Steam turbine fault diagnosis method and system based on knowledge graph

Abstract

The invention discloses a steam turbine fault diagnosis method and system based on a knowledge graph, and the method comprises the following steps: S1, synchronously collecting the vibration signal, the iron abrasive particle concentration, the oil viscosity and working condition parameters of a steam turbine bearing in a steam turbine, and relates to the technical field of steam turbine fault diagnosis. In order to solve the problems that a turbine bearing serving as a core rotating part is prone to abrasion faults, and early abrasion such as surface microcracks and oil film degradation concealment is high, vibration signals, iron abrasive particle concentration, oil viscosity and working condition parameters are synchronously collected to generate multi-source data streams with aligned timestamps, nonlinear feature extraction is combined, and a multi-source data stream with aligned timestamps is obtained. By constructing the initial knowledge graph, mapping the extracted features to the initial knowledge graph and generating the dynamic knowledge graph based on the graph neural network, time sequence reasoning of the bearing abrasion state is achieved, so that the early abrasion features of the turbine bearing can be obtained, and early warning is conducted according to the early abrasion features of the turbine bearing.

Description

Technical Field

[0001] The present invention relates to the technical field of steam turbine fault diagnosis, and in particular to a steam turbine fault diagnosis method and system based on a knowledge graph. Background Art

[0002] A steam turbine is a rotary power machine that converts the thermal energy of steam into mechanical energy. Its working principle is that steam with a certain pressure and temperature enters the turbine and flows through a series of annular nozzles and moving blades in sequence. The steam expands and accelerates in the nozzles, forming a high-speed steam flow, which impacts the moving blades and causes them to rotate, thereby converting the kinetic energy of the steam into mechanical energy of the rotor.

[0003] Publication No. CN108627345A discloses a method and system for diagnosing system-level faults in a steam turbine. The method comprises: dividing the steam turbine system into a set of components, units, and human actions with specified functions; performing a joint structure-function decomposition of the system based on the principle of functional clustering, based on the failure of component function clusters, unit function clusters, and human function clusters that lead to the failure of the steam turbine system; evaluating the fuzzy fault-symptom attributes of the steam turbine system based on the fuzziness, non-quantitativeness, or semi-quantitativeness of information transmission within the steam turbine system to obtain a fuzzy fault-symptom attribute evaluation criterion for the steam turbine system; and establishing a propagation model for the steam turbine system-level fault based on a causal model of the steam turbine system fault and the fuzzy fault-symptom attribute evaluation criterion. This method of establishing a model based on causal relationships improves the accuracy and effectiveness of the fault diagnosis method.

[0004] However, the above application still has the following problems: currently, turbine bearings are core rotating components, and their wear failures are also relatively common, but early wear such as surface microcracks and oil film degradation are highly concealed. Although vibration sensors can capture abnormal vibrations caused by wear, early wear signals are often masked by background noise. Oil analysis can detect the concentration of metal abrasive particles, but oil analysis is offline sampling, and there is an offline sampling cycle, which cannot track the wear dynamics of turbine bearings in real time. The above application can analyze the cause and effect of system-level failures, but only regards bearings as ordinary components of the main subsystem and is not designed for early bearing wear. The fuzzy fault and symptom attribute evaluation of the above application remains at the qualitative analysis level, and the diagnostic accuracy of early bearing wear is insufficient. Summary of the Invention

[0005] In order to solve the technical problems existing in the background technology, the present invention proposes a turbine fault diagnosis method and system based on knowledge graph.

[0006] The present invention proposes a method for diagnosing steam turbine faults based on a knowledge graph, comprising the following steps: In the steam turbine, synchronously collect the vibration signal of the turbine bearing, the iron wear particle concentration, the oil viscosity, and the operating parameters to generate a multi-source data stream with time stamp alignment; S2. Perform nonlinear feature extraction on the multi-source data streams to generate vibration feature vectors, oil feature vectors, and working condition feature vectors; S3. Build an initial knowledge graph, map the vibration feature vector, oil feature vector, and operating condition feature vector to the initial knowledge graph, update the node weights through the graph neural network, and generate a dynamic knowledge graph; S4. Perform temporal reasoning based on the dynamic knowledge graph to output the wear state vector of the steam turbine bearing. The wear state vector of the steam turbine bearing includes the bearing wear grade. S5. Generate a graded maintenance instruction according to the wear state vector of the turbine bearing.

[0007] Preferably, in S1, the vibration signal of the turbine bearing, the iron abrasive concentration and the oil viscosity, and the operating parameters are synchronously collected, including: The vibration signal of the bearing seat is collected by a three-axis acceleration sensor as the vibration signal of the turbine bearing; Real-time monitoring and collection of iron abrasive concentration and oil viscosity through online oil sensors; The operating parameters are synchronously obtained through the DCS system, including speed, load, and oil temperature.

[0008] Preferably, in S2, nonlinear feature extraction is performed on the multi-source data streams to generate vibration feature vectors as follows: Perform variational modal decomposition on the vibration signals in the multi-source data stream to generate the intrinsic modal components A; Obtain the bearing fault characteristic frequency band and filter out the eigenmode component B corresponding to the bearing fault characteristic frequency band; For each eigenmode component B, obtain the kurtosis value of the eigenmode component B within the bearing fault characteristic frequency band, obtain the energy entropy corresponding to the eigenmode component B in the bearing fault characteristic frequency band, and obtain the kurtosis value and energy entropy of the eigenmode component B; The vibration eigenvector is generated by combining the kurtosis value and energy entropy of the eigenmode component B in a preset order.

[0009] Preferably, in S2, nonlinear feature extraction is performed on the multi-source data streams to generate oil feature vectors as follows: Perform Kalman filtering on the iron wear particle concentration sequence and oil viscosity sequence in the multi-source data stream, generate smoothed iron wear particle concentration sequence and oil viscosity sequence, and obtain the mean value of the iron wear particle concentration sequence and the mean value of the oil viscosity sequence; Perform first-order difference calculations on the smoothed iron abrasive concentration series and oil viscosity series, respectively, to obtain the iron abrasive concentration change rate series and oil viscosity change rate series at adjacent time points, and obtain the mean, standard deviation, and maximum value of the iron abrasive concentration change rate series and the mean, standard deviation, and maximum value of the oil viscosity change rate series; Calculate the Lempel-Ziv complexity of the iron abrasive concentration change rate sequence and the oil viscosity change rate sequence respectively, and obtain the Lempel-Ziv complexity of the iron abrasive concentration change rate sequence and the Lempel-Ziv complexity of the oil viscosity change rate sequence; According to the mean of the iron abrasive concentration sequence, the mean of the oil viscosity sequence, the mean, standard deviation, and maximum value of the iron abrasive concentration change rate sequence, the mean, standard deviation, and maximum value of the oil viscosity change rate sequence, the Lempel-Ziv complexity of the iron abrasive concentration change rate sequence, and the Lempel-Ziv complexity of the oil viscosity change rate sequence, the oil feature vector is generated by combining them in a preset order.

[0010] Preferably, in S2, nonlinear feature extraction is performed on the multi-source data streams to generate the working condition feature vector as follows: Normalize the operating parameters in the multi-source data stream to generate a parameter sequence, reconstruct the parameter sequence into phase space, construct the phase space matrix by embedding the dimension and time delay parameters, and calculate the correlation dimension of the phase space matrix; According to the correlation dimension of the parameter sequence and the phase space matrix, the operating condition characteristic vector is generated by combining them in a preset order.

[0011] Preferably, S2 further includes constructing a vibration-oil coupling characteristic matrix, which is used to characterize the vibration-oil nonlinear coupling characteristics. The vibration-oil coupling characteristic matrix is ​​constructed as follows: Obtain the kurtosis value of the intrinsic mode component B within the bearing fault characteristic frequency band to form a kurtosis value set; Obtaining a sequence of iron abrasive particle concentration change rates to obtain a set of iron abrasive particle concentration change rates; The kurtosis values ​​in the kurtosis value set and the iron abrasive concentration change rates in the iron abrasive concentration change rate set are aligned according to timestamps, and are combined into a two-dimensional matrix in columns, where each row in the two-dimensional matrix corresponds to the kurtosis value and the iron abrasive concentration change rate at the same moment; The correlation of elements in the two-dimensional matrix is ​​analyzed by mutual information or Granger causality to form a vibration-oil coupling characteristic matrix; The vibration-oil coupling feature matrix breaks through the limitations of traditional single-modal feature analysis through matrix combination, and solves the problem of difficulty in capturing nonlinear coupling characteristics of multi-source data.

[0012] Preferably, in S3, the initial knowledge graph is constructed as follows: The initial knowledge graph contains the following node types: vibration node, oil node, and working condition node, among which: The vibration node corresponds to the vibration eigenvector, the oil node corresponds to the oil eigenvector, and the working condition node corresponds to the working condition eigenvector; Construct physical association rules between nodes, including vibration-oil association rules, working condition-vibration association rules, and working condition-oil association rules; Initialize the vibration-oil edge weights, working condition-vibration edge weights, and working condition-oil edge weights between nodes to form an initial knowledge graph containing rules driven by physical laws.

[0013] Preferably, in S3, the physical association rules between nodes are constructed as follows: Get the bearing material hardness, assume the bearing material hardness is H, the iron abrasive concentration change rate is , the kurtosis value of the eigenmode component B is K, the load is F, the load refers to the bearing load, and the oil temperature is T; middle, is the change in iron abrasive concentration, is the time variation; The vibration-oil association rules are constructed as follows: , where H≠0; is the base of the exponential function; 、 is the calibration coefficient, 、 The value of is obtained through machine learning algorithm; Vibration-Oil Association Rules The logic that the rate of change of iron abrasive concentration is proportional to the product of load and kurtosis value, inversely proportional to material hardness, and increases exponentially with oil temperature; Working condition-vibration association rules: Grubin oil film thickness formula is used to associate speed with energy entropy, load with energy entropy; negative correlation rules are defined; Assume that the energy entropy in the wear-free state is, the energy entropy of the eigenmode component B is, and the speed is n. The working condition-vibration association rules are constructed as follows: ; Where, is the calibration coefficient, and the value of is obtained through machine learning algorithm; The relationship between the speed n to the power of 0.68 and the load F to the power of 0.13 is used, which is consistent with the mathematical model of oil film thickness in elastic hydrodynamic lubrication theory. It is a quantitative representation of the oil film thickness; Energy entropy decreases with increasing oil film thickness, reflecting the impact of improved lubrication conditions on vibration signals; Working condition-oil correlation rules: Use the wear thermodynamic model to correlate the load with the rate of change of iron abrasive particle concentration, and the oil temperature with the rate of change of iron abrasive particle concentration; define positive correlation rules; Assume the base temperature of oil is; The working condition-oil association rules are constructed as follows: ; In the formula 、 is the calibration coefficient, 、 The value of is obtained through machine learning algorithm; The vibration-oil correlation rules, operating condition-vibration correlation rules, and operating condition-oil correlation rules all meet physical dimension consistency. Calibration coefficients are used to eliminate unit differences, ensuring that the calculation results are consistent with engineering practice. In initializing the relationship weights between nodes and forming the initial knowledge graph containing physical law-driven rules, the relationship weights between nodes are initialized as follows: For the vibration-oil edge weight in the initial knowledge graph , ; For the working condition-vibration edge weight in the initial knowledge graph , ; For the working condition-oil edge weight in the initial knowledge graph , ; 、 、 The calculation equations all meet the physical dimension consistency, and the unit difference is eliminated through the calibration coefficient to ensure that the calculation results are consistent with engineering practice.

[0014] Preferably, in S4, temporal reasoning is performed based on the dynamic knowledge graph to output the wear state vector of the turbine bearing as follows: A gated recurrent unit is used to perform temporal modeling on the dynamic knowledge graph. The retention ratio of historical information is controlled by updating the gate, and the degree of fusion of the current input and the historical state is controlled by resetting the gate, thus generating a temporal feature vector reflecting the evolution of the wear state. The time series feature vectors are weighted and aggregated through the attention pooling layer to identify several time windows that contribute most to the wear state prediction; The vibration feature vectors, oil feature vectors, and operating condition feature vectors corresponding to several time windows that contribute most to wear state prediction are extracted, and combined with the dynamic knowledge graph to output the wear state vector of the turbine bearing. The wear state vector of the turbine bearing includes the bearing wear grade.

[0015] Preferably, in S4, the bearing wear levels include primary wear, secondary wear, and tertiary wear; The settings of primary wear, secondary wear and tertiary wear can be set by setting thresholds for vibration characteristic vectors, oil characteristic vectors and working condition characteristic vectors; In S5, a graded maintenance instruction is generated according to the wear state vector of the turbine bearing, as follows: When the bearing wear level in the wear state vector is level one, the instructions are frequency-increasing monitoring of the lubrication system and oil particle spectral analysis; When the bearing wear level in the wear state vector is level 2, the instruction is to plan maintenance within 72 hours and measure the bearing oil film thickness; When the bearing wear level in the wear state vector is level three, the instruction is to stop immediately.

[0016] A steam turbine fault diagnosis system based on knowledge graph, comprising: Data acquisition module: Synchronously collects turbine bearing vibration signals, iron wear particle concentration, oil viscosity, and operating parameters to generate multi-source data streams with time stamp alignment; Feature extraction module: performs nonlinear feature extraction on multi-source data streams to generate vibration feature vectors, oil feature vectors, and working condition feature vectors; Knowledge graph generation module: Builds an initial knowledge graph, maps vibration feature vectors, oil feature vectors, and operating condition feature vectors to the initial knowledge graph, updates node weights through a graph neural network, and generates a dynamic knowledge graph; Bearing wear level output module: This module performs temporal reasoning based on a dynamic knowledge graph and outputs the wear state vector of the turbine bearing. The wear state vector of the turbine bearing includes the bearing wear level. Maintenance instruction generation module: Generates graded maintenance instructions based on the wear state vector of the turbine bearing.

[0017] The present invention proposes a knowledge graph-based steam turbine fault diagnosis method and system, which has the following beneficial technical effects: 1. This application addresses the problem that turbine bearings, as core rotating components, are prone to wear failures, and early wear such as surface microcracks and oil film degradation is highly concealed. By synchronously collecting vibration signals, iron abrasive concentration, oil viscosity, and operating parameters to generate a timestamp-aligned multi-source data stream, and combining nonlinear feature extraction to generate vibration feature vectors, oil feature vectors, and operating feature vectors, the application constructs an initial knowledge graph, maps the vibration feature vectors, oil feature vectors, and operating feature vectors to the initial knowledge graph, and dynamically updates the knowledge graph node weights based on a graph neural network to generate a dynamic knowledge graph, thereby realizing time-series reasoning of the bearing wear state, thereby obtaining the early wear characteristics of the turbine bearings, and thus issuing early warnings based on the early wear characteristics of the turbine bearings. This can significantly improve the early warning capability of the bearings, effectively solving the defects in traditional technologies that the early vibration signals are easily masked by background noise and offline oil sampling cannot track the wear dynamics in real time, thereby avoiding sudden failures caused by hidden wear.

[0018] 2. By constructing a vibration-oil coupling feature matrix, aligning the kurtosis value and the iron abrasive concentration change rate by timestamp and performing mutual information or Granger causality analysis, this method breaks through the limitation of traditional single-modal feature analysis that relies only on a single data dimension, effectively captures the nonlinear coupling characteristics between vibration and oil data, improves the characterization ability of early wear characteristics of turbine bearings, and solves the technical problem of the difficulty in capturing nonlinear coupling relationships of multi-source data, thereby improving the sensitivity and accuracy of early wear diagnosis, reducing missed detections and false alarms, and providing a more reliable basis for early warning and precise maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of a method for diagnosing steam turbine faults based on a knowledge graph according to the present invention; Figure 2 This is a principle block diagram of a knowledge graph-based turbine fault diagnosis system of the present invention. DETAILED DESCRIPTION

[0020] The following describes embodiments of the present invention in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar symbols throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention, and are not to be construed as limiting the present invention.

[0021] like Figure 1 A steam turbine fault diagnosis method based on a knowledge graph is shown, comprising the following steps: In the steam turbine, synchronously collect the vibration signal of the turbine bearing, the iron wear particle concentration, the oil viscosity, and the operating parameters to generate a multi-source data stream with time stamp alignment; S2. Perform nonlinear feature extraction on the multi-source data streams to generate vibration feature vectors, oil feature vectors, and working condition feature vectors; S3. Build an initial knowledge graph, map the vibration feature vector, oil feature vector, and operating condition feature vector to the initial knowledge graph, update the node weights through the graph neural network, and generate a dynamic knowledge graph; S4. Perform temporal reasoning based on the dynamic knowledge graph to output the wear state vector of the steam turbine bearing. The wear state vector of the steam turbine bearing includes the bearing wear grade. S5. Generate a graded maintenance instruction according to the wear state vector of the turbine bearing.

[0022] In an optional embodiment, in S1, the vibration signal of the turbine bearing, the iron abrasive concentration and the oil viscosity, and the operating parameters are synchronously collected, including: The vibration signal of the bearing seat is collected by a three-axis acceleration sensor as the vibration signal of the turbine bearing; Real-time monitoring and collection of iron abrasive concentration and oil viscosity through online oil sensors; Synchronously obtain operating parameters through the DCS system, including speed, load, and oil temperature; By synchronously collecting the vibration signals, iron abrasive concentration, oil viscosity, and operating parameters of the turbine bearings, the feature matching error caused by traditional asynchronous collection can be resolved.

[0023] In an optional embodiment, in S2, nonlinear feature extraction is performed on the multi-source data streams to generate vibration feature vectors as follows: Perform variational modal decomposition on the vibration signals in the multi-source data stream to generate the intrinsic modal components A; Obtain the bearing fault characteristic frequency band and filter out the eigenmode component B corresponding to the bearing fault characteristic frequency band; For each eigenmode component B, obtain the kurtosis value of the eigenmode component B within the bearing fault characteristic frequency band, obtain the energy entropy corresponding to the eigenmode component B in the bearing fault characteristic frequency band, and obtain the kurtosis value and energy entropy of the eigenmode component B; Generate a vibration eigenvector based on the kurtosis value and energy entropy of the eigenmode component B in a preset order; By filtering the characteristic frequency bands of bearing faults, irrelevant components in intrinsic mode component A are eliminated, and only the kurtosis and energy entropy of intrinsic mode component B related to the bearing fault are obtained, ultimately forming a vibration feature vector with diagnostic value. This method can effectively improve the signal-to-noise ratio of fault characteristics and enhance the accuracy and generalization ability of the diagnostic model.

[0024] In an optional embodiment, in S2, nonlinear feature extraction is performed on the multi-source data streams to generate oil feature vectors as follows: Perform Kalman filtering on the iron wear particle concentration sequence and oil viscosity sequence in the multi-source data stream, generate smoothed iron wear particle concentration sequence and oil viscosity sequence, and obtain the mean value of the iron wear particle concentration sequence and the mean value of the oil viscosity sequence; Perform first-order difference calculations on the smoothed iron abrasive concentration series and oil viscosity series, respectively, to obtain the iron abrasive concentration change rate series and oil viscosity change rate series at adjacent time points, and obtain the mean, standard deviation, and maximum value of the iron abrasive concentration change rate series and the mean, standard deviation, and maximum value of the oil viscosity change rate series; Calculate the Lempel-Ziv complexity of the iron abrasive concentration change rate sequence and the oil viscosity change rate sequence respectively, and obtain the Lempel-Ziv complexity of the iron abrasive concentration change rate sequence and the Lempel-Ziv complexity of the oil viscosity change rate sequence; The oil feature vector is generated by combining the mean of the iron abrasive concentration sequence, the mean of the oil viscosity sequence, the mean, standard deviation, and maximum value of the iron abrasive concentration change rate sequence, the mean, standard deviation, and maximum value of the oil viscosity change rate sequence, the Lempel-Ziv complexity of the iron abrasive concentration change rate sequence, and the Lempel-Ziv complexity of the oil viscosity change rate sequence in a preset order; Lempel-Ziv complexity is an existing algorithm for quantitatively analyzing the complexity of time series.

[0025] In an optional embodiment, in S2, nonlinear feature extraction is performed on the multi-source data streams to generate a working condition feature vector as follows: Normalize the operating parameters in the multi-source data stream to generate a parameter sequence, reconstruct the parameter sequence into phase space, construct the phase space matrix by embedding the dimension and time delay parameters, and calculate the correlation dimension of the phase space matrix; According to the correlation dimension of the parameter sequence and the phase space matrix, the operating condition characteristic vector is generated by combining them in a preset order.

[0026] S2 also includes constructing a vibration-oil coupling characteristic matrix, which is used to characterize the vibration-oil nonlinear coupling characteristics. The vibration-oil coupling characteristic matrix is ​​constructed as follows: Obtain the kurtosis value of the intrinsic mode component B within the bearing fault characteristic frequency band to form a kurtosis value set; Obtaining a sequence of iron abrasive particle concentration change rates to obtain a set of iron abrasive particle concentration change rates; The kurtosis values ​​in the kurtosis value set and the iron abrasive concentration change rates in the iron abrasive concentration change rate set are aligned according to timestamps, and are combined into a two-dimensional matrix in columns, where each row in the two-dimensional matrix corresponds to the kurtosis value and the iron abrasive concentration change rate at the same moment; The correlation of elements in the two-dimensional matrix is ​​analyzed by mutual information or Granger causality to form a vibration-oil coupling characteristic matrix; By constructing a vibration-oil coupling feature matrix, aligning the kurtosis value and the iron abrasive concentration change rate by timestamp and performing mutual information or Granger causality analysis, the limitation of traditional single-modal feature analysis that only relies on a single data dimension is broken through, and the nonlinear coupling characteristics between vibration and oil data are effectively captured, the characterization ability of early wear characteristics of turbine bearings is improved, and the technical difficulty of capturing the nonlinear coupling relationship of multi-source data is solved, thereby improving the sensitivity and accuracy of early wear diagnosis, reducing missed detections and false alarms, and providing a more reliable basis for early warning and precise maintenance.

[0027] In S3, construct the initial knowledge graph as follows: The initial knowledge graph contains the following node types: vibration node, oil node, and working condition node, among which: The vibration node corresponds to the vibration eigenvector, the oil node corresponds to the oil eigenvector, and the working condition node corresponds to the working condition eigenvector; Construct physical association rules between nodes, including vibration-oil association rules, working condition-vibration association rules, and working condition-oil association rules; In S3, the physical association rules between nodes are as follows: Get the bearing material hardness, assume the bearing material hardness is H, the iron abrasive concentration change rate is , the kurtosis value of the eigenmode component B is K, the load is F, the load refers to the bearing load, and the oil temperature is T; middle, is the change in iron abrasive concentration, is the time variation; The vibration-oil association rules are constructed as follows: , where H≠0; is the base of the exponential function; 、 is the calibration coefficient, 、 The value of is obtained through machine learning algorithm; Vibration-Oil Association Rules The logic that the rate of change of iron abrasive concentration is proportional to the product of load and kurtosis value, inversely proportional to material hardness, and increases exponentially with oil temperature; Working condition-vibration association rules: Grubin oil film thickness formula is used to associate speed with energy entropy, load with energy entropy; negative correlation rules are defined; Assume that the energy entropy in the wear-free state is, the energy entropy of the eigenmode component B is, and the speed is n. The working condition-vibration association rules are constructed as follows: ; Where, is the calibration coefficient, The value of is obtained through machine learning algorithm; The relationship between the speed n to the power of 0.68 and the load F to the power of 0.13 is used, which is consistent with the mathematical model of oil film thickness in elastic hydrodynamic lubrication theory. It is a quantitative representation of the oil film thickness; Energy entropy decreases with increasing oil film thickness, reflecting the impact of improved lubrication conditions on vibration signals; Working condition-oil correlation rules: Use the wear thermodynamic model to correlate the load with the rate of change of iron abrasive particle concentration, and the oil temperature with the rate of change of iron abrasive particle concentration; define positive correlation rules; Assume the base temperature of oil is; The working condition-oil association rules are constructed as follows: ; In the formula 、 is the calibration coefficient, 、 The value of is obtained through machine learning algorithm; The vibration-oil correlation rules, operating condition-vibration correlation rules, and operating condition-oil correlation rules all meet physical dimension consistency. Calibration coefficients are used to eliminate unit differences, ensuring that the calculation results are consistent with engineering practice. Initialize the relationship weights between nodes to form an initial knowledge graph containing rules driven by physical laws; Initialize the relationship weights between nodes as follows: For the vibration-oil edge weight in the initial knowledge graph , ; For the working condition-vibration edge weight in the initial knowledge graph , ; For the working condition-oil edge weight in the initial knowledge graph , ; 、 、 The calculation equations all meet the physical dimension consistency, and the unit difference is eliminated through the calibration coefficient to ensure that the calculation results are consistent with the actual engineering practice; By constructing an initial knowledge graph—a completely new construction rather than directly acquiring an existing graph—we embed the vibration-oil association rules, operating condition-vibration association rules, and operating condition-oil association rules for turbine bearings into the physical association rules between nodes in the initial knowledge graph, making the graph a digital representation of physical mechanisms and avoiding the domain knowledge loss caused by directly using a general-purpose graph. This newly constructed initial knowledge graph allows for initialization of relationship weights between nodes based on bearing material hardness, load, and oil temperature. This distinguishes it from the black-box nature of traditional data-driven models and significantly enhances its relevance for early wear scenarios in turbine bearings. By independently designing node types and association rules, we can specifically address the hidden nature of early wear in turbine bearings, a problem that existing graphs cannot adapt to the multidimensional feature fusion requirements of such complex industrial scenarios. By constructing an initial knowledge graph, we make it suitable for industrial scenarios where turbine bearings are sensitive to safety and maintenance costs.

[0028] In S4, temporal reasoning is performed based on the dynamic knowledge graph to output the wear state vector of the turbine bearing as follows: A gated recurrent unit is used to perform temporal modeling on the dynamic knowledge graph. The retention ratio of historical information is controlled by updating the gate, and the degree of fusion of the current input and the historical state is controlled by resetting the gate, thus generating a temporal feature vector reflecting the evolution of the wear state. The time series feature vectors are weighted and aggregated through the attention pooling layer to identify several time windows that contribute most to the wear state prediction; The vibration feature vectors, oil feature vectors, and operating condition feature vectors corresponding to several time windows that contribute most to wear state prediction are extracted, and combined with the dynamic knowledge graph to output the wear state vector of the turbine bearing. The wear state vector of the turbine bearing includes the bearing wear grade.

[0029] The gated recurrent unit is an existing recurrent neural network unit for processing sequence data; the attention pooling layer is an existing pooling operation that dynamically weights features using the attention mechanism.

[0030] In S4, the bearing wear levels include primary wear, secondary wear, and tertiary wear; The settings of primary wear, secondary wear and tertiary wear can be set by setting thresholds for vibration characteristic vectors, oil characteristic vectors and working condition characteristic vectors; In S5, a graded maintenance instruction is generated according to the wear state vector of the turbine bearing, as follows: When the bearing wear level in the wear state vector is level one, the instructions are frequency-increasing monitoring of the lubrication system and oil particle spectral analysis; When the bearing wear level in the wear state vector is level 2, the instruction is to plan maintenance within 72 hours and measure the bearing oil film thickness; When the bearing wear level in the wear state vector is level three, the instruction is to stop immediately.

[0031] This application addresses the problem that turbine bearings, as core rotating components, are prone to wear failures, and early wear such as surface microcracks and oil film degradation is highly concealed. By synchronously collecting vibration signals, iron abrasive concentration, oil viscosity, and operating parameters to generate a timestamp-aligned multi-source data stream, the traditional data fragmentation problem is solved. Vibration feature vectors, oil feature vectors, and operating feature vectors are generated by combining nonlinear feature extraction. By constructing an initial knowledge graph, the vibration feature vectors, oil feature vectors, and operating feature vectors are mapped to the initial knowledge graph. The knowledge graph node weights are dynamically updated based on a graph neural network to generate a dynamic knowledge graph, realizing time-series reasoning of the bearing wear status, thereby obtaining early wear characteristics of turbine bearings, and thus issuing early warnings based on the early wear characteristics of turbine bearings. This can significantly improve the early warning capability of bearing early wear, effectively solving the defects in traditional technologies that vibration signals are easily masked by background noise in the early stages and offline oil sampling cannot track wear dynamics in real time, thereby avoiding sudden failures caused by hidden wear.

[0032] like Figure 2 The steam turbine fault diagnosis system shown in FIG. 1 is based on a knowledge graph and includes: Data acquisition module: Synchronously collects turbine bearing vibration signals, iron wear particle concentration, oil viscosity, and operating parameters to generate multi-source data streams with time stamp alignment; Feature extraction module: performs nonlinear feature extraction on multi-source data streams to generate vibration feature vectors, oil feature vectors, and working condition feature vectors; Knowledge graph generation module: Builds an initial knowledge graph, maps vibration feature vectors, oil feature vectors, and operating condition feature vectors to the initial knowledge graph, updates node weights through a graph neural network, and generates a dynamic knowledge graph; Bearing wear level output module: This module performs temporal reasoning based on a dynamic knowledge graph and outputs the wear state vector of the turbine bearing. The wear state vector of the turbine bearing includes the bearing wear level. Maintenance instruction generation module: Generates graded maintenance instructions based on the wear state vector of the turbine bearing.

[0033] Meanwhile, the contents not described in detail in this specification belong to the prior art known to those skilled in the art.

[0034] In the embodiments provided by the present invention, it should be understood that the disclosed systems or methods can be implemented in other ways. For example, the embodiments of the invention described above are merely illustrative. For example, the division of modules is only a logical function division, and other division methods may be used in actual implementation.

[0035] Modules described as separate components may or may not be physically separate, and components shown as modules may or may not be physical modules, and may be located in one place or distributed across multiple network modules. Some or all of these modules may be selected to achieve the objectives of this embodiment based on actual needs.

[0036] In addition, the functional modules in various embodiments of the present invention may be integrated into a single processing module, each module may exist physically separately, or two or more modules may be integrated into a single module. The aforementioned integrated modules may be implemented in the form of hardware or hardware plus software functional modules.

[0037] It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the basic characteristics of the present invention.

[0038] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A steam turbine fault diagnosis method based on knowledge graph, characterized in that: The following steps are involved: In the steam turbine, synchronously collect the vibration signal of the turbine bearing, the iron wear particle concentration, the oil viscosity, and the operating parameters to generate a multi-source data stream with time stamp alignment; S2. Perform nonlinear feature extraction on the multi-source data streams to generate vibration feature vectors, oil feature vectors, and working condition feature vectors; S3. Build an initial knowledge graph, map the vibration feature vector, oil feature vector, and operating condition feature vector to the initial knowledge graph, update the node weights through the graph neural network, and generate a dynamic knowledge graph; S4. Perform temporal reasoning based on the dynamic knowledge graph to output the wear state vector of the steam turbine bearing. The wear state vector of the steam turbine bearing includes the bearing wear grade. S5. Generate a graded maintenance instruction according to the wear state vector of the turbine bearing.

2. A method for diagnosing steam turbine faults based on knowledge graph according to claim 1, characterized in that: In S1, the vibration signal of the turbine bearing, the iron wear particle concentration and oil viscosity, and the operating parameters are collected synchronously, including: The vibration signal of the bearing seat is collected by a three-axis acceleration sensor as the vibration signal of the turbine bearing; Real-time monitoring and collection of iron abrasive concentration and oil viscosity through online oil sensors; The operating parameters are synchronously obtained through the DCS system, including speed, load, and oil temperature.

3. The method for diagnosing steam turbine faults based on knowledge graph according to claim 1, characterized in that: In S2, nonlinear feature extraction is performed on the multi-source data streams to generate vibration feature vectors as follows: Perform variational modal decomposition on the vibration signals in the multi-source data stream to generate the intrinsic modal components A; Obtain the bearing fault characteristic frequency band and filter out the eigenmode component B corresponding to the bearing fault characteristic frequency band; For each eigenmode component B, obtain the kurtosis value of the eigenmode component B within the bearing fault characteristic frequency band, obtain the energy entropy corresponding to the eigenmode component B in the bearing fault characteristic frequency band, and obtain the kurtosis value and energy entropy of the eigenmode component B; The vibration eigenvector is generated by combining the kurtosis value and energy entropy of the eigenmode component B in a preset order.

4. A method for diagnosing steam turbine faults based on knowledge graph according to claim 3, characterized in that: In S2, nonlinear feature extraction is performed on the multi-source data streams to generate oil feature vectors as follows: Perform Kalman filtering on the iron wear particle concentration sequence and oil viscosity sequence in the multi-source data stream, generate smoothed iron wear particle concentration sequence and oil viscosity sequence, and obtain the mean value of the iron wear particle concentration sequence and the mean value of the oil viscosity sequence; Perform first-order difference calculations on the smoothed iron abrasive concentration series and oil viscosity series, respectively, to obtain the iron abrasive concentration change rate series and oil viscosity change rate series at adjacent time points, and obtain the mean, standard deviation, and maximum value of the iron abrasive concentration change rate series and the mean, standard deviation, and maximum value of the oil viscosity change rate series; Calculate the Lempel-Ziv complexity of the iron abrasive concentration change rate sequence and the oil viscosity change rate sequence respectively, and obtain the Lempel-Ziv complexity of the iron abrasive concentration change rate sequence and the Lempel-Ziv complexity of the oil viscosity change rate sequence; According to the mean of the iron abrasive concentration sequence, the mean of the oil viscosity sequence, the mean, standard deviation, and maximum value of the iron abrasive concentration change rate sequence, the mean, standard deviation, and maximum value of the oil viscosity change rate sequence, the Lempel-Ziv complexity of the iron abrasive concentration change rate sequence, and the Lempel-Ziv complexity of the oil viscosity change rate sequence, the oil feature vector is generated by combining them in a preset order.

5. The method for diagnosing steam turbine faults based on knowledge graph according to claim 4 is characterized in that: In S2, nonlinear feature extraction is performed on the multi-source data streams to generate the working condition feature vector as follows: Normalize the operating parameters in the multi-source data stream to generate a parameter sequence, reconstruct the parameter sequence into phase space, construct the phase space matrix by embedding the dimension and time delay parameters, and calculate the correlation dimension of the phase space matrix; According to the correlation dimension of the parameter sequence and the phase space matrix, the operating condition characteristic vector is generated by combining them in a preset order.

6. A method for diagnosing steam turbine faults based on knowledge graph according to claim 5, characterized in that: S2 also includes constructing a vibration-oil coupling characteristic matrix, which is used to characterize the vibration-oil nonlinear coupling characteristics. The vibration-oil coupling characteristic matrix is ​​constructed as follows: Obtain the kurtosis value of the intrinsic mode component B within the bearing fault characteristic frequency band to form a kurtosis value set; Obtaining a sequence of iron abrasive particle concentration change rates to obtain a set of iron abrasive particle concentration change rates; The kurtosis values ​​in the kurtosis value set and the iron abrasive concentration change rates in the iron abrasive concentration change rate set are aligned according to timestamps, and are combined into a two-dimensional matrix in columns, where each row in the two-dimensional matrix corresponds to the kurtosis value and the iron abrasive concentration change rate at the same moment; The correlation of elements in the two-dimensional matrix is ​​analyzed by mutual information or Granger causality to form a vibration-oil coupling characteristic matrix.

7. The method for diagnosing steam turbine faults based on knowledge graph according to claim 1, characterized in that: In S3, construct the initial knowledge graph as follows: The initial knowledge graph contains the following node types: vibration node, oil node, and working condition node, among which: The vibration node corresponds to the vibration eigenvector, the oil node corresponds to the oil eigenvector, and the working condition node corresponds to the working condition eigenvector; Construct physical association rules between nodes, including vibration-oil association rules, working condition-vibration association rules, and working condition-oil association rules; Initialize the vibration-oil edge weights, working condition-vibration edge weights, and working condition-oil edge weights between nodes to form an initial knowledge graph.

8. The method for diagnosing steam turbine faults based on knowledge graph according to claim 1 or 7, characterized in that: In S4, temporal reasoning is performed based on the dynamic knowledge graph to output the wear state vector of the turbine bearing as follows: Use gated recurrent units to perform temporal modeling on dynamic knowledge graphs and generate temporal feature vectors; The time series feature vectors are weighted and aggregated through the attention pooling layer to identify several time windows that contribute most to the wear state prediction; The vibration feature vectors, oil feature vectors, and operating condition feature vectors corresponding to several time windows that contribute most to wear state prediction are extracted, and combined with the dynamic knowledge graph to output the wear state vector of the turbine bearing. The wear state vector of the turbine bearing includes the bearing wear grade.

9. The method for diagnosing steam turbine faults based on knowledge graph according to claim 8, characterized in that: In S4, the bearing wear levels include primary wear, secondary wear, and tertiary wear; In S5, a graded maintenance instruction is generated according to the wear state vector of the turbine bearing, as follows: When the bearing wear level in the wear state vector is level one, the instructions are frequency-increasing monitoring of the lubrication system and oil particle spectral analysis; When the bearing wear level in the wear state vector is level 2, the instruction is to plan maintenance within 72 hours and measure the bearing oil film thickness; When the bearing wear level in the wear state vector is level three, the instruction is to stop immediately.

10. A knowledge graph-based steam turbine fault diagnosis system, used for using the knowledge graph-based steam turbine fault diagnosis method according to any one of claims 1 to 9, characterized in that: include: Data acquisition module: Synchronously collects turbine bearing vibration signals, iron wear particle concentration, oil viscosity, and operating parameters to generate multi-source data streams with time stamp alignment; Feature extraction module: performs nonlinear feature extraction on multi-source data streams to generate vibration feature vectors, oil feature vectors, and working condition feature vectors; Knowledge graph generation module: Builds an initial knowledge graph, maps vibration feature vectors, oil feature vectors, and operating condition feature vectors to the initial knowledge graph, updates node weights through a graph neural network, and generates a dynamic knowledge graph; Bearing wear level output module: This module performs temporal reasoning based on a dynamic knowledge graph and outputs the wear state vector of the turbine bearing. The wear state vector of the turbine bearing includes the bearing wear level. Maintenance instruction generation module: Generates graded maintenance instructions based on the wear state vector of the turbine bearing.

Images