Self-diagnosis method for turbine guide vane faults based on deep learning and graph neural networks

Abstract

The present invention discloses a method for self-diagnosing faults of a turbine guide vane based on deep learning and graph neural networks, including: S1, constructing a multi-sensor data set; S2, using a deep learning algorithm to identify key patterns and features in the multi-sensor data set; S3, constructing a graph neural network model; S4, inputting the features extracted in step S2 into the graph neural network model constructed in step S3 to analyze the overall operating state of the turbine guide vane; S5, according to the output of the graph neural network model, monitoring and analyzing the performance indicators of the turbine guide vane in real time; S6, applying a pre-trained fault diagnosis model to interpret the abnormal signals identified in step S5 and diagnose potential fault causes and fault types; S7, automatically generating a fault warning according to the diagnosis result of step S6. Through deep learning and graph neural network technologies, the present invention can process and analyze a large amount of data from multiple sensors in real time to ensure that the operating condition of the guide vane is monitored in real time and accurately.

Description

Technical Field

[0001] The present invention relates to the technical field of turbine diagnosis, and particularly to a self-diagnosis method for turbine guide vane faults based on deep learning and graph neural networks. Background Art

[0002] In a hydropower station, as a core device, the stable and efficient operation of a water turbine is directly related to the power generation efficiency and safety of the power station. The guide vane of the water turbine, as a key component, has a significant impact on the operation efficiency and stability of the entire unit. During operation, the guide vane bears complex hydrodynamic loads and is often at risk of various forms of structural damage such as wear, corrosion, and cracks. Therefore, timely and accurate diagnosis of the operating state and faults of the guide vane has become a key technical requirement for improving operation efficiency and ensuring equipment safety.

[0003] Traditional fault diagnosis techniques mainly rely on regular physical inspections or sensor data monitoring based on simple thresholds. These methods usually cannot handle the problems of large and complex amounts of data collected, especially when involving multiple sensors and data types. For example, sensor data such as vibration data, pressure indicators, temperature readings, etc. need to be comprehensively analyzed to accurately judge the health status of the guide vane. However, traditional methods have obvious deficiencies in data integration and real-time analysis and cannot effectively process and interpret these complex data relationships.

[0004] In addition, most existing fault diagnosis systems rely on predefined fault patterns and empirical rules, which not only limit the flexibility and adaptability of fault detection, but also often require professional maintenance teams for data analysis and fault diagnosis, resulting in high maintenance costs and delayed response times. When faced with unknown or atypical fault patterns, the effects of these systems are usually not good. Therefore, how to provide a self-diagnosis method for turbine guide vane faults based on deep learning and graph neural networks is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0005] An object of the present invention is to propose a self-diagnosis method for turbine guide vane faults based on deep learning and graph neural networks. Through deep learning and graph neural network technologies, the present invention can process and analyze a large amount of data from multiple sensors in real time to ensure that the operating condition of the guide vane is monitored in real time and accurately.

[0006] A self-diagnosis method for turbine guide vane faults based on deep learning and graph neural networks according to an embodiment of the present invention includes the following steps:

[0007] S1. Collect vibration data, pressure indicators, eddy current signal data, acoustic emission signal data, electromagnetic field change data, and 3D laser scan data generated during the operation of the turbine guide vane, and construct a multi-sensor data set;

[0008] S2. Extract features from the collected multi-sensor dataset using deep learning algorithms to identify key patterns and features in the multi-sensor dataset;

[0009] S3. Construct a graph neural network model, treat each sensor as a node in the graph, and define the edges between nodes according to the physical positions and functional connections between sensors and the correlation of data types;

[0010] S4. Input the features extracted in step S2 into the graph neural network model constructed in step S3, process the complex dependencies and interactions between data through the graph neural network model, and analyze the overall operating state of the turbine guide vane;

[0011] S5. According to the output of the graph neural network model, monitor and analyze the performance indicators of the turbine guide vane in real time, and identify signals deviating from the normal operating state, including structural integrity, wear degree, and potential crack development;

[0012] S6. Apply a pre-trained fault diagnosis model to interpret the abnormal signals identified in step S5, and diagnose potential fault causes and fault types, including structural damage, surface wear, or mechanical faults;

[0013] S7. According to the diagnosis results of step S6, automatically generate a fault warning and send a notification to the operation and maintenance personnel.

[0014] Optionally, the eddy current signal data is used to detect minor wear or defects on the surface of the guide vane, the acoustic emission signal data is used to monitor the abnormal noise generated during the operation of the guide vane, detect early signs of cracks or fractures, the electromagnetic field change data analyzes the motion state and structural instability by sensing the changes in the electromagnetic field generated during the movement of the guide vane, and the 3D laser scan data analyzes the morphological deformation and stress accumulation during operation by capturing the three-dimensional morphological changes of the guide vane in real time.

[0015] Optionally, the S1 includes:

[0016] S11. Install accelerometers at key positions of the turbine guide vane, record the vibration data V(t, p) in real time, and enhance the signal quality through modulation and demodulation components. The expression is:

[0017]

[0018] where A and B are amplitude modulation parameters, f m and f c are the modulation frequency and the carrier frequency respectively, is the phase, N(t) is the noise, and the time stamp t is jointly recorded with the guide vane position p;

[0019] S12. Fix a pressure sensor on the surface of the guide vane at the water flow contact point, continuously monitor and record the dynamic pressure index P(t, p), and enhance data analysis using a non-linear dynamic model. The formula is:

[0020] P(t, p) = P 0 +β(sin(ωt)+sin 2 (αt))+∈(t);

[0021] where P 0 is the base pressure value, ω is the fluctuation frequency, and β and α introduce non-linear effects to simulate the dynamic pressure fluctuation;

[0022] S13. Deploy eddy current sensors on the surface of the guide vane, and continuously collect eddy current signal data E(t, p) using a frequency-dependent impedance model:

[0023]

[0024] where Z(f) is the frequency-dependent impedance, Z represents the eddy current impedance caused by the surface state of the guide vane, and ζ(t) is the noise introduced by the environment and equipment;

[0025] S14. Install acoustic emission sensors on the guide vane structure and record acoustic emission signal data A(t, p) using signal decomposition technology:

[0026]

[0027] where S k (f, t) is the component of the acoustic signal in different frequency bands k, H k (f) is the filtering function corresponding to the frequency, and η(t) is the background noise;

[0028] S15. Configure electromagnetic field sensors around the movement area of the guide vane, and introduce a time-varying magnetic permeability model to record electromagnetic field change data M(t, p):

[0029]

[0030] where H(t, p) is the magnetic field intensity generated by the movement of the guide vane, μ is the magnetic permeability of the guide vane material, ξ(t) is the electromagnetic noise, and μ(t) is the time-varying magnetic permeability;

[0031] S16. Use a 3D laser scanning device to perform three-dimensional shape scanning on the guide vane and record 3D laser scanning data L(t, p):

[0032]

[0033] where x(t), y(t), z(t) are the guide vane scanning coordinate points at time t, x 0 、y0 , z 0 is the reference coordinate, and N(0, σ 2 ) indicates that the noise follows a Gaussian distribution, and σ 2 is the variance of the noise;

[0034] S17. Synchronize the time and calibrate the positions of the collected V, P, E, A, M, and L data to construct a comprehensive multi-sensor dataset D. Each data record includes the data value, timestamp t, sensor type, and position p, and adopt the synchronous interval model:

[0035]

[0036] Optionally, the S2 includes:

[0037] S21. Process the data collected in the comprehensive multi-sensor dataset D using a convolutional neural network, and apply a convolution operation to each type of sensor data:

[0038]

[0039] where V conv (t, p) represents the output of the vibration data after convolution processing, H k (s) represents the value of the k-th convolution kernel at the time offset s, ReLU represents the rectified linear unit function, which is used to increase the non-linearity of the network, S represents the time range covered by the convolution kernel, that is, the window size of the convolution operation, and K represents the total number of convolution kernels;

[0040] S22. Perform automatic feature engineering on each type of sensor data S i ∈ {V, P, E, A, M, L}, and calculate the statistical characteristics of the sensor data, including the mean μ, standard deviation σ, skewness, and kurtosis;

[0041] S23. Apply the short-time Fourier transform to further identify the periodic patterns and abnormal changes in the data:

[0042]

[0043] where h(t) is a time-dependent and adaptively adjusted window function;

[0044] S24. Combine machine learning techniques for feature extraction, and use principal component analysis to screen and synthesize key features from the above statistical characteristics and time-frequency analysis results:

[0045]

[0046] S25. The key features of all sensors Integrated into a comprehensive feature vector F for input into the graph neural network model for fault mode recognition and analysis:

[0047]

[0048] Among them, Concat represents the operation of concatenating different sensor feature vectors along a specific dimension.

[0049] Optionally, the mean μ, standard deviation σ, skewness, and kurtosis include:

[0050]

[0051]

[0052] Among them, represents the mean of the sensor data S i , represents the standard deviation of the sensor data S i , represents the skewness of the sensor data S i , indicating the degree of asymmetry of the data distribution, represents the kurtosis of the sensor data S i , indicating the sharpness of the data distribution, and N represents the number of samples.

[0053] Optionally, the S3 includes the following steps:

[0054] S31. Initialize sensor nodes in the graph neural network model, and each node n i represents a specific sensor type S i ;

[0055] S32. Define the edge e ij between nodes. The existence and weight of the edge are determined by the physical positions, functional connections, and data type correlations between sensors. For the edge between sensors S i and S j , its weight w ij is calculated as follows:

[0056]

[0057] Among them, pos i and pos j represent the physical position coordinates of sensors S i and S j , and ρ(type i , type j ) represents a relationship function based on data type similarity, used to adjust the connection strength between different types of sensors;

[0058] S33. Utilize edge e ij and node n i characteristics, apply the message passing mechanism in the graph neural network, enabling the node to update its own state based on the features of its neighbor nodes and the weights of the edges. The state update formula for node n i is as follows:

[0059]

[0060] where, is the state of node n i at time t, N(i) is the set of adjacent nodes of node n i , attr ij is the attribute of edge e ij , the distance and type relationship between sensors;

[0061] S34. Repeat the process of step S33 until the network reaches a stable state, and finally output the optimized state of each node.

[0062] Optionally, the S4 includes the following steps:

[0063] S41. Take the comprehensive feature vector F obtained in step S25 as the input and load it into the graph neural network model;

[0064] S42. In the graph neural network, each node n i receives the feature vector of its corresponding sensor

[0065] The feature vector includes the key patterns and features extracted from each sensor S i ;

[0066] S43. Apply the forward propagation algorithm of the graph neural network, use the feature vector of the node and the weight w of the edge ij to aggregate and update information, and introduce an improved attention mechanism to fuse information from different sensors. The state update formula for each node is:

[0067]

[0068] where, Concat represents concatenating the outputs of all attention heads, W O is the weight matrix of the output layer, used to transform the output of the multi-head attention to a dimension suitable for subsequent processing, head m is the output of the m-th attention head, defined as:

[0069]

[0070] is the attention weight of the m-th head between nodes i and j, calculated as follows:

[0071]

[0072] S44. Repeat step S43 until the network converges, so that the state of each node reflects the interaction between the current node and other nodes and the comprehensive analysis result of sensor data.

[0073] Optionally, the S5 includes the following steps:

[0074] S51. Receive the final state output of each node from the graph neural network model where each state reflects the guide vane performance characteristics corresponding to the corresponding sensor at a specific location;

[0075] S52. Evaluate the state of each node to determine the performance indicators related to structural integrity, wear degree, and potential crack development, and calculate the following performance indicators

[0076]

[0077] where the softmax function is used to convert the linear combination of node states into a probability distribution, representing the probabilities of various fault types;

[0078] S53. According to the performance indicators Real-time monitor the key performance parameters of the turbine guide vane, and identify signals deviating from the normal operating state by setting a predefined threshold θ:

[0079]

[0080] If the performance indicator of a certain sensor S i exceeds the threshold θ, a warning signal alert is generated, indicating the existence of potential structural integrity problems, excessive wear, or crack development.

[0081] Optionally, the S6 includes the following steps:

[0082] S61. Receive the warning signal alert, and each signal is associated with the abnormal performance indicators of one or more sensors S i of

[0083] S62. Use a fault diagnosis model pre-trained with historical fault data and known fault cases to identify structural damage, surface wear, or mechanical faults;

[0084] S63. The abnormal performance indicators in the warning signal Input into the fault diagnosis model, and the model outputs the prediction results of the fault causes and types. The prediction function F of the fault diagnosis model diag is expressed as:

[0085]

[0086] where W represents the weight matrix of the fault diagnosis model;

[0087] S64. According to the output of the fault diagnosis model, determine and explain the potential fault causes and specific fault types of the abnormal signals identified in step S5, including structural damage, surface wear, or mechanical faults.

[0088] Optionally, the establishment of the fault diagnosis model includes:

[0089] Define the online learning function of the fault diagnosis model, and update the parameters θ of the model through a continuous learning mechanism While receiving new data, the online learning mechanism adopts an incremental learning strategy, and the parameter update formula is:

[0090]

[0091] where η is the learning rate, is the gradient of the loss function calculated according to the newly received fault data D new Calculated;

[0092] Introduce an adaptive learning rate adjustment mechanism to automatically adjust the learning rate according to the performance of the fault diagnosis model on real-time fault data. The adjustment of the learning rate is based on the prediction performance of the fault diagnosis model and is dynamically adjusted using the following formula:

[0093] η new =η old ·exp(-αΔP);

[0094] where α is the adjustment coefficient and ΔP is a measure of the change in model performance.

[0095] The beneficial effects of the present invention are:

[0096] (1) Through deep learning and graph neural network technologies, the present invention can process and analyze a large amount of data from multiple sensors in real time. The deep learning algorithm effectively extracts key features from high-dimensional data, while the graph neural network strengthens the interactive analysis of data by simulating the complex relationships between sensors, ensuring that the operating conditions of the guide vanes are monitored in real time and accurately.

[0097] (2) The present invention utilizes a pre-trained fault diagnosis model to quickly identify and classify abnormal behaviors and potential faults of guide vanes. This not only reduces the dependence on manual intervention but also can identify complex fault patterns in the first instance, thus significantly improving the efficiency and timeliness of fault handling. BRIEF DESCRIPTION OF THE DRAWINGS

[0098] The accompanying drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention, but do not constitute a limitation to the present invention. In the drawings:

[0099] Figure 1 is a flowchart of a self-diagnosis method for turbine guide vane faults based on deep learning and graph neural network proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0100] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only showing the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0101] Reference Figure 1 , a self-diagnosis method for turbine guide vane faults based on deep learning and graph neural network, includes the following steps:

[0102] S1. Collect vibration data, pressure indicators, eddy current signal data, acoustic emission signal data, electromagnetic field change data, and 3D laser scanning data generated during the operation of the turbine guide vane, and construct a multi-sensor data set;

[0103] In this embodiment, the eddy current signal data is used to detect minor wear or defects on the surface of the guide vane, the acoustic emission signal data is used to monitor abnormal noises generated during the operation of the guide vane to detect early signs of cracks or fractures, the electromagnetic field change data analyzes the motion state and structural instability by sensing the changes in the electromagnetic field generated during the movement of the guide vane, and the 3D laser scanning data analyzes the morphological deformation and stress accumulation during operation by capturing the three-dimensional morphological changes of the guide vane in real time.

[0104] In this embodiment, S1 includes:

[0105] S11. Install accelerometers at key positions of the turbine guide vane to record vibration data V(t, p) in real time, and enhance the signal quality through modulation and demodulation components. The expression is:

[0106]

[0107] where A and B are amplitude modulation parameters, f m and f c are the modulation frequency and carrier frequency respectively, is the phase, N(t) is the noise, and the timestamp t is jointly recorded with the guide vane position p;

[0108] S12. Fix a pressure sensor on the surface of the guide vane at the water flow contact point, continuously monitor and record the dynamic pressure index P(t, p), and use a nonlinear dynamic model to enhance data analysis. The formula is:

[0109] P(t, p) = P 0 + β(sin(ωt) + sin 2 (αt)) + ∈(t);

[0110] where P 0 is the base pressure value, ω is the fluctuation frequency, and β and α introduce nonlinear effects to simulate the dynamic pressure fluctuation;

[0111] S13. Deploy eddy current sensors on the surface of the guide vane and continuously collect eddy current signal data E(t, p) using a frequency-dependent impedance model:

[0112]

[0113] where Z(f) is the frequency-dependent impedance, Z represents the eddy current impedance caused by the surface state of the guide vane, and ζ(t) is the noise introduced by the environment and equipment;

[0114] S14. Install acoustic emission sensors on the guide vane structure and record acoustic emission signal data A(t, p) using signal decomposition technology:

[0115]

[0116] where S k (f, t) is the component of the acoustic signal in different frequency bands k, H k (f) is the filtering function corresponding to the frequency, and η(t) is the background noise;

[0117] S15. Configure electromagnetic field sensors around the movement area of the guide vane and introduce a time-varying magnetic permeability model to record electromagnetic field change data M(t, p):

[0118]

[0119] where H(t, p) is the magnetic field intensity generated by the movement of the guide vane, μ is the magnetic permeability of the guide vane material, ξ(t) is the electromagnetic noise, and μ(t) is the time-varying magnetic permeability;

[0120] S16. Use a 3D laser scanning device to perform three-dimensional morphology scanning on the guide vane and record 3D laser scanning data L(t, p):

[0121]

[0122] Among them, x(t), y(t), and z(t) are the guide vane scanning coordinate points at time t, and x 0 , y 0 , z 0 are reference coordinates, and N(0, σ 2 ) indicates that the noise follows a Gaussian distribution, and σ 2 is the variance of the noise;

[0123] S17. Synchronize the time and calibrate the position of the collected V, P, E, A, M, and L data to construct a comprehensive multi-sensor dataset D. Each data record includes a data value, a timestamp t, a sensor type, and a position p, and adopt a synchronous interval model:

[0124]

[0125] S2. Use a deep learning algorithm to extract features from the collected multi-sensor dataset and identify key patterns and features in the multi-sensor dataset;

[0126] In this embodiment, S2 includes:

[0127] S21. Process the data collected in the comprehensive multi-sensor dataset D using a convolutional neural network and apply a convolution operation to each type of sensor data:

[0128]

[0129] Among them, V conv (t, p) represents the output of the vibration data after convolution processing, H k (s) represents the value of the kth convolution kernel at a time offset s, ReLU represents the rectified linear unit function, which is used to increase the non-linearity of the network, S represents the time range covered by the convolution kernel, that is, the window size of the convolution operation, and K represents the total number of convolution kernels;

[0130] S22. Perform automatic feature engineering on each type of sensor data S i ∈ {V, P, E, A, M, L}, calculate the statistical characteristics of the sensor data, including the mean μ, standard deviation σ, skewness, and kurtosis;

[0131] S23. Apply the short-time Fourier transform to further identify periodic patterns and abnormal changes in the data:

[0132]

[0133] Among them, h(t) is a time-dependent and adaptively adjusted window function;

[0134] S24. Combine machine learning techniques for feature extraction, and use principal component analysis to screen and synthesize key features from the above statistical characteristics and time-frequency analysis results:

[0135]

[0136] S25. Integrate the key features of all sensors into a comprehensive feature vector F for input into the graph neural network model for fault mode recognition and analysis:

[0137]

[0138] Among them, Concat represents the operation of concatenating different sensor feature vectors along a specific dimension.

[0139] In this embodiment, the mean μ, standard deviation σ, skewness, and kurtosis include:

[0140]

[0141]

[0142] Among them, represents the mean of the sensor data S i , represents the standard deviation of the sensor data S i , represents the skewness of the sensor data S i , indicating the degree of asymmetry of the data distribution, represents the kurtosis of the sensor data S i , indicating the sharpness of the data distribution, and N represents the number of samples.

[0143] S3. Build a graph neural network model, regard each sensor as a node in the graph, and define the edges between nodes according to the physical positions, functional connections, and data type correlations between sensors;

[0144] In this embodiment, S3 includes the following steps:

[0145] S31. Initialize the sensor nodes in the graph neural network model, and each node n i represents a specific sensor type S i ;

[0146] S32. Define the edges e ij between nodes. The existence and weight of the edges are determined by the physical positions, functional connections, and data type correlations between sensors. For the edge between sensors S i and S j , its weight w ij is calculated as follows:

[0147]

[0148] Among them, pos i and pos j represent the physical position coordinates of sensors S i and S j , and ρ(type i , type j ) represents a relational function based on data type similarity, which is used to adjust the connection strength between different types of sensors;

[0149] S33. Utilize the characteristics of edge e ij and node n i to apply the message passing mechanism in the graph neural network, enabling the node to update its own state according to the features of its neighbor nodes and the weights of the edges. The state update formula of node n i is as follows:

[0150]

[0151] Among them, is the state of node n i at time t, N(i) is the set of adjacent nodes of node n i , attr ij is the attribute of edge e ij , the distance and type relationship between sensors;

[0152] S34. Repeat the process of step S33 until the network reaches a stable state, and finally output the optimized state of each node.

[0153] S4. Input the features extracted in step S2 into the graph neural network model constructed in step S3, and analyze the overall operating state of the turbine guide vane by processing the complex dependencies and interactions between data through the graph neural network model;

[0154] In this embodiment, S4 includes the following steps:

[0155] S41. Take the comprehensive feature vector F obtained in step S25 as the input and load it into the graph neural network model;

[0156] S42. In the graph neural network, each node n i receives the feature vector of its corresponding sensor

[0157] The feature vector includes the key patterns and features extracted from each sensor S i ;

[0158] S43. Apply the forward propagation algorithm of the graph neural network, using the feature vectors of the nodes and the weights w of the edges ij to aggregate and update information, and introduce an improved attention mechanism to fuse information from different sensors. The state update formula for each node is as follows:

[0159]

[0160] where Concat means concatenating the outputs of all attention heads, and W O is the weight matrix of the output layer, used to transform the output of the multi-head attention to a dimension suitable for subsequent processing, and head m is the output of the m-th attention head, defined as:

[0161]

[0162] is the attention weight of the m-th head between nodes i and j, calculated as follows:

[0163]

[0164] S44. Repeat step S43 until the network converges, so that the state of each node reflects the interaction between the current node and other nodes and the comprehensive analysis result of the sensor data.

[0165] S5. According to the output of the graph neural network model, monitor and analyze the performance indicators of the turbine guide vane in real time, and identify signals deviating from the normal operating state, including structural integrity, wear degree, and potential crack development;

[0166] In this embodiment, S5 includes the following steps:

[0167] S51. Receive the final state output of each node from the graph neural network model where each state reflects the guide vane performance characteristics corresponding to the corresponding sensor at a specific position;

[0168] S52. Evaluate the state of each node to determine the performance indicators related to structural integrity, wear degree, and potential crack development, and calculate the following performance indicators

[0169]

[0170] where the softmax function is used to transform the linear combination of the node states into a probability distribution, representing the probabilities of various fault types;

[0171] S53. According to the performance indicators Monitor the key performance parameters of the turbine guide vane in real time and identify signals deviating from the normal operating state by setting a predefined threshold θ:

[0172]

[0173] If the performance index of a certain sensor S i exceeds the threshold θ, a warning signal alert is generated, indicating potential structural integrity problems, excessive wear, or crack development.

[0174] S6. Apply a pre-trained fault diagnosis model to interpret the abnormal signals identified in step S5, diagnose potential fault causes and fault types, including structural damage, surface wear, or mechanical failures;

[0175] In this embodiment, S6 includes the following steps:

[0176] S61. Receive the warning signal alert, and each signal is associated with one or more abnormal performance indices of the sensor S i

[0177] S62. Use a fault diagnosis model pre-trained with historical fault data and known fault cases to identify structural damage, surface wear, or mechanical failures;

[0178] S63. Input the abnormal performance indices in the warning signal into the fault diagnosis model, and the model outputs the prediction results of the fault cause and type. The prediction function F of the fault diagnosis model diag is expressed as:

[0179]

[0180] where W represents the weight matrix of the fault diagnosis model;

[0181] S64. According to the output of the fault diagnosis model, determine and interpret the potential fault causes and specific fault types of the abnormal signals identified in step S5, including structural damage, surface wear, or mechanical failures.

[0182] In this embodiment, the establishment of the fault diagnosis model includes:

[0183] Define the online learning function of the fault diagnosis model, and update the model parameters θ through a continuous learning mechanism while receiving new data. The online learning mechanism adopts an incremental learning strategy, and the parameter update formula is:

[0184]

[0185] where η is the learning rate, ​is the gradient of the loss function calculated based on the newly received fault data D new ;

[0186] An adaptive learning rate adjustment mechanism is introduced to automatically adjust the learning rate according to the performance of the fault diagnosis model on real-time fault data. The adjustment of the learning rate is based on the prediction performance of the fault diagnosis model and is dynamically adjusted using the following formula:

[0187] η new = η old ·exp(-αΔP);

[0188] where α is the adjustment coefficient and ΔP is a measure of the change in model performance.

[0189] Through deep learning and graph neural network technologies, the present invention can process and analyze a large amount of data from multiple sensors in real time. The deep learning algorithm effectively extracts key features from high-dimensional data, while the graph neural network strengthens the interactive analysis between data by simulating the complex relationships between sensors, ensuring that the operating conditions of the guide vanes are monitored in real time and accurately.

[0190] S7. According to the diagnosis result in step S6, automatically generate a fault warning and send a notice to the operation and maintenance personnel.

[0191] Embodiment 1:

[0192] In a hydropower station in Yunnan Province, the operation and maintenance team faced performance problems and faults frequently occurring in the turbine guide vanes. The turbine guide vanes are key components affecting the overall power generation efficiency, and their performance directly affects the power generation efficiency and the safe operation of the equipment. In this scenario, we applied the self-diagnosis technology for turbine guide vane faults based on deep learning and graph neural networks to improve the accuracy and efficiency of fault diagnosis.

[0193] The water turbine of this hydropower station consists of multiple guide vanes, and each guide vane is equipped with vibration sensors, pressure sensors, eddy current sensors, acoustic emission sensors, electromagnetic field sensors, and 3D laser scanning devices during operation. These sensors collect data on the operating status of the guide vanes in real time, including vibration frequency and amplitude, surface pressure changes, minor surface wear or defects, abnormal noise signals, electromagnetic field changes, and three-dimensional shape changes of the guide vanes.

[0194] In this embodiment, the operation and maintenance team collected a set of data related to the normal operation and abnormal situations of the guide vanes. These data were first integrated and input into the deep learning model for automatically extracting and identifying key features and patterns in the data. Then, these features were fed into the graph neural network model, where the nodes represent the respective sensors, and the edges between the nodes reflect the physical and functional connections between the sensors.

[0195] During a monitoring in April 2024, the system detected a sudden increase in the vibration data of a guide vane. At the same time, the acoustic emission sensor also captured abnormal high-frequency noise. These signals were immediately identified and analyzed by the graph neural network model, and the model predicted that this might be due to the appearance of cracks on the surface of the guide vane. Subsequently, the data from 3D laser scanning also showed minor deformations on the surface of the guide vane, which was consistent with the analysis of the graph neural network model. The system then issued a fault warning, prompting the operation and maintenance team to conduct an inspection.

[0196] The operation and maintenance team conducted a detailed inspection of the guide vane and confirmed the existence of cracks. Without this system, such cracks might expand unnoticed, resulting in damage to the guide vane and potentially causing safety accidents in severe cases. After adopting this technology, not only were potential large-scale faults detected in a timely manner, but the downtime and maintenance costs that might be caused were also greatly reduced. According to the operation and maintenance records, compared with the same period last year, the fault diagnosis time was reduced by 44.54%, the maintenance time was shortened by 30.75%, and the availability of the equipment was increased by 15.1%, demonstrating the effectiveness of the fault self-diagnosis system.

[0197] The present invention utilizes a pre-trained fault diagnosis model, which can quickly identify and classify the abnormal behaviors and potential faults of the guide vane. It not only reduces the dependence on manual intervention but also can identify complex fault patterns in the first place, thus greatly improving the efficiency and timeliness of fault handling.

[0198] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution of the present invention and its inventive concept, making equivalent substitutions or changes should be covered within the protection scope of the present invention.

Claims

1. A self-diagnosis method for turbine guide vane fault based on deep learning and graph neural network, characterized in that: The following steps are involved: S1. Collect vibration data, pressure index, eddy current signal data, acoustic emission signal data, electromagnetic field change data and 3D laser scanning data generated during the operation of turbine guide vanes, and construct a multi-sensor data set; S2, using deep learning algorithms to extract features from the collected multi-sensor data sets and identify key patterns and features in the multi-sensor data sets; S3 , build a graph neural network model, regard each sensor as a node in the graph, and define the edges between nodes according to the physical location and functional connection between sensors and the correlation of data types; S4, inputting the features extracted in step S2 into the graph neural network model constructed in step S3, processing the complex dependencies and interactions between the data through the graph neural network model, and analyzing the overall operating status of the turbine guide vanes; S5. Based on the output of the graph neural network model, the performance indicators of the turbine guide vanes are monitored and analyzed in real time to identify signals that deviate from normal operating conditions, including structural integrity, wear level, and potential crack development; S6. Applying the pre-trained fault diagnosis model to interpret the abnormal signal identified in step S5, diagnosing potential fault causes and fault types, including structural damage, surface wear or mechanical failure; S7. Automatically generate a fault warning based on the diagnosis result of step S6 and send a notification to the operation and maintenance personnel; The S2 specifically includes step S25: S25, the feature vector F of all sensors Si Integrate into a comprehensive feature vector F, which is used to input into the graph neural network model for fault mode identification and analysis: Concat represents the operation of connecting different sensor feature vectors along a specific dimension. Includes key patterns and features extracted from each sensor data; The S4 comprises the following steps: S41, taking the comprehensive feature vector F obtained in step S25 as input and loading it into the graph neural network model; S42. In the graph neural network, each node n i Receive the feature vector of its corresponding sensor S43. Apply the forward propagation algorithm of the graph neural network and use the feature vector of the node and the edge weight w ij Aggregate and update information, and introduce an improved attention mechanism to fuse information from different sensors. The state update formula for each node is: Among them, Concat means connecting the outputs of all attention heads, W O is the weight matrix of the output layer, which is used to convert the output of multi-head attention to a dimension suitable for subsequent processing. m is the output of the mth attention head and is defined as: is the attention weight of the mth head between nodes i and j, calculated as follows: S44, repeating step S43 until the network converges, so that the state of each node reflects the interaction between the current node and other nodes and the comprehensive analysis result of the sensor data; The S5 comprises the following steps: S51. Receive the final state output of each node from the graph neural network model Each state reflects the guide vane performance characteristics corresponding to the corresponding sensor at a specific position; S52: Status of each node Conduct an assessment to determine the performance indicators related to structural integrity, wear level and potential crack development, and calculate the following performance indicators PI Si : Among them, the softmax function is used to convert the linear combination of node states into a probability distribution to characterize the probability of various fault types; S53, according to performance indicators The key performance parameters of the turbine guide vanes are monitored in real time, and signals that deviate from normal operating conditions are identified by setting a predefined threshold value θ: alert={S i |PI Si >θ}; If a sensor S i If the performance indicator exceeds the threshold θ, an early warning signal alert is generated, indicating the presence of potential structural integrity problems, excessive wear or crack development.

2. A method for self-diagnosis of turbine guide vane faults based on deep learning and graph neural network according to claim 1, characterized in that: The eddy current signal data is used to detect minor wear or defects on the surface of the guide vane, the acoustic emission signal data is used to monitor the unconventional noise generated during the operation of the guide vane and detect early signs of cracks or fractures, the electromagnetic field change data analyzes the motion state and structural instability by sensing the changes in the electromagnetic field generated during the movement of the guide vane, and the 3D laser scanning data is used to analyze the morphological deformation and stress accumulation during operation by capturing the three-dimensional morphological changes of the guide vane in real time.

3. A method for self-diagnosis of turbine guide vane faults based on deep learning and graph neural network according to claim 2, characterized in that: The S1 includes: S11. Install accelerometers at key positions of turbine guide vanes to record vibration data V(t,p) in real time. Enhance signal quality through modulation and demodulation components. The expression is: Where A and B are the amplitude modulation parameters, f m and f c are the modulation frequency and the carrier frequency, is the phase, N(t) is the noise, and the timestamp t is recorded jointly with the guide vane position p; S12. Use pressure sensors fixed on the guide vane surface and water flow contact points to continuously monitor and record the dynamic pressure index P(t,p), and use nonlinear dynamic models to enhance data analysis. The formula is: P(t,p)=P0+β(sin(ωt)+sin 2 (αt))+∈(t); Among them, P0 is the basic pressure value, ω is the fluctuation frequency, and β and α introduce nonlinear effects to simulate the pressure fluctuation dynamics; S13, by deploying eddy current sensors on the guide vane surface, using a frequency-dependent impedance model to continuously collect eddy current signal data E(t,p): Where Z(f) is the frequency-dependent impedance, Z represents the eddy current impedance caused by the surface state of the guide vane, and ζ(t) is the noise introduced by the environment and equipment; S14. Install the acoustic emission sensor on the guide vane structure and use the signal decomposition technology to record the acoustic emission signal data A(t,p): Among them, S k (f, t) is the component of the acoustic signal in different frequency bands k, H k (f) is the filter function of the corresponding frequency, η(t) is the background noise; S15. Configure electromagnetic field sensors to surround the guide vane motion area, and introduce a time-varying magnetic permeability model to record electromagnetic field variation data M(t,p): Where H(t,p) is the magnetic field intensity generated by the guide vane movement, μ is the magnetic permeability of the guide vane material, ξ(t) is the electromagnetic noise, and μ(t) is the time-varying magnetic permeability; S16. Use a 3D laser scanning device to perform a three-dimensional morphological scan on the guide vane and record the 3D laser scanning data L(t,p): Among them, x(t), y(t), z(t) are the guide vane scanning coordinate points at time t, x0, y0, z0 are reference coordinates, N(0,σ 2 ) indicates that the noise follows a Gaussian distribution, σ 2 is the variance of the noise; S17, the collected V, P, E, A, M and L data are time synchronized and position calibrated to construct a comprehensive multi-sensor data set D. Each data record includes data value, timestamp t, sensor type and position p, using a synchronization interval model:

4. A method for self-diagnosis of turbine guide vane faults based on deep learning and graph neural network according to claim 3, characterized in that: The S2 includes: S21. Use a convolutional neural network to process the data collected in the comprehensive multi-sensor dataset D, applying a convolution operation to each sensor data: Among them, V conv (t,p) represents the vibration data output after convolution processing, H k (s) represents the value of the kth convolution kernel at time offset s, ReLU represents the linear rectification function, which is used to increase the nonlinearity of the network, S represents the time range covered by the convolution kernel, that is, the window size of the convolution operation, and K represents the total number of convolution kernels; S22. Perform automatic feature engineering on each type of sensor data and calculate the statistical characteristics of the sensor data, including mean μ, standard deviation σ, skewness, and kurtosis; S23. Apply short-time Fourier transform to further identify periodic patterns and irregular changes in the data: Where h(t) is a time-dependent, adaptively adjusted window function; S24. Combine machine learning technology to extract features, and use principal component analysis to filter and synthesize feature vectors from the above statistical characteristics and time-frequency analysis results:

5. A method for self-diagnosis of turbine guide vane faults based on deep learning and graph neural network according to claim 4, characterized in that: The mean μ, standard deviation σ, skewness and kurtosis include: in, represents the mean of the sensor data, represents the standard deviation of the sensor data, Represents the skewness of sensor data, indicating the degree of asymmetry of data distribution. represents the kurtosis of sensor data, which indicates the sharpness of data distribution, and N represents the number of samples.

6. A method for self-diagnosis of turbine guide vane faults based on deep learning and graph neural network according to claim 4, characterized in that: The S3 comprises the following steps: S31. Initialize sensor nodes in the graph neural network model. Each node n i Represents a specific sensor type S i ; S32. Define the edge e between nodes ij The existence and weight of the edge are determined by the physical location, functional connection and data type correlation between sensors. i and S j The edge between them has a weight w ij The calculation formula is as follows: Among them, pos i and pos j Indicates sensor S i and S j The physical location coordinates, ρ(type i ,type j ) represents a relationship function based on the similarity of data types, which is used to adjust the connection strength between sensors of different types; S33, using edge e ij and node n i The message passing mechanism is applied in the graph neural network to enable nodes to update their own states according to the characteristics of their neighbor nodes and the weights of the edges. i The state update formula is: in, is node n at time t i The state of node n i The adjacent node set of attr ij It is edge ij Attributes, distance and type relationships between sensors; S34. Repeat the process of step S33 until the network reaches a stable state, and finally output the optimized state of each node.

7. A method for self-diagnosis of turbine guide vane faults based on deep learning and graph neural network according to claim 6, characterized in that: The S6 comprises the following steps: S61, receiving warning signals alert, each signal is associated with one or more sensors S i Abnormal performance indicators S62. Identify structural damage, surface wear or mechanical failure using a fault diagnosis model pre-trained with historical fault data and known fault cases; S63. Abnormal performance indicators in the early warning signal Input into the fault diagnosis model, the model outputs the prediction results of the fault cause and type, and the prediction function F of the fault diagnosis model diag It is expressed as: Where W represents the weight matrix of the fault diagnosis model; S64. According to the output of the fault diagnosis model, determine and explain the potential fault cause and specific fault type of the abnormal signal identified in step S5, including structural damage, surface wear or mechanical failure.

8. The method for self-diagnosis of turbine guide vane fault based on deep learning and graph neural network according to claim 7, characterized in that: The fault diagnosis model establishment includes: Define the online learning function of the fault diagnosis model through a continuous learning mechanism While receiving new data, the model parameters θ are updated. The online learning mechanism adopts an incremental learning strategy, and the parameter update formula is: Where η is the learning rate, According to the newly received fault data D new Calculate the gradient of the loss function; An adaptive learning rate adjustment mechanism is introduced to automatically adjust the learning rate according to the performance of the fault diagnosis model on real-time fault data. The learning rate is adjusted dynamically based on the predictive performance of the fault diagnosis model using the following formula: or new =the old ·exp(-αΔP); Where α is the adjustment coefficient and ΔP is a measure of the change in model performance.

Images