Gas turbine fault detection method based on simulation priori and gas path communication mapping

Abstract

This invention discloses a gas turbine fault detection method based on simulation prior and gas path connectivity mapping, which relates to the field of gas turbine technology. By constructing the abstract gas path structure of the gas turbine and the input tensor structure of the neural network in a targeted manner, the method uses 3D-CNN to extract spatiotemporal coupling features and output fault diagnosis results. This method is used to solve the problems of traditional deep learning methods in extracting spatial features under the complex topology of gas turbines and the large number of early fault samples required.

Description

Technical Field

[0001] This invention relates to the field of gas turbine technology, and in particular to a gas turbine fault detection method based on simulation priors and gas path connectivity mapping. Background Technology

[0002] The propulsion system of medium and large vessels (such as ocean-going transport ships and special-purpose vessels) is crucial for ensuring navigation safety and operational efficiency. Among these systems, gas turbine propulsion systems have become a mainstream configuration due to their high power density and environmental friendliness. However, this type of propulsion system is complex, comprising two main subsystems: a gas turbine cycle and a steam turbine cycle. It involves dozens of key components, including compressors, combustion chambers, turbines, and heat exchangers, all of which are physically interconnected through pipelines and valves. Under long-term high-load and variable operating conditions, even minor faults in any component (such as blade wear, pipeline leaks, or decreased heat exchange efficiency) can trigger a chain reaction of failures if not detected in time, leading to power outages or even safety accidents. Therefore, accurate and efficient fault detection, especially early fault identification, of gas turbine propulsion systems for medium and large vessels has become a core requirement in the field of marine engineering.

[0003] Currently, fault detection in marine gas turbines mainly relies on a technical approach of "real-time sensor monitoring + data processing and analysis," with key supporting technologies falling into two categories: One approach is gas turbine thermodynamic simulation technology. This technology constructs component-level thermal models of the power system, based on the core principles of "energy conservation," "mass conservation," and "energy cascade utilization," to iteratively simulate parameters such as temperature, pressure, and flow rate under standard operating conditions. It outputs theoretical standard values ​​for each sensor monitoring point, providing a reference for determining whether the system deviates from its normal state. Currently, thermodynamic simulation is widely used in the design verification and operating condition prediction of gas turbines. For example, in component performance analysis, high-precision parameter simulation can be achieved by breaking down the cyclic process and refining the heat loss calculation for each stage. Zuming et al.'s gas turbine performance substitution model based on HDMR-ANN can reproduce key correction curves such as power and exhaust temperature under any external environment within ±2% error using only operating data. This provides a lightweight data-driven engine that can be directly embedded for real-time fusion of "simulation-measurement," thus overcoming the limitation of traditional thermodynamic models requiring offline calibration. However, this technology often operates independently, and its integration with real-time sensor data is still in its early stages.

[0004] Secondly, machine learning techniques are used in fault detection. With the development of artificial intelligence, deep learning models such as Convolutional Neural Networks (CNNs), due to their powerful spatial feature extraction capabilities, are increasingly being used for fault detection in industrial equipment. This involves converting time-series data such as temperature and pressure collected by sensors into feature vectors, which are then input into the CNN model to classify fault types. Zhong et al., addressing the problem of large individual differences and few fault samples in gas turbines, proposed a transfer framework of "CNN pre-training + SVM reclassification," mapping spatial features learned from large datasets under normal operating conditions to the fault space. This improved diagnostic accuracy by more than 15% under small sample conditions, directly verifying the transferability of CNNs in the "sample-scarce" scenario of marine gas turbines. Simultaneously, grey relational analysis, capable of handling grey systems where "partial information is known and partial information is unknown," is also used to quantify the correlation between parameters, assisting in fault location. Furthermore, traditional fault detection also combines threshold judgment and trend analysis methods to issue alarms for parameters that significantly exceed the normal range.

[0005] However, in practical scenarios of fault detection in gas turbine power systems of medium and large ships, the following challenges remain: 1) Complex physical connectivity of the system leads to weak data correlation; 2) The scarcity of publicly available marine gas turbine fault classification data severely impacts model performance; 3) Existing CNNs are mainly used to process Euclidean spatial data (grid data) such as images; however, gas turbine sensor networks are essentially non-Euclidean graph structures (graph data), with complex physical connections and feedback loops between nodes; traditional neural network schemes directly input sensor data as one-dimensional vectors into the neural network, disrupting the physical gas path connectivity between nodes; 4) Data processing methods have limitations. Traditional methods directly input one-dimensional time-series sensor data into the model without structural reorganization based on the system's physical connectivity and time-series characteristics, failing to construct spatially correlated feature data, thus hindering the performance of CNNs. 5) Insufficient model convergence and generalization ability: Existing models lack the "pre-existing knowledge of standard states" provided by thermodynamic simulation and rely solely on measured data for training. When fault data is scarce in the early stages, it is impossible to establish physical correlation constraints between parameters through grey relational analysis, resulting in slow model convergence and poor ability to identify fault types not present in the training data. 6) Poor synergy between simulation and measured data: Thermodynamic simulation is often used as an independent tool for predicting operating conditions and does not form a closed-loop feedback with real-time sensor data. It is impossible to accurately correspond the standard parameters obtained from simulation with the measured data to form "twin data" and it is also difficult to optimize the model training process through simulation results, resulting in insufficient adaptability of fault detection under multiple operating conditions. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, this invention provides a gas turbine fault detection method based on simulation prior and gas path connectivity mapping, which includes the following steps: S1. Construct a thermodynamic simulation prior model: Construct a thermodynamic simulation model corresponding to the gas turbine power system. Based on real-time collected environmental and control parameters, output the theoretical benchmark values ​​of each sensor node under standard operating conditions through iterative calculation. S2. Establishing gas path connectivity weights based on thermodynamic simulation perturbation method: In the simulation model constructed in S1, physical quantity perturbations are applied based on standard operating condition inputs, and the changes in the operating state of each node output by the simulation model are calculated; and the gas path connectivity weights are established. S3. Gas path diagram loop decoupling and serialization based on gas path connectivity weight: The sensor nodes, core components and environmental nodes of the gas turbine power system are abstracted into a directed cyclic graph. The depth-first search (DFS) algorithm is used to detect loops. When a loop is found, the connection relationship with the smallest weight in the loop is destroyed to decouple it and obtain a cyclic directed gas path graph structure. Starting from the gas turbine inlet node, the downstream node with the largest connection weight is always selected first to enter the main sequence, so as to retain the strongest main gas path feature transmission logic of the system and generate the gas path node main sequence mapping L. S4. Constructing a 3D Structured Feature Tensor: Based on the master sequence, the local direct connectivity between nodes, and the temporal relationship of sensor sampling, a 3D data mapping rule is constructed, including "time dimension - gas path influence dimension - local direct connectivity dimension". The prior reference data, calculated in real time by the thermodynamic simulation prior model described in S1 based on the gas turbine input conditions, is used as the prior reference data. It is then combined with the real-time sensor sampling data according to the mapping rule to form a 3D structured feature tensor. Differentiation, normalization, and fusion processing are then performed to obtain the final 3D structured feature tensor. ; S5. Extract fault features and output fault type and diagnosis results: Input the three-dimensional structured feature tensor described in S4 into the constructed three-dimensional convolutional neural network (3D-CNN) model, and use the translation operation of the convolution kernel in the time and space dimensions to extract fault features and output fault type and diagnosis results.

[0007] Furthermore, S2 also includes defining a connection weight between any two nodes with a direct physical connection relationship. This weight is quantified based on the sensitivity of the transmission of the state change of the upstream node to the state change of the downstream node. This weight reflects the importance of the gas path connection relationship and is used for subsequent loop decoupling and construction of master sequence mapping rules.

[0008] Furthermore, the specific calculation method for the gas path connectivity weight in S2 is as follows: Let the node... Pointing to node The connection relationship is When applying perturbation Then, any gas path node Changes in state include pressure changes Temperature changes and flow changes Gas path node Changes in state include pressure changes Temperature changes and flow changes Then the weight of the connectivity relationship for: in, For nodes pressure, For nodes temperature, For nodes Traffic, Furthermore, the specific execution process of loop decoupling and sequence generation in S3 is as follows: Loop decoupling: Using the DFS algorithm to traverse the gas path diagram, when a slave node is detected... Departing via several paths and returning To close the loop, backtrack all edges involved in the loop and compare the weights of each edge. Remove the edge with the smallest weight value until there are no closed cycles in the graph; Sequence generation: Initialize an empty sequence, first add the air inlet node to the sequence; then search all direct backward neighbor nodes of the current node, and prioritize adding the unmapped node with the largest connection weight to the sequence; if the current path has no successor node, backtrack to the node in the sequence that has an unmapped neighbor, and select the branch with the largest connection weight to continue expanding until all nodes are included in the sequence. .

[0009] Furthermore, the three-dimensional structured feature tensor in S4 The dimensions are defined as follows: Depth D: corresponds to the number of time steps in historical sampling, used to capture the temporal evolution characteristics of the fault; Height H: Corresponds to the master sequence of each gas path node generated in S3. Length; Width W: corresponds to the local spatial correlation degree. For each central node in the main sequence, it and its predecessor or successor nodes directly connected in the original physical pipeline are arranged in the same row. W is the maximum number of associated neighbors set. aisle The data channels corresponding to different physical properties should include at least three channels: temperature (T), pressure (P), and flow rate (M). Elements in a tensor Indicates time Located in the main order position The sensor and its first Physically associated neighbors in the channel The value on.

[0010] Furthermore, in step S4, when constructing the three-dimensional structured feature tensor, the sensor samples data in real time. Compared with the prior baseline data obtained from simulation The following differential fusion strategy is adopted: in This indicates a normalization operation.

[0011] Furthermore, the three-dimensional convolutional neural network in S5 includes convolutional layers and pooling layers, wherein: Convolution operations use a size of The 3D convolution kernel slides over the input tensor; where Coverage time window, Covering local sections of the main sequence of the gas pathway, The physical neighbors of the covered nodes are used to extract joint spatiotemporal coupling features spanning "time evolution-gas path transmission-local topology"; the pooling operation uses max pooling to reduce the dimensionality of the feature map, retaining the most significant fault response features within the local spatiotemporal receptive field.

[0012] Furthermore, in step S5, before outputting the fault type and diagnostic results, a fully connected layer processing step is included: flattening the feature map after convolution and pooling into a one-dimensional feature vector; normalizing the external discrete operating condition variables of the gas turbine and concatenating them with the one-dimensional feature vector; inputting the concatenated vector into a fully connected neural network (FCN), and outputting the corresponding... Probability distribution of various fault types; Furthermore, the fault type and diagnostic results output in S5 include: Output definition: Sets the total number of fault state types that the system needs to detect. The output layer contains One neuron; Probability distribution: The Softmax layer outputs an n-dimensional probability distribution vector. ,in This indicates that the current state belongs to the first... The probability of each fault state type, and ; 0-1 State Output Steps: Use the Argmax function or thresholding method to transform the probability vector into an n-dimensional 0-1 state indicator vector. : in for The elements in the final output In this context, the dimension index with a value of 1 corresponds to the specific fault state type of the gas turbine at the current moment, thereby achieving end-to-end intelligent diagnosis.

[0013] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: 1) In this invention, a simple iterative thermodynamic simulation model is constructed. In the simulation of standard operating conditions, the sensitivity of mutual interference between gas path connection nodes in the system is solved by setting small perturbations in the input parameters as the weight of the connection relationship of each node. A "loop breaking-connection mapping-reorganization" mechanism is proposed to construct a two-dimensional mapping of gas path connection of the gas turbine system. The loop breaking mechanism based on the minimum weight edge is used to eliminate loops. The adjacent connection weights are maximized by greedily maximizing and the main sequence of gas path in the system reflecting the influence of gas path is generated. A second-dimensional mapping is constructed based on local direct connection relationships. It transforms the complex cyclic graph structure into a linear sequence that conforms to the transmission logic of gas path operation characteristics. While adapting to the input requirements of convolutional neural networks, it preserves the main logic of energy transfer in the system to the greatest extent and avoids the loss of spatiotemporal causal relationships caused by random sorting.

[0014] 2) In this invention, when constructing the three-dimensional structured feature tensor, the sensor samples data in real time. Compared with the prior baseline data obtained from simulation Differential fusion is adopted. On the one hand, by introducing the residual channel of the simulation benchmark value, the common influence of the environment and variable operating conditions on the monitoring parameters is eliminated, which significantly enhances the model's sensitivity to early weak fault characteristics. On the other hand, the absolute value channel of the original data is retained, which prevents the loss of basic system state information due to excessive differentiating, thereby improving the robustness of the model under all operating conditions. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram illustrating the process of constructing three-dimensional data based on one-dimensional data collected by sensors according to the present invention; Figure 2 This is an abstract structure of the combined combustion and steam gas turbine power system of the present invention; Figure 3This is a schematic diagram illustrating the specific method by which the present invention constructs two-dimensional data from one-dimensional data collected by a sensor; Figure 4 The training results of the 3D CNN neural network constructed in this invention and the traditional fully connected neural network model are shown in the figure.

[0017] The attached diagram includes the following sensor labels: A1, Combustion chamber directional outlet sensor of the compressor; A2, Combustion turbine directional outlet sensor of the compressor; A3, Air inlet sensor of the compressor; B1, Combustion chamber gas inlet; B2, Combustion chamber exhaust gas outlet; C1, Combustion turbine high-pressure airflow inlet sensor; C2, Combustion turbine gas exhaust gas inlet sensor; C3, Combustion turbine power generation sensor; C4, Combustion turbine waste gas outlet sensor; D1, Waste heat boiler gas inlet sensor; D2, Waste heat boiler total exhaust port sensor; D3, Waste heat boiler total steam outlet. Sensors: D4, Waste heat boiler medium inlet sensor; E1, Steam turbine main steam inlet sensor; E2, Steam turbine power generation sensor; E3, Steam turbine condensate outlet sensor; E4, Steam turbine reheat steam outlet sensor; F1, Feedwater heater steam inlet sensor; F2, Feedwater heater condensate inlet sensor; F3, Feedwater heater total return outlet sensor; G1, Condenser steam inlet sensor; G2, Condenser condensate outlet sensor; Axis 4 represents the node dimension constructed through direct connectivity, axis 5 represents the node dimension containing trunk connectivity constructed through maximum weighted gas path serialization, and axis 3 represents the time dimension constructed through temporal sequence. Detailed Implementation

[0018] like Figure 1-4 As shown, the present invention provides a gas turbine fault detection method based on simulation prior and gas path connectivity mapping; including the following steps: S1. Construct a thermodynamic simulation prior model. Specifically, construct a thermodynamic simulation model corresponding to the gas turbine power system. Based on real-time collected environmental and control parameters, output the theoretical benchmark values ​​of each sensor node under standard operating conditions through iterative calculation. To obtain theoretical baseline values ​​under standard conditions, a connectivity analysis of the system's gas path was performed. Taking a combined gas turbine and gas turbine as an example, a thermodynamically coupled simulation model of a combined gas-steam cycle based on the equation of state of the actual working fluid was constructed using the open-source physical simulation library Coolprop. The system takes as input operating parameters and outputs the operating status of each gas path node, including pressure P, temperature T, and flow rate M. This coupled simulation model incorporates the Helmholtz energy state equation kernel and generates standardized state data containing multi-dimensional physical properties such as pressure, temperature, and flow rate through a closed-loop iteration of "forward calculation on the gas side" and "reverse derivation on the steam-water side." Specific steps include: S1.1 A real thermophysical property mapping model of the working fluid is constructed, and a thermophysical property calculation kernel is used to replace the ideal gas law. For any node in the system, the working fluid state is accurately calculated through the mapping function to ensure that the model can accurately capture the nonlinear property changes of steam phase change and gas high temperature and high pressure region; Where P is pressure and T is temperature. The working fluid type is air, fuel gas, water, or steam. The output state vector contains specific enthalpy h, specific entropy s, and density. and specific heat capacity at constant pressure ; S1.2. Construct a top-level gas turbine cycle simulation model. Based on the design operating condition input, calculate the state of each component sequentially along the working fluid flow direction: Compressor Model: Based on Pressure Ratio and isentropic efficiency Calculate the outlet specific enthalpy of the non-isentropic compression process. ; Combustion chamber model: Based on the principle of energy conservation, fuel flow rate is solved iteratively. Satisfies the equation: ;in For the lower heating value of fuel, For combustion efficiency, Design temperature for turbine inlet, For airflow, The specific enthalpy of the air at the combustion chamber inlet. Design temperature based on turbine inlet The specific enthalpy of the combustion chamber outlet gas is determined; Gas turbine model: Calculating exhaust temperature based on expansion ratio This parameter serves as the heat source boundary for subsequent waste heat utilization; S1.3 Construct a simulation model of the waste heat boiler and the bottom steam cycle, build an energy coupling interface connecting the top and bottom cycles, and calculate the total heat released on the gas side based on dynamic heat balance. , ,in For heat exchange efficiency, This represents the gas flow rate.

[0019] Based on the law of energy conservation, the steam flow rate is derived in reverse. , ;in The enthalpy of the main steam outlet. This refers to the enthalpy value at the water inlet. Constructing a Rankine cycle and regenerative model: based on steam flow rate The non-isentropic extraction expansion process of a steam turbine was simulated, and a dual mass and energy balance equation for the feedwater heater was established to calculate the actual feedwater state after regenerative mixing.

[0020] S1.4 A global multivariable closed-loop iterative solution is used to solve for the flue gas temperature on the gas side. Combustion chamber load and steam-water side feedwater temperature The presence of strong coupling nonlinear characteristics, Residual definition: Define the error vector Includes the energy balance residual of the feedwater heater. Sum of end difference constraint residuals ; Iterative correction: During the calculation process, the estimated value of the flue gas temperature is dynamically adjusted. and the guessed value of water supply temperature If the calculated regenerated water temperature is higher than the predicted value, then the damping factor will be used. Revised Guessing Baseline: Here, the superscript (k) of the variable indicates that the number of iterations is k. This is the temperature value calculated in the current iteration step calc; Convergence criterion: When the system energy residual When the value is less than a preset tolerance (e.g., 0.0001), the simulation is considered converged, and the steady-state data of all sensor nodes in the entire system are output: .

[0021] S2. Establishing Gas Path Connectivity Weights Based on Thermodynamic Simulation Perturbation Method: In the simulation model constructed in S1, physical quantity perturbations are applied based on the standard operating condition input, and the changes in the operating state of each node in the simulation model output are calculated; the connectivity weight between any two nodes with direct physical connectivity is defined. This weight is quantified based on the sensitivity of the upstream node's state change to the downstream node's state change; this weight reflects the importance of the gas path connectivity relationship and is used for subsequent loop decoupling and construction of the main sequence mapping rule. The specific steps are as follows: S2.1 Obtain the reference state of the gas path node: Using existing gas turbine thermodynamic simulation modules ( The input is defined as operating condition parameters, which include environmental parameters and control parameters. The output is... This includes all nodes under this working condition. thermodynamic state : , Representative node Pressure at the point , Representative node Temperature at the location Representative node flow rate at the location Under standard operating conditions Run the simulation to obtain the baseline state of the gas path nodes: .

[0022] S2.2 Perturbation Application and Sensitivity Calculation: Input under standard operating conditions of the target gas turbine Based on this, apply a small physical quantity perturbation. Running the simulation after disturbance, i.e., input And calculate the change in the output data. : The change It includes the state fluctuations of all nodes: for any node Its state change vector is ,in For nodes Pressure changes, For nodes Temperature changes, For nodes Changes in traffic flow.

[0023] S2.3 Weighting of gas path connectivity: For every physical connectivity relationship (i.e., directed edge) that exists in the system. Define its weights This weight represents the upstream node Fluctuations on downstream nodes The ability to transmit influence is calculated using the following formula: .

[0024] Compared to traditional fault detection methods based on fully connected networks, the above method can dynamically capture the actual thermodynamic coupling strength between components under the current operating conditions, accurately reflect the main relationship of gas path feature transmission, and thus provide more physically meaningful topological priors for subsequent feature extraction.

[0025] S3. Gas path diagram loop decoupling and serialization based on gas path connectivity weight: The sensor nodes, core components and environmental nodes of the gas turbine power system are abstracted into a directed cyclic graph. The depth-first search (DFS) algorithm is used to detect loops. When a loop is found, the connection relationship with the smallest weight in the loop is destroyed to decouple it and obtain a cyclic directed gas path graph structure. Starting from the gas turbine inlet node, the downstream node with the largest connection weight is always selected first to enter the main sequence, so as to retain the strongest main gas path feature transmission logic of the system and generate the gas path node main sequence mapping L. The specific steps include: S3.1 System Diagram Structure Abstraction: To construct a structured mapping of the gas turbine gas path state, the gas turbine gas path system is defined as a directed graph. : where the set of nodes Represents the various physical components in the gas path (such as compressor stages, combustion chambers, turbine stages, etc.); a set of node connection relationships (edges). : E represents the directional direct connections between the physical components in the gas path, and the weights of all connections in E form a set of weights. : .

[0026] S3.2 loop damage: The gas turbine's gas path structure is complex, containing physical feedback loops. To adapt to the input of deep learning models, these loops must be broken through algorithms.

[0027] This invention proposes a loop disruption technique based on the minimum gas turbine gas path weight: Detect all closed loops in the diagram, and for each loop, identify and disconnect it. The smallest edge (i.e., the feedback path with the weakest impact) is used to transform the cyclic graph into a directed acyclic graph (DAG). Cycle detection and minimum weight edge removal are implemented as follows: The following algorithm is used to implement the input directed weighted graph. and weight set Output an acyclic directed graph. : S3.2.1 Initialization Definition: Let the input image be Define the target graph for decoupling. The initial state is A complete copy: ,in ; S3.2.2 Iterative decoupling process: The evolution of graph structure is achieved through edge sets. The iterative correction is implemented by selectively removing edges based on the weights of the gas path node connections during the iteration process, and defining loop state indicator variables for the graph. The iterative process is as follows: Step A: In the current In a given context, if a closed path exists, the set of edges formed by that path is defined as follows: : like Then let the state variable Proceed to step B; if Then let the state variable The iteration terminates.

[0028] Step B: In the detected set of loops In the middle, find the edge with the smallest weight; Define minimum weight value for: Define the minimum weight edge to be removed. For edge elements that satisfy the above minimum value: Step C: Perform the edge removal operation, removing the identified edges from the current connectivity set. ,renew : After the update is complete, return to step A to continue the detection until the system reaches a loop-free state.

[0029] S3.2.3 Final state output: When the iterative process meets the termination condition, that is... Output the final acyclic directed graph. : at this time, It satisfies the prerequisites for topological sorting and retains all main airway features except for the minimum weight feedback edge. S3.3 Construction of a one-dimensional sequence mapping based on greedy sorting of gas turbine gas path correlation: In acyclic graph Based on this, and according to the primary and secondary relationships of gas path connectivity, a greedy strategy is used to flatten the topology into a one-dimensional sequence. ; An algorithm for constructing the main sequence mapping of the gas pathway is implemented using an input acyclic directed graph. Output a one-dimensional ordered sequence L, as shown in the following implementation: S3.3.1 Definition of Sets and Variables set up This is the decoupled acyclic directed graph; : Represents the final generated ordered sequence of nodes; : Represents the set of candidate nodes to be processed (Candidate Set); : Represents the pointer node currently being traversed; : Represents a directed edge The connectivity weight.

[0030] S3.3.2 Initialization process The initial state for sequence generation is defined as follows: Initialization sequence ,in For the gas turbine inlet node; initialize the candidate set. Set the current node pointer ;Will Marked as "Mapped".

[0031] S3.3.3 Sequence Iteration Construction Process The sequence is expanded through iterative computation until all nodes are included. In each iteration, perform the following steps: Step A: Candidate set update based on the current node Define its direct backward node set : Update candidate set ,Will Not yet included Add nodes to the set: Step B: Determining the Optimal Successor Node and State Transitioning. Define the set of feasible downstream nodes for the current node. : according to The state of the set is calculated in two cases for the next time step. Target node: (1) If Prioritize selecting the connection with the highest weight from the current node's direct downstream nodes; define the optimal successor node. for: Execution status update: (2) If When the current path is interrupted, find the strongest associated breakpoint between the mapped nodes and the candidate nodes to be processed globally; if Define the backtracking start node satisfy: The above expression represents finding a pair of nodes. , making the edge The weight is the largest among all possible connections, and this is taken. node.

[0032] Execution status update: Repeat steps A and B above until... Output the final sequence ; Through the above steps, the complex gas turbine topology was successfully transformed into a one-dimensional gas path master sequence. : in: Let be the i-th node after sorting; H is the sequence length, which is the total number of gas turbine gas path nodes; for any edge in In the sequence middle Must be located in Before.

[0033] S4. Constructing a 3D Structured Feature Tensor: Based on the master sequence, the local direct connectivity between nodes, and the temporal relationship of sensor sampling, a 3D data mapping rule is constructed, including "time dimension - gas path influence dimension - local direct connectivity dimension". The prior reference data, calculated in real time by the thermodynamic simulation prior model described in S1 based on the gas turbine input conditions, is used as the prior reference data. It is then combined with the real-time sensor sampling data according to the mapping rule to form a 3D structured feature tensor. Differentiation, normalization, and fusion processing are then performed to obtain the final 3D structured feature tensor. ; The specific steps are as follows: For each node in sequence L Retrieve the set of its directly related nodes in the original node graph G, including: Self-node: It itself; upstream node: direct inflow Downstream node: Nodes that flow out directly; Based on the above node set, a local neighborhood matrix is ​​constructed, specifically a two-dimensional spatial feature matrix at a certain time t. : Where the parameters are: : Represents the spatial characteristic matrix at time t. It is a structured data of the operating state of the entire system at that time. Each row of the matrix represents a "local physical system". : Corresponding to the aforementioned sequence The length of , i.e., the total number of system nodes; The local association width is the maximum number of neighbors. Column index j represents the master node. The (jp)th gas path association node; row index i corresponds to the i-th node in the topological sort sequence. ; The master node itself, i.e., the sequence. The i-th node The sampled value at time t; the present invention stipulates that the p-th column j=p stores the node's own data, which is specially designed in consideration of the situation that the convolution kernel of the subsequent convolutional neural network may lose features for edge data; : Feature values ​​of associated neighboring nodes; i.e., nodes The The normalized residual value of the first direct gas path connection node (upstream or downstream node) at time t; if there is no first direct gas path connection node (upstream or downstream node); A direct gas path connection node, then .

[0034] This matrix will include the "gas system backbone" (i.e., sequence) The distribution along the vertical axis H) is combined with the "local component coupling" (distributed along the horizontal axis W). This allows the convolution kernel to capture both the feature transfer trend of the entire system and the interaction between individual components and their upstream and downstream components as it slides across the matrix.

[0035] Sensor sampling data is stored in the database with a constant 100 samples. For each sampling moment, new data is saved while old data is removed. These 100 samples are then used to construct a three-dimensional structured dataset based on their temporal sequence. .

[0036] The three-dimensional structured feature tensor in S4 The dimensions are defined as follows: Depth D: corresponds to the number of time steps in historical sampling, used to capture the temporal evolution characteristics of the fault; Height H: Corresponds to the master sequence of each gas path node generated in S3. Length; Width W: corresponds to the local spatial correlation degree. For each central node in the main sequence, it and its predecessor or successor nodes directly connected in the original physical pipeline are arranged in the same row. W is the maximum number of associated neighbors set. aisle The data channels corresponding to different physical properties should include at least three channels: temperature (T), pressure (P), and flow rate (M). Elements in a tensor Indicates time Located in the main order position The sensor and its first Physically associated neighbors in the channel The value on.

[0037] When constructing the three-dimensional structured feature tensor, real-time sampling data from the sensor is used. Compared with the prior baseline data obtained from simulation The following differential fusion strategy is adopted: in This indicates a normalization operation.

[0038] The above fusion, on the one hand, eliminates the common influence of environment and variable operating conditions on monitoring parameters by introducing the residual channel of simulation benchmark value, and significantly enhances the model's sensitivity to early weak fault characteristics; on the other hand, it retains the absolute value channel of the original data, preventing the loss of basic system state information due to excessive differencing, thereby improving the robustness of the model under all operating conditions.

[0039] S5. Extract fault features and output fault type and diagnosis results: Input the three-dimensional structured feature tensor described in S4 into the constructed three-dimensional convolutional neural network (3D-CNN) model, and use the translation operation of the convolution kernel in the time and space dimensions to extract fault features and output fault type and diagnosis results.

[0040] The specific steps are as follows: S5.1 Extract spatiotemporal coupling features; Convolution operations use a size of The 3D convolutional kernel slides across the input tensor; this sliding kernel performs computation on the input tensor to extract joint features, where... Coverage time window, Covering local sections of the main sequence of the gas pathway, The physical neighbors of the covered nodes are used to extract joint spatiotemporal coupling features spanning "temporal evolution-gas path transmission-local topology"; S5.2 Feature Dimensionality Reduction and Vectorization; Max pooling is used to reduce the dimensionality of the convolutional feature map, preserving the most significant fault response features within the local receptive field and suppressing noise interference. Subsequently, a Flatten layer flattens the multidimensional feature map into a one-dimensional deep feature vector. This is to prepare for subsequent classification.

[0041] S5.3 Multi-source operating condition feature fusion; To eliminate the masking of fault characteristics by varying operating conditions, this model introduces a fully connected layer for feature fusion; the external discrete operating condition variables of the gas turbine (including ambient temperature, commanded load, speed setpoint, etc.) are normalized to obtain the operating condition vector. ; Compare this vector with the deep feature vector output by S5.2 The features are concatenated to form a fused feature vector. : The above fusion uses a fully connected layer to post-fuse discrete external operating condition parameters with deep spatiotemporal features, effectively providing an "operational background" context for fault identification. This enables the model to distinguish between parameter drift caused by changes in operating conditions and actual fault deviations, thereby reducing the false alarm rate under varying operating conditions.

[0042] S5.4 Fault probability calculation and status output; fuse feature vectors Input a fully connected classification network (FCN), with the ends connected to a Softmax normalized exponential function, which maps to the fault category space; Output definition: Sets the total number of fault state types that the system needs to detect. (For example: no faults, compressor fouling, turbine blade erosion, sensor misalignment, etc.), then the output layer contains One neuron.

[0043] Probability distribution: The Softmax layer outputs an n-dimensional probability distribution vector. ,in This indicates that the current state belongs to the first... The probability of each fault state type, and .

[0044] 0-1 State Output Steps: To provide a clear diagnostic conclusion, the probability vector is transformed into an n-dimensional 0-1 state indicator vector using the Argmax function or a threshold method. : in: in for The elements in the final output In this context, the dimension index with a value of 1 corresponds to the specific fault state type of the gas turbine at the current moment, thereby achieving end-to-end intelligent diagnosis.

[0045] In this specification, the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the descriptions of the embodiments described later are relatively simple, and relevant parts can be referred to the descriptions of the foregoing embodiments.

[0046] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A gas turbine fault detection method based on simulation priors and gas path connectivity mapping, characterized in that, Includes the following steps: S1. Construct a thermodynamic simulation prior model: Construct a thermodynamic simulation model corresponding to the gas turbine power system. Based on real-time collected environmental and control parameters, output the theoretical benchmark values ​​of each sensor node under standard operating conditions through iterative calculation. S2. Establishing gas path connectivity weights based on thermodynamic simulation perturbation method: In the simulation prior model constructed in S1, physical quantity perturbation is applied on the basis of standard operating condition input, and the change in the operating state of each node output by the simulation model is calculated. And establish the weighting of gas path connections; S3. Gas path diagram loop decoupling and serialization based on gas path connectivity weight: The sensor nodes, core components and environmental nodes of the gas turbine power system are abstracted into a directed cyclic graph. The depth-first search (DFS) algorithm is used to detect loops. When a loop is found, the connection relationship with the smallest weight in the loop is destroyed to decouple it and obtain a cyclic directed gas path graph structure. Starting from the gas turbine inlet node, the downstream node with the largest connection weight is always selected first to enter the main sequence, so as to retain the strongest main gas path feature transmission logic of the system and generate the gas path node main sequence mapping L. S4. Constructing a 3D Structured Feature Tensor: Based on the master sequence, the local direct connectivity between nodes, and the temporal relationship of sensor sampling, a 3D data mapping rule is constructed, including "time dimension - gas path influence dimension - local direct connectivity dimension". The prior reference data, calculated in real time by the thermodynamic simulation prior model described in S1 based on the gas turbine input conditions, is used as the prior reference data. It is then combined with the real-time sensor sampling data according to the mapping rule to form a 3D structured feature tensor. Differential processing, separate normalization, and fusion processing are then performed to obtain the final 3D structured feature tensor. ; S5. Extract fault features and output fault type and diagnosis results: Input the three-dimensional structured feature tensor described in S4 into the constructed three-dimensional convolutional neural network (3D-CNN) model, and use the translation operation of the convolution kernel in the time and space dimensions to extract fault features and output fault type and diagnosis results.

2. The method according to claim 1, characterized in that, S2 further includes defining a connection weight between any two nodes with a direct physical connection relationship. This weight is quantified based on the sensitivity of the transmission of the state change of the upstream node to the state change of the downstream node. This weight reflects the importance of the gas path connection relationship and is used for subsequent loop decoupling and construction of master sequence mapping rules.

3. The method according to claim 2, characterized in that, The specific calculation method for the gas path interconnection weight in S2 is as follows: Let the node... Pointing to node The connection relationship is When applying perturbation Then, any gas path node Changes in state include pressure changes Temperature changes and flow changes Gas path node Changes in state include pressure changes Temperature changes and flow changes Then the weight of the connectivity relationship for: , in, For nodes pressure, For nodes temperature, For nodes Traffic, 4. The method according to claim 1, characterized in that, The specific execution process of loop decoupling and sequence generation in S3 is as follows: Loop decoupling: The gas path diagram is traversed using the DFS algorithm. When a slave node is detected... Departing via several paths and returning To close the loop, backtrack all edges involved in the loop and compare the weights of each edge. Remove the edge with the smallest weight value until there are no closed cycles in the graph; Sequence generation: Initialize an empty sequence, first add the air inlet node to the sequence; then search all direct backward neighbor nodes of the current node, and prioritize adding the unmapped node with the largest connection weight to the sequence; if the current path has no successor node, backtrack to the node in the sequence that has an unmapped neighbor, and select the branch with the largest connection weight to continue expanding until all nodes are included in the sequence. .

5. The method according to claim 4, characterized in that, The three-dimensional structured feature tensor in S4 The dimensions are defined as follows: Depth D: corresponds to the number of time steps in historical sampling, used to capture the temporal evolution characteristics of the fault; Height H: Corresponds to the master sequence of each gas path node generated in S3. Length; Width W: corresponds to the local spatial correlation degree. For each central node in the main sequence, it and its predecessor or successor nodes directly connected in the original physical pipeline are arranged in the same row. W is the maximum number of associated neighbors set. aisle The data channels corresponding to different physical properties should include at least three channels: temperature (T), pressure (P), and flow rate (M). Elements in a tensor Indicates time Located in the main order position The sensor and its first Physically associated neighbors in the channel The value on.

6. The method according to claim 1, characterized in that, In step S4, when constructing the three-dimensional structured feature tensor, the sensor samples data in real time. Compared with the prior baseline data obtained from simulation The following differential fusion strategy is adopted: , in This indicates a normalization operation.

7. The method according to claim 1, characterized in that, The three-dimensional convolutional neural network in S5 includes convolutional layers and pooling layers, wherein: Convolution operations use a size of The 3D convolution kernel slides over the input tensor; where Coverage time window, Covering local sections of the main sequence of the gas pathway, The physical neighbors of the covered nodes are used to extract joint spatiotemporal coupling features spanning "time evolution-gas path transmission-local topology"; the pooling operation uses max pooling to reduce the dimensionality of the feature map, retaining the most significant fault response features within the local spatiotemporal receptive field.

8. The method according to claim 7, characterized in that, In step S5, before outputting the fault type and diagnosis result, a fully connected layer processing step is also included: flattening the feature map after convolution and pooling into a one-dimensional feature vector; normalizing the external discrete operating condition variables of the gas turbine and concatenating them with the one-dimensional feature vector. The concatenated vector is input into a fully connected neural network (FCN), and the corresponding vector is output through the Softmax layer. The probability distribution of various fault types.

9. The method according to claim 8, characterized in that, The fault types and diagnostic results output in S5 include: Output definition: Sets the total number of fault state types that the system needs to detect. The output layer contains One neuron; Probability distribution: The Softmax layer outputs an n-dimensional probability distribution vector. ,in This indicates that the current state belongs to the first... The probability of each fault state type, and ; 0-1 state output steps: Use the Argmax function or threshold method to transform the probability vector into an n-dimensional 0-1 state indicator vector. : , ， for The elements in the final output In this context, the dimension index with a value of 1 corresponds to the specific fault state type of the gas turbine at the current moment, thereby achieving end-to-end intelligent diagnosis.

Images