guide vane fault comprehensive risk index vector.

[0016] Beneficial Effects: This invention proposes a fault diagnosis method and system for turbine guide vanes. By employing a guide vane attention-enhanced cyclic feature identification algorithm, it fully exploits the implicit dependencies between various physical parameters at the temporal evolution level, enabling dynamic adjustment of feature representation according to operating conditions. This avoids false alarms and missed alarms caused by single-parameter threshold alarms in high-head, variable-load scenarios. An adaptive optimization model for physical prior parameters is introduced, embedding turbine operating boundary constraints into the feature correction process. This ensures that diagnostic results remain within a reasonable physical space, improving diagnostic robustness under varying operating conditions. A graph-structured fault association inference function is used to dynamically construct an association graph based on the mutual information of measurement points. The prior central nodes of fault modes are incorporated into the graph topology for structured inference, achieving quantified ranking of membership metrics under multiple fault modes and enhancing the sensitivity of identifying early, weak faults. Finally, a comprehensive risk index vector is calculated and output through joint criteria and stored and displayed on a turbine fault time-series data intelligent analysis platform. This provides hydropower station operation and maintenance personnel with intuitive and hierarchical risk-oriented decision-making basis, improving the accuracy and engineering practicality of guide vane fault diagnosis. Attached Figure Description

[0017] Figure 1This is a flowchart illustrating the overall process of the method of the present invention. Figure 2 This is a flowchart of method step S2 of the present invention; Figure 3 This is a flowchart of method step S3 of the present invention; Figure 4 This is a flowchart of method step S4 of the present invention; Figure 5 This is a flowchart of step S5 of the method of the present invention; Figure 6 This is a visualization of the system interface of the present invention. Detailed Implementation

[0018] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] like Figure 1 As shown, the fault diagnosis method for turbine guide vanes includes the following steps: S1, collecting turbine guide vane bearing temperature, guide vane opening, upstream and downstream water level difference, flow channel pressure pulsation, guide vane torque, vibration acceleration, sway displacement, and volute inlet velocity to form a multi-dimensional state parameter set; S2, based on the multi-dimensional state parameter set, calling the guide vane attention-enhanced cyclic feature identification algorithm to mine the implicit dependencies between parameters at the temporal evolution level, so as to form a guide vane operating condition latent feature sequence at each time point after attention reweighting; S3, inputting the guide vane operating condition latent feature sequence into the physical prior parameter adaptive optimization model, based on the turbine operating boundary constraints. Adaptive optimization is performed on the adjustable prior coefficients within the model to output the guide vane dynamic characteristic parameters corrected for physical consistency; S4, using a graph-structured fault association inference function, a fault association graph is constructed with each measuring point of the guide vane as a graph node and the mutual information between measuring points as edge weights, and the state membership vector of the current guide vane operating state belonging to each preset fault mode is inferred at the graph topology level; S5, based on the state membership vector and the corrected guide vane dynamic characteristic parameters, a joint criterion calculation is performed to obtain the guide vane fault comprehensive risk index vector; S6, the guide vane fault comprehensive risk index vector is pushed to the turbine fault time series data intelligent analysis platform for storage and display.

[0020] In step S1, the turbine guide vane body and adjacent flow channel area are the monitoring objects. A platinum resistance temperature sensor is embedded in the outer wall of the guide vane bearing to continuously collect the bearing temperature at a sampling frequency of 100Hz. A magnetostrictive displacement sensor is installed on the side of the guide vane servo actuator stroke rod to obtain the guide vane opening at a sampling frequency of 200Hz. Piezoresistive pressure transmitters are arranged at the inlet end of the volute and the outlet end of the tailrace pipe to synchronously collect the upstream and downstream water level difference at a sampling frequency of 50Hz. A high-frequency piezoelectric pressure sensor is installed on the flow channel wall at the guide vane outlet to pick up the flow channel pressure pulsation at a sampling frequency of 500Hz. A strain gauge torque sensor is connected in series at the guide vane pivot end to measure the guide vane torque at a sampling frequency of 150Hz. Piezoelectric accelerometers are fixed vertically and horizontally on the guide vane bearing housing to obtain vibration acceleration at a sampling frequency of 1kHz. An eddy current displacement sensor is installed at the guide vane journal to detect swing displacement at a sampling frequency of 200Hz. An ultrasonic time-of-flight flow meter is installed on the straight pipe section of the volute inlet to measure the volute inlet velocity at a sampling frequency of 60Hz. The analog voltage or current signals output by all sensors are synchronously triggered for sampling by a data acquisition card and converted into digital quantities. The sampling time covers at least three complete rotation cycles of the turbine. The sampling data from each channel are aligned with a unified timestamp and assembled into a multi-dimensional state parameter set including eight physical dimensions for each moment. This parameter set serves as the common input basis for subsequent steps.

[0021] In step S2, the 8-dimensional observation vectors from all time points in the multidimensional state parameter set are arranged in ascending order to form the original input sequence. For each observation vector at the current time point in the sequence, the guide leaf attention-enhanced cyclic feature identification algorithm is invoked to first calculate the unnormalized attention energy between the current time point's cyclic hidden state and the observation vectors from previous time points. The value of this energy represents the degree of influence of previous time points on the feature representation of the current time point. Subsequently, a global normalization operation is performed on the unnormalized attention energy of all historical time points, so that the attention weight values ​​of each historical time point are distributed between 0 and 1 and sum to 1. The attention weights obtained after normalization are used to evaluate the original observation vectors at each historical time point. The measured vectors are weighted and summed to obtain the context feature vector at the current moment. This context feature vector is essentially a condensed expression of all historical moment information after reweighting based on attention weights. Then, the context feature vector, the current moment observation vector, and the previous moment's cyclic hidden state are input into the gated cyclic unit. The update gate and reset gate inside the gated cyclic unit control the forgetting ratio of historical information and the fusion ratio of the current input, respectively. After nonlinear mapping by the gated cyclic unit, the hidden features of the guide vane condition at that moment are output. The above attention weight calculation and gating update process are repeated for each moment in the sequence, and finally a sequence of hidden features of the guide vane condition covering the entire time window is formed.

[0022] When implementing step S3, the turbine's operating boundary constraints are pre-stored. These constraints include an upper limit of 100% and a lower limit of 0% for the guide vane opening, an upper limit of 70℃ for the bearing temperature safety threshold, and an upper limit of 1.2MPa and a lower limit of 0.2MPa for the flow channel pressure. These boundary values ​​are combined into a physical reference vector. The hidden characteristic sequence of the guide vane operating condition output from S2 is substituted into the physical consistency mapping function in the physical prior parameter adaptive optimization model. This mapping function includes multiple adjustable prior coefficients, each corresponding to an empirical correction factor for a physical law. Initially, each prior coefficient is assigned a value of 1.0. The mapping function output under the current prior coefficient value is then compared with the physical parameters. The test vector is compared dimension by dimension and the deviation between the two is calculated. The prior coefficients are iteratively adjusted according to the gradient direction of the deviation relative to each prior coefficient. The step size of each iteration is set to 0.01. At the same time, the sum of the squares of the deviations between the current value and the initial value of 1.0 of each prior coefficient is added to the iterative objective function as a regularization penalty term to prevent the prior coefficients from deviating too much from the physical initial value. After every 100 iterations, the current deviation is recalculated and compared with the preset convergence threshold of 0.001. If the deviation is less than the threshold, the iteration stops. The optimized value of each prior coefficient is substituted into the mapping function and recalculated. The output of the mapping function is the guide vane dynamic characteristic parameter after physical consistency correction.

[0023] In step S4, eight measuring points—guide vane bearing temperature, guide vane opening, upstream and downstream water level difference, flow channel pressure pulsation, guide vane torque, vibration acceleration, sway displacement, and volute inlet velocity—are used as graph nodes. Initially, each node is independent and unconnected. Mutual information is calculated for the state time series between any two nodes. The mutual information value reflects the statistically significant degree of information sharing between the two measuring points. An association threshold of 0.3 is set. Undirected edges are established between node pairs with mutual information values ​​exceeding 0.3, using the mutual information value as the initial weight of the edge. For four predefined fault modes—bearing wear fault, guide vane opening jamming fault, flow channel foreign object blockage fault, and hydraulic imbalance fault—the prior feature vector of each fault mode is... Mapped to the feature space of graph nodes, four prior center nodes for each fault mode are formed in the feature space. Each prior center node is paired with eight measurement point nodes one by one. The transmission entropy between each pair of nodes is calculated as the correlation strength coefficient. This correlation strength coefficient quantitatively describes the information transmission efficiency of the dynamic feature changes of the measurement points to the center node of a specific fault mode. At the same time, the joint probability distribution entropy between the feature value of each measurement point node at the current time and the typical feature value of each fault mode is calculated as the mutual information matching coefficient. The correlation strength coefficient and the mutual information matching coefficient are multiplied by the measurement point dimension and then summed over all measurement points. The summation result is then converted into a probability value between 0 and 1 by a normalization mapping function. This probability value is the state membership degree of the fault mode. After calculation for the four preset fault modes, a state membership degree vector of length 4 is obtained.

[0024] In step S5, a membership alarm threshold of 0.5 is preset. Each membership value in the state membership vector obtained in S4 is compared with 0.5, and all fault modes with a membership greater than 0.5 are selected to form a candidate fault mode set. The candidate fault mode set may include one or more fault modes. For each fault mode in the candidate set, feature components related to the physical mechanism of the fault mode are extracted from the guide vane dynamic characteristic parameters output in S3 after physical consistency correction to form a fault feature subset. The feature components pre-associated to each fault mode are specified according to the fault physical model during system initialization. For example, bearing wear faults are associated with bearing temperature and vibration acceleration feature components, and opening jamming faults are associated with guide vane opening and guide vane torque feature components. For each fault mode in the candidate set, the fault mode is... The relative deviation between the current amplitude of each feature component in the fault feature subset and its respective normal operating amplitude baseline value is calculated as a percentage. This relative deviation is then weighted and fused with the state membership value of the fault mode. The weighting coefficient for the relative deviation is set to 0.4, and the weighting coefficient for the state membership value is set to 0.6. The result of the weighted fusion is used as the individual risk score for the fault mode. The higher the individual risk score, the greater the risk level of the fault mode. The individual risk scores of all fault modes in the candidate set are arranged in ascending order of fault mode number. For fault modes that are not included in the candidate set, their individual risk scores are set to 0. Finally, a guide vane fault comprehensive risk index vector of length 4 is formed. The value of each element in this vector directly represents the risk severity of the corresponding fault mode at the current moment.

[0025] In step S6, the comprehensive risk index vector of guide vane failure calculated in S5 is encapsulated into a data transmission message. This message includes a timestamp field, a fault mode number field, and a corresponding risk score field. The message is pushed to the turbine fault time-series data intelligent analysis platform via the industrial Ethernet protocol. This platform is deployed on an industrial server equipped with a graphics processing unit (GPU) and runs a real-time Linux system. The platform contains a real-time time-series database, which uses the timestamp as the primary key, the fault mode number as the tag, and the risk score as the numerical field. The received comprehensive risk index vector of guide vane failure is written into the database according to the timestamp order. Data storage files are created on a daily rolling basis to control the size of a single file. The platform front-end queries historical risk indicator data from a real-time time-series database according to a time range through a web visualization interface. The query results are plotted as line graphs on the same coordinate system to show the evolution curves of the risk scores of each fault mode over time. The curve colors are configured according to the fault mode type for easy differentiation. The latest risk score value at the current moment is highlighted in numerical form in the dashboard area at the top of the interface. Operation and maintenance personnel can use this platform to view the risk change trends and current risk levels of each fault mode in real time. The entire storage and display process is continuously executed in a loop until the user manually terminates the diagnostic task.

[0026] Preferably, the expression for the guide vane attention-enhanced cyclic feature identification algorithm is: , ,and Attention weight , and It presents a two-way nested dependency relationship. The generation process needs to rely on Participation in Unnormalized Attention Energy The solution operation, after exponential normalization, yields the normalized attention score. This value is further multiplied by the final attention weights and then incorporated into the summation term. The iterative calculation formula is used to deduce the current cyclic hidden state. Newly generated It will then serve as the next stage of computation. Re-enter the next moment The process of obtaining the attention score involves a cyclical exchange of computational parameters between the two entities. The normalized attention score is the core intermediate parameter connecting the cyclical hidden states at different times, and the iterative computation of attention weights and temporal hidden states is completed through the repeated transmission of these parameters. This indicates the cyclic hidden state at time 1. express The hidden state of the cycle at any moment, Represents a non-linear activation function. This represents the hidden state transition weight matrix. This represents the input mapping weight matrix. A multidimensional vector of state parameters input at any given time. Indicates the total length of the timing window. Let represent the original input vector at time i. Let represent the normalized attention score for the j-th input position at time t. This represents the unnormalized attention energy at time t for the j-th input position. This represents the activation function. This represents the hidden state mapping weight matrix in attention computation. This represents the input mapping weight matrix in attention calculation. This represents the input vector at time j. Let represent the final attention weight for the i-th input position at time t. Let represent the normalized attention score for the m-th input position at time t. This represents the product operator.

[0027] Specifically, the guide vane attention-enhanced cyclic feature identification algorithm is based on the physical fact that the influence of historical information on the current state is uneven in time-series modeling. During the operation of the turbine guide vane, parameter fluctuations closer to the current moment usually contain more direct fault symptom information. However, under specific operating conditions, the accumulation of anomalies from distant historical moments cannot be ignored. Therefore, an attention mechanism is needed to assign differentiated weights to each historical moment. In this algorithm, the current cyclic hidden state is jointly mapped from the previous hidden state, the current input vector, and the historical context features weighted by attention weights. The attention weights are not directly specified but calculated. The algorithm calculates the unnormalized attention energy and the matching score between the current hidden state and the inputs from each historical moment. The higher the score, the more important the historical moment is to the expression of the current feature. Then, the scores of all historical moments are exponentially normalized to obtain the attention score distribution. Finally, the attention weights are obtained by multiplying the attention scores of all positions before the current moment using a product operation. This ensures that the weight decay is smoother and more reasonable as the historical moment progresses. In the specific implementation, the dimension of the recurrent hidden state is set to 64, the dimension of the input mapping weight matrix is ​​64 rows and 8 columns, the dimension of the hidden state transition weight matrix is ​​64 rows and 64 columns, the activation function is the hyperbolic tangent function, the dimension of the mapping matrix in the attention calculation is 64 rows and 64 columns, the initial values ​​of all matrix elements are initialized randomly with uniform distribution, the update gate and reset gate in the gated recurrent unit are configured with weight matrices independently, and the loss function is trained on historical fault data through the time-series backpropagation algorithm until the loss function converges, and then the parameters of each matrix are fixed for the actual diagnosis process.

[0028] Preferably, the expression for the adaptive optimization model of the physical prior parameters is: ,in ,in This represents the vector of highest priority test coefficients after adaptive optimization. This represents the vector of adjustable prior coefficients to be optimized. This represents a mapping function based on physical priors. This represents the input latent feature sequence of guide vane operating conditions. This represents the physical reference vector derived from the boundary constraints of the turbine operation. Represents the regularization coefficient. This represents the regularization constraint term on the prior coefficient vector. Describe the domain space of the prior coefficient vector. This represents the gradient operator for calculating the prior coefficient vector. Represents the physical consistency loss function. This represents the integration operator. The square operation represents the 2-norm of a vector. This represents the summation operation of the deviations of each component of the prior coefficients from their initial values, where... This represents the total dimension of the prior coefficient vector. This represents the k-th prior coefficient component. This represents the initial reference value of the k-th prior coefficient.

[0029] Specifically, the physical prior parameter adaptive optimization model is based on the fundamental principle that the physical laws governing the operation of turbine guide vanes should be mutually constrained by data-driven features. Simply relying on data feature extraction may lead to diagnostic results deviating from the true physical state. Therefore, the optimization objective must simultaneously consider data fitting error and physical consistency penalty terms. The objective of solving the highest priority a priori coefficient vector is to find the coefficient combination within the priori coefficient domain space that minimizes the sum of squared deviations between the mapping function output and the physical reference vector while simultaneously satisfying regularization constraints. The sum of squared deviations measures the degree to which the current model output deviates from the physical boundary, while the regularization constraint prevents the prior coefficients from over-adapting to a specific operating condition and losing generalization ability. The regularization constraint is specifically defined as the integral of the square of the gradient norm 2 of the loss function within the priori coefficient vector domain space. This integral operation represents the total variation of the loss function over the entire domain as the prior coefficients change. A larger variation indicates that the model output is more sensitive to coefficient changes. Minimizing this integral value forces the optimization process to tend towards a physically more stable coefficient region, improving diagnostic stability under different head conditions. In the specific implementation, the prior coefficient vector dimension is set to 12-dimensional, corresponding to 12 kinds of guide vane physical experience correction factors. The initial value of each factor is 1.0. The domain space is a 12-dimensional hypercube ranging from 0.5 to 2.0. The regularization coefficient is 0.01. The gradient descent method is used for iterative optimization. The integral operation is approximated by Monte Carlo numerical integration in the domain space. The mean of the square of the gradient 2 norm of the loss function is calculated by sampling 1000 points and multiplied by the volume of the domain to obtain the approximate integral value.

[0030] Preferably, the expression for the graph structure fault association inference function is: ,in This indicates the membership degree of the current guide vane operating state to the r-th preset fault mode. Represents the normalized mapping function. This represents the total number of nodes in the graph. This represents the product operator. This represents the number of edges included in the path between two nodes in the graph topology. This represents the correlation strength coefficient on the nth edge between the r-th failure mode prior center node and the q-th measurement point node. Let represent the mutual information matching coefficient between the r-th fault mode and the q-th measurement point, where the measurement point node corresponds to the physical measurement point position of each guide vane, the fault mode prior center node corresponds to the cluster center of each preset fault type in the feature space, the association strength coefficient is calculated by the transfer entropy between nodes, and the mutual information matching coefficient is derived from the joint probability distribution between the current feature of the measurement point and the typical feature of the fault mode.

[0031] Specifically, the graph-structured fault association inference function is based on the theoretical framework that information transmission between nodes in a graph topology must simultaneously consider the topological path strength and the degree of node feature matching. In a multi-measuring-point system of a turbine guide vane, the occurrence of a certain fault mode usually generates a chain response at multiple measuring points. However, the sensitivity of different measuring points to the fault mode varies. Therefore, it is necessary to establish a complete inference channel from the measuring point node through the topological path to the prior center node of the fault mode. The physical meaning expressed by this formula is that the state membership degree of a certain preset fault mode is equal to the sum of the contributions of all measuring point nodes to the fault mode after normalization mapping. The contribution of each measuring point is calculated by multiplying the association strength on all topological paths between the measuring point and the prior center node of the fault mode by the mutual information matching coefficient between the measuring point and the fault mode. The product operation is for the association strength coefficients of each edge traversed by the topological path between two nodes. This is equivalent to the total path strength being equal to the cumulative product of the edge strengths, which conforms to the physical law of edge-by-edge attenuation when information propagates along a multi-hop path. In the specific implementation, the number of preset fault modes is set to 4, the total number of graph nodes is 12, including 8 measurement point nodes and 4 fault mode prior center nodes, the association strength threshold is set to 0.2, and paths below this threshold are ignored in the calculation to reduce noise interference, the embedding dimension is set to 3 when calculating the transfer entropy, the time delay is set to 2 sampling intervals, the mutual information matching coefficient is derived after calculating the joint probability distribution entropy by the kernel density estimation method, the normalization mapping function is selected as the flexible maximum function to map the contribution of all measurement points to a probability distribution between 0 and 1, and the sum of all elements in the final state membership vector is equal to 1.

[0032] Preferably, the multi-dimensional state parameter set includes a guide vane bearing temperature sampling frequency of 100Hz, a guide vane opening obtained by a displacement sensor at a sampling frequency of 200Hz, an upstream and downstream water level difference collected by a differential pressure transmitter at a sampling frequency of 50Hz, a flow channel pressure pulsation picked up by a high-frequency pressure sensor at a sampling frequency of 500Hz, a guide vane torque measured by a torque sensor at a sampling frequency of 150Hz, a vibration acceleration obtained by a piezoelectric accelerometer at a sampling frequency of 1kHz, a swing displacement collected by an eddy current displacement sensor at a sampling frequency of 200Hz, and a volute inlet flow velocity obtained by an ultrasonic flow meter at a sampling frequency of 60Hz. The turbine fault time-series data intelligent analysis platform has a built-in industrial server with a graphics processor acceleration unit. The industrial server is equipped with a real-time time-series database for storing intermediate features and final indicators generated in each step. The graphics processor acceleration unit is used to accelerate the attention weight calculation process in the guide vane attention enhancement cyclic feature identification algorithm in parallel. The real-time time-series database persistently stores the guide vane fault comprehensive risk index vector with timestamp as the primary key and supports historical trend backtracking query.

[0033] Preferred, such as Figure 2As shown, S2 includes the following sub-steps: S21, arranging all parameters of each time step in the multidimensional state parameter set in chronological order to form the original input sequence; S22, for the input vector at each time step in the original input sequence, calculating the unnormalized attention energy between the hidden state at that time step and the input vectors at each historical time step; S23, performing a normalization operation on all unnormalized attention energies to obtain the attention weight distribution, and using the attention weights to weight and aggregate the input vectors at each historical time step to form the context feature vector; S24, inputting the context feature vector, the current time step input vector, and the previous time step hidden state into the cyclic update gating unit, and outputting the guide vane working condition hidden state features at that time step after attention reweighting.

[0034] Specifically, S21 arranges the 8-dimensional observation vectors from all time points in the multi-dimensional state parameter set according to the sampling timestamps, forming an original input sequence of 3000 time points. Each time point's vector includes values ​​for eight physical dimensions: bearing temperature, guide vane opening, upstream and downstream water level difference, channel pressure pulsation, guide vane torque, vibration acceleration, sway displacement, and volute inlet velocity. S22, for the input vector at time point t in the original input sequence, performs a dot product operation between the cyclic hidden state at that time point and the input vectors from all historical time points to obtain unnormalized attention energy. This energy ranges from -10 to 10, representing the degree of matching between the current state and each historical time point. S23 inputs the unnormalized attention energy from all historical time points into an exponential normalization function, distributing the normalized attention score for each historical time point between 0 and 1, with a sum of 1. This score is then used to weight and sum the input vectors from each historical time point to obtain a context feature vector, which has the same 8-dimensional dimension as the input vector. S24 concatenates the context feature vector, the current input vector, and the previous time-cycle hidden state into an 80-dimensional vector, which is then input into the gated loop unit. The activation thresholds of the update gate and the reset gate inside the gated loop unit are set to 0.5 and 0.3, respectively. After nonlinear mapping, a 64-dimensional hidden state feature of the guide vane working condition is output. The above process is executed sequentially for all 3000 time steps to form a hidden state feature sequence of 3000 rows and 64 columns.

[0035] Preferred, such as Figure 3As shown, step S3 includes the following sub-steps: S31, reading the upper and lower limits of guide vane opening, the safe threshold of bearing temperature, and the allowable range of flow channel pressure from the turbine operating boundary constraints as physical reference vectors; S32, substituting the hidden characteristic sequence of guide vane operating conditions into the mapping function of the physical prior parameter adaptive optimization model, and calculating the deviation between the model output under the current prior coefficients and the physical reference vector; S33, performing gradient descent iterative updates on the prior coefficient vector based on the deviation, and using the sum of the squares of the deviations of each component of the prior coefficients from their initial reference values ​​as regularization constraints to participate in iterative optimization; S34, terminating the iteration when the deviation is less than a preset convergence threshold, and outputting the guide vane dynamic characteristic parameters corrected for physical consistency under the current prior coefficients.

[0036] Specifically, S31 reads the turbine's operational boundary constraints from the system configuration file, including the upper limit of guide vane opening (100%) and the lower limit of 0%, the bearing temperature safety threshold of 70 degrees Celsius, and the upper limit of flow channel pressure (1.2 MPa) and the lower limit of 0.2 MPa. These six boundary values ​​are combined into a 6-dimensional physical reference vector in the order of upper and lower limits of opening, temperature threshold, and pressure. S32 inputs the 3000-row, 64-column latent feature sequence output from S2 into the mapping function of the physical prior parameter adaptive optimization model. This mapping function includes 12 adjustable prior coefficients, all initially set to 1.0. The output of the mapping function is a 6-dimensional vector. The deviation is obtained by subtracting the output from the 6-dimensional physical reference vector dimension by dimension and calculating the sum of the squares of the differences in each dimension. S33 calculates the gradient direction by taking the partial derivatives of the 12 prior coefficients with respect to the deviation. The prior coefficients are then updated in the opposite direction of the gradient with a learning rate of 0.01. Simultaneously, the sum of the squares of the deviations between the current value and the initial value of 1.0 for each prior coefficient is calculated as a regularization penalty term, multiplied by 0.01, and added to the objective function for iteration. S34 compares the current deviation with the convergence threshold of 0.001 after every 100 iterations. The iteration terminates when the deviation is less than 0.001, and the optimized values ​​of the 12 prior coefficients are substituted into the mapping function for recalculation, outputting 6-dimensional dynamic feature parameters of the guide vane.

[0037] Preferred, such as Figure 4As shown, S4 includes the following sub-steps: S41, using the guide vane bearing temperature measurement point, guide vane opening measurement point, upstream and downstream water level difference measurement point, flow channel pressure pulsation measurement point, guide vane torque measurement point, vibration acceleration measurement point, sway displacement measurement point, and volute inlet velocity measurement point as graph nodes, initialize all graph nodes to a disconnected state; S42, calculate the mutual information of the state sequence between every two graph nodes, establish an edge connection between node pairs with mutual information greater than the association threshold, and use the mutual information value as the initial weight of the edge; S43, map the prior feature vectors of each preset fault mode to the graph node feature space to form the prior center node of each fault mode, and calculate the transfer entropy between each measurement point node and each prior center node as the association strength coefficient; S44, based on the association strength coefficient between the measurement point node and the prior center node and the mutual information matching coefficient between the current feature of the measurement point node and the typical feature of the fault mode, calculate the state membership vector of each fault mode through the graph structure fault association inference function.

[0038] Specifically, in S41, eight measuring points—bearing temperature, guide vane opening, upstream and downstream water level difference, flow channel pressure pulsation, guide vane torque, vibration acceleration, sway displacement, and volute inlet velocity—are established as graph nodes. The initial state of each node does not include any edge connections, and each node stores 3000 time-series state data points for the corresponding measuring point. In S42, for the 28 pairs of nodes (two pairs in total), the mutual information between each pair of node state sequences is calculated. The mutual information is discretized into 10 intervals using an equal-width binning method, with the mutual information value ranging from 0 to 1. An association threshold of 0.3 is set, and undirected edges are established between node pairs with mutual information values ​​greater than 0.3, using this mutual information value as the edge weight. If the mutual information values ​​of all node pairs are less than 0.3, the node pair with the highest mutual information value is retained to ensure graph connectivity. S43 maps the prior feature vectors of the four preset fault modes to the graph node feature space to form four prior center nodes. The propagation entropy of 32 pairs of nodes between the eight measurement point nodes and the four prior center nodes is calculated, with an embedding dimension of 3 and a time delay of two sampling intervals. The propagation entropy value is used as the association strength coefficient. S44 multiplies the association strength coefficient between each pair of measurement point nodes and the prior center node by the mutual information matching coefficient, sums the results over the eight measurement points, and then normalizes the results using a flexible maximum function to obtain the membership degrees of the four fault modes. These are combined into a four-dimensional state membership vector.

[0039] Preferred, such as Figure 5As shown, step S5 includes the following sub-steps: S51, comparing each membership value in the state membership vector with a preset membership alarm threshold to filter out a set of candidate fault modes whose membership exceeds the alarm threshold; S52, for each fault mode in the candidate fault mode set, extracting the feature components associated with the fault mode from the corrected guide vane dynamic feature parameters to form a fault feature subset; S53, weighting and fusing the amplitude deviation of the fault feature subset with the state membership value to obtain a single risk score for each candidate fault mode; S54, arranging the single risk scores of all candidate fault modes according to the fault mode number to form a guide vane fault comprehensive risk index vector, where the numerical value of each element in the vector represents the severity of the risk of the corresponding fault mode at the current moment.

[0040] Specifically, S51 compares each membership value in the 4-dimensional state membership vector output by S4 with the preset membership alarm threshold of 0.5, and filters out the fault mode numbers with membership values ​​exceeding 0.5 to form a candidate fault mode set. If all membership values ​​are less than 0.5, the candidate set is empty, and a 4-dimensional risk index vector of all zeros is directly output and subsequent sub-steps are skipped. S52 extracts the pre-associated feature components of each fault mode from the 6-dimensional guide vane dynamic feature parameters output by S3. The bearing wear fault is associated with temperature and vibration acceleration components (2-dimensional), the opening jamming fault is associated with opening and torque components (2-dimensional), the flow channel blockage fault is associated with water level difference and pressure pulsation components (2-dimensional), and the hydraulic imbalance fault is associated with swing and velocity components (2-dimensional). The feature components associated with each fault mode are pre-stored in the configuration table during system initialization. S53 For each fault mode in the candidate set, extract the relative deviation between the current amplitude of its associated feature component and the baseline amplitude of normal operation. The baseline value is determined by the average value of each feature component within the first 600 seconds of stable turbine operation. Multiply the relative deviation by a weighting coefficient of 0.4 and add the membership value multiplied by a weighting coefficient of 0.6 to obtain a single-item risk score, with a score range between 0 and 1. S54 Arrange all four fault modes in order of number 1 to 4. Fault modes belonging to the candidate set are filled with the single-item risk score at the corresponding number position, and fault modes not belonging to the candidate set are filled with 0 at that position, ultimately forming a 4-dimensional guide vane fault comprehensive risk index vector.

[0041] A fault diagnosis system for turbine guide vanes includes: a guide vane multi-dimensional state parameter synchronous acquisition unit, used to synchronously sample and assemble guide vane bearing temperature, guide vane opening, upstream and downstream water level difference, flow channel pressure pulsation, guide vane torque, vibration acceleration, sway displacement, and volute inlet velocity into a multi-dimensional state parameter set; the output of this unit is connected to a guide vane temporal latent feature extraction unit; a guide vane temporal latent feature extraction unit, which incorporates a guide vane attention-enhanced cyclic feature identification algorithm, used to receive the multi-dimensional state parameter set and mine the implicit dependencies between temporal parameters to form a guide vane operating condition latent feature sequence; the output of this unit is connected to a physical consistency dynamic feature correction unit; and a physical consistency dynamic feature correction unit, which incorporates a physical prior parameter adaptive optimization model, used to receive the guide vane operating condition latent feature sequence and adjust the adjustable prior coefficients according to the operating boundary constraints. The unit performs adaptive optimization to output guide vane dynamic characteristic parameters corrected for physical consistency. Its output is connected to the fault graph reasoning membership calculation unit. This unit has a built-in graph-structured fault association reasoning function, used to construct a fault association graph with each measuring point of the guide vane as a graph node and infer the state membership vector of the current operating state belonging to each preset fault mode. Its output is connected to the joint criterion risk comprehensive assessment unit. This unit receives the state membership vector and the corrected guide vane dynamic characteristic parameters and performs joint criterion calculation to obtain the guide vane fault comprehensive risk index vector. Its output is connected to the turbine fault time-series data intelligent analysis platform interface unit. This interface unit receives, stores, and displays the guide vane fault comprehensive risk index vector.

[0042] like Figure 6The diagram shows a user interface visualization of a fault diagnosis system for turbine guide vanes. As the terminal display and interactive platform for the comprehensive risk index vector of guide vane faults, it relies on an industrial server with a built-in graphics processor acceleration unit and a real-time time-series database to achieve persistent data storage, historical backtracking queries, and visualization output. A vertical navigation bar is positioned on the left side of the interface, containing seven major functional modules: Operation Overview, Real-time Monitoring, Data Analysis, Fault Diagnosis, Historical Records, Alarm Management, and System Settings. This allows maintenance personnel to switch between different monitoring and analysis dimensions as needed. The main display area at the top of the interface features a multi-dimensional real-time monitoring panel for status parameters. This panel presents real-time values ​​for eight core status parameters: guide vane opening, guide vane bearing temperature, upstream and downstream water level difference, flow channel pressure pulsation, guide vane torque, vibration acceleration, sway displacement, and volute inlet velocity. The corresponding normal range thresholds for each parameter are also labeled, providing a clear picture of the guide vane's current operating condition. The middle section of the interface includes a comprehensive fault risk analysis area. One side displays the current status membership of preset fault modes such as guide vane jamming, guide vane wear, and guide vane deformation in a bar chart, quantifying the probability of each fault mode occurring. The other side displays the evolution trend of risk indicators over a continuous date range using a time-series line chart, aligning with the design of a real-time time-series database that supports historical trend backtracking. The right side of the interface features a real-time alarm list, recording the trigger times of abnormal events such as high vibration acceleration and flow channel pressure pulsation approaching the upper limit, providing real-time push notifications of fault symptoms. The system status bar at the bottom of the interface displays operational information such as equipment operation evaluation conclusions, data update time, and collection frequency. The overall interface integrates multi-source monitoring data, fault membership, risk trends, and alarm information, providing hydropower station operation and maintenance personnel with an intuitive and hierarchical basis for guide vane fault diagnosis and decision-making.

[0043] A fault diagnosis method and system for turbine guide vanes is proposed, constructing a complete progressive chain from synchronous acquisition of multi-dimensional parameters to temporal feature extraction, physical correction, graph topology reasoning, and joint evaluation. Each stage forms an orderly data flow and feature progression relationship, avoiding the fragmented processing mode where feature extraction and fault identification are disconnected in traditional schemes. The temporal-level attention weighting mechanism dynamically focuses on the historical evolution information of key measurement points according to the current operating conditions, enabling feature representation to possess both time memory and operating condition adaptability. The introduction of physical boundary constraints ensures that feature correction during the optimization process always follows the actual operating laws of the turbine, suppressing feature distortion in data-driven methods under non-design conditions. Graph topology reasoning uses prior knowledge of fault modes as structured nodes to participate in association strength calculation, enabling membership evaluation to have an interpretable reasoning path.

[0044] This invention addresses the problem of fragmented signal processing at various measurement points and lack of physical constraints in traditional schemes. It employs a temporal attention mechanism to place all measurement point parameters within a unified temporal framework for joint feature mining. Simultaneously, an adaptive optimization model using physical prior parameters embeds operational boundary conditions as regularization constraints into the feature correction stage. This effectively captures the deep coupling relationships between multiple parameters, and the corrected features always adhere to physical consistency criteria, improving diagnostic robustness under high head and variable load conditions. Furthermore, addressing the issue of ambiguous membership determination in concurrent multi-fault scenarios caused by existing static graph topology construction, this invention establishes edge weight connections between measurement points using dynamic mutual information and introduces prior center nodes of fault modes to participate in graph structure reasoning. By calculating the correlation strength and matching coefficient between the current features of the measurement points and each prior center node, it outputs the state membership vector of each fault mode. This achieves quantitative differentiation and effective ranking of risk membership in concurrent multi-fault scenarios, eliminating the inherent limitation of ambiguous discrimination in overlapping fault mode areas in traditional static graphs.

[0045] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," "link," and "fix" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0046] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A fault diagnosis method for turbine guide vanes, characterized in that, Includes the following steps: S1 collects a multi-dimensional set of state parameters, including turbine guide vane bearing temperature, guide vane opening, upstream and downstream water level difference, flow channel pressure pulsation, guide vane torque, vibration acceleration, sway displacement, and volute inlet velocity. S2, based on the multi-dimensional state parameter set, calls the guide vane attention-enhanced cyclic feature identification algorithm to mine the implicit dependencies between parameters at the temporal evolution level, so as to form the guide vane working condition latent feature sequence at each time point after attention reweighting; S3, input the hidden feature sequence of the guide vane working condition into the physical prior parameter adaptive optimization model, perform adaptive optimization on the adjustable prior coefficients inside the model according to the turbine operating boundary constraints, and output the guide vane dynamic feature parameters after physical consistency correction. S4. Using the graph structure fault association reasoning function, a fault association graph is constructed with each measuring point of the guide vane as a graph node and the mutual information between measuring points as the edge weight. The current guide vane operating state belongs to the state membership vector of each preset fault mode at the graph topology level. S5. Based on the state membership vector and the modified guide vane dynamic characteristic parameters, a joint criterion calculation is performed to obtain the guide vane fault comprehensive risk index vector. S6 pushes the comprehensive risk index of guide vane failure to the turbine failure time sequence data intelligent analysis platform for storage and display.

2. The fault diagnosis method for turbine guide vanes according to claim 1, characterized in that, The expression for the guide vane attention-enhanced cyclic feature identification algorithm is as follows: , ,and Attention weight ,in This indicates the cyclic hidden state at time 1. express The hidden state of the cycle at any given moment. Represents a non-linear activation function. This represents the hidden state transition weight matrix. This represents the input mapping weight matrix. A multidimensional vector of state parameters input at any given time. Indicates the total length of the timing window. Let represent the original input vector at time i. Let represent the normalized attention score for the j-th input position at time t. This represents the unnormalized attention energy at time t for the j-th input position. This represents the activation function. This represents the hidden state mapping weight matrix in attention computation. This represents the input mapping weight matrix in attention calculation. This represents the input vector at time j. Let represent the final attention weight for the i-th input position at time t. Let represent the normalized attention score for the m-th input position at time t. This represents the product operator.

3. The fault diagnosis method for turbine guide vanes according to claim 1, characterized in that, The expression for the adaptive optimization model of the physical prior parameters is: ,in ,in This represents the vector of highest priority test coefficients after adaptive optimization. This represents the vector of adjustable prior coefficients to be optimized. This represents a mapping function based on physical priors. This represents the input latent feature sequence of guide vane operating conditions. This represents the physical reference vector derived from the boundary constraints of the turbine operation. Represents the regularization coefficient. This represents the regularization constraint term on the prior coefficient vector. Describe the domain space of the prior coefficient vector. This represents the gradient operator for calculating the prior coefficient vector. Represents the physical consistency loss function. This represents the integration operator. The square operation represents the 2-norm of a vector. This represents the summation operation of the deviations of each component of the prior coefficients from their initial values, where... This represents the total dimension of the prior coefficient vector. This represents the k-th prior coefficient component. This represents the initial reference value of the k-th prior coefficient.

4. The fault diagnosis method for turbine guide vanes according to claim 1, characterized in that, The expression for the graph structure fault association inference function is: ,in This indicates the membership degree of the current guide vane operating state to the r-th preset fault mode. Represents the normalized mapping function. This represents the total number of nodes in the graph. This represents the product operator. This represents the number of edges included in the path between two nodes in the graph topology. This represents the correlation strength coefficient on the nth edge between the r-th failure mode prior center node and the q-th measurement point node. Let represent the mutual information matching coefficient between the r-th fault mode and the q-th measurement point, where the measurement point node corresponds to the physical measurement point position of each guide vane, the fault mode prior center node corresponds to the cluster center of each preset fault type in the feature space, the association strength coefficient is calculated by the transfer entropy between nodes, and the mutual information matching coefficient is derived from the joint probability distribution between the current feature of the measurement point and the typical feature of the fault mode.

5. The fault diagnosis method for turbine guide vanes according to claim 1, characterized in that, The multidimensional state parameters are as follows: guide vane bearing temperature sampling frequency is 100Hz; guide vane opening is obtained by a displacement sensor at a sampling frequency of 200Hz; upstream and downstream water level difference is collected by a differential pressure transmitter at a sampling frequency of 50Hz; flow channel pressure pulsation is picked up by a high-frequency pressure sensor at a sampling frequency of 500Hz; guide vane torque is measured by a torque sensor at a sampling frequency of 150Hz; vibration acceleration is obtained by a piezoelectric accelerometer at a sampling frequency of 1kHz; sway displacement is collected by an eddy current displacement sensor at a sampling frequency of 200Hz; and volute inlet flow velocity is obtained by an ultrasonic flow meter at a sampling frequency of 60Hz. The turbine fault time-series data intelligent analysis platform has a built-in industrial server with a graphics processor acceleration unit. The industrial server is equipped with a real-time time-series database to store intermediate features and final indicators generated in each step. The graphics processor acceleration unit is used to accelerate the attention weight calculation process in the guide vane attention enhancement cyclic feature identification algorithm in parallel. The real-time time-series database persistently stores the guide vane fault comprehensive risk index vector with timestamp as the primary key and supports historical trend backtracking query.

6. The fault diagnosis method for turbine guide vanes according to claim 1, characterized in that, S2 includes the following steps: Arrange all parameters at each time step in the multidimensional state parameter set in chronological order to form the original input sequence; for the input vector at each time step in the original input sequence, calculate the unnormalized attention energy between the hidden state at that time step and the input vectors at each historical time step. Normalize all unnormalized attention energies to obtain the attention weight distribution, and use the attention weights to weight and aggregate the input vectors at each historical time point to form the context feature vector; The context feature vector, the current input vector, and the hidden state from the previous time step are input together into the cyclically updated gating unit, and the hidden features of the guide vane working condition at that time step are output after attention reweighting.

7. The fault diagnosis method for turbine guide vanes according to claim 1, characterized in that, S3 includes: The upper and lower limits of guide vane opening, the safe threshold of bearing temperature, and the allowable range of flow channel pressure in the boundary constraints of turbine operation are read as physical reference vectors. Substitute the latent feature sequence of the guide vane working condition into the mapping function of the adaptive optimization model of the physical prior parameters, and calculate the deviation between the model output and the physical reference vector under the current prior coefficients; The prior coefficient vector is updated iteratively by gradient descent based on the deviation, and the sum of the squares of the deviations of each component of the prior coefficient from its initial reference value is used as a regularization constraint to participate in the iterative optimization. The iteration terminates when the deviation is less than the preset convergence threshold, and the dynamic characteristic parameters of the guide vane corrected for physical consistency under the current prior coefficients are output.

8. The fault diagnosis method for turbine guide vanes according to claim 1, characterized in that, S4 includes the following steps: Using the guide vane bearing temperature measurement point, guide vane opening measurement point, upstream and downstream water level difference measurement point, flow channel pressure pulsation measurement point, guide vane torque measurement point, vibration acceleration measurement point, sway displacement measurement point, and volute inlet velocity measurement point as graph nodes, all graph nodes are initialized to a disconnected state; the mutual information of the state sequence between every two graph nodes is calculated, and edge connections are established between node pairs with mutual information greater than the association threshold, with the mutual information value as the initial weight of the edge; The prior feature vectors of each preset fault mode are mapped to the feature space of the graph node to form the prior center node of each fault mode. The transfer entropy between each measurement point node and each prior center node is calculated as the association strength coefficient. Based on the association strength coefficient between the measurement point node and the prior center node and the mutual information matching coefficient between the current feature of the measurement point node and the typical feature of the fault mode, the state membership vector of each fault mode is calculated through the graph structure fault association inference function.

9. The fault diagnosis method for turbine guide vanes according to claim 1, characterized in that, S5 includes: The membership values ​​in the state membership vector are compared with the preset membership alarm threshold to filter out a set of candidate fault modes whose membership exceeds the alarm threshold. For each fault mode in the candidate fault mode set, the feature components associated with the fault mode in the corrected guide vane dynamic feature parameters are extracted to form a fault feature subset. The magnitude deviation of the fault feature subset is weighted and fused with the state membership value to obtain a single risk score for each candidate fault mode. The individual risk scores of all candidate failure modes are arranged according to the failure mode number to form a comprehensive risk index vector for guide vane failure. The numerical value of each element in this vector represents the severity of the risk of the corresponding failure mode at the current moment.

10. A fault diagnosis system for turbine guide vanes, characterized in that, This system is applied to the fault diagnosis method for turbine guide vanes as described in claim 1, comprising: The guide vane multidimensional state parameter synchronous acquisition unit is used to synchronously sample and assemble the guide vane bearing temperature, guide vane opening, upstream and downstream water level difference, flow channel pressure pulsation, guide vane torque, vibration acceleration, swing displacement and volute inlet velocity into a multidimensional state parameter set. The output of this unit is connected to the guide vane temporal hidden state feature extraction unit. The guide vane temporal latent feature extraction unit has a built-in guide vane attention-enhanced cyclic feature identification algorithm, which is used to receive a multi-dimensional state parameter set and mine the implicit dependencies between temporal parameters to form a guide vane working condition latent feature sequence. The output of this unit is connected to the physical consistency dynamic feature correction unit. The physical consistency dynamic feature correction unit has a built-in physical prior parameter adaptive optimization model. It is used to receive the hidden feature sequence of the guide vane working condition and perform adaptive optimization on the adjustable prior coefficients according to the operating boundary constraints to output the guide vane dynamic feature parameters after physical consistency correction. The output of this unit is connected to the fault graph reasoning membership calculation unit. The fault diagram reasoning membership calculation unit has a built-in graph structure fault association reasoning function, which is used to construct a fault association graph with each measuring point of the guide vane as a graph node and reason the state membership vector of the current operating state belonging to each preset fault mode. The output of this unit is connected to the joint criterion risk comprehensive assessment unit. The joint criterion risk comprehensive assessment unit is used to receive the state membership vector and the corrected guide vane dynamic characteristic parameters and perform joint criterion calculation to obtain the guide vane fault comprehensive risk index vector. The output of this unit is connected to the interface unit of the intelligent analysis platform for turbine fault time series data. The interface unit of the intelligent analysis platform for time-series data of turbine failures is used to receive, store, and display the comprehensive risk index vector of guide vane failures.