Gas turbine power plant operation abnormity monitoring and early warning system based on artificial intelligence

Abstract

The invention relates to the technical field of gas turbine power plant abnormity monitoring, in particular to a gas turbine power plant operation abnormity monitoring and early warning system based on artificial intelligence. The system comprises a data acquisition unit, and the data acquisition unit is used for acquiring multi-dimensional operation data of a gas turbine power plant, and obtaining a global feature expression vector by introducing a multi-modal feature fusion network based on an attention mechanism; the working condition identification unit constructs a dynamic evolution model of the gas turbine based on a long short-term memory network and a hidden Markov model, and performs time sequence coding on global feature expression vectors; and the anomaly recognition unit constructs an anomaly recognition model based on the combination of the deep auto-encoder and the time sequence convolutional network, calculates a comprehensive anomaly score, and then judges an operation anomaly state. The operation feature vectors of different subsystems are mapped to the unified dimension, invalid modals are shielded, missing modals are compensated, and the technical problem that the global state characterization precision is affected due to incomplete fusion information caused by modal failure or missing is solved.

Description

Technical Field

[0001] This invention relates to the field of anomaly monitoring technology in gas turbine power plants, and more specifically, to an artificial intelligence-based system for monitoring and early warning of operational anomalies in gas turbine power plants. Background Technology

[0002] As the core power equipment in power plants, the operating status of gas turbines directly affects the stability of power supply, operation and maintenance cost control, and production safety. With the accelerated transformation of the energy structure, gas turbine generator sets are being upgraded towards higher parameters, larger capacity, and greater intelligence. However, their operating environment is constantly subjected to extreme conditions such as high temperature, high pressure, high speed, and strong vibration. The concealment and suddenness of failures have significantly increased. Faults such as blade breakage, bearing wear, and unstable combustion may lead to unplanned shutdowns of the unit, causing economic losses of hundreds of thousands of dollars, or even triggering safety accidents.

[0003] Gas turbine operation monitoring technology has evolved from traditional manual inspections to intelligent system monitoring. Although current mainstream intelligent monitoring solutions have incorporated machine learning technology, significant shortcomings remain. Sensors, as the core component of data acquisition, often experience signal drift, failure, or data loss due to extreme operating conditions such as high temperature, high pressure, and strong vibration. This leads to invalid data being mixed in during the multi-source feature fusion process, interfering with the accuracy of global feature expression and thus affecting the reliability of operating condition identification and anomaly warning. At the same time, gas turbine combustion status monitoring relies on combustion chamber pressure, turbine outlet temperature, and emission parameters. The lack of emission parameters will result in the incomplete characterization of combustion efficiency-related features, thereby reducing the accuracy of operating condition identification (such as unstable combustion conditions) and even causing the anomaly detection model to fail due to incomplete input features. In view of this, an artificial intelligence-based operation anomaly monitoring and early warning system for gas turbine power plants is designed. Summary of the Invention

[0004] The purpose of this invention is to provide an artificial intelligence-based monitoring and early warning system for gas turbine power plants, in order to solve the problems mentioned in the background art, such as signal drift, failure or data loss of sensors, which lead to invalid data mixed in during the multi-source feature fusion process, interfere with the accuracy of global feature expression, and thus affect the reliability of operating condition identification and anomaly early warning.

[0005] To achieve the above objectives, the present invention aims to provide an artificial intelligence-based monitoring and early warning system for operational anomalies in gas turbine power plants, comprising:

[0006] A data acquisition unit is used to collect multi-dimensional operating data of a gas turbine power plant and to preprocess the multi-dimensional operating data.

[0007] The multi-source feature extraction unit extracts features from the synchronously processed multi-dimensional operating signal to obtain the operating feature vectors of each subsystem. Furthermore, by introducing a multimodal feature fusion network based on an attention mechanism, the running feature vector is improved. Perform weighted aggregation to obtain the global feature representation vector. ;

[0008] The operating condition identification unit employs a gas turbine dynamic evolution model constructed based on a long short-term memory network and a hidden Markov model to analyze the global feature representation vector. Timing coding is performed to identify the current operating condition of the gas turbine;

[0009] An anomaly detection unit constructs an anomaly detection model based on a combination of a deep autoencoder and a temporal convolutional network, reconstructs and predicts multi-source features, calculates a comprehensive anomaly score, and then determines the abnormal operating state.

[0010] Among them, multi-source features include at least a global feature representation vector. and current gas turbine operating status .

[0011] As a further improvement to this technical solution, the multi-dimensional operating data includes at least the gas flow rate, compressor inlet temperature, combustion chamber pressure, turbine outlet temperature, bearing vibration, motor current, voltage, and emission parameters.

[0012] The multidimensional operational data preprocessing includes timestamp synchronization and adaptive Kalman filtering.

[0013] As a further improvement to this technical solution, the multi-source feature extraction unit includes a feature extraction module and a feature fusion module;

[0014] The feature extraction module is used to extract the time-domain, frequency-domain, and statistical features of the preprocessed multidimensional operational data to obtain the operational feature vectors of each subsystem. ;

[0015] The feature fusion module uses an attention-based multimodal feature fusion network to assign importance weights to features from each subsystem. The operational characteristic vectors of each subsystem Perform weighted aggregation to generate a global feature representation vector. ;

[0016] The multimodal feature fusion network includes a modal encoder, an attention fusion layer, and a residual normalization module, which is used to map the running feature vectors from different subsystems into a unified global feature representation vector.

[0017] As a further improvement to this technical solution, a global feature representation vector is generated through the attention-based multimodal feature fusion network. The specific steps involved are as follows:

[0018] For each running feature vector By performing a linear mapping transformation through a modal encoder, a uniform-dimensional embedding representation is obtained. ;

[0019] For each modal embedding The corresponding query, key, and value vectors are calculated, and a scaled dot product attention mechanism is used in the attention fusion layer to calculate the compatibility score. ;

[0020] If the attention fusion layer has missing modalities, a mask identifier is introduced. Masking invalid modes and compatibility scoring Make corrections;

[0021] Attention weights are calculated using the Softmax function. The value vectors are weighted and aggregated to generate a preliminary global fusion result. The global fusion result is then processed through a residual normalization module. Nonlinear mapping, residual connections, and layer normalization are performed to form the initial feature representation vector. ;

[0022] Initial feature representation vector Finally, the global feature representation vector is obtained after normalization. .

[0023] As a further improvement to this technical solution, the working condition identification unit includes a timing coding module and a working condition prediction module;

[0024] The gas turbine dynamic evolution model is coupled to a hidden Markov model via a long short-term memory network (LSTM) in series. The LSM represents the global feature vector. Perform temporal encoding to output a sequence of hidden layer state vectors;

[0025] The Hidden Markov Model is used to receive the hidden state vector sequence output by the Long Short-Term Memory Network, establish the state transition model of the gas turbine operating conditions, and calculate the optimal state path using the Viterbi algorithm.

[0026] As a further improvement to this technical solution, the operating condition identification unit identifies the current operating condition of the gas turbine, and the specific steps involved are as follows:

[0027] The received global feature representation vector is processed by the time-series coding module. According to time step Perform serialization to construct the input sequence;

[0028] The input sequence is encoded using a long short-term memory network, and the activation values ​​of the input gate, forget gate, and output gate are calculated respectively.

[0029] At each time step The hidden state vector is updated through the state transfer mechanism of the Long Short-Term Memory network, and the hidden state vector sequence is constructed.

[0030] The timing coding module calculates a weighted average of the hidden state vectors at each stage using a sliding time window mechanism to obtain the timing coding result. ;

[0031] The operating condition prediction module is used to receive the hidden state vector sequence output by the time-series coding module and input it into the hidden Markov model.

[0032] Define the set of potential operating conditions for gas turbine operation. And initialize the state transition probability matrix, observation probability matrix and initial state distribution;

[0033] Based on historical running datasets, the state transition probability matrix, observation probability matrix, and initial state distribution are iteratively optimized using a forward-backward algorithm.

[0034] The Viterbi algorithm is used to calculate the optimal state path for the state sequence of the Hidden Markov Model. The current gas turbine operating condition category is determined based on the terminal state of the obtained optimal state path, and the confidence level is evaluated based on the log probability difference of the optimal path.

[0035] The state transition path within a few future time steps is predicted using the Markov propagation algorithm.

[0036] As a further improvement to this technical solution, the anomaly identification unit includes an anomaly scoring module and an anomaly identification output module;

[0037] The anomaly scoring module performs a comprehensive analysis of the reconstruction and prediction results of the gas turbine operating characteristics based on the anomaly identification model to generate a comprehensive anomaly score.

[0038] The anomaly identification output module determines whether the gas turbine is in an abnormal state based on the anomaly score result according to the anomaly judgment rules by setting a preset anomaly score threshold.

[0039] As a further improvement to this technical solution, the anomaly identification model generates anomaly scores, and the specific steps involved are as follows:

[0040] The anomaly scoring module (41) is used to receive the hidden layer state vector sequence, the global feature expression vector sequence, and the current gas turbine operating status;

[0041] Current moment Next, the input hidden layer state vector sequence and global feature representation vector sequence are input into the deep autoencoder for feature reconstruction, generating a reconstructed vector. And based on the current moment Global feature representation vector With reconstructed vector The degree of deviation is used to calculate the reconstruction error;

[0042] The encoded reconstructed vector Input a temporal convolutional network to perform multi-step future state prediction and generate a prediction vector. Calculate the prediction error based on the deviation between the actual global feature vector and the prediction vector.

[0043] The reconstruction error and prediction error are weighted and fused to obtain a comprehensive anomaly score. .

[0044] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0045] 1. In the AI-based operation anomaly monitoring and early warning system of this gas turbine power plant, the multi-source feature extraction unit uses multi-modal feature fusion technology (including modal encoder, attention fusion layer, modal health detection, masking mechanism and modal compensation mechanism) to map the operation feature vectors of different subsystems to a unified dimension. It dynamically allocates weights according to feature importance, masks invalid modalities and compensates for missing modalities, and generates accurate global feature expression vectors. This solves the technical problems of existing technologies, such as one-sided information from a single modality, the inability of fixed-weight fusion to highlight key features, and incomplete fusion information caused by modal failure or missing features, which affect the accuracy of global state representation.

[0046] 2. In the AI-based operation anomaly monitoring and early warning system of this gas turbine power plant, the operating condition identification unit captures the long-term dependency relationship of global feature vectors by cascading and coupling LSTM and Hidden Markov Model, establishes an operating condition state transition model, accurately identifies the current operating condition and predicts the state transition path for several future time steps. This solves the technical problems of existing static modeling, which cannot adapt to the dynamic operating conditions of gas turbines, is difficult to capture the evolution law of operating conditions, resulting in a high misjudgment rate of operating condition identification and the inability to predict changes in operating conditions in advance. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating the overall process of the present invention.

[0048] The meanings of the labels in the diagram are as follows:

[0049] 1. Data acquisition unit;

[0050] 2. Multi-source feature extraction unit; 21. Feature extraction module; 22. Feature fusion module;

[0051] 3. Operating condition identification unit; 31. Timing coding module; 32. Operating condition prediction module;

[0052] 4. Anomaly identification unit; 41. Anomaly scoring module; 42. Anomaly identification output module. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] Please see Figure 1 As shown, an artificial intelligence-based monitoring and early warning system for operational anomalies in gas turbine power plants is provided, including a data acquisition unit 1. The data acquisition unit 1 is used to collect multi-dimensional operational data of the gas turbine power plant and preprocess the multi-dimensional operational data to achieve unified time-series alignment and signal smoothing for different sampling frequencies and data sources.

[0055] In this embodiment, the multidimensional operating data includes gas flow rate, compressor inlet temperature, combustion chamber pressure, turbine outlet temperature, bearing vibration, motor current, voltage, and emission parameters;

[0056] The multidimensional operational data preprocessing includes timestamp synchronization and adaptive Kalman filtering to remove measurement noise and missing anomalies, achieve unified time alignment and signal smoothing for different sampling frequencies and data sources, and ensure data consistency and reliability for subsequent feature extraction and analysis modules. Furthermore, the multidimensional operational data preprocessing also includes multidimensional Z-Score standardization. Timestamp synchronization and adaptive Kalman filtering are used to remove measurement noise and missing anomalies, achieve unified time alignment and signal smoothing for different sampling frequencies and data sources, and multidimensional Z-Score standardization is used to eliminate dimensional differences in various physical parameters (such as temperature, pressure, vibration, etc.).

[0057] The system for monitoring and early warning of operational anomalies also includes a multi-source feature extraction unit 2. This unit extracts features from the synchronously processed multi-dimensional operational signals (preprocessed dimensionless multi-dimensional operational signals) to obtain the operational feature vectors of each subsystem. Furthermore, by introducing a multimodal feature fusion network based on an attention mechanism, the running feature vector is improved. Perform weighted aggregation to obtain the global feature representation vector. To achieve cross-modal information fusion and global state representation of the operational characteristics of different subsystems;

[0058] The running feature vector is an intermediate layer vector obtained by feature extraction for a single data source or a single physical subsystem (such as combustion chamber, compressor, turbine section and cooling and lubrication system, etc.);

[0059] The operating feature vectors are derived from: time-domain statistical features (mean, variance, peak value, etc.); frequency-domain features (spectral energy, peak frequency, etc.); and state-derived features (pressure fluctuation rate, temperature gradient, proportion of gas composition, etc.).

[0060] In this embodiment, the multi-source feature extraction unit 2 includes a feature extraction module 21 and a feature fusion module 22;

[0061] Feature extraction module 21 is used to extract the time domain, frequency domain, and statistical features of the preprocessed multidimensional operating data to obtain the operating feature vectors of each subsystem. ;

[0062] Feature fusion module 22 uses an attention-based multimodal feature fusion network to assign importance weights to features from each subsystem. The operational characteristic vectors of each subsystem Perform weighted aggregation to generate a global feature representation vector. ;

[0063] Specifically, a global feature representation vector is generated through the attention-based multimodal feature fusion network. The specific steps involved are as follows:

[0064] For each running feature vector By performing a linear mapping transformation through a modal encoder, a uniform-dimensional embedding representation is obtained. ;

[0065] ;

[0066] In the formula, For the first Each operating feature vector (containing standardized multidimensional sensing parameters, such as pressure, temperature, flow rate, load, vibration amplitude, etc.). Indicates the first The encoding result of each running feature in a unified embedding space, i.e., the embedding representation, is used as input for subsequent context feature fusion or anomaly detection models. This represents the target embedding dimension after mapping by the modal encoder, used to unify the spatial dimension of different subsystems or modal features, and to achieve cross-modal feature fusion; This represents the linear mapping function corresponding to the modal encoder, used to transform the original feature vector... Mapped to a unified embedding space, the function parameters are determined by the weight matrix. With bias vector To be determined jointly; Represents the weight matrix; This represents the weight bias term; the weight matrix and bias term are used to eliminate dimensional differences between modal features.

[0067] For each modal embedding The corresponding query, key, and value vectors are calculated, and a scaled dot product attention mechanism is used in the attention fusion layer to calculate the compatibility score. ;

[0068] ; ; ;

[0069] Compatibility rating :

[0070] ;

[0071] In the formula, This represents the linear transformation weight matrix used to generate the query vector, which is used to embed the features. Mapped to the query space; This represents the linear transformation weight matrix used to generate the key vector, which is used to construct the correlation matching benchmark between features; This represents the linear transformation weight matrix used to generate the value vector, which is used for information transfer during feature weighted aggregation; Indicates by the first A query vector generated from embedded features is used to represent the active retrieval needs of this modality during fusion. Indicates by the first A key vector generated from embedded features is used to measure the relevance of its matching with other modalities; Indicates by the first The value vectors generated by the embedded features are used to be weighted and aggregated under the action of attention weights to achieve information fusion; Indicates the first query vectors With the Key vectors The compatibility score between modes reflects the modality With mode The degree of correlation between them is usually... Input the softmax value to obtain the attention weights; Indicates the scaling factor. Used to prevent the inner product value from increasing with the vector dimension. Increasing too much, while maintaining the numerical stability of the attention distribution, where Typically, the dimension of the query vector or key vector is taken; Indicates the transpose symbol;

[0072] If the attention fusion layer contains missing modalities or invalid data, a mask identifier is introduced. Masking invalid modalities (or invalid data) and scoring compatibility. To improve the model's interpretability and generalization performance, this embodiment optionally adds a regularization constraint term based on attention entropy to the multimodal feature fusion network during the training phase to constrain the sparsity of the attention distribution. This is to suppress excessively dispersed attention distribution and improve the weight recognition ability of key modalities;

[0073] The attention weights are obtained through softmax regularization and are the th . The query pair for the first Attention level of each modality; This represents the attention entropy regularization term; represents the regularization coefficient, a hyperparameter used to control the proportion of the entropy constraint term contributing to the overall loss function; its value typically ranges from 1 to 1. ;

[0074] in, ;

[0075] In the formula, Represents the mask identifier, with values... When the first When a modality or input data is valid When the corresponding mode is missing, invalid, or unavailable, ;in, A large positive number is used to suppress invalid modes (or invalid data) before Softmax normalization. This indicates an assignment update, that is, setting the variable... Overridden by the new correction value;

[0076] In this embodiment, a modal health detection and masking mechanism is further introduced into the multi-source feature extraction unit 2 to detect and dynamically mask the validity of each modal input signal during the multi-modal feature fusion process, so as to prevent the failure mode from affecting the global feature expression.

[0077] Specifically, the modal health detection submodule calculates a health score based on the temporal stability index, noise energy index, and signal integrity index of the modal input. :

[0078] ;

[0079] In the formula, Indicates signal mode The signal-to-noise ratio is affected when the signals from the gas turbine vibration sensor or pressure sensor contain strong noise (such as ignition interference or motor electromagnetic coupling), leading to... Decrease; when the sensor is operating stably, the signal energy is concentrated, and the noise is low, Close to 1; Indicates data continuity (the reciprocal of the sampling loss rate); This represents the normalized variance of modal features within the time window; , and These are the weighting coefficients for the corresponding indicators. When all three indicators have been normalized to the [0,1] interval, this ensures the health score is within a certain range. Comparability, preferred settings ; For modal The larger the value, the healthier the modality and the more reliable the data, in terms of its stability and reliability within the current time window.

[0080] By setting a threshold in advance When the health score of any modality Below the preset threshold When this occurs, the system automatically triggers a masking mechanism to block the modality in the attention fusion layer, preventing its abnormal data from interfering with the global feature representation.

[0081] In addition, to avoid information loss, a modal compensation mechanism is introduced simultaneously. The missing modal embedding based on time interpolation or autoencoder reconstruction is used to replace the shielded modes, thereby maintaining the continuity of multimodal fusion and global semantic consistency (for example, when the compressor vibration sensor fails or the cooling flow measurement point drifts, the system can automatically detect and shield abnormal modes, and still rely on modes such as temperature, load, and pressure to achieve stable feature fusion and state characterization).

[0082] Attention weights are calculated using the Softmax function. The value vectors are weighted and aggregated to generate a preliminary global fusion result. The global fusion result is then processed through a residual normalization module. Nonlinear mapping, residual connections, and layer normalization are performed to form the initial feature representation vector. It is used to reflect the overall operating characteristics of the gas turbine system at the current moment;

[0083] in, In the formula, Indicates the first The query pair for the first Attention allocation for each modality; This represents the total number of modalities or feature channels, i.e., the number of key vectors or input modalities in the current attention computation; For key indexing, For querying the index;

[0084] ;

[0085] In the formula, This represents a feedforward network module used to enhance the representation of nonlinear features (typically consisting of two fully connected layers and an intermediate nonlinear activation function (such as ReLU or GELU)). Presentation layer normalization operation; This is the preliminary global fusion result obtained through attention-weighted aggregation; This represents the initial feature representation vector after nonlinear mapping and layer normalization processing by the residual normalization module, used to reflect the global operating state of the gas turbine system;

[0086] Initial feature representation vector Finally, the global feature representation vector is obtained after normalization. ;

[0087] ;

[0088] In the formula, This represents a vector normalization function, used to standardize the input features to eliminate numerical differences between different feature dimensions; This represents the global feature representation vector after normalization, used to comprehensively characterize the overall operating state of the gas turbine system at the current moment. Specifically, the normalization function can adopt a normalization method based on the L2 norm.

[0089] In this embodiment, due to differences in sampling frequency, noise level, and data stability among different sensing channels (such as temperature, pressure, vibration, current, voltage, and fuel flow), the resulting operational data exhibits significant multi-source heterogeneous characteristics. Therefore, this embodiment provides a training method for a multimodal feature fusion network based on an attention mechanism:

[0090] The training phase first performs time synchronization and normalization preprocessing on various sensor signals to construct a multimodal input sample set. For each modal data, a corresponding modal encoder is set to extract the local operating feature embedding under that modality. Subsequently, the correlation and weight distribution between each modality are calculated through an attention fusion layer to achieve adaptive adjustment of the contribution of different modal features. This is used to dynamically capture the cross-modal dependency of the gas turbine under complex operating conditions, making the fusion result more consistent with the real change trend of the overall operating state.

[0091] This fusion network performs end-to-end learning through a joint objective function. If the system has clear labels (such as normal or abnormal states), a classification loss function is used to optimize recognition performance. If some samples lack labels, a self-supervised or contrastive learning mechanism is introduced to improve feature representation by maximizing semantic consistency between modalities.

[0092] Meanwhile, the reconstruction error term can be superimposed to constrain the model's fidelity in the reconstructed feature space, ensuring the model remains robust even with weak annotations or missing data. To avoid excessive concentration or imbalance in attention distribution, attention sparsity constraints are introduced during weight updates during training, allowing the model to retain supplementary information from secondary modalities while focusing on the primary modalities. Furthermore, in cases of sensor malfunctions or missing data, the system automatically masks failed channels through a masking mechanism or replaces them with learnable missing modal embeddings, thereby maintaining the continuity and stability of the overall feature fusion process.

[0093] An adaptive gradient optimization algorithm is preferably used for iterative updates, combined with a dynamic learning rate adjustment strategy to achieve rapid convergence. The system can periodically perform online fine-tuning and incremental training based on changes in gas composition, load fluctuations, or operational control strategies to maintain the model's adaptability and prediction accuracy over long-term operation. Through these steps, the trained multimodal feature fusion network can output a global feature representation vector characterizing the overall operating state of the gas turbine. .

[0094] In this embodiment, the multimodal feature fusion network includes a modal encoder, an attention fusion layer, and a residual normalization module, which is used to map the running feature vectors from different subsystems into a unified global feature representation vector. Furthermore, the multimodal feature fusion network may optionally adopt a multi-head attention structure, which divides the feature dimension into several subspaces and computes multiple sets of queries, keys, and values ​​in parallel to improve the precision of cross-modal interaction and nonlinear modeling capabilities.

[0095] Through the aforementioned weighted fusion process, the system can adaptively identify the contribution of different subsystem features to the global operating state, forming a high-dimensional and consistent global feature expression vector. This provides a unified input basis for subsequent operating condition identification and anomaly detection units, enabling semantic association and enhanced state perception among multi-source signals.

[0096] The abnormal operation monitoring and early warning system also includes an operating condition identification unit 3. This unit 3 uses a gas turbine dynamic evolution model constructed based on a Long Short-Term Memory (LSTM) network and a Hidden Markov Model (HMM) to analyze the global feature representation vector. Timing coding is performed to identify the current operating state of the gas turbine (start-up, steady state, variable load, shutdown, etc.).

[0097] In this embodiment, the working condition identification unit 3 includes a timing coding module 31 and a working condition prediction module 32;

[0098] The gas turbine dynamic evolution model is coupled to a hidden Markov model via a long short-term memory network (LSTM) in series. The LSM represents the global feature vector. Time-series encoding is performed to output a hidden state vector sequence, which reflects the time correlation, load change pattern and potential trend characteristics during the operation of the gas turbine;

[0099] The Hidden Markov Model is used to receive the hidden state vector sequence output by the Long Short-Term Memory Network, establish a state transition model for the gas turbine operating conditions, and calculate the optimal state path using the Viterbi algorithm to identify the current gas turbine operating condition category.

[0100] The operating condition identification unit 3 identifies the current operating condition of the gas turbine, and the specific steps involved are as follows:

[0101] The received global feature representation vector is processed by the timing coding module 31. According to time step Perform serialization to construct the input sequence It is used to reflect the characteristic change trajectory of the gas turbine over a continuous period of time;

[0102] The input sequence is encoded using a Long Short-Term Memory (LSTM) network. The activation values ​​of the input gate, forget gate, and output gate are calculated to control the weight ratio of historical features and current input, thereby capturing long-term dependent information and suppressing short-term disturbances. (The LSTM forget gate is used to filter short-term disturbances (such as load fine-tuning and gas source fluctuations), the input gate is used to introduce new observation features (such as temperature rise rate and fuel ratio change rate), and the output gate is used to retain key dynamic evolution information.)

[0103] Specifically, in this embodiment, the forget gate calculates the activation value of the forget gate based on the hidden state of the previous time step and the current input feature vector. This is used to filter out the part of the historical information that is useful for the current prediction and to filter out short-term disturbances, such as gas turbine load fine-tuning or gas source fluctuations, thereby avoiding short-term noise interference in the model's capture of long-term dependent features.

[0104] The input gate calculates its activation value by jointly encoding the current input feature vector and the previous hidden state, thereby controlling which new observed features (such as the rate of temperature rise and the rate of change of fuel ratio) are introduced into the cell state, thus achieving a dynamic response to changes in the current state.

[0105] During gas turbine operation, environmental conditions (such as inlet air temperature, humidity, and gas calorific value) significantly affect the combustion process, compressor efficiency, and turbine load distribution, thereby influencing the overall operating condition of the gas turbine. Therefore, an environmental feature input layer is introduced in the operating condition identification unit 3 to jointly model external environmental parameters during the LSTM time-series coding stage. The environmental feature input layer constructs an environmental feature vector from the collected environmental parameters. This environmental feature vector includes environmental parameters such as gas turbine inlet air temperature, inlet air humidity, and gas calorific value, which characterize the impact of external conditions on the gas turbine's combustion efficiency and output performance. After performing a normalization transformation on the above environmental feature vector (the environmental feature input layer performs normalization and embedding mapping processing on inlet air temperature, inlet humidity, and gas calorific value to generate an environmental embedding vector with the same dimension as the global feature expression vector), a linear embedding mapping function is used to... The vector is mapped to the same dimension as the global feature representation vector, and the environment embedding vector is concatenated with the global feature representation vector to form a joint input, which is then fed into the LSTM temporal coding network to achieve synchronous modeling of the gas turbine operating state and external environmental disturbances.

[0106] The output gate combines the encoding results of the current cell state and the hidden layer state at the previous time step to generate the output gate activation value, which is used to retain key dynamic evolution information and determine the output of the current hidden layer state vector for subsequent working condition prediction and state transition modeling.

[0107] At each time step The hidden state vector is updated through the state transfer mechanism of the long short-term memory network, and a hidden state vector sequence is constructed. This hidden state vector sequence constitutes the time-dependent expression of the gas turbine operation process.

[0108] The timing coding module 31 calculates the weighted average of the hidden state vectors at each stage using a sliding time window mechanism, extracts the long-term dependence features of the gas turbine performance parameters, and obtains the timing coding result. ;

[0109] The operating condition prediction module 32 is used to receive the hidden layer state vector sequence output by the timing coding module 31 and input it into the hidden Markov model.

[0110] Define the set of potential operating conditions for gas turbine operation. In the formula, This represents the total number of potential states, i.e., the number of hidden states into which the gas turbine operating conditions are divided. This parameter can be automatically selected and determined based on operating feature clustering, expert experience, or the model. It initializes the state transition probability matrix (describing the temporal transition rules between states), the observation probability matrix (describing the mapping relationship from hidden states to observed signals (such as temperature, speed, vibration value, etc.), and the initial state distribution (the initial state distribution is used to describe the state probability at the beginning of the model. The initialization adopts historical data statistics or expert experience assignment. When there are labeled historical samples, supervised estimation or EM algorithm is preferred to train the initial parameters. The initialization needs to cover typical operating conditions (start-up, steady state, load increase, load decrease, shutdown, etc.) and reserve special hidden states for rare but dangerous abnormal operating conditions to improve the identification sensitivity). This is used to describe the transition relationship between different operating conditions and their observation characteristics.

[0111] Specifically, the state transition probability matrix Observation probability matrix and initial state distribution ,in:

[0112] , Indicates from state Transition to state The probability of;

[0113] , Indicates the state Generate observations The probability of;

[0114] , This indicates the initial state of the system. The probability of;

[0115] In the formula, Indicates the first One potential operating condition; Represents the observed value, i.e., time. The observed sensor data of the output gas turbine (such as temperature, pressure, vibration, current, etc.); Indicates the first A number of potential operating conditions (such as startup, steady state, variable load, shutdown, etc.); Represents probability;

[0116] For each potential operating condition Define the observation probability matrix (That is, the emission probability, and the emission probability can be implemented using a Gaussian model).

[0117] Based on historical operating datasets (and expert-annotated operating condition samples), the state transition probability matrix, observation probability matrix, and initial state distribution are iteratively optimized using a forward-backward algorithm to obtain the optimal state transition structure.

[0118] The Viterbi algorithm is used to calculate the optimal state path for the state sequence of the Hidden Markov Model to identify the current operating condition category of the gas turbine. The current operating condition category of the gas turbine is determined based on the terminal state of the obtained optimal state path, and the confidence level of identification is evaluated based on the log probability difference or posterior probability of the optimal path.

[0119] Furthermore, by employing the Markov propagation algorithm, the state transition path within a few future time steps is predicted, thereby enabling the identification and determination of the current operating conditions of the gas turbine.

[0120] In this embodiment, the Viterbi algorithm is used for optimal path inference (i.e., the Viterbi algorithm is used to identify the optimal path in the time series). The most likely hidden state path under (where, Indicates at time The observed values, Given the total length of the time series, the specific steps involved in retaining only the optimal path probability up to the current time step at each step, and finding the globally optimal path through backtracking, are as follows:

[0121] For time series, at time... At that time, for each potential operating condition. Calculate its initial score:

[0122] ;

[0123] Concatenate backtrack pointer ;

[0124] In the formula, Represents the working state in a Hidden Markov Model (HMM) The initial probability represents the prior probability that the gas turbine may be in this state when it starts or the sequence begins; Represents the observation probability (i.e., the emission probability), and indicates the operating condition. Below, the first hidden state vector is observed. The probability; logarithmic form This is used to convert multiplication operations into addition to avoid underflow of consecutive multiplications; This indicates that the hidden state vector was observed in the first step. Under these circumstances, the gas turbine is most likely to start in state The logarithmic probability;

[0125] For each moment in the time series, from the second step to the last step, represent each potential operating condition. calculate:

[0126] ;

[0127] At the same time, record the index of the predecessor state that makes the above expression take its maximum value. ,in, A backtracking pointer, used to record at time steps. Under these conditions, the operating condition is reached. From which predecessor state is the optimal path? It was transferred over;

[0128] In the formula, Indicates the time At any given time, the operating condition is as follows: The optimal logarithmic path probability (i.e., the optimal cumulative probability to date). Represents the state transition probability, indicating the state from the previous time step. Transition to the current state The possibility; This means converting the state transition probability into a logarithmic probability and adding it to the previous optimal probability to obtain the result. Path score; This means choosing the option that maximizes the path to the current state from all possible preceding states. The path; Indicates the current state Lower observation The logarithm of the launch probability is used to update the path probability;

[0129] At any moment Select the final optimal state (used to determine the most likely state of the gas turbine at the end of the observation sequence, and as the endpoint of the backtracking):

[0130] ;

[0131] The final log probability of the optimal path is ;

[0132] In the formula, Indicated to At any given moment, the status of each operating condition The optimal cumulative log probability; This indicates that the state with the highest probability is chosen as the final optimal state. , Indicates based on time 1 to time 2. All observation data, calculated using the Viterbi algorithm, at the current moment The most likely hidden state (i.e. operating condition). This represents the logarithmic probability of the entire optimal path, used to quantify the confidence level of the entire path;

[0133] From the key state Starting from the beginning, use the backtracking pointer in reverse. Find the optimal predecessor state for each step, and backtrack from the end to the beginning to obtain the complete optimal path. (that is, to obtain the entire optimal state sequence) ).

[0134] Furthermore, the terminal state of the optimal path obtained through backtracking... This serves as the result of determining the current operating condition category of the gas turbine; among which... In the formula, A predefined set of operating condition categories for the HMM (such as normal, light load, overload, abnormal vibration, etc.). Indicates the time of the gas turbine The most likely operating condition category; This represents the last state in the set of states, i.e., the [number]th state. One state;

[0135] Calculate the posterior probability using the forward-backward algorithm. To quantify confidence levels, confidence metrics include:

[0136] Log probability difference between the optimal path and the second-best path ;

[0137] Posterior probability The deviation of the state duration from the historical average duration is obtained through normalization (such as normalization to ensure that the sum of the posterior probabilities of all working conditions is 1).

[0138] In the formula, Indicates time Logarithmic path probability of the optimal operating condition; This represents the observation sequence, a collection of observational data from the gas turbine over a continuous period. It includes encoded characteristics of sensor data such as temperature, pressure, and vibration, reflecting the gas turbine's performance from time 1 to time 2. The trajectory of its running status; This indicates the gas turbine at the last observation time. The actual working condition; This indicates all non-optimal operating conditions ( To remove Other operating condition indexes, Take from ) The maximum value represents the logarithmic probability corresponding to the second most reliable working condition path;

[0139] By setting a confidence threshold When the confidence level is lower than the confidence threshold When this occurs, the current operating condition is marked as a low-confidence category, triggering manual verification or a more conservative early warning strategy;

[0140] Furthermore, in this embodiment, after obtaining the current inferred state and path, the future is predicted based on the state transition matrix. The most likely sequence of states for the step;

[0141] Using the Markov propagation algorithm, the current state distribution is... (The most likely state can be represented by a posterior distribution or a unit vector.) Starting with this, iteratively apply the transition matrix:

[0142] ;

[0143] The maximum probability component at each step is used as the predicted state;

[0144] In the formula, This indicates that the gas turbine will be the next... Each time step (i.e., time point) The probability distribution of the working condition is a dimensionless distribution. vectors ( (This represents the total number of preset potential operating conditions, such as start-up, steady state, variable load, shutdown, etc.). Each element in the vector corresponds to the gas turbine at time [time value missing]. The probability of being in a certain potential working condition is such that the sum of all elements is 1. Represents the state transition matrix conduct The result obtained from the matrix multiplication operation, i.e., the gas turbine from time... working conditions Departure, Passing through The working condition transitions to the next time step after each time step. The probability of; Indicates the gas turbine at the current target time. The probability distribution of operating conditions;

[0145] By outputting the predicted path sequence and the state probability distribution at each step, it can be used to trigger early warnings or control recommendations based on operating conditions.

[0146] The system for monitoring and warning of operational anomalies also includes an anomaly identification unit 4. The anomaly identification unit 4 constructs an anomaly identification model based on a combination of a deep autoencoder and a temporal convolutional network (TCN), reconstructs and predicts multi-source features, calculates a comprehensive anomaly score, and then determines the operational anomaly status.

[0147] Among them, multi-source features include at least a global feature representation vector. and current gas turbine operating status .

[0148] In this embodiment, the anomaly identification unit 4 includes an anomaly scoring module 41 and an anomaly identification output module 42;

[0149] Among them, the anomaly scoring module 41 performs a comprehensive analysis on the reconstruction and prediction results of the gas turbine operation characteristics based on the anomaly identification model to generate a comprehensive anomaly score;

[0150] Specifically, the anomaly identification model generates anomaly scores through the following steps:

[0151] Anomaly scoring module 41 is used to receive the hidden layer state vector sequence. Global feature representation vector sequence and the current operating status of the gas turbine, among which, This represents the hidden state vector output by the gas turbine at the last time step of the observation period, after being time-coded by the Long Short-Term Memory (LSTM) network of the operating condition identification unit 3. This indicates the last time step (time) of the gas turbine observation period. After being fused by the multi-source feature extraction unit 2, the output is the global running state feature vector;

[0152] Current moment Next, the input hidden state vector sequence and global feature representation vector sequence are input into a deep autoencoder (DAE) for feature reconstruction, generating a reconstructed vector. And based on the current moment Global feature representation vector With reconstructed vector The degree of deviation is used to calculate the reconstruction error;

[0153] The encoded reconstructed vector or original input feature sequence Input a temporal convolutional network (TCN) to perform multi-step future state prediction, generate a prediction vector, and calculate the prediction error based on the deviation between the actual global feature vector and the prediction vector;

[0154] The reconstruction error and prediction error are weighted and fused together, and a comprehensive anomaly score is obtained according to the weights set based on the gas turbine operating conditions and historical experience. ;

[0155] in, In the formula, This indicates the reconfiguration error, reflecting the extent to which the instantaneous state of the gas turbine deviates from the historical pattern; This indicates the prediction error, reflecting the degree to which the gas turbine's operating trend deviates from the expected pattern; and This indicates the preset weights, set based on gas turbine operating conditions and historical experience; among them, the comprehensive anomaly score... A higher value indicates a more abnormal operating state of the gas turbine; the anomaly score also considers short-term disturbances (from...). (capture) and long-term trend deviation (by) (Capture), enabling a comprehensive assessment of the gas turbine's operating status;

[0156] In this embodiment, the depth autoencoder (DAE) includes an encoder and a decoder;

[0157] The encoder is used to transform the input global feature representation vector. (Global Feature Representation Vector) Multi-source heterogeneous data (including gas turbine temperature, pressure, speed, load, fuel ratio, etc.) are compressed into a low-dimensional latent space to generate latent vectors, thereby extracting key mode features of gas turbine operation; and the encoder layer adopts a layer-by-layer fully connected network to gradually extract high-order combined features and form latent vectors.

[0158] The decoder is used to reconstruct the latent vector into a vector with the same dimension as the input. , used to approximate the original features;

[0159] Deep autoencoders (DAEs) reconstruct features and generate reconstructed vectors. The specific steps are as follows:

[0160] Let the input global feature representation vector be... Furthermore, the depth autoencoder (DAE) consists of an L-layer encoder and an L-layer decoder.

[0161] Input global feature representation vector Mapping to low-dimensional latent vectors (generally );

[0162] The latent vector is decoded Map back to the original dimensions to generate a reconstructed vector. ;

[0163] The L2 norm is used to calculate the deviation between the input and the reconstructed data, forming the reconstruction error. ;

[0164] in, In the formula, This represents the L2 norm (Euclidean norm). Used to quantify the degree of deviation of the gas turbine's operating status from its historical normal mode.

[0165] Furthermore, in this embodiment, a multi-step future state prediction is performed using a Temporal Convolutional Network (TCN) to generate a prediction vector. The specific steps involved are as follows:

[0166] The reconstructed vector or original input feature sequence Inputting a multi-layer temporal convolutional network (TCN), the TCN employs causal convolution, dilated convolution, residual connections, and a multi-layer stacked structure to perform multi-step future state prediction and generate a prediction vector. ;

[0167] Based on the actual global feature vector With prediction vector Calculate the degree of deviation and the prediction error In the formula, Indicates time The actual global feature representation vector, i.e., the gas turbine in the future... The global running state vector is obtained by actually collecting data at each time step and processing it through the multi-source feature extraction unit 2. ( (The feature dimension includes features of multiple subsystems such as temperature, pressure, and load). This represents the time step offset for future predictions; Indicates the maximum number of prediction steps; Indicates time The predicted global feature representation vector.

[0168] Among them, the Temporal Convolutional Network (TCN) is trained using historical normal operating data to automatically learn the load change pattern, temperature rise trend and dynamic relationship of fuel ratio, and achieve multi-step prediction;

[0169] Reconstruct the vector sequence or original input feature sequence Constructing time series input For multi-step future state prediction in TCN, where, Indicates the gas turbine at the current time step The reconstructed feature vector generated after processing by the decoder of a deep autoencoder (DAE) Indicates the gas turbine at the current time step The global running state feature vector obtained after fusion processing by the multi-source feature extraction unit 2 Indicates time Global feature representation vector; input sequence Each time step contains multi-source heterogeneous characteristics of the gas turbine (such as temperature, pressure, speed, fuel ratio, load, etc.), reflecting the operating state trajectory of the gas turbine in a continuous period of time;

[0170] Furthermore, the Temporal Convolutional Network (TCN) architecture includes convolutional layers (ensuring that the current prediction relies solely on historical time step information, avoiding the leakage of future information), dilated convolutions (expanding the receptive field by setting a dilation factor; TCN can capture long-term operating trends of gas turbines, such as periodic load changes, fuel consumption patterns, and temperature rise / fall trends), residual connections (introducing residual connections into multi-layered stacked convolutional networks to avoid gradient vanishing or exploding problems during deep network training; especially for gas turbine scenarios, it can maintain the stability and continuity of the prediction sequence, ensuring prediction accuracy even if the input sequence contains sudden noise or abnormal perturbations), multi-layer stacking (TCN uses a multi-layered stacked convolutional structure to extract higher-order temporal features layer by layer), and an output layer (outputting future...). Global feature vectors at each time step, and each prediction vector Includes key operating parameters of the gas turbine, such as temperature, pressure, speed, load, and fuel ratio.

[0171] Specifically, in this embodiment, historical normal operation data is used. Or reconstruct vector Supervised training is performed, using feature vectors from future time steps as labels, to minimize the prediction error loss function; the optimization algorithm uses Adam, RMSProp, or other gradient descent methods, and a sliding window strategy is used to update the model parameters during training.

[0172] Furthermore, regarding the current time step The prediction error is calculated based on the degree of deviation between the actual global feature vector and the predicted vector. This prediction error is used to quantify the degree of deviation of the future operating state of the gas turbine from the historical pattern and trend, and is used by the subsequent anomaly scoring module 41 to generate a comprehensive anomaly score.

[0173] The anomaly identification output module 42 determines whether the gas turbine is in an abnormal state based on the anomaly judgment rules and the anomaly score result by setting a preset anomaly score threshold, and generates an anomaly identification output to trigger an early warning.

[0174] In this embodiment, the specific steps (i.e., anomaly judgment rules) involved in the anomaly recognition output module 42 generating the anomaly recognition output are as follows:

[0175] Set an abnormal scoring threshold and multiple level thresholds Used to distinguish mild cases ( ), moderate ( ) and serious ( Abnormal state;

[0176] The calculated comprehensive anomaly score Compared with the preset abnormal scoring threshold Comparison:

[0177] If the comprehensive abnormal score Less than the abnormal scoring threshold If the gas turbine is operating normally, no further action is required.

[0178] If the comprehensive abnormal score Greater than or equal to the anomaly scoring threshold If so, the anomaly identification output module 42 is triggered to classify the levels (subdivide the abnormal state according to the preset level threshold).

[0179] The detailed classification includes the following rules:

[0180] like If it is, it is judged as a mild abnormality (an abnormality has been triggered but has not yet reached the mild level threshold, which can be regarded as the initial stage of the abnormality).

[0181] like If so, it is judged as moderately abnormal;

[0182] like If so, it is judged as a serious abnormality;

[0183] like If so, it is judged as an extremely serious abnormality;

[0184] Anomaly identification output is generated based on the classification results, including the current gas turbine status category, comprehensive anomaly score, and triggered alarm or maintenance prompt information; finally, the anomaly identification output module 42 generates alarm, maintenance prompt and trend warning information according to the level.

[0185] Furthermore, in this embodiment, the abnormal operation monitoring and early warning system also involves an abnormal classification unit. The abnormal classification unit classifies and analyzes the abnormal features output by the abnormal identification unit, and uses the feature importance ranking and SHAP interpretation mechanism based on XGBoost to calculate the contribution of each input variable to the abnormal score and output the abnormal source (such as gas supply, lubrication system, cooling system, etc.).

[0186] Specifically, the anomaly classification and early warning unit receives the comprehensive anomaly score, reconstruction error vector, prediction error vector, and global feature vector output by the anomaly identification unit 4. ;

[0187] The above features are used to construct the input matrix. This is used to reflect the trajectory of the operating characteristics of the gas turbine's multiple subsystems over a continuous period of time. In the formula, The total number of features is equal to the sum of the original key parameters and derived features; each row corresponds to the state of the gas turbine at a certain time step, and each column corresponds to the key parameters or derived features of the gas turbine, including fuel flow rate, gas pressure, gas temperature, lubricating oil pressure, lubricating oil temperature, cooling water temperature, cooling water flow rate, speed, load, exhaust temperature, etc.

[0188] Input matrix Input a gradient boosting tree-based classification model (XGBoost, where, during training or prediction, the model learns anomaly category labels based on historical gas turbine operating data and labeled abnormal event data, including abnormal gas supply, abnormal lubrication system, abnormal cooling system, abnormal load regulation, etc.).

[0189] Each tree is iteratively accumulated for prediction to obtain the final anomaly category prediction. During model training or prediction, the tree depth, learning rate and subsampling ratio are adjusted in combination with parameters such as gas turbine load change rate, fuel ratio change rate and coolant temperature rise rate to ensure that the model is not sensitive to short-term disturbances and improve anomaly detection accuracy.

[0190] The contribution of each input feature to the anomaly category prediction is calculated using XGBoost internal metrics (gain, cover, frequency). These metrics are normalized and fused according to preset weights to obtain a preliminary contribution, which is then ranked.

[0191] The parameters of key sub-modules of the gas turbine are categorized and mapped. For example, fuel flow and gas pressure correspond to the gas supply system; lubricating oil pressure and lubricating oil temperature correspond to the lubrication system; and cooling water temperature and cooling water flow correspond to the cooling system. Features are mapped to system modules (gas, lubrication, cooling, control, measurement, etc.). Features within the same module are weighted and summarized (e.g., maximum value or weighted average) to obtain the module-level contribution, which facilitates the location of subsystems. Features with high abnormal contribution are sorted to provide a basis for subsequent anomaly source location.

[0192] The SHAP mechanism is applied to each anomaly category prediction result to calculate the Shapley value for each input feature. To clarify the local and global contributions of each feature to the anomaly score (specifically, the TreeSHAP algorithm is used to calculate the Shapley value of each input variable for samples at key time points and aggregated over the time window to obtain the local and global contributions; a feature is considered a high-contribution feature only when it simultaneously meets the XGBoost ranking threshold and the SHAP aggregation contribution threshold);

[0193] Based on the Shapley values, the key feature parameters with the greatest contribution were selected and mapped to specific subsystems of the gas turbine to determine the source of the anomaly.

[0194] Output information about the source of the anomaly (such as gas supply, lubrication system, cooling system, etc., and their corresponding contribution).

[0195] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An artificial intelligence-based monitoring and early warning system for operational anomalies in gas turbine power plants, characterized in that, include: Data acquisition unit (1), the data acquisition unit (1) is used to acquire multi-dimensional operation data of gas turbine power plant and preprocess the multi-dimensional operation data; The multi-source feature extraction unit (2) extracts features from the synchronized multi-dimensional operating signal to obtain the operating feature vectors of each subsystem. Furthermore, by introducing a multimodal feature fusion network based on an attention mechanism, the running feature vector is improved. Perform weighted aggregation to obtain the global feature representation vector. ; Operating condition identification unit (3), wherein the operating condition identification unit (3) adopts a gas turbine dynamic evolution model constructed based on long short-term memory network and hidden Markov model, and performs global feature expression vector analysis. Timing coding is performed to identify the current operating condition of the gas turbine; Anomaly identification unit (4) constructs an anomaly identification model based on the combination of deep autoencoder and temporal convolutional network (TCN), reconstructs and predicts multi-source features, calculates a comprehensive anomaly score, and then judges the abnormal running state. Among them, multi-source features include at least a global feature representation vector. and current gas turbine operating status .

2. The artificial intelligence-based operation anomaly monitoring and early warning system for gas turbine power plants according to claim 1, characterized in that: The multidimensional operating data includes at least the following: gas flow rate, compressor inlet temperature, combustion chamber pressure, turbine outlet temperature, bearing vibration, motor current, voltage, and emission parameters. The multidimensional operational data preprocessing includes timestamp synchronization and adaptive Kalman filtering.

3. The artificial intelligence-based operation anomaly monitoring and early warning system for gas turbine power plants according to claim 1, characterized in that: The multi-source feature extraction unit (2) includes a feature extraction module (21) and a feature fusion module (22); The feature extraction module (21) is used to extract the time domain, frequency domain, and statistical features of the preprocessed multidimensional running data to obtain the running feature vectors of each subsystem. ; The feature fusion module (22) uses an attention-based multimodal feature fusion network to assign importance weights to the features of each subsystem. The operational characteristic vectors of each subsystem Perform weighted aggregation to generate a global feature representation vector. ; The multimodal feature fusion network includes a modal encoder, an attention fusion layer, and a residual normalization module, which is used to map the running feature vectors from different subsystems into a unified global feature representation vector.

4. The artificial intelligence-based operation anomaly monitoring and early warning system for gas turbine power plants according to claim 3, characterized in that: The global feature representation vector is generated through the attention-based multimodal feature fusion network. The specific steps involved are as follows: For each running feature vector By performing a linear mapping transformation through a modal encoder, a uniform-dimensional embedding representation is obtained. ; For each modal embedding The corresponding query, key, and value vectors are calculated, and a scaled dot product attention mechanism is used in the attention fusion layer to calculate the compatibility score. ; If the attention fusion layer has missing modalities, a mask identifier is introduced. Masking invalid modes and compatibility scoring Make corrections; Attention weights are calculated using the Softmax function. The value vectors are weighted and aggregated to generate a preliminary global fusion result. The global fusion result is then processed through a residual normalization module. Nonlinear mapping, residual connections, and layer normalization are performed to form the initial feature representation vector. ; Initial feature representation vector Finally, the global feature representation vector is obtained after normalization. .

5. The artificial intelligence-based operation anomaly monitoring and early warning system for gas turbine power plants according to claim 1, characterized in that, The working condition identification unit (3) includes a timing coding module (31) and a working condition prediction module (32). The gas turbine dynamic evolution model is coupled to a hidden Markov model via a long short-term memory network (LSTM) in series. The LSM represents the global feature vector. Perform temporal encoding to output a sequence of hidden layer state vectors; The Hidden Markov Model is used to receive the hidden state vector sequence output by the Long Short-Term Memory Network, establish the state transition model of the gas turbine operating conditions, and calculate the optimal state path using the Viterbi algorithm.

6. The artificial intelligence-based operation anomaly monitoring and early warning system for gas turbine power plants according to claim 5, characterized in that: The operating condition identification unit (3) identifies the current operating condition of the gas turbine, and the specific steps involved are as follows: The received global feature representation vector is processed by the timing coding module (31). According to time step Perform serialization to construct the input sequence; The input sequence is encoded using a long short-term memory network, and the activation values ​​of the input gate, forget gate, and output gate are calculated respectively. At each time step The hidden state vector is updated through the state transfer mechanism of the Long Short-Term Memory network, and the hidden state vector sequence is constructed. The timing coding module (31) calculates the weighted average of the hidden state vectors at each stage through a sliding time window mechanism to obtain the timing coding result. ; The working condition prediction module (32) is used to receive the hidden layer state vector sequence output by the timing coding module (31) and input it into the hidden Markov model; Define the set of potential operating conditions for gas turbine operation. And initialize the state transition probability matrix, observation probability matrix and initial state distribution; Based on historical running datasets, the state transition probability matrix, observation probability matrix, and initial state distribution are iteratively optimized using a forward-backward algorithm. The Viterbi algorithm is used to calculate the optimal state path for the state sequence of the Hidden Markov Model. The current gas turbine operating condition category is determined based on the terminal state of the obtained optimal state path, and the confidence level is evaluated based on the log probability difference of the optimal path. The state transition path within a few future time steps is predicted using the Markov propagation algorithm.

7. The artificial intelligence-based operation anomaly monitoring and early warning system for gas turbine power plants according to claim 1, characterized in that: The anomaly identification unit (4) includes an anomaly scoring module (41) and an anomaly identification output module (42). The anomaly scoring module (41) performs a comprehensive analysis of the reconstruction and prediction results of the gas turbine operation characteristics based on the anomaly identification model to generate a comprehensive anomaly score. The anomaly identification output module (42) determines whether the gas turbine is in an abnormal state based on the anomaly judgment rule and the anomaly score result by using a preset anomaly score threshold.

8. The artificial intelligence-based operation anomaly monitoring and early warning system for gas turbine power plants according to claim 7, characterized in that: The anomaly identification model generates an anomaly score through the following specific steps: The anomaly scoring module (41) is used to receive the hidden layer state vector sequence, the global feature expression vector sequence, and the current gas turbine operating status; Current moment Next, the input hidden layer state vector sequence and global feature representation vector sequence are input into the deep autoencoder for feature reconstruction, generating a reconstructed vector. And based on the current moment Global feature representation vector With reconstructed vector The degree of deviation is used to calculate the reconstruction error; The encoded reconstructed vector Input a temporal convolutional network to perform multi-step future state prediction and generate a prediction vector. Calculate the prediction error based on the deviation between the actual global feature vector and the prediction vector. The reconstruction error and prediction error are weighted and fused to obtain a comprehensive anomaly score. .

Images