Gas turbine health degree evaluation system and method

Abstract

The invention provides a gas turbine health degree evaluation system and method, and relates to the technical field of intelligent operation and maintenance of industrial equipment. The system comprises a data acquisition and preprocessing module used for acquiring and preprocessing operation data; the iTransform core processing module is used for carrying out deep feature learning and pattern recognition on the multivariate time sequence data; and the task specific output layer is used for mapping the feature representation output by the iTransform core processing module to a final output result according to the application task. According to the system, the gas turbine fault diagnosis and prediction precision can be remarkably improved, the calculation efficiency of processing mass and long-time-series data of the gas turbine is improved, the characteristic information of the gas turbine multi-element sensor is more effectively utilized and understood, the learning ability of the independent dynamic mode and fault characteristics of each sensor of the gas turbine is enhanced, and the fault diagnosis accuracy is improved. And the interpretability and the auxiliary decision-making capability of fault diagnosis of the gas turbine are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent operation and maintenance of industrial equipment, and in particular to a gas turbine health assessment system and method. Background Art

[0002] As key power equipment in energy, transportation, and other fields, the safe and stable operation of gas turbines is crucial. Traditional gas turbine condition monitoring and fault diagnosis methods rely primarily on expert experience, physical models, or simple statistical methods (such as threshold-based alarms, ARIMA models, and support vector machines (SVMs). These methods often suffer from low accuracy, weak generalization, and difficulty detecting early faults when dealing with increasingly complex, multivariate, and nonlinear gas turbine operating data.

[0003] In recent years, with the advancement of artificial intelligence (AI), deep learning-based methods such as recurrent neural networks (RNNs), long short-term memory (LSTMs), gated recurrent units (GRUs), and convolutional neural networks (CNNs) have been applied to gas turbine time series data analysis. These methods have significantly improved the model's expressiveness and prediction accuracy. However, models like RNNs and LSTMs are prone to vanishing / exploding gradients when processing extremely long sequences, and their serial processing mechanism limits computational efficiency. CNNs, on the other hand, focus more on extracting local features.

[0004] The Transformer model, owing to its powerful parallel processing capabilities and ability to capture long-range dependencies, has achieved tremendous success in natural language processing and is gradually being introduced into the field of time series analysis. Traditional Transformer models based on time-step tokenization (TST) directly treat each time point in a time series as a token. When processing multivariate time series, this approach often embeds variables with different physical meanings at the same time point, potentially leading to information confusion. Furthermore, when the time series length T is very large, the computational complexity of its self-attention mechanism (O(T²)) becomes a bottleneck. Research has attempted to adapt the Transformer to the characteristics of time series, for example, by reducing complexity through sparse attention mechanisms or by combining it with other network structures.

[0005] However, existing Transformer variants for time series forecasting still have room for improvement in fully exploiting the inherent correlations between variables in multivariate time series while efficiently learning the unique temporal dynamic patterns of each variable. This is especially true in scenarios like gas turbines, which involve a large number of sensors, high-dimensional data, complex coupling relationships between variables, and the need to detect early, weak fault signals from extremely long historical data. A more efficient and targeted multivariate time series processing architecture is urgently needed.

[0006] Existing technologies in the fields of gas turbine fault diagnosis, prognosis, and health assessment suffer from the following major shortcomings: 1. Inadequate modeling of coupled relationships between multi-sensor data: Traditional statistical methods and some early deep learning models (such as those that process each sensor independently) struggle to effectively capture the complex nonlinear and time-delayed coupled relationships between different physical parameters of a gas turbine (such as temperature, pressure, vibration, and speed). These coupled relationships often contain critical information about the health status of the equipment and the evolution of faults. Traditional Transformers, which process multi-sensor information in a mixed manner across time steps, also struggle to directly and clearly model the specific dependencies between variables. 2. Inadequate learning of dynamic patterns in individual sensors: When processing multi-sensor information in a mixed manner, the model may struggle to fully learn and characterize each sensor's unique baseline behavior, trend changes, cyclical fluctuations, and specific abnormal patterns—critical for distinguishing normal from abnormal conditions and identifying specific fault types. 3. The conflict between capturing long-range dependencies and computational efficiency: Many gas turbine faults evolve over a long period of time, and early fault signals can be very weak, requiring analysis over a long time window to detect. RNNs / LSTMs face gradient issues and efficiency bottlenecks when processing extremely long sequences. When the time series length T is large, the computational complexity of traditional Transformers, O(T²), makes it difficult to efficiently process extremely long historical data. 4. Poor model interpretability: Many deep learning models are like "black boxes" with opaque decision-making processes. It is difficult to explain the basis of fault diagnosis results to operation and maintenance personnel, or to point out which key parameters lead to abnormal judgments. This limits the credibility and guiding role of the model in practical applications. 5. Insensitive to early and weak fault signals: Due to the above-mentioned limitations in modeling capabilities, efficiency, and information utilization, existing methods may find it difficult to effectively identify early and weak fault signals in the embryonic stage, resulting in missed optimal maintenance opportunities. Summary of the Invention

[0007] The purpose of the present invention is to provide a gas turbine health assessment system and method to address the above-mentioned deficiencies in the prior art, so as to solve the problem of gas turbine health assessment.

[0008] To achieve the above object, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides a gas turbine health assessment system, the system comprising: Data acquisition and preprocessing module, used to obtain operating data from multiple sensors of the gas turbine and perform data preprocessing; The iTransformer core processing module is used to perform deep feature learning and pattern recognition on multivariate time series data preprocessed by the data acquisition and preprocessing module. The iTransformer core processing module includes an input transposition unit, a variable label embedding layer, and a stacked Transformer block. The input transposition unit is used to convert time series data of each length T from multiple different sensors into multiple independent variable labels. The variable label embedding layer is used to embed each variable label, mapping each variable label from the original sequence of length T to a vector representation of a preset fixed dimension. The stacked Transformer block is composed of multiple inverted Transformer blocks stacked together, each of which includes a multi-head self-attention mechanism layer and a feedforward network layer. The task-specific output layer is used to map the feature representation output by the iTransformer core processing module to the final output result according to the application task.

[0009] Optionally, the system further comprises: The result display and alarm module is used to present the output results of the task-specific output layer to the user in the form of visualization or reports, and trigger alarms according to preset rules; The knowledge base and model library are used to store pre-trained model parameters, standardized parameters, fault feature library, HI construction rules, historical cases, and provide support for the iTransformer core processing module and task-specific output layer.

[0010] Optionally, a multi-head self-attention mechanism layer acts on the embedded representations of multiple variable labels to capture the interdependencies and correlation weights between different sensor data sequences; a feedforward network layer is independently applied to the representation of each variable label output by the multi-head self-attention mechanism layer, and a nonlinear transformation is performed on the representation of each sensor that has been fused with the information of other related sensors to learn the time series pattern and deep features of the sensor itself.

[0011] Optionally, the variable label representation input to the inverted Transformer block is first normalized by one layer and then enters the multi-head self-attention mechanism layer. The output of the multi-head self-attention mechanism layer passes through a residual connection and then undergoes normalization by one layer before entering the feedforward network layer. The output of the feedforward network layer passes through a residual connection and then undergoes normalization by one layer as the final output of the inverted Transformer block.

[0012] Optionally, the multiple sensors include a temperature sensor, a pressure sensor, a vibration sensor and a rotation speed sensor.

[0013] In a second aspect, the present invention further provides a method for evaluating the health of a gas turbine, which is used in the system according to the first aspect, and comprises the following steps: After the gas turbine health assessment system is started, the data acquisition and preprocessing module collects multivariate time series data within a specified time window from each sensor of the gas turbine, where the number of sensors is N; Preprocess the collected raw data, including cleaning, synchronization, and normalization operations; The iTransformer model is processed. The preprocessed N sensor time series are fed into the iTransformer core processing module. The complete sequence of each sensor is regarded as a variable label. Then, the variable label embedding layer embeds these N variable labels into N D_model-dimensional vectors. The inverted Transformer block learns features. These N embedded vectors pass through L stacked inverted Transformer blocks in sequence. In each block, the multi-head self-attention mechanism layer calculates the self-attention between the N variable labels, captures the mutual dependence between different sensors, and updates the representation of each variable label to incorporate information from other related variables. The feedforward network layer independently performs nonlinear transformation on each updated variable label representation to learn and refine the deep time series features of the sensor itself. After processing by L inverted Transformer blocks, the feature representation of the N final variable labels is obtained. According to the application task, the feature representation output by the iTransformer core processing module is mapped to the final output result.

[0014] Optionally, based on the application task, the feature representation output by the iTransformer core processing module is mapped to the final output result, specifically including: The feature vector output by the iTransformer core processing module is input into a pre-trained classifier to determine whether the gas turbine is currently faulty and identify the specific fault type. The output result is the fault category label and confidence level. The feature vector output by the iTransformer core processing module is input into the regression model to predict the value of one or more key sensor parameters in the future. The output result is a prediction curve or predicted value. The feature vector output by the iTransformer core processing module is input into the health assessment model to quantitatively assess the current overall health status of the gas turbine, and the output result is a health score or grade.

[0015] Optionally, the method also includes: result display and decision support, displaying the diagnostic results, prediction information, and health index to the operation and maintenance personnel through dashboards, trend charts or reports. If a serious fault is diagnosed, a parameter is predicted to exceed a limit, or the health index is lower than a preset threshold, the gas turbine health assessment system automatically triggers an alarm and notifies relevant personnel via SMS, email or system message; for fault diagnosis, the attention weights calculated between variables by the multi-head self-attention mechanism layer are used to highlight the key sensors that contribute most to the current fault judgment.

[0016] The beneficial effects of the present invention include: The gas turbine health assessment system provided by the present invention includes: a data acquisition and preprocessing module, which is used to obtain operating data from multiple sensors of the gas turbine and perform data preprocessing; an iTransformer core processing module, which is used to perform deep feature learning and pattern recognition on multivariate time series data preprocessed by the data acquisition and preprocessing module. The iTransformer core processing module includes an input transposition unit, a variable label embedding layer and a stacked Transformer block. The input transposition unit is used to convert time series data of each length T from multiple different sensors into multiple independent variable labels. The variable label embedding layer is used to perform an embedding operation on each variable label, mapping each variable label from the original sequence of length T to a vector representation of a preset fixed dimension. The stacked Transformer block is composed of multiple inverted Transformer blocks stacked together, and each inverted Transformer block includes a multi-head self-attention mechanism layer and a feedforward network layer; a task-specific output layer is used to map the feature representation output by the iTransformer core processing module to the final output result according to the application task. This system can significantly improve the accuracy of gas turbine fault diagnosis and prediction, improve the computational efficiency of processing massive, long-term time series data of gas turbines, more effectively utilize and understand the characteristic information of gas turbine multi-sensors, enhance the learning ability of independent dynamic modes and fault characteristics of each gas turbine sensor, and improve the interpretability and decision-making support capabilities of gas turbine fault diagnosis. It is expected to improve the generalization and robustness of the model for different operating conditions and early weak faults, and significantly enhance the ability to process extremely long historical data of gas turbines to support long-term health management. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 A schematic diagram of the iTransformer model structure provided by an embodiment of the present invention is shown; Figure 2 A schematic diagram of the process flow of a gas turbine intelligent monitoring system provided by an embodiment of the present invention is shown; Figure 3 A schematic structural diagram of a gas turbine bearing early fault diagnosis module provided by a first embodiment of the present invention is shown; Figure 4 The figure shows the overall flow chart of early fault diagnosis of gas turbine bearings provided by the first embodiment of the present invention; Figure 5 A schematic diagram of the structure of a gas turbine HI trend prediction and RUL evaluation module provided by a second embodiment of the present invention is shown; Figure 6 The overall flow chart of HI trend prediction and RUL evaluation of a gas turbine provided by the second embodiment of the present invention is shown. DETAILED DESCRIPTION

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0020] Existing Transformer variants for time series forecasting still have room for improvement in fully exploiting the inherent correlations between variables in multivariate time series while efficiently learning the unique temporal dynamic patterns of each variable. This is especially true in scenarios like gas turbines, which involve a large number of sensors, high-dimensional data, complex coupling relationships between variables, and the need to detect early, weak fault signals from extremely long historical data. A more efficient and targeted multivariate time series processing architecture is urgently needed.

[0021] This paper applies the iTransformer model to gas turbine health assessment. By adopting an "inverted" perspective, the model treats the complete time series of each variable as a token and applies the core Transformer mechanism to the variable dimension, providing a new approach to solving the aforementioned problems. The core concept of the proposed iTransformer-based gas turbine health assessment method and system is to use the iTransformer model to process multivariate time series data collected from various gas turbine sensors, enabling accurate analysis and intelligent decision-making of the gas turbine's operating status.

[0022] Figure 1A schematic diagram of the iTransformer model structure provided by an embodiment of the present invention is shown; Figure 2 The figure shows a flow chart of the gas turbine intelligent monitoring system provided by the embodiment of the present invention. Figure 1 and Figure 2 As shown, the gas turbine health assessment system provided by the present invention includes: a data acquisition and preprocessing module 201, an iTransformer core processing module 202 and a task-specific output layer 203.

[0023] The data acquisition and preprocessing module 201 is used to acquire operating data from multiple gas turbine sensors and perform data preprocessing. Specifically, the data acquisition and preprocessing module 201 is responsible for collecting operating data in real time or in batches from various gas turbine sensors (such as temperature sensors, pressure sensors, vibration sensors, and speed sensors). It also performs data preprocessing, including data cleaning (processing missing values ​​and outliers), data synchronization, data normalization / standardization, and, if needed, data downsampling or feature engineering.

[0024] The data acquisition and preprocessing module 201 can be connected to a commonly used data acquisition system for gas turbines (such as DCS, SCADA), and perform adaptive processing on the original data to meet the input requirements of the subsequent iTransformer model.

[0025] The iTransformer core processing module 202 is used to perform deep feature learning and pattern recognition on the multivariate time series data preprocessed by the data acquisition and preprocessing module. Figure 1 The iTransformer core processing module 202 includes an input transposition unit 101 , a variable token embedding layer 102 , and stacked Transformer blocks 100 .

[0026] The input transposition unit 101 is used to convert time series data of length T from multiple (e.g., N) different sensors into multiple (e.g., N) independent variable tokens. That is, the complete time series of each sensor is considered as a whole input unit. The input transposition unit 101 processes the complete time series of each sensor as an indivisible unit, which is fundamentally different from the traditional Transformer that treats each time point as a token. Optionally, the multiple sensors include a temperature sensor, a pressure sensor, a vibration sensor, and a speed sensor.

[0027] The variable tag embedding layer 102 is used to embed each variable tag, mapping it from its original sequence of length T to a vector representation of a preset fixed dimension (e.g., dimension D_model). This embedding process can employ a one-dimensional convolutional network (1D CNN), a multi-layer perceptron (MLP), or other suitable sequence encoding methods to extract global features from each sensor's complete sequence. The variable tag embedding layer 102 can extract a summarized feature representation from each sensor's entire time series segment, paving the way for subsequent modeling of inter-variable relationships and learning of intra-variable patterns.

[0028] The stacked Transformer block 100 is composed of multiple inverted Transformer blocks stacked together. The stacked Transformer block 100 is the core structure of the iTransformer model. Figure 1 As shown, each inverted Transformer block includes a multi-head self-attention layer 103 and a feedforward network layer 104. The multi-head self-attention (MHSA) layer 103 operates on the embedded representations of N "variable tags." Specifically, the attention mechanism is computed across the variable dimension (sensor dimension) to capture the interdependencies and correlation weights between different sensor data sequences. The key technology behind the MHSA layer 103's application to the variable dimension is that the self-attention mechanism is computed across N different "variable tags," directly modeling the interactions between different sensors (variables) rather than between time points. This enables the model to capture cross-variable dependencies, such as how the pattern of sensor B changes after a certain pattern in sensor A occurs. Its computational complexity is primarily related to N².

[0029] The Feed-Forward Network (FFN) layer 104 is applied independently to the representation of each variable label output by the MHSA layer. Specifically, each sensor's representation (which has already been fused with information from other related sensors) undergoes a further nonlinear transformation to learn the sensor's unique time series patterns and deep features. Each inverted Transformer block also includes layer normalization (LN) and residual connections to stabilize the training process and facilitate information flow. These are typically applied at the input and output of the MHSA and FFN. The FFN layer 104 is applied independently to each variable label. Its key technique is to apply a separate FFN to the representation of each variable (sensor) after the MHSA captures the interactive information between variables. This allows the FFN to focus on learning the time series dynamics of that specific sensor (such as trends, seasonality, and waveform characteristics under specific faults) without being directly influenced by the patterns of other variables, thus achieving a learning strategy of "interaction first, independent refinement later."

[0030] The multi-head self-attention mechanism layer acts on the embedded representation of multiple variable labels to capture the interdependencies and correlation weights between different sensor data sequences; the feedforward network layer is independently applied to the representation of each variable label output by the multi-head self-attention mechanism layer, and performs a nonlinear transformation on the representation of each sensor that has been fused with the information of other related sensors to learn the time series pattern and deep features of the sensor itself. Figure 1 The internal connection of the inverted Transformer block shown: the variable label representation input to the inverted Transformer block is first normalized by one layer and then enters the multi-head self-attention mechanism layer 103. The output of the multi-head self-attention mechanism layer 103 passes through a residual connection (added to the input of the MHSA) and then passes through another layer of normalization. Then it enters the feedforward network layer 104. The output of the feedforward network layer passes through a residual connection (added to the input of the FFN) and then passes through another layer of normalization as the final output of the inverted Transformer block.

[0031] The task-specific output layer 203 maps the feature representations output by the iTransformer core processing module to the final output, based on the application task (fault diagnosis, state prediction, or health assessment). For example, for state prediction, this might be a regression layer that outputs predicted values ​​for key parameters over a period of time. For health assessment, this might be a regression layer or scoring model that outputs a quantitative health index. The task-specific output layer 203 can be flexibly configured based on the specific needs of gas turbine operation and maintenance (diagnosis, prediction, or assessment), transforming the deep features learned by the iTransformer into meaningful decision-making support information.

[0032] Optionally, the system also includes a results display and alarm module 204, which presents the output results of the task-specific output layer (such as diagnostic conclusions, prediction trends, health scores, and alarm information) to users (such as operation and maintenance engineers) in the form of visualizations or reports, and triggers alarms based on preset rules. The results display and alarm module 204 can provide a user-friendly human-computer interaction interface and automatically trigger alarms based on the severity of the diagnostic or prediction results, prompting the attention of operation and maintenance personnel. A knowledge base and model library are used to store pre-trained model parameters, standardized parameters, fault signature libraries, health indicator (HI) construction rules, and historical cases, and provide support for the iTransformer core processing module and task-specific output layer. The knowledge base and model library can persistently store and manage trained iTransformer models, related configuration parameters, and gas turbine domain knowledge (such as failure modes, HI assessment criteria, degradation models, etc.), supporting rapid model deployment, version management, knowledge accumulation, and continuous optimization of model effects.

[0033] Figure 2 The system data flow shown is as follows: the gas turbine sensor data is first input into the data acquisition and preprocessing module 201. The processed multivariate time series data is fed into the iTransformer core processing module 202. Inside the module 202, the data is first processed by the variable tokenization and variable token embedding layers, and then passes through one or more stacked inverted Transformer blocks 100 in sequence. Inside each inverted Transformer block (100), the data first passes through the MHSA layer 103, and its output (after residual connection and layer normalization) is then input into the FFN layer 104, and the output of the FFN (also after residual connection and layer normalization) is used as the output of the block. The feature representation of the output of the last inverted Transformer block is fed into the task-specific output layer 203 for final analysis. The analysis results are finally presented to the user by the result display and alarm module 204.

[0034] In a second aspect, the present invention further provides a method for evaluating the health of a gas turbine, which is used in the system according to the above aspect, and comprises the following steps: Data collection and input. After the gas turbine health assessment system is started, the data collection and preprocessing module 201 collects multivariate time series data within a specified time window from various sensors of the gas turbine (for example, sensors installed in key locations such as the compressor, combustion chamber, turbine, and bearings). The number of sensors is N. For example, N sensor readings are collected once per second over the past 24 hours.

[0035] Data preprocessing: Preprocessing the collected raw data. Data preprocessing includes cleaning (such as filling missing values ​​and smoothing noise), synchronization (ensuring that all sensor data are aligned in time), and normalization (for example, scaling the data to a similar numerical range).

[0036] The iTransformer model is processed. The preprocessed N sensor time series (each with a length of T) are fed into the iTransformer core processing module 202. First, the complete sequence of each sensor (length T) is regarded as a variable tag. Then, the variable tag embedding layer 102 embeds these N variable tags into N vectors of D_model dimensions. The inverted Transformer block learns features. These N embedded vectors pass through L stacked inverted Transformer blocks 100 in turn. In each block, the multi-head self-attention mechanism (MHSA) layer 103 calculates The self-attention between N variable labels is calculated to capture the mutual dependencies between different sensors, and the representation of each variable label is updated to incorporate information from other related variables. The feedforward network (FFN) layer 104 independently performs nonlinear transformation on each updated variable label representation to learn and refine the deep time series features of the sensor itself. After processing by L inverted Transformer blocks, the feature representations of N final variable labels are obtained. These representations can be further aggregated (for example, by global average pooling or selecting representations of specific variables related to the task) to form an overall feature vector for downstream tasks.

[0037] Task execution and result output is used to map the feature representation (feature vector) output by the iTransformer core processing module 202 to the final output result according to the application task.

[0038] Optionally, it is used to map the feature representation output by the iTransformer core processing module to the final output result according to the application task, specifically including: Fault diagnosis: The feature vector output by the iTransformer core processing module 202 is input into a pre-trained classifier (e.g., a Softmax classifier) ​​to determine whether the gas turbine is currently faulty and identify the specific fault type (e.g., bearing wear, blade cracks, unstable combustion, etc.). The output result is the fault category label and confidence level. State prediction: The feature vector output by the iTransformer core processing module 202 is input into the regression model to predict the values ​​of one or more key sensor parameters within a certain period of time (e.g., exhaust temperature within the next hour, vibration amplitude within the next 24 hours). The output result is a prediction curve or predicted value. Health assessment: The feature vector output by the iTransformer core processing module 202 is input into a health assessment model (which can be a regression model that outputs a health index from 0 to 100, or a rule-based or model-based scoring system) to quantitatively assess the current overall health of the gas turbine and output a health score or grade.

[0039] Optionally, the method also includes: result display and decision support, the gas turbine health assessment system displays the diagnosis results, prediction information, and health index to the operation and maintenance personnel through a dashboard, trend chart or report. If a serious fault is diagnosed, a parameter is predicted to exceed the limit, or the health index is lower than a preset threshold, the gas turbine health assessment system automatically triggers an alarm and notifies relevant personnel through SMS, email or system message; for fault diagnosis, the attention weights calculated between variables by the multi-head self-attention mechanism (MHSA) layer 103 are used to highlight the key sensors that contribute most to the current fault judgment, providing explainability support for fault location and cause analysis.

[0040] Through the above steps, the present invention can comprehensively, accurately and efficiently monitor and analyze the operating status of the gas turbine, thereby achieving early fault warning, assisting operation and maintenance decision-making, optimizing maintenance plans, and ultimately improving the reliability, availability and economy of the gas turbine.

[0041] In order to more specifically illustrate the technical solution of the present invention, it is further described in detail below with reference to the accompanying drawings and embodiments.

[0042] Example 1: Early fault diagnosis of gas turbine bearings based on iTransformer.

[0043] 1.1 Technical issues: Gas turbine bearing failures are common, and their early signs are often subtle and easily overwhelmed by background noise. This example aims to use the iTransformer model to analyze multi-element sensor data related to gas turbine bearings, enabling accurate diagnosis and early warning of early bearing failures (such as pitting, wear, and crack initiation).

[0044] 1.2 Detailed technical route: (1) E1-201: Data acquisition and preprocessing module (reference Figure 3 and Figure 4 ) Input: Raw high-frequency vibration sensor signals on the gas turbine bearing seat (X, Y, Z three-axis acceleration signals with a sampling frequency of 10kHz-50kHz), lubricating oil temperature, lubricating oil pressure, shaft speed, and optional gas turbine load and other raw sensor data streams.

[0045] Technical route and neural network module: 1. Data acquisition: Get raw data through data acquisition card or DCS / SCADA system.

[0046] 2. Data Cleaning: Missing Value Handling: For slowly varying signals such as temperature and pressure, linear interpolation or forward / backward filling is used. For high-frequency vibration signals, if there are small missing values, interpolation based on adjacent points or marking as unavailable segments can be considered. Outlier Handling: For slowly varying signals, statistical threshold-based methods (such as the 3-sigma rule) or boxplots are used to identify, correct, or eliminate outliers. Vibration signals require careful processing to avoid removing true impact signals.

[0047] 3. Signal processing (mainly for vibration signals): Denoising: Wavelet Threshold Denoising can be used to remove background noise.

[0048] Filtering (optional): Based on the bearing fault characteristic frequency (such as BPFI, BPFO, BSF, FTF), a bandpass filter (such as Butterworth or Chebyshev filter) can be used to enhance the signal in the fault-related frequency band.

[0049] Segmentation: The continuous high-frequency vibration signal is segmented according to a fixed length (for example, the number of data points corresponding to 1-10 rotations of the shaft, or a fixed time length such as 0.1 second), and each segment is a component of a time series sample.

[0050] 4. Feature engineering (optional, alternatively or in combination with using the original sequence directly): Time domain features: Calculate the root mean square (RMS), kurtosis, skewness, peak-to-peak value, margin factor, impulse factor, etc. for each vibration signal segment.

[0051] Frequency domain features: Perform fast Fourier transform (FFT) on each vibration signal segment to extract the amplitude and energy at the characteristic frequency (BPFI, BPFO, etc.), or the energy ratio within a specific frequency band.

[0052] Time-frequency domain features: Short-time Fourier transform (STFT), wavelet transform (WT, such as continuous wavelet transform CWT or discrete wavelet transform DWT) or wavelet packet energy spectrum can be used to extract features that can simultaneously reflect time and frequency information.

[0053] If feature engineering is performed, each raw sensor sequence (especially vibration) is converted into one or more feature sequences.

[0054] 5. Data synchronization and alignment: Ensure that all sensor data (or their extracted feature sequences) are strictly aligned in terms of timestamps.

[0055] 6. Data normalization / standardization: All sequences input to iTransformer (whether they are original processed sequences or feature sequences) are Z-score normalized (making their mean 0 and standard deviation 1) or Min-Max normalized (scaling to the range [0, 1] or [-1, 1]) to facilitate stable training of neural networks.

[0056] Output: Time series data after road preprocessing ( ),in is the uniform length of the time series (for example, the length of a vibration signal after segmentation, or the number of sampling points of other signals in the time window). The road sequence will be used as the input of iTransformer.

[0057] (2) E1-202: iTransformer core processing module (reference Figure 1 and Figure 3 ) Input: From module E1-201 Time series after road preprocessing , each Length is .

[0058] Technical route and neural network module: 1. Variable token embedding layer 102: Purpose: To convert the complete time series (length) of each sensor into ) are independently mapped into a feature vector of fixed dimension (D_model dimension) as the “variable label” representation of the sensor.

[0059] Neural Network Selection: For the original or lightly processed vibration signal sequence ( ): A one-dimensional convolutional neural network (1DCNN) is typically used. A 1D CNN consists of several convolutional layers (`Conv1D`, using activation functions such as ReLU), pooling layers (`MaxPooling1D` or `AvgPooling1D`), and finally a global average pooling layer (`GlobalAveragePooling1D`) or a flattening layer (`Flatten`) followed by a fully connected layer (`Dense`), outputting a D_model-dimensional vector. 1D CNNs are effective in automatically learning and extracting local patterns, periodicity, and impact features from vibration sequences.

[0060] For the extracted time domain / frequency domain / time-frequency domain feature sequence ( ), or for sensor signals with relatively slow changes (such as temperature and pressure sequences): a multi-layer perceptron (MLP) can be used to transform the length The sequence is flattened and fed into one or more fully connected layers (`Dense`, using activation functions such as ReLU), which ultimately output a D_model-dimensional vector. Alternatively, a shallower 1D CNN can be used.

[0061] Operation: Yes The input sequences are applied with independent embedding networks, and we get variable label embeddings with D_model dimensions .

[0062] 2. L stacked Transformer blocks 100: Each block contains MHSA and FFN.

[0063] Multi-Head Self-Attention Mechanism (MHSA) 103: Input: Output of the previous layer Variable markers indicate .

[0064] Operation: Query, Key, Value matrix are all generated by this The variable labels are obtained by linear transformation. The algorithm calculates the interdependencies and influence weights between different sensor signals (e.g., abnormal vibration in a specific direction and slight changes in lubricant temperature). The output is N updated variable label representations, each of which incorporates information from other related variables.

[0065] Neural network module: A standard multi-head self-attention module implementation, consisting of multiple parallel "Scaled Dot-Product Attention" heads.

[0066] Feedforward Network (FFN) 104: Input: MHSA output The updated variable marker is indicated.

[0067] Action: Apply a FFN independently to each variable tag representation.

[0068] Neural network module: usually consists of two fully connected layers (`Dense`), the first fully connected layer expands the dimension from D_model to (Usually D ff =4*D_model), using ReLU or GELU activation function, the second fully connected layer then converts the dimension from Mapped back to D_mode.

[0069] Residual connection and layer normalization (Add & Norm): Used at both the input and output of MHSA and FFN to stabilize training. Layer Normalization is performed on the labeled feature dimension of each variable.

[0070] Output: After processing by L Transformer blocks, we get N final, more informative D_model-dimensional variable label representations.

[0071] (3) E1-203: Task-specific output layer (bearing fault classifier) ​​(reference Figure 3 and Figure 4 ) Input: N final variable token representations from module E1-202.

[0072] Technical route and neural network module: 1. Feature aggregation (optional): Global average pooling: Average N D_model-dimensional vectors in the variable dimension to obtain an aggregated feature vector of D_model dimension.

[0073] Concatenation: Concatenate N D_model-dimensional vectors directly into a feature vector of N*D_model dimension (if N is not large).

[0074] Attention weighted aggregation: A simple attention layer can be introduced to learn The importance weights of the variables are marked and then weighted summed.

[0075] 2. Classification Header: Neural network module: This is typically an MLP consisting of one or more fully connected (Dense) layers. Activation functions such as ReLU and dropout layers may be used in between to prevent overfitting. The number of neurons in the final fully connected layer is equal to the number of predefined bearing fault categories ("healthy," "early pitting," "mid-term wear," "outer race fault," "inner race fault," "rolling element fault," and so on). The Softmax activation function is used to output the probability of each category.

[0076] Output: Probability distribution of bearing fault categories corresponding to input samples.

[0077] (4) E1-205: Knowledge Base / Model Base (reference Figure 3 and Figure 4 ) Storage content: pre-trained iTransformer model parameters (including embedding layers, Transformer blocks, classification head weights and biases), normalization parameters, fault feature library, attention maps of historical fault cases, etc.

[0078] (5) E1-206: Explainability Analysis Module (reference Figure 3 and Figure 4 ) Input: The attention weight matrix calculated by each MHSA layer in the E1-202 module ( ).

[0079] Route: Analyze and visualize these attention weights. For a specific diagnosed fault, you can see which sensor variables play a key role in "attention" to other variables or the final decision. If a sensor variable receives a high weight in multiple attention heads or has strong interactions with other sensors associated with known faults, it is considered a key indicator.

[0080] Output: A visual map or list of the key sensors (variables) and their interactions that contribute most to the current fault diagnosis results.

[0081] (6) E1-204: Result display and alarm module (reference Figure 3 and Figure 4 ) The result display and alarm module can provide visual diagnostic results, alarm information, etc., prompting the attention of operation and maintenance personnel.

[0082] Example 2: Gas turbine health index (HI) trend prediction and remaining useful life (RUL) assessment based on iTransformer.

[0083] 2.1 Technical Issues: Gas turbine performance gradually degrades with increasing operating time. Accurately predicting future trends in the health index (HI) or assessing the remaining useful life (RUL) of key components is crucial for developing condition-based maintenance (CBM) strategies and avoiding unplanned downtime. This embodiment utilizes iTransformer to process long-term multi-sensor data to accurately predict HI trends and quantitatively assess RUL.

[0084] 2.2 Detailed technical route: (1) E2-201: Data collection, preprocessing and HI / RUL tag construction module (reference Figure 5 and Figure 6 ) Input: Sensor data from the long-term operation (months to years) of the gas turbine (such as performance parameters such as temperature, pressure, flow, and efficiency of the compressor, combustion chamber, and turbine; status parameters such as vibration trend values ​​and oil analysis data), as well as historical maintenance records, component replacement information, and known failure events.

[0085] Technical route and neural network module: 1. Data collection: Get data from historical databases (such as PIMS) or CMMS.

[0086] 2. Health Index (HI) label construction: Purpose: Assign a quantitative health label to historical data samples (usually a continuous value of 0-1 or 0-100, where 1 or 100 represents perfect health and 0 represents complete failure).

[0087] Construction method (you can select or combine one or more of the following): Method A: HI construction based on rules or expert knowledge: Select several key performance parameters (KPIs), such as efficiency, power output deviation from design value, and the distance of specific component temperature from threshold value.

[0088] Set normal ranges and deterioration thresholds for each KPI.

[0089] The current status of multiple KPIs is integrated into a comprehensive HI value through weighted summation or fuzzy logic system. For example, ,in is the weight, It is a single score obtained by mapping the KPI value.

[0090] Method B: Data-driven unsupervised / semi-supervised HI construction: Methods based on reconstruction error: Use data from normal operating conditions to train an autoencoder (AE) or variational autoencoder (VAE). For new data points, the magnitude of the reconstruction error can reflect the degree of deviation from the normal state. HI can be defined as the inverse of the reconstruction error or the value after conversion by a nonlinear function (such as an exponential decay function). For example, .

[0091] Distance-based methods: Use normal data to build one or more cluster centers (such as using K-Means). The greater the distance between the new data point and the nearest normal cluster center, the lower the healthiness.

[0092] Principal Component Analysis (PCA)-based methods extract the main direction of data variation. Under normal conditions, data points should be stably distributed in a low-dimensional principal component space. Data points that deviate from this distribution are considered less healthy. The degree of deviation can be quantified using the Mahalanobis distance or T² statistic.

[0093] Method C: Data-driven supervised HI construction (if partial label information exists): Use known maintenance events (such as HI=100 after overhaul and HI close to a certain threshold before component replacement) or clear performance degradation curves (such as the component efficiency attenuation curve fitted based on historical data) as "anchor points" to mark the HI values ​​of some data points.

[0094] A regression model (such as support vector regression SVR, Gaussian process regression GPR, or a simple neural network) can then be trained to fit these (sensor data -> HI label) mappings and interpolate or predict HI for unlabeled data.

[0095] It is also possible to assume that the HI decays linearly or nonlinearly (e.g., exponentially) from the new state (HI=1) to the failure state (HI=0), and estimate the HI based on the operating time or number of cycles of the component.

[0096] Smoothing of HI labels: The original HI sequence may contain noise, which can be smoothed using a moving average or Savitzky-Golay filter.

[0097] 3. RUL tag construction: For components with a clear history of failure, RUL = failure time - current time. During training, the remaining runtime or number of cycles from a starting point to failure in historical data is typically used as the label.

[0098] 4. Data cleaning and synchronization: Handle missing values ​​(e.g., interpolation using the time series decomposition model STL, machine learning-based interpolation such as the KNN interpolator), outliers, and ensure that all sensor data are temporally aligned with HI / RUL labels.

[0099] 5. Sequence segmentation and sample construction: Split long time series data into segments with fixed input length (look-back window, such as the past 30 days of data) and target forecast length (Predicting HI trends for the next 7 days) . yes Multi-sensor data, The corresponding future HI sequence of steps or a single RUL value.

[0100] 6. Data normalization / standardization: Each sensor sequence in is Z-score normalized or Min-Max normalized.

[0101] Output: Preprocessed multivariate time series input samples and the corresponding target HI sequence or RUL value .

[0102] (2) E2-202: iTransformer core processing module (reference Figure 1 and Figure 5 ) Input: From module E2-201 Long-term sensor sequence fragments after preprocessing (Dimension ).

[0103] Technical route and neural network module: 1. Variable label embedding layer 102: Purpose: To convert each sensor’s long-term history sequence fragment (length ) are independently mapped to -dimensional feature vector.

[0104] Neural Network Selection: For performance parameter sequences with strong trends, an MLP (flattening the sequence before input) or a simpler 1D CNN (using a small number of convolutional layers to capture local trends followed by pooling) can be used. Alternatively, a more complex sequence encoder can be employed, using an RNN-based encoder (LSTM or GRU) followed by a fully connected layer that outputs a D_model-dimensional vector to capture longer-term temporal dependencies. However, to maintain the efficiency advantage of the iTransformer, the embedding layer is typically designed to be relatively lightweight.

[0105] Operation: Yes The input sequences are applied with independent embedding networks, and we get Variable token embeddings with D_model dimensions.

[0106] 2. L stacked Transformer blocks 100: (Same as the MHSA and FFN structure in Example 1, but in this scenario) MHSA 103: You will learn how different performance and condition parameters evolve over long-term degradation processes. For example, how a slow decrease in compressor efficiency can be correlated with a gradual increase in turbine exhaust temperature and a slight decrease in power output.

[0107] FFN 104: will learn the nonlinear dynamic patterns of each parameter itself during long-term degradation.

[0108] Output: After After processing by Transformer blocks, we get The final D_model dimensional variable label representations contain deep features for predicting HI / RUL.

[0109] (3) E2-203: Task-specific output layer (HI / RUL regressor) (Ref. Figure 5 and Figure 6 ) Input: From module E2-202 The final variable is marked.

[0110] Technical route and neural network module: 1. Feature aggregation (same as in Example 1, such as global average pooling).

[0111] 2. Return to the head: HI trend prediction: The aggregated D_model dimension feature vector is input into an MLP. The output dimension of the last layer of the MLP is (If you predict the future time steps), or the output dimension is The activation function is usually a linear activation function or selected according to the HI range (such as Sigmoid followed by scaling and translation to [0,100]).

[0112] RUL evaluation: The aggregated feature vector is fed into an MLP with the last layer having an output dimension of 1 (predicting a single RUL value) and using a linear activation function or ReLU (since RUL is non-negative).

[0113] Output: A sequence of predicted future HI values, or a current RUL estimate.

[0114] (4) E2-205: Knowledge Base / Model Library (reference Figure 5 and Figure 6 ) Storage content: pre-trained iTransformer model parameters, HI / RUL definition rules and construction models (such as pre-trained AE models), historical degradation curves, component life distribution data, maintenance thresholds, etc.

[0115] (5) E2-206: Explainability Analysis Module (reference Figure 5 and Figure 6 ) Input: Attention weights of MHSA in module E2-202.

[0116] Technical route: Analyze which sensor parameters (variables) contribute most to the current HI decline trend or RUL assessment results, or which interaction patterns between parameters are closely related to the degradation process.

[0117] Output: Visualization of key parameters and their interactions that affect HI / RUL predictions.

[0118] (6) E2-204: Results display and decision support module (reference Figure 5 and Figure 6 ) The result display and decision support module can provide visual HI / RUL predictive maintenance decision support to operation and maintenance managers.

[0119] The core of this invention lies in applying the unique architecture and processing mechanism of the iTransformer model to the field of gas turbine health assessment (including fault diagnosis and state prediction), thereby achieving more accurate, efficient, and explainable intelligent monitoring. Compared with existing technologies (such as traditional time-step-labeled Transformer models, ARIMA models, RNN / LSTM models, and traditional gas turbine monitoring methods), this invention (the iTransformer-based gas turbine health assessment method and system) has the following significant benefits: 1. Significantly Improves the Accuracy of Gas Turbine Fault Diagnosis and Prediction: Through its unique "inverted perspective" (using each sensor's complete time series as a "variable tag"), this invention more effectively captures the complex dependencies and independent dynamic patterns of each sequence in gas turbine multi-sensor data. Better Representation of Coupling Relationships between Multi-Sensors: The Multi-Head Self-Attention Mechanism (MHSA) directly computes correlations between N "sensor variable tags," thereby capturing more global and deeper dependencies between sensors (such as the coordinated changes or delayed responses of multiple parameters under specific fault modes), which is crucial for accurately identifying complex gas turbine faults. Accurate Learning of Independent Temporal Patterns for Each Sensor: A Feedforward Network (FFN) is independently applied to the representation of each "sensor variable tag" (after integrating multi-variable interaction information), enabling it to focus on learning the unique complex temporal dynamic patterns of each sensor (such as characteristic waveforms, trend drift, and periodic anomalies under specific fault conditions). This avoids the interference of mixed information in traditional time-step tags on single-sensor pattern learning. Quantitative results demonstrate this: Drawing on iTransformer's performance on common time series forecasting tasks, for example, compared to the classic LSTM model, the iTransformer model reduced the mean squared error (MSE) by approximately 24.78% and the mean absolute error (MAE) by approximately 14.17% on multivariate time series forecasting tasks. Applying this efficient architecture to gas turbine data is expected to significantly improve fault diagnosis accuracy and key parameter prediction precision.

[0120] 2. Improve the computational efficiency of processing massive and long time series data of gas turbines: Gas turbines usually generate a large amount of sensor data with a long time span. When the time series length T is much larger than the number of sensors N (`T >> N`), the present invention has obvious advantages in computational complexity and memory consumption. The traditional Transformer's self-attention complexity is mainly O( ). This invention applies the self-attention mechanism to the sensor dimension, and its core complexity is transformed into O( For gas turbines (usually where N is much smaller than T), Much smaller than This makes the present invention more efficient when processing extremely long time series. This allows the model to review a longer window of historical data to capture early weak fault signals or long-term degradation trends, overcoming the computational bottleneck of traditional Transformers when processing such data.

[0121] 3. More Effective Utilization and Understanding of Gas Turbine Multi-sensor Feature Information: Through the core innovation of "sensor variable tokenization," the complete time series of each sensor is treated as an independent unit, laying the foundation for directly modeling the interactions between sensors. This avoids the information loss and noise interference caused by the mixed embedding of sensor data with different physical meanings and dynamic characteristics in traditional methods. The application of MHSA at the sensor level enables the attention weight matrix to directly and clearly reflect the mutual influence strength and dependency relationships between different sensors, which helps to understand fault propagation paths or multi-parameter collaborative anomaly patterns.

[0122] 4. Enhanced learning of independent dynamic patterns and fault signatures for each gas turbine sensor: By applying the FFN independently to each "sensor variable tag" (after the MHSA captures the interactions between sensors), the model can more specifically and deeply learn and characterize the time series characteristics of that specific sensor, such as its unique baseline behavior, trends, periodicity, and characteristic manifestations under specific faults (for example, the change in the spectral signature of a vibration sensor under a bearing fault). This "interaction first, independent refinement later" strategy is more effective than the traditional FFN model, which operates on mixed information time step tags.

[0123] 5. Improving the interpretability and decision-making support of gas turbine fault diagnosis: This method calculates attention weights across N "sensor variable labels." The resulting attention map directly and intuitively reveals which sensor parameters (or specific combinations thereof) play a key role in assessing a specific fault mode or health status. This provides operators with a more intuitive and reliable diagnostic basis, pinpointing the sensors or components requiring particular attention, and enhancing the interpretability of model decisions and their guiding value in actual operations and maintenance.

[0124] 6. Potentially Improved Model Generalization and Robustness for Different Operating Conditions and Early-Stage, Minor Faults: By decoupling global learning of inter-sensor relationships (MHSA) and local learning of individual sensor internal patterns (FFN), this method potentially improves generalization of gas turbine data patterns under diverse operating conditions. This modular learning approach more readily adapts to situations where sensors have varying scales, units, or dynamic characteristics. The ability to efficiently utilize long historical data, combined with deep learning of subtle inter-sensor correlations and individual sensor patterns, makes this method more likely to capture early, subtle fault signals in their infancy, enabling more sensitive fault warnings.

[0125] 7. Significantly Enhanced Capabilities for Processing Extremely Long Historical Data of Gas Turbines to Support Long-Term Health Management: As discussed in the computational efficiency section, this invention shifts the primary computational bottleneck from the time series length T to the number of sensors N, effectively processing extremely long time series that traditional Transformers struggle to handle. This is crucial for analyzing long-term performance degradation trends of gas turbines, accurately predicting remaining useful life (RUL), and developing optimized condition-based maintenance (CBM) strategies. It overcomes the limitations of the ARIMA model in modeling complex nonlinear dependencies, as well as the gradient and memory bottlenecks of RNNs / LSTMs for extremely long time series.

[0126] In summary, by applying the innovative architecture of iTransformer to the field of gas turbines, the present invention can bring significant beneficial effects in fault diagnosis and prediction accuracy, massive long time series data processing efficiency, multivariate feature utilization and understanding, model interpretability, and early weak fault detection capabilities, providing strong technical support for achieving smarter, more reliable, and more economical operation and maintenance of gas turbines.

[0127] The above embodiments are only for illustrating the technical concept and features of the present invention. Their purpose is to enable ordinary technicians in this field to understand the content of the present invention and implement it. They are not intended to limit the scope of protection of the present invention. Any equivalent changes or modifications made according to the spirit of the present invention should be included in the scope of protection of the present invention.

Claims

1. A gas turbine health assessment system, characterized in that: The system comprises: Data acquisition and preprocessing module, used to obtain operating data from multiple sensors of the gas turbine and perform data preprocessing; The iTransformer core processing module is used to perform deep feature learning and pattern recognition on multivariate time series data preprocessed by the data acquisition and preprocessing module. The iTransformer core processing module includes an input transposition unit, a variable label embedding layer, and a stacked Transformer block. The input transposition unit is used to convert time series data of each length T from multiple different sensors into multiple independent variable labels. The variable label embedding layer is used to embed each variable label, mapping each variable label from the original sequence of length T to a vector representation of a preset fixed dimension. The stacked Transformer block is composed of multiple inverted Transformer blocks stacked together, each of which includes a multi-head self-attention mechanism layer and a feedforward network layer. The task-specific output layer is used to map the feature representation output by the iTransformer core processing module to the final output result according to the application task.

2. The gas turbine health assessment system according to claim 1, characterized in that: The system further comprises: The result display and alarm module is used to present the output results of the task-specific output layer to the user in the form of visualization or reports, and trigger alarms according to preset rules; The knowledge base and model library are used to store pre-trained model parameters, standardized parameters, fault feature library, HI construction rules, historical cases, and provide support for the iTransformer core processing module and task-specific output layer.

3. The gas turbine health assessment system according to claim 1, wherein: The multi-head self-attention mechanism layer acts on the embedded representation of multiple variable labels to capture the interdependencies and correlation weights between different sensor data sequences; the feedforward network layer is independently applied to the representation of each variable label output by the multi-head self-attention mechanism layer, and performs a nonlinear transformation on the representation of each sensor that has been fused with the information of other related sensors to learn the time series pattern and deep features of the sensor itself.

4. The gas turbine health assessment system according to claim 3, characterized in that: The variable label representation input to the inverted Transformer block first passes through a layer of normalization, then enters the multi-head self-attention mechanism layer. The output of the multi-head self-attention mechanism layer passes through a residual connection, then passes through a layer of normalization, and then enters the feedforward network layer. The output of the feedforward network layer passes through a residual connection, then passes through a layer of normalization, as the final output of the inverted Transformer block.

5. The gas turbine health assessment system according to claim 1, wherein: The plurality of sensors include a temperature sensor, a pressure sensor, a vibration sensor, and a rotation speed sensor.

6. A gas turbine health assessment method, characterized in that: The method is used in a system according to any one of claims 1 to 5, and the method comprises the following steps: After the gas turbine health assessment system is started, the data acquisition and preprocessing module collects multivariate time series data within a specified time window from each sensor of the gas turbine, where the number of sensors is N; Preprocess the collected raw data, including cleaning, synchronization, and normalization operations; The iTransformer model is processed. The preprocessed N sensor time series are fed into the iTransformer core processing module. The complete sequence of each sensor is regarded as a variable label. Then, the variable label embedding layer embeds these N variable labels into N D_model-dimensional vectors. The inverted Transformer block learns features. These N embedded vectors pass through L stacked inverted Transformer blocks in sequence. In each block, the multi-head self-attention mechanism layer calculates the self-attention between the N variable labels, captures the mutual dependence between different sensors, and updates the representation of each variable label to incorporate information from other related variables. The feedforward network layer independently performs nonlinear transformation on each updated variable label representation to learn and refine the deep time series features of the sensor itself. After processing by L inverted Transformer blocks, the feature representation of the N final variable labels is obtained. According to the application task, the feature representation output by the iTransformer core processing module is mapped to the final output result.

7. The gas turbine health assessment method according to claim 6, characterized in that: Based on the application task, the feature representation output by the iTransformer core processing module is mapped to the final output result, specifically including: The feature vector output by the iTransformer core processing module is input into a pre-trained classifier to determine whether the gas turbine is currently faulty and identify the specific fault type. The output result is the fault category label and confidence level. The feature vector output by the iTransformer core processing module is input into the regression model to predict the value of one or more key sensor parameters in the future. The output result is a prediction curve or predicted value. The feature vector output by the iTransformer core processing module is input into the health assessment model to quantitatively assess the current overall health status of the gas turbine, and the output result is a health score or grade.

8. The gas turbine health assessment method according to claim 7, characterized in that: The method also includes: result display and decision support, which displays the diagnosis results, prediction information, and health index to operation and maintenance personnel through dashboards, trend charts, or reports. If a serious fault is diagnosed, a parameter is predicted to exceed a limit, or the health index is lower than a preset threshold, the gas turbine health assessment system automatically triggers an alarm and notifies relevant personnel via SMS, email, or system message. For fault diagnosis, the attention weights calculated between variables by the multi-head self-attention mechanism layer are used to highlight the key sensors that contribute most to the current fault judgment.

Images