Intelligent online abnormality detection method and system for gas turbine

Abstract

This invention provides an intelligent online anomaly detection method and system for gas turbines. The method includes acquiring historical operating data and real-time multi-source sensor data for typical operating conditions throughout the entire lifecycle; constructing a refined semantic space and physical causal graph for operating conditions; identifying operating conditions based on a lightweight classifier; and modeling causal relationships between components through dynamic routing of the causal graph. It dynamically calibrates heterogeneous features using a multi-head attention mechanism and residual gating units, and performs graph neural network inference along directed causal paths. Anomaly identification and localization are performed using a combination of data-driven and knowledge-guided approaches, and a consistency constraint mechanism is used to fuse decisions, ensuring high-confidence output. The anomaly attribution stage employs gradient-weighted activation mapping and causal counterfactual reasoning, ultimately generating a structured anomaly cause explanation text. This invention possesses high real-time performance, strong causal explanation, and closed-loop optimization capabilities, effectively supporting the intelligent fault diagnosis process for the operation and maintenance of gas turbines in power plants.

Description

Technical Field

[0001] This invention relates to the field of intelligent condition monitoring and anomaly detection technology for gas turbines, and in particular to an intelligent online anomaly detection method and system for gas turbines. Background Technology

[0002] Existing intelligent online anomaly detection systems for gas turbines are widely used in the power industry to improve the automation and intelligence level of gas turbine operation and maintenance. Current mainstream technologies mostly employ multi-source data fusion, deep learning, or data-driven feature analysis methods to achieve status monitoring and anomaly alarms for key gas turbine components. Common solutions include sliding window feature extraction based on time-series signals, weighted fusion decision-making (such as Pearson correlation coefficient weighting and information entropy weighting), single-channel or multi-channel convolutional long short-term memory networks (CNN-LSTM), and dynamic Bayesian networks. These methods can process multimodal operating data of gas turbines (including vibration, temperature, pressure, current, etc.) in real time and output anomaly alarms by setting thresholds, anomaly probabilities, or confidence indices. On the other hand, with the advancement of knowledge graph and causal reasoning technologies, some technologies have introduced equipment status knowledge graphs, directed fault propagation graphs, and frequency band sensitivity matrices to improve the intelligence level and accuracy of anomaly detection.

[0003] Representative technologies primarily focus on the fusion performance and discrimination accuracy of anomaly detection models. For example, publicly available dynamic Bayesian networks and hypersphere single-classification models can, to some extent, achieve joint analysis of multi-source heterogeneous data, enabling preliminary anomaly discrimination for key gas turbine operating conditions (such as cold / hot start-up, load changes, and transition anomalies). For each operating stage of the gas turbine, some solutions have established knowledge representation structures for equipment status and basic mapping rules with sensor signals. However, the application scope of these methods is still mainly concentrated on "is it abnormal?" alarms and preliminary fault location, with limited ability to provide in-depth explanations of anomaly causes and decision support.

[0004] Analysis of existing technologies reveals the following key issues: First, existing anomaly detection systems typically employ static weighted fusion or black-box deep learning decision-making, and their results mainly output anomaly probabilities or anomaly regions, lacking the ability to trace and explain the causes of anomalies. Second, given the dynamic coupling of multi-source heterogeneous sensor data and the significant differences between operating conditions, existing weighted fusion schemes (such as mutual information weighting, Pearson weighting, etc.) are difficult to automatically adapt to changes in operating conditions, resulting in weak model transferability and insufficient robustness. Third, the knowledge injection path relies heavily on the disclosed equipment status map or fault propagation map, but its causal reasoning process is limited by the static topology and cannot be dynamically attributed in combination with the actual operating status of the gas turbine. Fourth, existing deep learning and data fusion technologies generally suffer from the "black box" problem, that is, the anomaly identification results lack a transparent explanation of the relationship between the original signal characteristics and physical mechanisms, making it difficult to effectively support operation and maintenance personnel in making accurate fault diagnosis and maintenance decisions.

[0005] Therefore, there is an urgent need to develop a gas turbine anomaly identification method that can achieve adaptive operating conditions, physical causal constraints, dual-path reasoning, consistency fusion, and highly interpretable attribution. Summary of the Invention

[0006] This invention provides a method and system for intelligent online anomaly detection of gas turbines, aiming to solve the problems existing in the prior art mentioned in the background section.

[0007] In a first aspect, the present invention provides an intelligent online anomaly detection method for gas turbines, comprising the following steps: S1: Obtain historical operating data of the gas turbine; and predict the current operating condition category label and its confidence level; S2: Dynamically route based on operating condition category labels and load a dedicated physical causal sub-graph that matches the current operating condition; S3: Collect real-time multi-source heterogeneous time-series data of the gas turbine, and input the extracted local feature representation into the gated multi-head attention mechanism. Initialize the attention weight distribution of each sensor channel according to the PageRank centrality index of the causal graph nodes. S4: Based on the initial attention weights, the multi-source features are dynamically calibrated through the residual gating unit to generate a weighted joint feature representation of the time series; and the local feature representation is mapped to the causal sub-graph node space. The abnormal observation signal is propagated along the directed causal edge using the graph neural network to generate the causal influence score vector of each component node. S5: Perform anomaly probability prediction on the weighted time series joint feature representation to obtain anomaly discrimination results; identify key component nodes with significantly increased scores from the causal influence score vector to form anomaly localization results for the knowledge-guided path; calculate the KL divergence between the output distributions of the two paths and determine whether it is less than a preset threshold. S6: If satisfied, the outputs of the data-driven path and the knowledge-guided path are merged to generate the final anomaly identification decision; otherwise, the adversarial perturbation regularization feedback mechanism is triggered to adjust the weight update strategy in the gated multi-head attention mechanism, suppress the false activation of non-causal features, and re-execute steps S4 to S5 until the consistency constraint condition is satisfied.

[0008] Preferably, the method further includes: S7: Once an abnormal event is confirmed, the attribution inversion engine is activated to record the change in the causal impact score of the target abnormal node. S8: Based on the score changes after each causal edge is masked in step S7, extract the Top-3 causal path chains that most significantly reduce the probability of anomalies, and combine them with the spatiotemporal distribution map of key features to generate a structured explanation text of the cause of anomalies containing physical meaning.

[0009] Preferably, in step S3, local time-frequency feature extraction is performed on the time-series segments of the multi-source heterogeneous time-series data: for the vibration signal, short-time Fourier transform is used to calculate its time-varying spectral energy density matrix, and the energy envelope curves of multiple key sub-frequency bands in the 0-2kHz frequency band are extracted; for the temperature and pressure signals, discrete wavelet transform is applied to decompose the signals, and the db6 wavelet basis is selected to obtain the low-frequency trend component and high-frequency detail coefficients, and a multi-scale gradient feature vector is constructed; for the current signal, Hilbert-Huang transform is applied, and the first three intrinsic mode functions are obtained through empirical mode decomposition, and their instantaneous amplitude and frequency variance are calculated to form a non-stationary electrical feature representation; finally, the local time-frequency feature tensor is output.

[0010] Preferably, in step S5, a lightweight fully connected network is used to perform nonlinear mapping processing to project high-dimensional temporal features onto the anomaly probability space and output the anomaly discrimination result of the data-driven path. The causal impact score vector is normalized using Softmax to obtain the standardized impact distribution Q of the knowledge guidance path; The KL divergence is calculated based on two probabilistic output distributions: the normalized anomaly probability distribution P and the standardized influence distribution Q.

[0011] Secondly, the present invention provides an intelligent online anomaly detection system for gas turbines, wherein the system uses the above-mentioned method to realize intelligent online anomaly detection of gas turbines.

[0012] The beneficial technical effects of this invention are as follows: 1. This invention achieves the organic unity of dynamic fusion strategy and interpretable diagnostic capability in the anomaly identification process by constructing a collaborative architecture of working condition semantic space and dedicated physical causal graph; 2. This invention adopts a graph routing mechanism driven by operating condition labels, dynamically loading causal structures that match the current operating state online to ensure that the inference logic always follows the real physical constraints of the gas turbine under specific operating conditions. At the same time, attention initialization is performed by combining the PageRank centrality index of sensor nodes in the causal graph, and residual gating units are introduced to achieve secondary calibration of data-driven features. This significantly improves the model's sensitivity and response accuracy to key abnormal signals, especially showing stronger robustness and stability in transient processes such as cold start-up and rapid load shedding, avoiding misjudgments caused by parameter drift or noise interference. 3. This invention proposes a dual-path consistency constraint mechanism and an attribution inversion engine, realizing the parallel processing of original time-series signals and physical causal propagation processes from anomaly detection to causal tracing. The data path and knowledge path are used to process the original time-series signals and physical causal propagation processes in parallel. The KL divergence constraint forces the output distribution of the two to maintain semantic consistency, suppresses false activation of non-causal features, and improves the physical rationality of decision results. The adversarial perturbation regularization term is introduced during the training phase to further enhance the model's resistance to uncertain perturbations and ensure the reliability of high-confidence judgments. 4. When an abnormal event is identified, the system automatically initiates attribution inversion. On the one hand, Grad-CAM is used to locate key time windows and frequency bands to reveal the local time-frequency domain contribution area. On the other hand, counterfactual masking experiments are performed on the causal subgraph to quantify the influence of each causal path on the abnormal score, generate the Top-3 causal chain sequence, and finally fuse them to form a structured explanatory text. Attached Figure Description

[0013] Figure 1 A flowchart illustrating the method of Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the process for generating attention weight distribution in an embodiment of the present invention; Figure 3 This is a schematic diagram of the process for generating causal influence score vectors in an embodiment of the present invention. Detailed Implementation

[0014] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0015] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0016] like Figure 1 As shown in the figure, this embodiment provides a method for intelligent online anomaly detection of gas turbines, which specifically includes the following steps: S1: Acquire historical operating data of the gas turbine under various typical operating conditions, including cold start, hot start, 50% load ramp-up, 90% load steady state, rapid load shedding, and combustion oscillation transition state. Label key control parameter combinations for each operating condition, construct a full life cycle operating condition semantic space based on the parameter combinations, and train a lightweight operating condition state classifier to output the current operating condition category label and its confidence level. S2: Based on the operating condition category label output by S1, dynamically route and load a dedicated physical causal sub-graph matching the current operating condition from the pre-stored physical causal graph library. The sub-graph nodes are key components of the gas turbine and their observable state variables. The edge relationships are defined by thermodynamic equations, rotor dynamics models and historical fault root cause analysis reports, and are assigned causal strength weights to form a structured causal reasoning path constrained by physical laws. S3: Synchronously acquire real-time multi-source heterogeneous time-series data of the gas turbine, including vibration, temperature, pressure and current signals. Extract local time-frequency features from each type of signal to generate local feature representations corresponding to four types of original time-series segments. Input the feature representations into the gated multi-head attention mechanism and initialize the attention weight distribution of each sensor channel according to the PageRank centrality index of the caustic graph nodes in step S2. S4: Based on the initial attention weights generated in step S3, the multi-source features are dynamically calibrated through the residual gating unit, and the contribution of each channel is fused to generate a weighted joint feature representation of the time series; at the same time, the original features extracted in S3 are mapped to the causal sub-graph node space loaded in S2, and the abnormal observation signal is propagated along the directed causal edge using the graph neural network to generate the causal influence score vector of each component node. S5: Perform anomaly probability prediction on the weighted time series joint feature representation generated in step S4, and output the anomaly discrimination result of the data-driven path; at the same time, normalize the causal influence score vector generated in S4, identify key component nodes with significantly increased scores, and form the anomaly localization result of the knowledge-guided path; based on the dual-path consistency constraint mechanism, calculate the KL divergence between the output distributions of the two paths, and determine whether it is less than the preset threshold of 0.15; S6: If the KL divergence in step S5 satisfies the condition of being less than 0.15, then the output results of the data-driven path and the knowledge-guided path are fused to generate the final anomaly identification decision; otherwise, the adversarial perturbation regularization feedback mechanism is triggered to adjust the weight update strategy in the gated multi-head attention mechanism, suppress the false activation of non-causal features, and re-execute steps S4 to S5 until the consistency constraint condition is met. S7: Once an abnormal event is confirmed, the attribution inversion engine is activated: Based on the weighted time series joint feature representation in step S4, the gradient weighted class activation mapping method is used to locate the time window and frequency band region with the highest contribution in each sensor signal, and generate a spatiotemporal distribution map of key features at the data level; at the same time, counterfactual reasoning is performed on the causal subgraph loaded in step S2, causal edges are masked one by one and the graph neural network is run again, and the change in the causal influence score of the target abnormal node is recorded; S8: Based on the score changes after each causal edge is masked in step S7, extract the Top-3 causal path chains that most significantly reduce the probability of anomalies. Combine the spatiotemporal distribution map of key features to generate a structured anomaly cause explanation text containing physical meaning. Push the explanation text to the visualization module to support power plant operation and maintenance personnel in making traceable fault diagnosis decisions.

[0017] In this embodiment, step S1 involves acquiring historical operating data of the gas turbine under various typical operating conditions, including cold start, hot start, 50% load ramp-up, 90% load steady state, rapid load shedding, and combustion oscillation transition state. Key control parameter combinations are labeled for each operating condition. A full lifecycle operating condition semantic space is constructed based on these parameter combinations, and a lightweight operating condition state classifier is trained to output the current operating condition category label and its confidence level. This includes the following steps: S1.1: Obtain historical operating data sequences of the gas turbine under six typical operating conditions: cold start, hot start, 50% load ramp-up, 90% load steady state, rapid load shedding, and combustion oscillation transition state. The data sequences include time series signals of key control parameters such as speed, fuel valve opening, exhaust temperature, and compressor outlet pressure, which serve as the original input dataset for constructing the operating condition semantic space. S1.2: The time series signals of key control parameters under various operating conditions obtained in step S1.1 are segmented and normalized. The statistical feature vectors of each type of operating condition are extracted based on the sliding window, including the mean, standard deviation, rate of change slope, spectral entropy value and higher-order moment index, to generate a set of distinctive operating condition feature representations to characterize the dynamic behavior patterns of different operating stages. S1.3: Based on the set of operating condition feature representations generated in step S1.2, hierarchical clustering algorithm combined with expert annotation information is used to align the samples into categories, define 12 categories of refined operating state labels (including sub-modes), construct a discrete operating condition semantic space covering the entire life cycle of the gas turbine, and form a state mapping benchmark system from original observation to high-level semantics; S1.4: Based on the working condition semantic space constructed in step S1.3, a labeled training sample set is constructed, and a lightweight working condition state classifier network structure is designed: a three-layer multilayer perceptron (MLP) with a GELU activation function is used. The input is the normalized working condition feature vector, and the output is the probability distribution of each working condition category. The model parameters are optimized through the cross-entropy loss function to ensure that the number of classifier parameters is less than 8K, meeting the real-time requirements of online inference; the cross-entropy loss function... express: In the formula, The total number of samples; For the first The true label vector of each sample; This represents the corresponding predicted probability distribution.

[0018] S1.5: Validate and deploy the lightweight operating condition classifier trained in step S1.4. Use an independent test set to evaluate its classification accuracy and confidence calibration performance in real-world scenarios such as noise disturbance and missing data. Output the current operating condition category label and its corresponding confidence score as the control signal input for subsequent dynamic routing of the physical causal graph.

[0019] In this embodiment, in step S2, based on the operating condition category label output in S1, a dedicated physical causal sub-graph matching the current operating condition is dynamically routed and loaded from a pre-stored physical causal graph library. The sub-graph nodes are key components of the gas turbine and their observable state variables. The edge relationships are defined by thermodynamic equations, rotor dynamics models, and historical fault root cause analysis reports, and are assigned causal strength weights to form a structured causal reasoning path constrained by physical laws. Specifically, the steps include the following: S2.1: Based on the full life cycle operating condition semantic space definition constructed in S1, obtain standardized operating condition identification codes corresponding to 12 typical operating conditions (cold start, hot start, 50% load ramp-up, 90% load steady state, rapid load shedding, combustion oscillation transition state, etc.). Based on the combination of key control parameters for each operating condition (speed deviation rate, fuel valve opening change slope, exhaust temperature gradient, compressor outlet pressure fluctuation spectrum entropy value), use historical operating datasets to train dedicated physical causal graph generation rules to establish the mapping relationship between operating condition identification codes and causal structures. Based on the full lifecycle operating condition semantic space definition constructed by S1, the corresponding standardized operating condition identifier code is generated by hash coding method, so as to map the discretized operating state of the semantic space into a unique coding sequence and realize the digital expression of operating conditions.

[0020] Furthermore, a sample set that meets the current working condition identification code is selected from the historical operating dataset through a feature combination matching algorithm, and the joint distribution of each parameter is fitted using the kernel density estimation method to form a parameter feature probability model of the working condition.

[0021] Furthermore, a rule generation engine is used to construct a physical causal graph generation rule set based on the probabilistic model of operating condition parameters. This rule set consists of parameter condition constraints and causal structure templates, which are used to describe the potential physical dependencies between observable variables under the current operating condition.

[0022] Furthermore, the association pattern mining method is used to extract high-frequency causal patterns between component state variables that meet the operating conditions from historical fault samples, embed them into the edge template of the physical causal graph generation rule, and set the initial causal strength weight in combination with the value range of the operating condition parameters.

[0023] S2.2: Based on the operating condition-causal mapping rules established in step S2.1, node-based modeling is performed on each key component of the gas turbine (compressor, combustion chamber, turbine, bearing, fuel nozzle) and its observable state variables (vibration amplitude, exhaust temperature, combustion pulsation frequency, lubricating oil temperature). The directional influence relationship between components is extracted using thermodynamic equations, rotor dynamics models and historical fault root cause analysis reports, generating a set of directional causal edges, and assigning a weight value (0.1~0.9) to each causal edge to form the initial physical causal graph primitives.

[0024] S2.3: Based on the physical causal graph primitives generated in step S2.2, perform causal structure pruning and intensity parameter tuning for each type of operating condition: according to the dynamic coupling characteristics of each component under the condition, retain the dominant causal path, remove inactive connection edges, and adjust the causal intensity weight distribution according to the energy transfer law unique to the condition to form 12 dedicated physical causal subgraphs. Each subgraph corresponds to a type of operating condition and is stored in the pre-built physical causal graph library. S2.4: During the online operation phase of the system, receive the real-time operating condition category label and its confidence level output in step S1, retrieve the matching dedicated physical causal sub-graph from the pre-stored physical causal graph library based on the label, perform graph routing operation, load the successfully matched sub-graph into the current inference context environment, and freeze the parameter update channels of the remaining inactive sub-graphs to isolate irrelevant causal path interference. S2.5: Perform topological verification and dynamic initialization processing on the dedicated physical causal subgraph loaded in S2.4, check whether its node and edge relationships completely cover the currently collected multi-source sensor observation variables. If there is a mapping missing, trigger an alarm mechanism and enable the default steady-state subgraph as a backup scheme. Finally, output a structured causal reasoning path that is constrained by physical laws and consistent with the current running state for subsequent dual-path feature processing module to call.

[0025] In this embodiment, in step S3, real-time multi-source heterogeneous time-series data of the gas turbine are simultaneously acquired, including vibration, temperature, pressure, and current signals. Local time-frequency features are extracted for each type of signal, generating local feature representations corresponding to four types of original time-series segments. These feature representations are then input into a gated multi-head attention mechanism. The attention weight distribution of each sensor channel is initialized based on the PageRank centrality index of the causal graph nodes in S2, such as... Figure 2 As shown, the specific steps include the following: S3.1: Acquire the real-time multi-source heterogeneous time-series data stream of the gas turbine under the current operating conditions. The data stream includes the axial vibration acceleration sequence output by the vibration sensor, the exhaust temperature time sequence collected by the thermocouple, the compressor outlet pressure sequence measured by the pressure transmitter, and the generator stator current waveform recorded by the current transformer. Based on the differences in physical dimensions and sampling frequencies of each sensor, perform mode alignment preprocessing, and unify the timestamps to 1kHz equal time intervals through a resampling interpolation algorithm to generate a standardized set of original time-series segments as input data for subsequent local time-frequency transformation. S3.2: Perform local time-frequency feature extraction on each type of standardized original time-series segment generated in step S3.1: For vibration signals, use Short Time Fourier Transform (STFT) to calculate their time-varying spectral energy density matrix, and extract the energy envelope curves of 5 key sub-frequency bands (such as the blade passing frequency and its harmonics) within the 0-2kHz frequency band; For temperature and pressure signals, apply Discrete Wavelet Transform (DWT) to decompose to the 4th level, select the db6 wavelet basis to obtain the low-frequency trend component and high-frequency detail coefficients, and construct a multi-scale gradient feature vector; For current signals, perform Hilbert-Huang Transform (HHT), obtain the first 3 intrinsic mode functions (IMF) through Empirical Mode Decomposition (EMD), and calculate their instantaneous amplitude and frequency variance to form a non-stationary electrical feature representation; Finally, output the local time-frequency feature tensors corresponding to the four types of sensor modes; S3.3: Based on the four types of local time-frequency feature tensors output in step S3.2, construct a multi-channel feature matrix, where each row corresponds to a high-dimensional feature vector of a sensor modality; input this feature matrix into a gated multi-head attention mechanism (GMHA), which consists of K=4 parallel attention heads, each head independently calculates the query-key-value mapping relationship, captures long-range dependencies between cross-modal features, and generates an unweighted attention response matrix as a pre-representation for dynamic weight calibration; S3.4: Obtain the dedicated physical causal graph structure loaded in step S2, and calculate the topological importance measure of the observable state variable nodes directly associated with the four types of sensors in the graph: Based on the directed edge connection relationship of the graph, use the PageRank algorithm to iteratively solve the steady-state influence score of each node, set the damping coefficient α=0.85, the convergence threshold ε=1e-6, and output the PageRank centrality index values ​​of the compressor outlet pressure fluctuation node, the exhaust temperature anomaly node, the bearing vibration amplitude node, and the generator current harmonic distortion node to form a four-dimensional initial weight vector to characterize the prior contribution of each sensor in the physical causal chain under different operating conditions; S3.5: Based on the four-dimensional PageRank centrality index vector generated in step S3.4, perform channel-level weight initialization on the unweighted attention response matrix output by the gated multi-head attention mechanism in step S3.3: After normalizing the centrality index into a probability distribution, multiply it element by element into the attention head output corresponding to each sensor channel to realize attention bias injection based on prior knowledge of physical causal structure, and generate an initial weighted feature representation matrix with working condition adaptability, which serves as the input for dynamic calibration of the residual gated unit in the next stage.

[0026] In this embodiment, in step S4, based on the initial attention weights generated in step S3, the multi-source features are dynamically calibrated using a residual gating unit, and the contribution of each channel is fused to generate a weighted joint feature representation of the time series. Simultaneously, the original features extracted in S3 are mapped to the causal sub-graph node space loaded in S2, and the abnormal observation signal is propagated along the directed causal edges using a graph neural network to generate the causal influence score vector for each component node, such as... Figure 3 As shown, the specific steps include the following: S4.1: Obtain the initial attention weight distribution of each sensor channel in the gated multi-head attention mechanism output by S3. This weight distribution is initialized based on the PageRank centrality index of each observable state variable node in the physical causal graph loaded by S2. Based on this initial attention weight, the local time-frequency feature representations corresponding to the four types of time-series signals of vibration, temperature, pressure and current are weighted respectively to highlight the representation priority of high centrality variables under the current working condition and generate a multi-source feature set with initial weights. A weighted mapping method is used to realize the correspondence between the initial weights and the local time-frequency characteristics of four types of time-series signals: vibration, temperature, pressure, and current.

[0027] Furthermore, channel-level multiplication operations of various time-series feature tensors are implemented through a feature weighting processing algorithm to obtain a weighted set of local feature tensors.

[0028] Furthermore, a high centrality priority screening method is used to prioritize the retention of feature channels above the threshold in the weighted feature tensor, and to generate a feature subset that highlights the high centrality variables.

[0029] S4.2: Input the multi-source feature set with initial weights generated in step S4.1 into the residual gating unit. This unit consists of two fully connected layers and a GELU activation function, and introduces a cross-layer identity mapping structure. Based on this residual gating unit, perform nonlinear transformation and gradient stabilization optimization to dynamically calibrate the attention weights of each channel feature and output the calibrated attention weight vector, thereby suppressing false salience responses caused by noise interference. Furthermore, the output of each fully connected network layer is processed by the GELU activation function, which utilizes its smooth nonlinearity to enhance the response stability in regions of small changes and maintain the differentiability of the gradient in the near-zero interval.

[0030] Furthermore, a cross-layer identity mapping structure is introduced, which directly adds the input multi-source feature set with initial weights to the output of the second fully connected network element by element to form a residual path, thereby alleviating the gradient vanishing phenomenon in deep feature mapping.

[0031] Furthermore, based on the combined results of the residual path output and the nonlinear transformation output, a weight normalization method is used to adjust them into a uniform attention weight vector, ensuring the comparability of the contributions of different channels in the subsequent weighted summation stage.

[0032] Furthermore, to suppress spurious responses caused by noise interference, a gradient stabilization optimization strategy is introduced to apply smoothing constraints to the attention weight gradient during backpropagation, thereby reducing overshoot and oscillations in weight updates.

[0033] S4.3: Based on the dynamically calibrated attention weight vector output in step S4.2, a weighted summation operation is performed on the four types of local time-frequency feature representations to generate a unified time series joint feature representation. This joint feature representation integrates the dynamic contribution of multi-source heterogeneous signals under the current operating conditions and serves as the input of the high-level semantic representation of the data-driven path to the subsequent anomaly probability prediction module. S4.4: Obtain the local feature representations corresponding to the four types of original time segments extracted in S3, and map them to the node space of the dedicated physical causal graph dynamically loaded in S2, where each observable state variable (such as vibration amplitude and exhaust temperature) corresponds to a node in the graph; based on this node-feature mapping relationship, construct a graph signal input matrix with physical semantic alignment capability as the starting point for information propagation of the graph neural network; S4.5: Based on the graph signal input matrix constructed in step S4.4, the graph neural network inference process is executed. This process consists of a cascaded GCN convolutional layer and an EdgeConv edge enhancement layer, and message passing is performed along the directed edges of the physical causal subgraph. Based on the causal intensity weights defined by the thermodynamic equation and the rotor dynamics model, the abnormal observation signals are propagated and aggregated layer by layer, and finally the causal influence score vector of each key component node is output, which is used to identify the core affected units on the potential fault propagation path. In the cascaded structure of GCN and EdgeConv, causal strength weights defined by thermodynamic equations and rotor dynamics models are introduced, and a weighted adjacency matrix is ​​used for information transmission, namely: ; In the formula, This is the normalized causal strength weight matrix. This is the original topological adjacency matrix. This is the element-wise multiplication operation for Hadamard.

[0034] The node feature update is achieved using a layer-by-layer feature aggregation formula, namely: In the formula, For the first Layer node feature matrix; for Layer trainable weight matrix; It is a nonlinear activation function that ensures efficient propagation and nonlinear combination of anomalous observation signals under causal topological constraints.

[0035] In this embodiment, in step S5, anomaly probability prediction is performed on the weighted time series joint feature representation generated in step S4, and the anomaly discrimination result of the data-driven path is output; at the same time, the causal influence score vector generated in S4 is normalized to identify key component nodes with significantly increased scores, forming the anomaly localization result of the knowledge-guided path; based on the dual-path consistency constraint mechanism, the KL divergence between the output distributions of the two paths is calculated, and it is determined whether it is less than a preset threshold of 0.15, specifically including the following steps: S5.1: Based on the weighted time series joint feature representation generated in S4, a lightweight fully connected network is used to perform nonlinear mapping processing to project high-dimensional time series features onto the anomaly probability space and output the anomaly discrimination result of the data-driven path, i.e., the normalized anomaly occurrence probability distribution P. S5.2: Softmax normalization is performed on the causal influence score vector generated in S4 to eliminate the scale difference of node activation amplitude under different working conditions, obtain the standardized influence distribution Q of the knowledge guidance path, and characterize the abnormal contribution ranking of each key component under the current physical causal structure. S5.3: Calculate the Kullback-Leibler divergence based on the two probabilistic output distributions P and Q. This study aims to generate consistency evaluation indicators by quantifying the semantic deviation between data-driven reasoning paths and knowledge-guided causal paths in anomaly localization results. .

[0036] S5.4: Consistency evaluation indicators The consistency evaluation index is compared with the preset threshold of 0.15. If the value is less than 0.15, the dual-path outputs are deemed to have sufficient consistency, and the final decision fusion stage is entered; otherwise, a reasoning conflict is determined, and a feedback correction mechanism needs to be activated to suppress the activation of false features.

[0037] In this embodiment, in step S6, if the KL divergence in step S5 satisfies the condition of being less than 0.15, then the output results of the data-driven path and the knowledge-guided path are fused to generate the final anomaly identification decision; otherwise, the adversarial perturbation regularization feedback mechanism is triggered to adjust the weight update strategy in the gated multi-head attention mechanism, suppress the spurious activation of non-causal features, and steps S4 to S5 are re-executed until the consistency constraint condition is met. Specifically, the steps include the following: S6.1: Obtain the anomaly probability distribution of the data-driven path and the normalized distribution of the causal influence score of the knowledge-guided path output in S5, and use them as input for dual-path consistency verification; calculate the difference measure between the two distributions based on Kullback-Leibler divergence to quantify the semantic deviation between the data path and the knowledge path in anomaly detection, and generate KL divergence results for subsequent decision branch control. S6.2: Determine whether the KL divergence calculated in S6.1 is less than the preset threshold of 0.15; if yes, determine that the dual-path outputs are consistent and enter the abnormal decision fusion stage; if no, determine that there is a significant semantic conflict and activate the adversarial perturbation regularization feedback mechanism to correct the weight allocation deviation in the attention mechanism and ensure that subsequent reasoning conforms to physical causal logic. S6.3: Under the condition that the KL divergence does not meet the threshold, an adversarial perturbation regularization term is introduced based on the attention weight distribution of each sensor channel in the current gated multi-head attention mechanism. Specifically, by applying gradient direction perturbation to the attention weight space, the false response mode when non-causal features are activated is simulated, and backpropagation is used to adjust the parameter update strategy of the residual gated unit to enhance the model's ability to suppress physically irrelevant features. S6.4: Based on the optimized attention weight update strategy in step S6.3, the multi-source feature dynamic calibration and graph neural network causal propagation process in S4 are re-executed to generate the updated weighted time series joint feature representation and node-level causal influence score vector, which serve as the result of a new round of dual-path inference and provide new input for verifying consistency again. S6.5: Re-input the updated dual-path output from step S6.4 into S5 for anomaly detection and KL divergence re-evaluation, forming a closed-loop iterative optimization process; continue executing S6.3 to S6.5 until the KL divergence is below 0.15, ensuring that the final output anomaly identification result has both high confidence and strong causal consistency, meeting the dual technical requirements of interpretability and stability in power plant operation and maintenance scenarios.

[0038] In this embodiment, in step S7, after confirming the existence of an abnormal event, the attribution inversion engine is activated: based on the weighted time series joint feature representation in S4, the gradient weighted class activation mapping method is used to locate the time window and frequency band region with the highest contribution in each sensor signal, generating a spatiotemporal distribution map of key features at the data level; simultaneously, counterfactual reasoning is performed on the causal subgraph loaded in S2, causal edges are masked one by one and the graph neural network is run again, recording the change in the causal influence score of the target abnormal node, specifically including the following steps: S7.1: Based on the weighted time series joint feature representation generated in S4, input the gradient weighted class activation map (Grad-CAM) calculation framework, perform backpropagation operation on the time-frequency feature map corresponding to each sensor channel, calculate the gradient response intensity of the final anomaly discrimination category relative to each time-frequency unit, so as to quantify the contribution of different signal segments to anomaly decision-making; S7.2: Based on the gradient response intensity distribution output in step S7.1, spatial aggregation is performed along the time axis and frequency axis to generate a spatiotemporal distribution map of key features of each sensor signal, and the time window interval and dominant frequency band with the highest contribution in vibration, temperature, pressure and current signals are identified as abnormal sensitive areas at the data level. S7.3: On the dedicated physical causal subgraph dynamically loaded in step S2, for the target anomalous node (such as concentrated vibration energy of turbine blades or abnormal exhaust temperature gradient), execute the counterfactual reasoning mechanism: sequentially mask any directed causal edge in the graph, keep the rest of the structure unchanged, rerun the graph neural network (GCN+EdgeConv) to calculate the anomalous signal propagation, and record the change in the causal influence score of the target node after each masking. ;Right now: ; In the formula, The target node's causal impact score under baseline conditions; To block the The causal influence score of the target node after a directed edge.

[0039] S7.4: The change in score after masking each causal edge, accumulated in step S7.3. Arrange the top-3 causal edge combinations that cause the most significant drop in the abnormal score of the target node in descending order, construct the causal path chain set with the highest cause weight, and form the core fault propagation path output at the knowledge level. S7.5: Semantically align and structurally fuse the key feature spatiotemporal distribution map generated in step S7.2 with the Top-3 causal path chains extracted in S7.4: Map the time-frequency band regions with high contribution to the causal map nodes connected to the corresponding sensors, and combine the causal chains obtained by counterfactual reasoning to generate a structured explanation text of the cause of anomalies containing the logical relationship of 'observation anomaly → component state evolution → system-level impact', which is then output to the subsequent visualization module as a traceable diagnostic basis.

[0040] In this embodiment, in step S8, based on the score changes after each causal edge is masked in step S7, the Top-3 causal path chains that most significantly reduce the probability of anomaly occurrence are extracted. Combined with the spatiotemporal distribution map of key features, a structured anomaly cause explanation text containing physical meaning is generated, and this explanation text is pushed to the visualization module to support power plant operation and maintenance personnel in making traceable fault diagnosis decisions. Specifically, the steps include the following: S8.1: Based on the change in causal influence score of the target abnormal node recorded by counterfactual reasoning on the causal subgraph in step S7, all the masked causal edges are quantitatively sorted according to the significance of their leading to a decrease in the probability of anomaly occurrence. The attribution contribution index of each causal edge is calculated, and an ordered sequence of causal edge importance is generated as the input basis for subsequent path extraction. S8.2: Extract the Top-3 causal edges that cause the most significant decrease in the score of the target abnormal node from the causal edge importance sequence generated in S8.1, and use them as the core mechanism that dominates the anomaly propagation; based on the pre-stored physical causal subgraph topology, trace the upstream source node forward and the affected target node backward along the directed causal edge, reconstruct the complete anomaly propagation path chain, and generate three candidate sets of anomaly paths with clear physical causal logic. S8.3: Semantic parsing is performed on each anomaly propagation path chain generated in S8.2, mapping nodes and edges to physical component state descriptions and interaction relationships that can be read in natural language (such as 'combustion pulsation frequency ↑ → combustion chamber pressure fluctuation ↑'). Combined with power plant equipment terminology standards, a preliminary structured causal sentence is generated to form an anomaly cause expression prototype with domain consistency. S8.4: Align the spatiotemporal distribution map of key data features generated in S7 with the three structured causal sentences output in S8.3 across modalities: Based on the time window matching mechanism, associate the high contribution frequency bands and time regions located by Grad-CAM with the physical variable nodes to which the corresponding sensor signals belong, verify their consistency with the nodes involved in the causal path, and select the final Top-3 causal path chains supported by both data evidence and causal logic; S8.5: Integrate the three causal path chains with dual verification of data and knowledge confirmed in S8.4 into a unified structured anomaly cause explanation text, encapsulate it using a predefined template format (including anomaly type, starting source component, propagation path, terminal effect and confidence level), and push it to the visualization module through a standardized interface for operation and maintenance personnel to perform interactive browsing and diagnostic tracing.

[0041] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0042] Unless otherwise defined, the technical or scientific terms used herein should be understood in their ordinary sense by one of ordinary skill in the art to which this invention pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” involved in the embodiments of this invention refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0043] The above description is merely an exemplary embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An intelligent online anomaly detection method for a gas turbine, characterized in that, Includes the following steps: S1: Obtain historical operating data of the gas turbine; and predict the current operating condition category label and its confidence level; S2: Dynamically route based on operating condition category labels and load a dedicated physical causal sub-graph that matches the current operating condition; S3: Collect real-time multi-source heterogeneous time-series data of the gas turbine, and input the extracted local feature representation into the gated multi-head attention mechanism. Initialize the attention weight distribution of each sensor channel according to the PageRank centrality index of the causal graph nodes. S4: Based on the initial attention weights, the multi-source features are dynamically calibrated through the residual gating unit to generate a weighted joint feature representation of the time series; and the local feature representation is mapped to the causal sub-graph node space. The abnormal observation signal is propagated along the directed causal edge using the graph neural network to generate the causal influence score vector of each component node. S5: Perform anomaly probability prediction on the weighted time series joint feature representation to obtain anomaly discrimination results; identify key component nodes with significantly increased scores from the causal influence score vector to form anomaly localization results for the knowledge-guided path; calculate the KL divergence between the output distributions of the two paths and determine whether it is less than a preset threshold. S6: If satisfied, the output results of the data-driven path and the knowledge-guided path are combined to generate the final anomaly identification decision; Otherwise, the adversarial perturbation regularization feedback mechanism is triggered, the weight update strategy in the gated multi-head attention mechanism is adjusted, the spurious activation of non-causal features is suppressed, and steps S4 to S5 are re-executed until the consistency constraint condition is met.

2. The intelligent online anomaly detection method for a gas turbine according to claim 1, characterized in that, The method further includes: S7: Once an abnormal event is confirmed, the attribution inversion engine is activated to record the change in the causal impact score of the target abnormal node. S8: Based on the score changes after each causal edge is masked in step S7, extract the Top-3 causal path chains that most significantly reduce the probability of anomalies, and combine them with the spatiotemporal distribution map of key features to generate a structured explanation text of the cause of anomalies containing physical meaning.

3. The intelligent online anomaly detection method for a gas turbine according to claim 1, characterized in that, In step S1, historical operating data sequences of the gas turbine under six typical operating conditions are obtained: cold start, hot start, 50% load ramp-up, 90% load steady state, rapid load shedding, and combustion oscillation transition state. The data sequences include time series signals of key control parameters such as speed, fuel valve opening, exhaust temperature, and compressor outlet pressure.

4. The intelligent online anomaly detection method for a gas turbine according to claim 1, characterized in that, In step S1, a lightweight operating condition classifier is used to predict the current operating condition category label and its confidence level. The lightweight operating condition classifier uses a three-layer multilayer perceptron superimposed with the GELU activation function.

5. The method and system for intelligent online anomaly detection of a gas turbine of claim 1, wherein, In step S2, a matching dedicated physical causal subgraph is retrieved from the pre-stored physical causal graph library, a graph routing operation is performed, the successfully matched subgraph is loaded into the current inference context environment, and the parameter update channels of the remaining inactive subgraphs are frozen to isolate interference from irrelevant causal paths. The loaded dedicated physical causal subgraph is subjected to topological verification and dynamic initialization. The node and edge relationships are checked to ensure that they completely cover the currently collected multi-source sensor observation variables. If there is a mapping gap, an alarm mechanism is triggered and the default steady-state subgraph is used as a backup. The structured causal reasoning path that is constrained by physical laws and consistent with the current operating state is output.

6. The intelligent online anomaly detection method for gas turbines according to claim 1, characterized in that, In step S3, local time-frequency feature extraction is performed on the time-series segments of the multi-source heterogeneous time-series data: For the vibration signal, short-time Fourier transform is used to calculate its time-varying spectral energy density matrix, and the energy envelope curves of multiple key sub-frequency bands in the 0-2kHz frequency band are extracted; for the temperature and pressure signals, discrete wavelet transform is applied to decompose them, and the db6 wavelet basis is selected to obtain the low-frequency trend component and high-frequency detail coefficients, and a multi-scale gradient feature vector is constructed; for the current signal, Hilbert-Huang transform is applied, and the first three intrinsic mode functions are obtained through empirical mode decomposition, and their instantaneous amplitude and frequency variance are calculated to form a non-stationary electrical feature representation; finally, the local time-frequency feature tensor is output.

7. The intelligent online anomaly detection method for gas turbines according to claim 1, characterized in that, In step S3, a multi-channel feature matrix is ​​constructed using the local time-frequency feature tensor and input into a gated multi-head attention mechanism. Each head independently calculates the query-key-value mapping relationship to capture long-range dependencies between cross-modal features and generate an unweighted attention response matrix. Obtain the loaded dedicated physical causal graph structure, calculate the topological importance measure of the nodes with observable state variables in the graph, use the PageRank algorithm to iteratively solve the steady-state influence score of each node, and output the PageRank centrality index values ​​of the compressor outlet pressure fluctuation node, exhaust temperature anomaly node, bearing vibration amplitude node, and generator current harmonic distortion node. Based on the PageRank centrality index vector, channel-level weight initialization is performed on the unweighted attention response matrix to obtain the initial attention weight distribution.

8. The intelligent online anomaly detection method for gas turbines according to claim 1, characterized in that, In step S4, the residual gating unit is composed of a fully connected network and a GELU activation function, and a cross-layer identity mapping structure is introduced; based on the residual gating unit, nonlinear transformation and gradient stabilization optimization are performed to dynamically calibrate the attention weights of each channel feature and output the calibrated attention weight vector. A joint time series feature representation is generated by performing a weighted summation operation on four types of local time-frequency feature representations based on the calibrated attention weight vector; Local feature representations are mapped to the node space of a dedicated physical causal graph; based on the node-feature mapping relationship, a graph signal input matrix with physical semantic alignment capability is constructed. The graph neural network inference process is executed, which consists of a cascaded GCN convolutional layer and an EdgeConv edge enhancement layer, and message passing is performed along the directed edges of the physical causal subgraph to obtain the causal influence score vector.

9. The intelligent online anomaly detection method for gas turbines according to claim 1, characterized in that, In step S5, a lightweight fully connected network is used to perform nonlinear mapping processing to project high-dimensional temporal features onto the anomaly probability space and output the anomaly discrimination result of the data-driven path. The causal impact score vector is subjected to Softmax normalization to obtain the standardized impact distribution of the knowledge guidance path. ; Anomaly probability distribution based on normalization Distribution of Standardization Effects Given two probabilistic output distributions, calculate their KL divergence.

10. A gas turbine intelligent online anomaly detection system, characterized in that, The system described herein utilizes the method described in any one of claims 1-9 to achieve intelligent online anomaly detection of gas turbines.

Images