Neural monitoring and maintenance planning optimization system for hydraulic turbine set diagram

Abstract

The invention discloses a hydraulic turbine set diagram neural monitoring and maintenance planning optimization system, belongs to the technical field of intelligent operation and maintenance of hydraulic power generation equipment, and aims to solve the technical problem that the intelligent degree of maintenance planning of a hydraulic turbine set is insufficient. The system comprises a data acquisition module, a graph structure modeling module, a graph neural network processor, a maintenance optimization calculation module and a man-machine interaction output module. Compared with a traditional method, the maintenance decision accuracy can be effectively improved, the unplanned shutdown time is shortened, and the service life of key components is prolonged.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for hydropower equipment, specifically to a turbine group neural monitoring and maintenance planning optimization system. Background Technology

[0002] Traditional hydro-generator monitoring systems often employ threshold-based statistical analysis methods or expert-experience-based rule engines, resulting in three significant architectural flaws. First, equipment status analysis relies heavily on the independent processing of single-point sensor data, failing to effectively model the dynamic coupling relationships between key components such as the turbine, generator, governor, and excitation system, leading to insufficient accuracy in system-level fault warnings. Second, maintenance decision-making logic is fixed in a pre-defined rule base, lacking the ability to model the spatiotemporal correlation of equipment status degradation paths, making it difficult to adapt to the multi-condition switching characteristics of hydro-generator units. Third, existing systems only consider single equipment parameter thresholds when optimizing maintenance plans, failing to establish a quantitative model of the relationship between equipment topology and performance degradation transmission, resulting in inaccurate allocation of maintenance resources.

[0003] Current mainstream technical solutions suffer from the following technical bottlenecks:

[0004] 1) Time series-based prediction models cannot capture the structured correlation features between devices in non-Euclidean space;

[0005] 2) Traditional neural networks lack explicit encoding mechanisms for graph structure features when processing device topology relationships;

[0006] 3) The dynamic maintenance planning did not consider the cascading impact of equipment node state propagation on the overall system reliability. This often leads to a situation in practical applications where over-maintenance of some equipment coexists with under-maintenance of critical components.

[0007] Existing technical solutions suffer from a lack of state representation dimensions and insufficient correlation reasoning ability when dealing with complex system failures in hydro turbine units. In particular, the misjudgment rate is very high when dealing with coupled fault scenarios of generator winding overheating and cooling system failure. Summary of the Invention

[0008] To address the technical problem of insufficient intelligence in the maintenance planning of hydro-generator units, this invention provides a hydro-generator unit diagram neural network monitoring and maintenance planning optimization system, comprising:

[0009] The data acquisition module is equipped with a multi-source sensor group to collect real-time vibration, temperature, and pressure data of the turbine unit and generate pre-processed data.

[0010] The graph structure modeling module is used to receive the preprocessed data and dynamically generate a device topology diagram. ,in Represents a set of device nodes. Represents the set of connecting edges;

[0011] A graph neural network processor is used to extract inter-device correlation features based on the device topology graph through graph convolution operations and output state prediction results.

[0012] The maintenance optimization calculation module generates maintenance decision schemes based on the state prediction results and feeds them back to the human-computer interaction output module.

[0013] The human-computer interaction output module is used to provide a visual interface based on the maintenance decision scheme, display the maintenance decision scheme and support user interaction.

[0014] Furthermore, the data acquisition module includes:

[0015] Vibration sensor arrays are used to be installed in the turbine runner chamber, top cover, and guide vane bearings of the unit;

[0016] Temperature sensor arrays are used for installation in stator windings, thrust bearings, and cooling system piping;

[0017] Pressure transmitters are used in hydraulic systems for volutes, tailrace pipes, and governors.

[0018] The time synchronization unit adopts the IEEE 1588 precision clock protocol;

[0019] The signal preprocessing unit includes a bandpass filter and an AD conversion circuit.

[0020] Furthermore, the graph structure modeling module includes:

[0021] The node definition unit defines each sensor as a graph node. ;

[0022] The edge weight calculation unit establishes edge sets based on the device's physical topology and signal correlation. ;

[0023] Dynamic adjacency matrix generation unit, based on time window Update the adjacency matrix This allows for the precise capture of the physical connections between equipment components as time and load change. Simultaneously, a node feature matrix is ​​constructed by fusing multi-source sensor data; the formula is as follows:

[0024]

[0025] In the formula, The node feature matrix, Indicates the first Time-domain peak-to-peak value of a vibration sensor, For the first The gradient change at each temperature measuring point For the first The pulsating components of each pressure measurement point; the node feature matrix maps real-time monitoring data into structured features of graph nodes.

[0026] Furthermore, the graph neural network processor includes:

[0027] The spatiotemporal feature extraction layer employs gated graph convolutional units to achieve deep feature extraction of device status through multi-layer node feature propagation. Its computation process is represented as follows:

[0028]

[0029] In the formula, In the diagrammatic neural network, the first Layer nodes The hidden state, These are the edge weight coefficients based on the attention mechanism. For the first Layer trainable weight matrix;

[0030] Status prediction head, outputs device health index ;

[0031] Fault mode classifier generates fault probability vectors And trigger a tiered alarm through an anomaly detection formula, wherein the anomaly detection formula is:

[0032]

[0033] in The deviation of node features. For nodes Real-time feature vectors, For nodes The baseline eigenvector.

[0034] Furthermore, the maintenance optimization calculation module includes: [The module contains:]

[0035] Multi-objective optimization unit, used to establish constraints:

[0036]

[0037] in Represents the maintenance cost function. For risk accumulation function, For expected power generation losses;

[0038] The Pareto front solver uses the NSGA-II algorithm to generate nondominated solution sets.

[0039] The decision tree reasoning unit selects the optimal maintenance plan based on the expert rule base.

[0040] Furthermore, the human-computer interaction output module includes:

[0041] A 3D visualization unit is used to display stress cloud diagrams of key components of the unit;

[0042] The alarm priority sorting unit is used to dynamically adjust the alarm order based on the device risk score to ensure that high-risk events are handled first.

[0043] The maintenance work order automatic generation unit is used to convert the maintenance decision plan into a standardized executable work order.

[0044] The beneficial effects of this invention are:

[0045] (1) A method for representing the structure of a water turbine assembly is proposed, which breaks through the limitation of isolated analysis of components in traditional monitoring systems;

[0046] (2) Design a dynamically updated graph neural network model to realize the coupled analysis of equipment status and related influences;

[0047] (3) Construct a multi-objective optimization model that includes economy, safety and reliability, and quantitatively evaluate the comprehensive benefits of the maintenance plan;

[0048] (4) Develop an expert knowledge base with self-learning capabilities and continuously optimize model parameters through an online feedback mechanism;

[0049] (5) Achieve deep integration of monitoring data, physical models and expert experience to support the visualization and simulation of maintenance decisions. Attached Figure Description

[0050] Figure 1 This is a diagram showing the system module architecture of the present invention. Detailed Implementation

[0051] The technical solution of the present invention will be further described below with reference to embodiments, but it is not limited thereto. Any modifications or equivalent substitutions to the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention should be covered within the protection scope of the present invention. In the following embodiments, the process equipment or apparatus not specifically specified are all conventional equipment or apparatus in the art. Unless otherwise specified, the raw materials used in the embodiments of the present invention are all commercially available; unless otherwise specified, the technical means used in the embodiments of the present invention are all conventional means well known to those skilled in the art.

[0052] Example 1, combined with Figure 1 This embodiment describes a turbine group diagram neural monitoring and maintenance planning optimization system, which includes:

[0053] The data acquisition module 100 is equipped with a multi-source sensor group to collect real-time vibration, temperature and pressure data of the turbine unit and generate pre-processed data.

[0054] The graph structure modeling module 200 is used to receive the preprocessed data and dynamically generate a device topology diagram. ,in Represents a set of device nodes. Represents the set of connecting edges;

[0055] The graph neural network processor 300 is used to extract inter-device correlation features based on the device topology graph through graph convolution operations and output state prediction results.

[0056] The maintenance optimization calculation module 400 generates a maintenance decision scheme based on the state prediction results and feeds it back to the human-computer interaction output module.

[0057] The human-computer interaction output module 500 is used to provide a visual interface based on the maintenance decision scheme, display the maintenance decision scheme and support user interaction.

[0058] Specifically, this invention collects unit operation data in real time through a multi-source sensor group, transforms it into a dynamic equipment topology map to accurately characterize the physical connections and state coupling relationships between components, and then uses the spatiotemporal feature extraction capability of graph neural networks to deeply mine the fault propagation patterns between equipment. Finally, based on a multi-objective optimization algorithm, it generates maintenance decisions that balance economy and reliability, realizing a leap from isolated data monitoring to system-level intelligent operation and maintenance.

[0059] The data acquisition module 100 includes:

[0060] Vibration sensor array 1.1 is used to be arranged in the turbine runner chamber, top cover and guide vane bearing area of ​​the unit;

[0061] Temperature sensor group 1.2 is used for installation in stator windings, thrust bearings, and cooling system piping;

[0062] Pressure transmitter 1.3 is used for mounting in the volute, tailrace pipe and governor hydraulic system;

[0063] The time synchronization unit 101 adopts the IEEE 1588 precision clock protocol;

[0064] The signal preprocessing unit 102 includes a bandpass filter and an AD conversion circuit.

[0065] The graph structure modeling module 200 includes:

[0066] Node definition unit 201 defines each sensor as a graph node. ;

[0067] Edge weight calculation unit 202 establishes edge sets based on device physical topology and signal correlation. ;

[0068] Dynamic adjacency matrix generation unit 203, according to time window Update the adjacency matrix This allows for the precise capture of the physical connections between equipment components as time and load change. Simultaneously, a node feature matrix is ​​constructed by fusing multi-source sensor data; the formula is as follows:

[0069]

[0070] In the formula, The node feature matrix, Indicates the first Time-domain peak-to-peak value of a vibration sensor, For the first The gradient change at each temperature measuring point For the first The pulsating components of each pressure measurement point; the node feature matrix maps real-time monitoring data into structured features of graph nodes.

[0071] The graph structure modeling module 200 accurately represents the topological connections and state coupling relationships between equipment components through mathematical modeling. Specifically, this module maps each sensor as a graph node through the node definition unit 201, establishing a digital twin of the equipment components; the edge weight calculation unit 202 constructs an edge set based on the physical connections of the equipment and signal correlations, quantifying the interaction strength between components; and the dynamic adjacency matrix generation unit 203 captures the dynamic changes in equipment connection relationships with operating loads in real time, thus overcoming the limitations of traditional static topology models. In this process, the node feature matrix reduces multi-source monitoring data such as vibration, temperature, and pressure to a unified mathematical expression, so that the real-time state of each node is encoded into a machine-readable feature vector.

[0072] The graph neural network processor 300 includes:

[0073] The spatiotemporal feature extraction layer 301 employs gated graph convolutional units to achieve deep feature extraction of device status through multi-layer node feature propagation. Its computation process is as follows:

[0074]

[0075] In the formula, In the diagrammatic neural network, the first Layer nodes The hidden state, These are the edge weight coefficients based on the attention mechanism. For the first Layer trainable weight matrix;

[0076] Status prediction head 302, outputs device health index ;

[0077] Fault mode classifier 303 generates fault probability vectors. And trigger a tiered alarm through an anomaly detection formula, wherein the anomaly detection formula is:

[0078]

[0079] in The deviation of node features. For nodes Real-time feature vectors, For nodes The baseline eigenvector.

[0080] The graph neural network processor 300 deeply integrates device topology relationships with real-time monitoring data, achieving intelligent mapping from multi-source information to state assessment. The spatiotemporal feature extraction layer 301 employs a gated graph convolution formula, dynamically quantifying the correlation strength between device nodes through attention weights. Combined with multi-layer propagation of trainable matrices, it gradually transforms initial node features into high-dimensional hidden states, thereby explicitly modeling the spatial propagation path and temporal evolution of faults in the device topology network.

[0081] The state prediction head 302 maps the final layer features to a health index, enabling continuous quantitative assessment of the degree of equipment degradation; the fault mode classifier 303 outputs the probability of occurrence of specific fault types through probability vectors.

[0082] The maintenance optimization calculation module 400 includes: [components]

[0083] Multi-objective optimization unit 401 is used to establish constraints:

[0084]

[0085] in Represents the maintenance cost function. For risk accumulation function, For expected power generation losses;

[0086] The Pareto front solver 402 uses the NSGA-II algorithm to generate nondominated solution sets.

[0087] Decision tree reasoning unit 403 selects the optimal maintenance plan based on the expert rule base.

[0088] Specifically, the maintenance optimization calculation module 400 transforms equipment condition prediction results into maintenance decisions through multi-objective optimization technology. Its fundamental purpose is to solve the problem of balancing cost and risk in traditional maintenance planning. This module quantifies maintenance costs through mathematical formulas established by the multi-objective optimization unit 401.

[0089] The Pareto front solver 402 uses the NSGA-II algorithm to generate nondominated solution sets, ensuring that the output solution set covers the optimal trade-off between cost and risk.

[0090] The decision tree reasoning unit 403 combines the expert rule base to select the maintenance plan that best fits the actual working conditions from the Pareto solution set, forming a complete decision chain from mathematical optimization to engineering implementation.

[0091] The human-computer interaction output module 500 includes:

[0092] The 3D visualization unit 501 is used to display stress cloud diagrams of key components of the unit;

[0093] The alarm priority sorting unit 502 is used to dynamically adjust the alarm order based on the device risk score to ensure that high-risk events are handled first.

[0094] The maintenance work order automatic generation unit 503 is used to convert the maintenance decision plan into a standardized executable work order.

Claims

1. A hydro-turbine assembly neural network monitoring and maintenance planning optimization system, characterized in that, include: The data acquisition module (100) is equipped with a multi-source sensor group to collect real-time vibration, temperature and pressure data of the turbine unit and generate pre-processed data. The graph structure modeling module (200) is used to receive the preprocessed data and dynamically generate a device topology diagram. ,in Represents a set of device nodes. Represents the set of connecting edges; Graph Neural Network Processor (300) is used to extract inter-device correlation features based on the device topology graph through graph convolution operation and output state prediction results; The maintenance optimization calculation module (400) generates a maintenance decision scheme based on the state prediction results and feeds it back to the human-computer interaction output module; The human-computer interaction output module (500) is used to provide a visual interface based on the maintenance decision scheme, display the maintenance decision scheme and support user interaction.

2. The turbine group diagram neural monitoring and maintenance planning optimization system according to claim 1, characterized in that, The data acquisition module (100) includes: Vibration sensor array (1.1) is used to be arranged in the turbine runner chamber, top cover and guide vane bearing area of ​​the unit; Temperature sensor array (1.2) for installation in stator windings, thrust bearings and cooling system piping; Pressure transmitter (1.3) is used to be installed in the spiral casing, tailrace pipe and governor hydraulic device; The time synchronization unit (101) adopts the IEEE 1588 precision clock protocol; The signal preprocessing unit (102) includes a bandpass filter and an AD conversion circuit.

3. The turbine group diagram neural monitoring and maintenance planning optimization system according to claim 1, characterized in that, The graph structure modeling module (200) includes: Node definition unit (201) defines each sensor as a graph node. ; Edge weight calculation unit (202) establishes edge sets based on device physical topology and signal correlation. ; Dynamic adjacency matrix generation unit (203), by time window Update the adjacency matrix This allows for the precise capture of the physical connections between equipment components as time and load change. Simultaneously, a node feature matrix is ​​constructed by fusing multi-source sensor data; the formula is as follows: In the formula, The node feature matrix, Indicates the first Time-domain peak-to-peak value of a vibration sensor For the first The gradient change at each temperature measuring point For the first The pulsating components of each pressure measurement point; the node feature matrix maps real-time monitoring data into structured features of graph nodes.

4. The turbine group diagram neural monitoring and maintenance planning optimization system according to claim 1, characterized in that, The graph neural network processor (300) include: The spatiotemporal feature extraction layer (301) employs gated graph convolutional units to achieve deep feature extraction of the device state through multi-layer node feature propagation. Its computation process is as follows: In the formula, In the diagrammatic neural network, the first Layer nodes The hidden state, These are the edge weight coefficients based on the attention mechanism. For the first Layer trainable weight matrix; Status prediction head (302) outputs device health index ; Fault mode classifier (303) generates fault probability vectors And trigger a tiered alarm through an anomaly detection formula, wherein the anomaly detection formula is: in The deviation of node features. For nodes Real-time feature vectors, For nodes The baseline eigenvector.

5. The turbine group diagram neural monitoring and maintenance planning optimization system according to claim 1, characterized in that, The maintenance optimization calculation module (400) includes: The multi-objective optimization unit (401) is used to establish constraints: in Represents the maintenance cost function. For risk accumulation function, For expected power generation losses; The Pareto front solver (402) uses the NSGA-II algorithm to generate nondominated solution sets; The decision tree reasoning unit (403) selects the optimal maintenance plan based on the expert rule base.

6. The turbine group diagram neural monitoring and maintenance planning optimization system according to claim 1, characterized in that, The human-computer interaction output module (500) includes: The three-dimensional visualization unit (501) is used to display the stress cloud diagram of key components of the unit; The alarm priority sorting unit (502) is used to dynamically adjust the alarm order based on the device risk score to ensure that high-risk events are handled first. The maintenance work order automatic generation unit (503) is used to convert the maintenance decision plan into a standardized executable work order.

Images