Gas turbine exhaust temperature sensor fault diagnosis system and method

Abstract

The invention provides a gas turbine exhaust temperature sensor fault diagnosis system and method, and relates to the technical field of industrial equipment state monitoring and diagnosis. The system comprises a sensor signal acquisition module for acquiring an original temperature signal; the signal preprocessing module is used for acquiring and preprocessing original signals and working condition parameters; the feature extraction and windowing module is used for acquiring data and calculating features; the data driving diagnosis module is used for receiving the time sequence characteristics and evaluating the health state; and the decision fusion module is used for comprehensively analyzing the multi-source information and making a final fault diagnosis judgment. According to the system, the accuracy and the reliability of fault diagnosis can be remarkably improved, the false alarm rate and the missing report rate are reduced, early warning and accurate identification of early weak faults of the sensor are realized, the adaptability and the robustness of the diagnosis system to variable working conditions of the gas turbine are improved, and the safety, the economical efficiency and the operation and maintenance efficiency of operation of the gas turbine are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial equipment status monitoring and diagnosis, and in particular to a gas turbine exhaust temperature sensor fault diagnosis system and method. Background Art

[0002] Gas turbines are core power equipment in power generation and drive systems. Their exhaust gas temperature (EGT) is a key parameter for monitoring combustion status, operating performance, and safety. EGT sensors (typically thermocouples) operate in harsh environments such as high temperature, high pressure, and vibration over a long period of time, making them susceptible to various faults such as drift, sticking, increased noise, open circuits, and short circuits. Failure to promptly and accurately diagnose these faults can lead to misjudgments in the control system, resulting in turbine overheating, reduced efficiency, unplanned downtime, and even major safety incidents.

[0003] Existing methods for sensor fault diagnosis primarily include threshold-based methods, signal analysis-based methods (such as Fourier transform and wavelet analysis), statistical methods (such as principal component analysis (PCA) and independent component analysis (ICA), and traditional machine learning methods (such as support vector machines (SVMs), decision trees, and artificial neural networks (ANNs). Threshold-based methods are simple but insensitive to early, weak faults and prone to false alarms due to changes in operating conditions. Signal analysis and statistical methods can extract fault features to a certain extent, but their ability to process complex, nonlinear, and non-stationary sensor signals is limited, and feature extraction often relies on expert experience. While traditional machine learning methods have seen some improvements, they still struggle with handling long-term dependencies, capturing subtle dynamic changes, and coping with data imbalance. Furthermore, their model generalization and adaptability to unknown operating conditions need to be improved. In recent years, deep learning has demonstrated great potential in time series analysis. For example, recurrent neural networks (RNNs) and their variants, long short-term memory networks (LSTMs) and gated recurrent units (GRUs), have been applied to fault diagnosis. However, they still suffer from vanishing / exploding gradients and low computational efficiency when processing extremely long sequences. The Transformer model and its variants (such as Informer) have achieved breakthroughs in processing long sequence dependencies through the self-attention mechanism. However, in actual industrial applications, directly applying the original Informer model may still face problems such as insufficient capture of local features and high computational complexity, and insufficient adaptability optimization for specific industrial objects (such as EGT sensors).

[0004] Therefore, how to effectively utilize the advantages of deep learning to develop a diagnostic system and method that can accurately and robustly diagnose various failure modes of gas turbine EGT sensors, especially one that can achieve early warning and adapt to complex and changeable operating conditions, is a technical problem that needs to be solved urgently. Summary of the Invention

[0005] The object of the present invention is to provide a gas turbine exhaust temperature sensor fault diagnosis system and method to solve the gas turbine exhaust temperature sensor fault diagnosis problem in view of the above-mentioned deficiencies in the prior art.

[0006] 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 exhaust temperature sensor fault diagnosis system, the system comprising: A sensor signal acquisition module is used to establish an interface with the exhaust temperature sensor array of the gas turbine and the gas turbine control system or data acquisition system, and to collect the original temperature signal of the exhaust temperature sensor in real time or periodically, and to synchronously obtain relevant gas turbine operating parameters; A signal preprocessing module is used to obtain the original temperature signal and gas turbine operating parameters from the sensor signal acquisition module and perform preprocessing operations; The feature extraction and windowing module is used to obtain the data after the preprocessing operation of the signal preprocessing module, cut the data into a series of multidimensional time series segments according to the preset window length L_win and sliding step size S_step, record the multidimensional time series segments as windows, and calculate the statistical features or frequency domain features within each window; The data-driven diagnosis module integrates one or more pre-trained improved informer models. This module receives windowed multidimensional time series features from the feature extraction and windowing module. Using the deep learning capabilities of the improved informer model, it assesses the health status of the exhaust temperature sensor within the current time window and outputs preliminary diagnostic results. The decision fusion module is used to obtain the preliminary diagnostic results from the data-driven diagnosis module and the signals of other preset independent diagnostic channels, and to perform a comprehensive analysis of the multi-source information obtained, and to make a final fault diagnosis judgment based on the preset fusion logic. The output of the decision fusion module is the final fault diagnosis result, which includes the fault type, the time when the fault occurred, and the confidence level of the diagnosis. The preset other independent diagnostic channels include an alarm system based on threshold comparison and a judgment system based on an expert rule base.

[0007] Optionally, the gas turbine operating parameters include turbine load, compressor outlet pressure, ambient temperature and fuel flow rate.

[0008] Optionally, the preprocessing operations specifically include: performing filtering operations to remove high-frequency noise and power frequency interference; performing normalization or standardization operations; performing resampling operations to unify data with different sampling rates to a fixed time interval according to actual needs; performing data cleaning operations to interpolate short-term missing data or remove bad pixels.

[0009] Optionally, the system further comprises: The physical model residual calculation module is used to predict the theoretical output value of the exhaust temperature sensor at the current moment based on the current gas turbine operating parameters and / or historical exhaust temperature sensor data from the signal preprocessing module or the feature extraction and windowing module. Then, the physical model residual calculation module calculates the residual between the actually measured exhaust temperature sensor data from the signal preprocessing module and the predicted theoretical output value, and inputs the residual as an additional feature into the data-driven diagnosis module, or transmits the statistical characteristics of the residual to the decision fusion module.

[0010] Optionally, the system further comprises: The alarm output module is used to obtain the fault diagnosis results from the decision fusion module and, when a confirmed fault is detected, generate a corresponding alarm signal and send it to the upper system through a predetermined interface, or directly notify the relevant operation and maintenance personnel via SMS or email; The human-machine interface module is used to display all relevant information of the system, including real-time data, diagnostic results and alarm information on the system.

[0011] Optionally, the improved Informer model includes: Input embedding layer, which consists of a fully connected layer or a one-dimensional convolutional layer with superimposed position encoding; Improved Informer encoder, which consists of multiple identical encoder sublayers, each of which contains a ProbSparse multi-head self-attention module; Classification head,The classification head includes a pooling layer and a fully connected classification layer.

[0012] In a second aspect, the present invention further provides a method for diagnosing a gas turbine exhaust temperature sensor fault, which is applied to the system according to the first aspect, and the method comprises the following steps: The data-driven diagnosis module receives a windowed multidimensional time series feature X_win from the feature extraction and windowing module; The input embedding layer converts the received X_win into the embedded representation X_emb inside the model; The improved Informer encoder performs complex spatiotemporal feature extraction on X_emb and outputs a deep feature representation H_enc; Convert H_enc to a fixed-length vector h_final; The classification head maps h_final to the probability P_fault of each predefined fault category; The calculated probability vector P_fault is output to the decision fusion module for subsequent comprehensive judgment.

[0013] In a third aspect, the present invention further provides a method for diagnosing a gas turbine exhaust temperature sensor fault, which is applied to the system according to the first aspect, and the method comprises the following steps: S301: The data-driven diagnosis module receives a windowed multidimensional time series feature X_win from the feature extraction and windowing module; S302: The improved Informer anomaly detection model reconstructs or predicts the input feature sequence X_win to obtain an output sequence X_rec; S303: The anomaly scoring unit inside the module calculates the reconstruction error between X_win and X_rec to obtain AnomalyScore; S304: The system compares AnomalyScore with the preset anomaly threshold Th_anomaly. If AnomalyScore is less than or equal to Th_anomaly, it is determined to be normal and the process proceeds to step S308. However, if AnomalyScore is greater than Th_anomaly, it is determined to be abnormal and the process proceeds to step S305. S305: The system first outputs a preliminary abnormal signal. At the same time, the feature construction unit is activated, which constructs a feature vector f_clf for subsequent classification based on the window X_win currently judged to be abnormal. S306: The pre-trained lightweight fault classifier receives the feature vector f_clf, identifies the specific fault type, and outputs the probability P_fault_specific of each fault category; S307: The probability vector P_fault_specific containing the specific fault category information is output to the decision fusion module for it to make the final comprehensive decision; S308: Outputting a "normal" state or an abnormality score less than a preset value, the processing flow of the current window ends.

[0014] The beneficial effects of the present invention include: The gas turbine exhaust temperature sensor fault diagnosis system provided by the present invention includes: a sensor signal acquisition module, which is used to establish an interface with the exhaust temperature sensor array of the gas turbine and the gas turbine control system or data acquisition system, and to collect the original temperature signal of the exhaust temperature sensor in real time or periodically, and synchronously obtain related gas turbine operating parameters; a signal preprocessing module, which is used to obtain the original temperature signal and gas turbine operating parameters from the sensor signal acquisition module and perform preprocessing operations; a feature extraction and windowing module, which is used to obtain data after preprocessing operations by the signal preprocessing module, cut the data into a series of multidimensional time series segments according to a preset window length L_win and sliding step size S_step, record the multidimensional time series segments as windows, and calculate the statistical features or frequency domain features in each window; a data-driven diagnosis module, which internally integrates One or more pre-trained improved Informer models are formed, which are used to receive windowed multidimensional time series features from the feature extraction and windowing module, and then evaluate the health status of the exhaust temperature sensor in the current time window through the deep learning ability of the improved Informer model, and output a preliminary diagnosis result; a decision fusion module is used to obtain the preliminary diagnosis results from the data-driven diagnosis module and the signals of other preset independent diagnosis channels, and to conduct a comprehensive analysis of the obtained multi-source information, and to make a final fault diagnosis judgment according to the preset fusion logic. The output of the decision fusion module is the final fault diagnosis result, which includes the fault type, the time when the fault occurred, and the confidence of the diagnosis. The preset other independent diagnosis channels include an alarm system based on threshold comparison and a judgment system based on an expert rule base. This system can significantly improve the accuracy and reliability of fault diagnosis, effectively reduce the false alarm rate and missed alarm rate, realize early warning and accurate identification of early weak sensor faults, enhance the coverage and identification accuracy of various complex fault modes, improve the adaptability and robustness of the diagnostic system to the variable operating conditions of gas turbines, and improve the safety, economy and operation and maintenance efficiency of gas turbines. The system design has good flexibility and scalability. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] 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.

[0016] Figure 1 A schematic structural diagram of a gas turbine exhaust temperature sensor fault diagnosis system according to an embodiment of the present invention is shown; Figure 2A schematic diagram showing the internal logical structure of a data-driven diagnosis module provided by the first embodiment of the present invention is shown; Figure 3 A schematic diagram showing the internal workflow of the data-driven diagnosis module provided by the first embodiment of the present invention is shown; Figure 4 A schematic diagram showing the internal logical structure of a data-driven diagnosis module provided by a second embodiment of the present invention is shown; Figure 5 A schematic diagram of the internal workflow of a data-driven diagnosis module provided by the second embodiment of the present invention is shown. DETAILED DESCRIPTION

[0017] 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.

[0018] Existing technologies for gas turbine EGT sensor fault diagnosis have the following major shortcomings: 1. Inadequate ability to diagnose early, subtle faults, resulting in a high false-positive rate: Traditional methods based on thresholds or simple statistical features struggle to effectively identify subtle degradation or drift in sensor performance in the early stages. They often only detect faults after they have become more severe, missing the optimal opportunity for early maintenance and intervention. 2. Low accuracy in recognizing complex, nonlinear fault patterns, resulting in a high false-positive rate: EGT sensor signals are influenced by multiple complex factors, including combustion dynamics, operating conditions, and environmental factors, exhibiting highly nonlinear and non-stationary characteristics. Existing methods have limited capabilities in feature extraction and pattern recognition, making it difficult to accurately distinguish between true faults and normal operating fluctuations. This is especially true when multiple fault signatures are similar or overlap, leading to false positives or misidentification of one fault as another. 3. Lack of adaptability and robustness to the variable operating conditions of gas turbines: Gas turbines operate under diverse operating conditions, such as startup and shutdown, load fluctuations, and varying ambient temperatures, causing the normal output characteristics of EGT sensors to vary accordingly. Many fault models built using existing diagnostic methods are sensitive to changes in operating conditions. When operating conditions change significantly, diagnostic performance degrades significantly, leading to false positives or missed negatives. 4. Fault feature extraction relies on expert experience, limiting generalization: Some methods based on signal analysis or statistics require manual design and selection of fault features, which is not only time-consuming and labor-intensive, but the extracted features may not be universally applicable, resulting in poor generalization to new operating conditions or unseen fault modes outside the training data. 5. Traditional machine learning models have limited ability to handle long-term dependencies and capture subtle dynamic changes: The onset and evolution of EGT sensor faults is often a cumulative process, and their characteristics may be embedded in data spanning long time periods. Traditional machine learning models (including some shallow neural networks) have limitations in effectively learning long-range temporal dependencies. 6. Fault samples are scarce in industrial data, and class imbalance is difficult to effectively address: In real industrial applications, EGT sensors have far more normal operating data than fault data, and the amount of data for different types of faults can also vary significantly. This results in trained models being biased towards the majority class (normal state), resulting in poor recognition performance for minority class faults.

[0019] Therefore, how to effectively utilize the advantages of deep learning to develop a diagnostic system and method that can accurately and robustly diagnose various failure modes of gas turbine EGT sensors, especially one that can achieve early warning and adapt to complex and changeable operating conditions, is a technical problem that needs to be solved urgently.

[0020] Example 1: A gas turbine exhaust temperature sensor fault diagnosis system based on an improved Informer model.

[0021] This embodiment discloses a gas turbine exhaust temperature sensor fault diagnosis system. The purpose of this system is to use advanced data-driven technology and optionally combine physical model prior knowledge to achieve early and accurate detection and classification of various possible failure modes of gas turbine exhaust temperature (EGT) sensors. The overall structure of the system is shown in the figure below. Figure 1 shown.

[0022] The gas turbine exhaust temperature sensor fault diagnosis system provided in this embodiment includes: a sensor signal acquisition module 10, a signal preprocessing module 20, a feature extraction and windowing module 30, a data-driven diagnosis module 50, and a decision fusion module 60.

[0023] The sensor signal acquisition module 10 is used to establish an interface with the gas turbine's exhaust temperature sensor array and the gas turbine control system or data acquisition system. It collects raw temperature signals from the exhaust temperature sensors in real time or periodically, and simultaneously obtains relevant gas turbine operating parameters. Specifically, the sensor signal acquisition module 10 is responsible for establishing an interface with the gas turbine's exhaust temperature sensor array (e.g., multiple K-type or N-type thermocouples, or platinum resistance thermocouples (RTDs)) and the gas turbine control system or data acquisition system. Through these interfaces, the module 10 can collect raw temperature readings from each exhaust temperature sensor in real time or periodically, and simultaneously obtain relevant gas turbine operating parameters, such as gas turbine load L, compressor outlet pressure P2, ambient temperature Tamb, fuel flow rate Ff, etc. This data can be analog signals requiring A / D conversion or digital signals directly acquired via a fieldbus protocol (such as HART, Modbus, OPC UA, etc.). The collected raw data is then transmitted to the signal preprocessing module 20.

[0024] The signal preprocessing module 20 is used to obtain the raw temperature signal and gas turbine operating parameters from the sensor signal acquisition module and perform preprocessing operations. Specifically, the signal preprocessing module 20 is a software algorithm module that integrates a digital filter (such as a median filter, Butterworth low-pass filter, or Kalman filter), a normalization / standardization unit, a resampling unit, and a data cleaning unit. The latter is used to handle missing values ​​or obvious outliers in the data. The signal preprocessing module 20 performs a series of necessary preprocessing operations on the raw exhaust temperature signal and operating parameters output by the sensor signal acquisition module 10. Specifically, it performs filtering to remove high-frequency noise and power frequency interference; performs normalization or standardization, such as using minimum-maximum normalization (X_norm = (X - X_min) / (X_max - X_min) or Z-score standardization (X_std = (X - μ) / σ), where μ is the mean and σ is the standard deviation. This unifies exhaust temperature data and operating parameters of different dimensions and numerical ranges into similar numerical ranges, which is crucial for the stability of subsequent model processing; performs resampling, unifying data with different sampling rates to a fixed time interval, such as one data point per second or every five seconds, as required; and performs data cleaning, interpolating short-term missing data (e.g., linear interpolation or spline interpolation) or removing bad points. It is worth noting that data augmentation strategies (details of which will be detailed in the subsequent description of the data-driven diagnosis module) can also be logically performed at this stage or later to expand the training sample, especially for sparse fault samples. The preprocessed signal flows to the feature extraction and windowing module 30.

[0025] The feature extraction and windowing module 30 is used to obtain data preprocessed by the signal preprocessing module and, according to a preset window length L_win and sliding step size S_step, segment the data into a series of multidimensional time series segments, each of which is referred to as a window. The module then calculates the statistical features or frequency domain features within each window. The feature extraction and windowing module 30 is also a software algorithm module, internally comprising a time window generation unit and a feature calculation unit. Its core function is to segment the continuous time series data output by the signal preprocessing module 20 into a series of multidimensional time series segments, or windows, that may or may not overlap, according to a preset window length L_win (e.g., 60 sampling points, corresponding to the past 5 minutes of data) and sliding step size S_step (e.g., 1 sampling point or 10 sampling points). Each such window contains L_win time steps, and at each time step, the feature vector x_t has a dimension of d_feat. This feature vector x_t can be directly composed of preprocessed exhaust temperature sensor readings (such as the reading of the currently monitored sensor, readings of other sensors in the same group, or the average or weighted average of multiple sensor readings) and relevant real-time operating parameters. Furthermore, as an optional enhancement, the module can also calculate supplementary features, such as window-wide statistical features (such as mean, variance, rate of change, kurtosis, and peaks) or frequency-domain features (such as energy spectrum features extracted via short-time Fourier transforms (STFTs) or wavelet transforms). These supplementary features can be added to the feature vector x_t at each time step or used as additional global features for the entire window.

[0026] The data-driven diagnosis module 50 integrates one or more pre-trained improved Informer models. This module receives windowed multi-dimensional time series features from the feature extraction and windowing module 30 and, leveraging the deep learning capabilities of the improved Informer models, evaluates the health status of the exhaust temperature sensor within the current time window and outputs a preliminary diagnostic result. Specifically, the data-driven diagnosis module 50 is the core module of the system. It integrates one or more pre-trained "improved Informer models with separable attention convolution." The detailed internal architecture, data augmentation methods, and training process of this model will be described in more detail in subsequent embodiments, in conjunction with specific application scenarios (e.g., direct fault classification or anomaly detection). Its general function is to receive windowed multi-dimensional time series features from the feature extraction and windowing module 30 (which may also include residual features from the optional physical model residual calculation module 40), leverage the powerful deep learning capabilities of the improved Informer models to perform a detailed evaluation of the health status of the exhaust temperature sensor within the current time window, and output a preliminary diagnostic result. Depending on the specific embodiment configuration, its output can be in the form of a probability distribution of each predefined fault category, or an anomaly score that quantifies the degree to which the current state deviates from the normal mode. The preliminary diagnosis results are then sent to the decision fusion module 60. The improved Informer model includes: an input embedding layer, which is composed of a fully connected layer or a one-dimensional convolutional layer and superimposed with a positional encoding; an improved Informer encoder, which is composed of multiple layers of identical encoder sublayers, each of which contains a ProbSparse multi-head self-attention module; and a classification head, which includes a pooling layer and a fully connected classification layer.

[0027] The decision fusion module 60 is used to obtain the preliminary diagnostic results from the data-driven diagnosis module and signals from other pre-defined independent diagnostic channels. It then performs a comprehensive analysis of the acquired multi-source information and makes a final fault diagnosis decision based on pre-defined fusion logic. The output of the decision fusion module is the final fault diagnosis result, which includes the fault type, fault occurrence time, and diagnosis confidence level. The pre-defined independent diagnostic channels include an alarm system based on threshold comparison and a judgment system based on an expert rule base. Specifically, the decision fusion module 60 is a software algorithm module that may include a weighted averaging unit, a logic rule engine, a Bayesian reasoning unit, or a more complex fusion unit based on DS evidence theory. This module's function is to receive and integrate the preliminary diagnostic results (e.g., fault category probability or anomaly score) output by the data-driven diagnosis module 50. Furthermore, if the physical model residual calculation module 40 is enabled, it also receives residual information or physical model-based alarm signals output by that module. Furthermore, this module can selectively receive signals from other independent diagnostic channels (e.g., an alarm system based on simple threshold comparison and a judgment system based on an expert rule base). The decision fusion module 60 comprehensively analyzes this multi-source information and makes a final fault diagnosis based on pre-set fusion logic (for example, if the data-driven model is highly confident in a particular fault type and the physical model residuals also significantly exceed limits, the fault is confirmed with high confidence; or if the data-driven model output is abnormal, but the physical model can reasonably explain the fluctuation as caused by a dramatic change in operating conditions, the alarm level may be lowered or the system may be deemed normal; a voting mechanism or a confidence-based weighted fusion strategy may also be used). Its output is the final fault diagnosis result, which typically includes key information such as the fault type, the time of occurrence, and the confidence level of the diagnosis.

[0028] Gas turbine operating parameters include turbine load, compressor outlet pressure, ambient temperature, and fuel flow. Preprocessing operations can optionally include filtering to remove high-frequency noise and power frequency interference; normalization or standardization; resampling to unify data with different sampling rates to a fixed time interval based on actual needs; and data cleaning to interpolate short-term missing data or remove bad pixels.

[0029] The system also includes a physical model residual calculation module 40, which is configured to predict the theoretical output value of the exhaust temperature sensor at the current moment based on the current gas turbine operating parameters and / or historical exhaust temperature sensor data from the signal preprocessing module or the feature extraction and windowing module. The physical model residual calculation module then calculates the residual between the actual measured exhaust temperature sensor data from the signal preprocessing module and the predicted theoretical output value, and inputs the residual as an additional feature into the data-driven diagnosis module, or transmits the statistical characteristics of the residual to the decision fusion module. Specifically, the physical model residual calculation module 40 integrates a simplified thermal response model of the gas turbine exhaust temperature sensor, or an empirical or semi-empirical physical model describing the exhaust temperature of the gas turbine. This model can be implemented using transfer functions, state-space equations, or a lookup table. Its function is to predict the theoretical output value of the exhaust temperature sensor at the current moment, or to provide an expected normal operating range, based on the current gas turbine operating parameters (which can be provided by the feature extraction and windowing module 30 or directly obtained from the signal preprocessing module 20) and / or historical exhaust temperature data. The module then calculates the residual between the actual measured exhaust temperature (derived from the signal preprocessing module 20) and the theoretical exhaust temperature predicted by the module itself, i.e., e_phy = T_measured - T_predicted. This calculated residual signal, e_phy, can be input as an additional feature into the subsequent data-driven diagnosis module 50, or its statistical characteristics (such as the residual mean, variance, or whether it exceeds preset limits) can be directly used by the decision fusion module 60. The advantage of introducing a physical model residual calculation module is that the physical model can, to a certain extent, explain normal exhaust temperature fluctuations caused by known operating conditions, thereby allowing the residual signal to more significantly highlight abnormal conditions caused by sensor failures or unknown external disturbances.

[0030] Optionally, the system also includes an alarm output module 70 and a human-machine interface module 80. The alarm output module 70 is used to obtain the fault diagnosis results from the decision fusion module and, upon detecting a confirmed fault, generates a corresponding alarm signal. This signal is sent to a higher-level system via a predetermined interface, or the relevant operations and maintenance personnel are notified directly via text message or email. Specifically, the alarm output module 70 typically includes an interface module with a higher-level monitoring system (such as a distributed control system (DCS) or supervisory control and data acquisition (SCADA) system), or a message notification module (such as a text message gateway interface or an email server interface). Based on the final diagnostic judgment output by the decision fusion module 60, the alarm output module generates a corresponding alarm signal upon detecting a confirmed fault. This signal is sent to a higher-level system via a predetermined interface, or the relevant operations and maintenance personnel are notified directly via other means (such as text message or email). The alarm signal may include detailed information such as the faulty sensor number, the identified fault type, the exact time of fault occurrence, and recommended maintenance measures. The human-machine interface module 80 is used to display all relevant system information, including real-time data, diagnostic results, and alarm information. Specifically, all relevant information of the system, including real-time data, diagnostic results and alarms, will be displayed on the human-machine interface (HMI) module 80. This module is a graphical user interface (GUI) software that can be implemented based on Web technology, or it can be a desktop application, or it can be integrated into the existing plant monitoring platform. Its main functions are to display the temperature curve of each exhaust temperature sensor and the relevant gas turbine operating parameters in real time; to display historical data trends, fault diagnosis results and alarm records; and, depending on the specific implementation, it can also provide advanced functions such as model parameter configuration, diagnostic threshold adjustment, diagnostic report generation and export. In some advanced implementations, the human-machine interface module can even visualize some status information within the improved Informer model, such as the distribution diagram of attention weights, which helps operation and maintenance personnel understand the specific basis for the model to make a certain diagnostic judgment.

[0031] Regarding the connection relationship between the modules of the system, as mentioned above Figure 1As shown, the raw data collected by the sensor signal acquisition module 10 is transmitted to the signal preprocessing module 20. The data processed by the signal preprocessing module 20 is transmitted to the feature extraction and windowing module 30. The windowed feature sequence generated by the feature extraction and windowing module 30 is primarily transmitted to the data-driven diagnosis module 50. At the same time, some raw or preprocessed signals or features may also be transmitted to the physical model residual calculation module 40 (if this optional module is enabled). The residual information calculated by the physical model residual calculation module 40 (if enabled) can be transmitted to the data-driven diagnosis module 50 as additional input features or directly to the decision fusion module 60. The preliminary diagnosis result output by the data-driven diagnosis module 50 is transmitted to the decision fusion module 60. After integrating various information, the decision fusion module 60 transmits the final diagnosis decision to the alarm output module 70 and the human-machine interface module 80. Based on this diagnosis decision, the alarm output module 70 sends an alarm message to the upper system or operation and maintenance personnel. The human-machine interface module 80 comprehensively displays all relevant information to the user and accepts possible user commands, such as data query and parameter configuration.

[0032] The gas turbine exhaust temperature sensor fault diagnosis system disclosed in this embodiment generally operates in the following phases. First, the system initiates the initialization phase, loading all pre-trained model parameters (including parameters of the data-driven diagnosis model (i.e., the improved informer model) and physical model parameters (if enabled), decision fusion rules, and key configuration information such as alarm thresholds. The human-machine interface module 80 will display the system in standby mode. Subsequently, the system enters the online monitoring and diagnosis cycle. This phase proceeds sequentially through a series of steps: First, data acquisition: the sensor signal acquisition module 10 obtains the latest exhaust temperature sensor readings and related operating parameters from the gas turbine at a preset period (e.g., once per second). Next, signal preprocessing occurs: the signal preprocessing module 20 performs filtering, normalization, resampling, and data cleaning on the collected raw data. Next, feature extraction and windowing occur: the feature extraction and windowing module 30 constructs the preprocessed data stream into a multidimensional time series window of fixed length L_win. If the physical model residual calculation module 40 is enabled, the optional physical model residual calculation step is executed. This module calculates the residual between the theoretical and actual measured exhaust temperatures based on the current operating conditions and / or historical exhaust temperature data. The core diagnostic step is data-driven diagnosis. The data-driven diagnosis module 50 receives a windowed feature sequence (which may include physical model residuals) and uses its internal improved Informer model to perform forward propagation calculations, outputting preliminary diagnostic results (such as a fault probability or anomaly score). Next, the decision fusion module 60 integrates the outputs from module 50 and module 40 (if enabled) and, based on a pre-defined fusion strategy, makes a final fault diagnosis decision. This decision determines the presence of a fault, the specific fault type, and the confidence level of the diagnosis. Alarms and outputs are then generated. If the decision fusion module 60 confirms a fault, the alarm output module 70 generates a corresponding alarm signal and transmits it via a designated channel. Simultaneously, the human-machine interface module 80 updates the sensor status, diagnostic results, and alarm information on the display in real time. Finally, the system automatically returns to data collection and continuously performs online monitoring and diagnosis.

[0033] In addition to the real-time online workflow, this system also features an important offline training and model update phase, which is crucial for maintaining and improving system performance. System operators will periodically, or as needed (for example, when turbine characteristics change significantly or new, unknown failure modes emerge), use newly accumulated historical data (this data should include data from normal operation as well as confirmed failure case data, and may require data augmentation) to retrain or incrementally learn the improved Informer model in the data-driven diagnosis module 50. In this way, the model's performance can be continuously optimized, enabling it to better adapt to changes in turbine characteristics and identify new failure modes. Similarly, the parameters of the physical model and the rules in the decision fusion module can also be regularly adjusted and optimized based on actual operational results.

[0034] Example 2: Detailed implementation of the direct fault classification diagnosis module based on the improved Informer model.

[0035] This embodiment will elaborate on a specific implementation of the data-driven diagnosis module 50 under the system framework of the aforementioned embodiment 1. In this implementation, the "improved informer model" is used to directly perform end-to-end, multi-category classification on multiple predefined fault types of the exhaust temperature sensor. In this specific application scenario, the "improved informer model" used can be called an "improved informer classification model." The internal logical structure diagram of this module is shown in FIG. Figure 2 shown.

[0036] The gas turbine exhaust temperature sensor fault diagnosis method provided in this embodiment is applied to the system according to the above aspects. The method corresponds to the internal working process of the data driven diagnosis module 50, such as Figure 3 As shown, the method includes the following steps: S201 receiving features: the data driven diagnosis module receives a windowed multi-dimensional time series feature X_win from the feature extraction and windowing module; S202 Embedding: The input embedding layer converts the received X_win into the embedded representation X_emb inside the model; S203 encoding: Improve the Informer encoder 51 to perform complex spatiotemporal feature extraction on X_emb and output a deep feature representation H_enc; S204 Pooling (optional): Convert H_enc into a fixed-length vector h_final; S205 Classification: The classification head 53 maps h_final (or directly a specific part of H_enc, such as the output of the [CLS] tag) to the probability P_fault of each predefined fault category; S206 output: The calculated probability vector P_fault is output to the decision fusion module 60 for subsequent comprehensive judgment.

[0037] The input interface of the data-driven diagnosis module 50 is responsible for receiving the windowed multidimensional time series feature X_win from the feature extraction and windowing module 30, whose dimension is R^(L_win × d_feat), where L_win represents the length of the window and d_feat represents the feature dimension at each time step.

[0038] Its core part is to improve the Informer classification model, which logically corresponds to the above Figure 2 The combined structure shown. This model first includes an input embedding layer. This layer consists of a fully connected layer or a one-dimensional convolutional layer, with position encoding and optional timestamp embedding superimposed. Its function is to map each d_feat-dimensional feature vector x_t in the input X_win sequence to the representation dimension d_model inside the model (for example, in this embodiment, d_model can be set to 512), and add position and time context information to obtain the embedded sequence X_emb, whose dimension is R^(L_win × d_model). In terms of specific parameters, for example, d_model is set to 512; the position encoding is generated using a standard sine / cosine function; if timestamp embedding is used, the time information such as year, month, day, hour, and minute can be one-hot encoded (one-hot encoding) or concatenated after processing through a learnable embedding layer, and then mapped to the d_model dimension through a linear projection layer.

[0039] The embedding layer is followed by the improved Informer encoder, which logically corresponds to Figure 2Module 51 in the encoder. This encoder consists of N_e layers (for example, in this embodiment, N_e can be set to 3 layers) stacked with identical encoder sublayers. Each encoder sublayer contains a ProbSparse multi-head self-attention module. In this module, for example, h = 8 attention heads can be set, and the query, key, and value dimensions of each attention head, d_k = d_v = d_model / h = 64. The sampling factor c in the ProbSparse mechanism can be set to 5. The encoder sublayer also contains a separable convolution module. The depthwise convolution can use a one-dimensional kernel, such as kernel_size = 7, with padding = 'causal' to ensure that the calculation of the current time step does not depend on future information. The subsequent pointwise convolution (i.e., 1x1 convolution) is responsible for adjusting the number of channels back to d_model. The activation function in this module can be GeLU. After each self-attention module and separable convolution module, residual connections and layer normalization are applied to stabilize the training process. To handle longer sequences and improve computational efficiency, the encoder also employs self-attention distillation. For example, a one-dimensional convolutional layer (e.g., Conv1D(filters=d_model, kernel_size=3, strides=2, padding='causal')) can be applied after each of the first and second encoder sublayers, followed by a max pooling layer (e.g., MaxPooling1D(pool_size=3, strides=2, padding='same')). This effectively halves the sequence length. After the entire encoder process, the input embedding sequence X_emb is converted into a highly abstracted contextual representation sequence H_enc with dimensions R^(L_out × d_model), where L_out is the shortened sequence length after distillation.

[0040] The output of the encoder is then fed into the Classification Head, which logically corresponds to Figure 2The pooling layer and classifier 53 in the classification head may first include an optional pooling layer. This pooling layer can perform global average pooling on the encoder output sequence H_enc, or directly take the output of the last time step of the sequence to obtain a fixed-length feature vector h_final with a dimension of R^(d_model). If a [CLS] tag (classification token) strategy similar to that in the BERT model is adopted, the output corresponding to the [CLS] tag position can be directly extracted as h_final. The resulting feature vector h_final is then fed into one or more fully connected layers. For example, it can be a sequence containing the structure Linear(d_model, d_hidden_clf) -> ReLU -> Dropout(p=0.3) -> Linear(d_hidden_clf, N_fault_classes). Here, d_hidden_clf is the dimension of the classifier's hidden layer (for example, it can be set to 256), and N_fault_classes is the number of predefined fault categories (for example, it can be set to 7 categories, including: normal state, high sensor output linear drift, low linear drift, signal sticking (unchanged output value), signal open circuit (maximum or minimum output value / specific abnormal value), signal short circuit (output close to zero value / specific abnormal value), and increased sensor output noise). The function of these fully connected layers is to map the deep features extracted by the encoder into the probability space of predefined fault categories. Finally, through a Softmax activation function, an N_fault_classes-dimensional probability vector P_fault is output. Each element P_fault[i] in this vector represents the probability that the current input window data belongs to the i-th fault category. The Softmax function ensures that the sum of the probabilities of all categories is 1.

[0041] Finally, the output interface of this module is responsible for outputting the calculated fault category probability vector P_fault to the decision fusion module 60 for subsequent comprehensive decision making.

[0042] Before training the improved Informer classification model in this embodiment, a series of data augmentation strategies are typically applied to the training data obtained from the signal preprocessing module 20 (especially those belonging to minority fault categories) to improve the model's generalization and robustness. For example, Gaussian noise injection can be used to inject random Gaussian noise with a mean of zero and a standard deviation between 0.01 and 0.05 times the standard deviation of the original exhaust temperature signal into the exhaust temperature signal channel of all samples. Time warping techniques can also be used, such as using cubic spline interpolation to perform a smooth, nonlinear, random warp on the time axis of the original signal. The degree of warping (i.e., the range of variation of the time factor) can be controlled to within ±10% of the original duration. To address class imbalance, particularly when the number of samples for some fault categories is significantly smaller than that for other categories (e.g., less than 10% of the total sample size), the SMOTE (Synthetic Minority Over-sampling Technique) algorithm or its variants can be applied. SMOTE synthesizes new minority class samples that are similar but not identical to real samples by interpolating between minority class samples and their nearest neighbors, thereby achieving a relatively balanced number of samples across all fault categories in the training dataset. Alternatively, a method for synthetic specific fault mode injection can be employed. This method randomly injects simulated typical fault features such as "drift," "stuckness," and "increased noise" into a portion of normal operating data windows, based on pre-set parameter ranges (e.g., drift rate range, signal stuck value range, injected noise amplitude range, etc.), and labels these synthesized samples with corresponding fault labels.

[0043] The training process of the improved Informer classification model in this embodiment generally follows the following steps. First, in the data preparation phase, historical gas turbine operating data is collected and the aforementioned signal preprocessing and feature windowing operations are performed. The aforementioned data augmentation strategy is then applied to expand and balance the training dataset. The entire dataset is then divided into training, validation, and test sets in a ratio of, for example, 7:1.5:1.5. Next, in the model initialization phase, the specific network architecture of the improved Informer classification model is constructed according to the aforementioned description, and the network weight parameters are initialized using methods such as He initialization. Regarding the choice of loss function, considering that fault data in actual industrial scenarios often suffers from class imbalance, a weighted cross-entropy loss function is employed. The weight of each class can be set inversely proportional to the number of samples in the training set. For example, the weight w_i for class i can be calculated using the formula (total number of samples / N_fault_classes) / (number of samples for class i), thereby giving classes with fewer samples a higher loss weight. For optimizer and learning rate settings, use the AdamW optimizer with an initial learning rate of 1e-4 and weight decay of 1e-2. Also, use the Cosine Annealing Learning Rate Scheduler, with the T_max parameter (representing the number of iterations per half a cosine period) set to approximately one-third of the total training rounds.

[0044] In the subsequent training phase, the batch size (e.g., batch_size) is set (e.g., 64), and training is performed for max_epochs (e.g., 100). After each training round, model performance metrics, such as the weighted average F1 score, are calculated on the validation set. To prevent overfitting and select the optimal performing model, early stopping is often employed. Specifically, if a key performance metric on the validation set (such as the weighted average F1 score in this case) does not improve for, say, 10 consecutive rounds, training is terminated early, and the model parameters that achieved the best performance on the validation set are saved as the final model. Finally, in the evaluation phase, the final model is comprehensively evaluated on a separate test set that was not used in training or validation, using various classification performance metrics, including accuracy, precision, recall, F1 score, and confusion matrix.

[0045] Example 3: Detailed implementation of a two-stage anomaly detection and fault identification and diagnosis module based on the improved Informer model.

[0046] This embodiment also provides a method for diagnosing a gas turbine exhaust temperature sensor fault, which is applied to the system according to the above aspects. The method corresponds to the internal working process of the data-driven diagnosis module 50, such as Figure 5 As shown, the method includes the following steps: S301 Receiving features: The data-driven diagnosis module receives a windowed multi-dimensional time series feature X_win from the feature extraction and windowing module. This step is the same as the corresponding step in Example 2.

[0047] S302 Reconstruction / Prediction: The improved Informer anomaly detection model 54 reconstructs or predicts the input feature sequence X_win to obtain the output sequence X_rec, such as Figure 4 shown.

[0048] S303 Anomaly Scoring: The anomaly scoring unit 54c inside the module calculates the reconstruction error between X_win and X_rec to obtain AnomalyScore; S304 Abnormality determination: The system compares AnomalyScore with the preset abnormality threshold Th_anomaly. If AnomalyScore is less than or equal to Th_anomaly, it is determined to be normal and the process proceeds to step S308. However, if AnomalyScore is greater than Th_anomaly, it is determined to be abnormal and the process proceeds to step S305. S305 triggers preliminary alarm and feature construction: the system first outputs a preliminary abnormality signal. At the same time, the feature construction unit 55 is activated, which constructs a feature vector f_clf for subsequent classification for the window X_win currently determined to be abnormal (which can be marked as X_win_anomaly at this time); S306 Fault classification: The pre-trained lightweight fault classifier 56 receives the feature vector f_clf, identifies the specific fault type, and outputs the probability P_fault_specific of each fault category; S307 outputs the classification result: the probability vector P_fault_specific containing the specific fault category information is output to the decision fusion module for it to make the final comprehensive decision; S308: Outputting a “normal” state or an abnormality score less than a preset value, the processing flow of the current window ends.

[0049] This embodiment will elaborate on another alternative specific implementation of the data-driven diagnosis module 50 under the system framework of the aforementioned embodiment 1. Unlike the direct fault classification in embodiment 2, this embodiment adopts a two-stage strategy. The core of stage 1 is to use the "improved Informer model" (the model is configured as a reconstruction or prediction model at this time, so it is called the "improved Informer anomaly detection model") to perform unsupervised time series anomaly detection. Its goal is to identify data segments that deviate significantly from the normal operating mode. Then, in stage 2, when and only when a significant anomaly is detected, the system will activate a lightweight fault classifier, which specifically analyzes the characteristics of the data segments that have been identified as abnormal to identify the specific fault type. The internal logical structure diagram of this module is as follows: Figure 4 shown.

[0050] The input interface of the data-driven diagnosis module 50 is the same as that of Example 2, and is responsible for receiving the windowed multidimensional time series feature X_win from the feature extraction and windowing module 30, whose dimension is R^(L_win×d_feat).

[0051] The core of Phase 1 is the improved Informer anomaly detection model, which logically corresponds to the above Figure 4 Modules 54 and 54c in this model. This model first includes an improved Informer reconstruction model core (corresponding to Figure 4 54a and 54b in Example 2). The structure and function of the input embedding layer are similar to those described in Example 2, which converts the input X_win into an embedded representation X_emb. This is followed by an improved Informer encoder (corresponding to Figure 4 54a), its structure is also similar to the encoder in Example 2, for example, it can include N_e = 3 layers of encoder sublayers, the model internal representation dimension d_model can be set to 256, the number of attention heads h = 4, and the depth convolution kernel size k_d = 5 in the separable convolution module. The encoder also includes a self-attention distillation operation to process long sequences and extract key information, and its output is a context representation sequence H_enc. Unlike Example 2, the anomaly detection model in this embodiment also requires an improved Informer decoder (corresponding to Figure 454b in [54b]. The decoder is composed of N_d layers (for example, N_d can be set to 2 layers) of identical decoder sub-layers stacked together. Each decoder sub-layer contains a masked ProbSparse multi-head self-attention module (to ensure that the output of the current time step does not depend on future information), a multi-head cross-attention module (where the query comes from the output of the previous module of the decoder itself, and the key and value come from the output of the last layer of the encoder H_enc, so as to integrate the encoded information into the decoding process), and a separable convolution module or a standard feedforward network. The input to the decoder is usually a "target" version of the encoder input X_win. For reconstruction tasks, this target version can be a sequence with the same shape as X_win but with some elements randomly masked out, or a simple starting token sequence (such as a sequence of all zeros or a specific learnable embedding sequence). The function of the decoder is to gradually generate a reconstruction sequence of the original input X_win (or, if it is a prediction task, a prediction of the future sequence) based on the context information H_enc output by the encoder and its own input. It is denoted as X_rec, which has the same dimension as X_win, that is, R^(L_win × d_feat). To achieve this, the decoder is connected to an output layer (also called a reconstruction head), which is usually a linear layer responsible for mapping the d_model-dimensional vector sequence output by the last layer of the decoder back to the dimension d_feat of the original input features, thereby obtaining the reconstructed sequence X_rec.

[0052] After obtaining the reconstructed sequence X_rec, the abnormality score and threshold judgment unit (corresponding to Figure 454c) in

[15] . This unit consists of an error calculation unit and a threshold storage and comparison unit. Its core function is to calculate the reconstruction error between the original input window X_win and its corresponding reconstructed version X_rec. Common error metrics are mean squared error (MSE) or mean absolute error (MAE), typically averaged over timesteps, for example, AnomalyScore = mean_over_timesteps(||X_win - X_rec||^2) or AnomalyScore = mean_over_timesteps(|X_win - X_rec|). The calculated AnomalyScore is then compared with a pre-set anomaly threshold Th_anomaly. The setting of this threshold Th_anomaly is crucial and is typically based on the statistical distribution of reconstruction errors generated when the model is run on a large amount of normal operating data (e.g., an independent validation dataset). For example, the 99th percentile of this error distribution or the error mean plus three standard deviations can be used as the threshold. If the AnomalyScore of the current window exceeds this Th_anomaly, the system determines that the data in the window is abnormal. The output of this unit is a Boolean preliminary abnormality signal IsAnomaly (True indicates abnormality, False indicates normality) and the currently calculated AnomalyScore value.

[0053] If and only if the IsAnomaly output by the anomaly scoring and threshold judgment unit in stage 1 is True, the lightweight fault classification process in stage 2 will be activated, which logically corresponds to the above Figure 4 Modules 55 and 56 in the first step are the feature construction unit (corresponding to Figure 455 in ). This unit's input is the raw data window identified as anomaly, called X_win_anomaly, and its corresponding reconstructed error sequence E_rec = X_win_anomaly - X_rec_anomaly. It may also include the AnomalyScore value calculated in stage 1. The function of the feature construction unit is to extract or construct a set of more discriminative feature vectors f_clf from this input information, specifically for subsequent fault classification. Its dimension can be set to d_clf. These features can be diverse and may include, for example: statistical characteristics of the reconstructed error sequence E_rec, such as the AnomalyScore value itself, the variance, kurtosis (kurtosis), and skewness of the error sequence; statistical characteristics of the original abnormal signal window X_win_anomaly itself, such as the mean, variance, and energy of specific frequency bands of each signal channel within the window (for example, calculated by fast Fourier transform (FFT)); in more advanced implementations, it is even possible to consider utilizing some internal state information of the Informer model when processing the abnormal segment, such as the pooled representation of the encoder output H_enc in the abnormal segment, or the weight distribution characteristics of a specific attention head (however, the calculation of such features may be more complex and may not be preferred in scenarios where lightweightness is sought). For example, a d_clf-dimensional feature vector f_clf may contain the following elements: [mean(E_rec), std(E_rec), skewness(E_rec), kurtosis(E_rec), mean(X_win_anomaly_ch1),std(X_win_anomaly_ch1),spectral_energy_band1(X_win_anomaly_ch1),...], where ch1 represents the first sensor channel.

[0054] The constructed feature vector f_clf is then fed into the lightweight fault classifier (corresponding to Figure 456 in). The classifier is a pre-trained simple classification model with relatively low computational overhead. Available models include: support vector machine (SVM), such as SVM using radial basis function (RBF) kernel; decision tree or random forest composed of multiple decision trees; gradient boosting machine (Gradient Boosting Machine), such as LightGBM or XGBoost and other efficient implementations; or a small multi-layer perceptron (MLP) with a very simple structure, for example, only containing 1 to 2 hidden layers, each hidden layer containing dozens of neurons. The input of the lightweight fault classifier is the feature vector f_clf output by the feature construction unit (55). Its function is to classify the abnormal segment based on these features and output the specific fault category probability P_fault_specific. Its output form is similar to the classifier in Example 2, which is a probability vector of N_fault_classes dimension.

[0055] The output interface behavior of the data-driven diagnosis module 50 of this embodiment is as follows: If no anomaly is detected in stage 1 (i.e., IsAnomaly is False), the module may simply output a status signal indicating "normal" or a low anomaly score. However, if an anomaly is detected in stage 1 (i.e., IsAnomaly is True), the module will output the specific fault category probability P_fault_specific obtained by the lightweight fault classifier (56) in stage 2 to the decision fusion module 60 for its final decision. At the same time, the preliminary anomaly signal IsAnomaly and the corresponding AnomalyScore generated in stage 1 can also be output as auxiliary reference information.

[0056] Regarding the data enhancement strategy in this embodiment, its application is also divided into two stages. For the training of the improved Informer anomaly detection model in stage 1, since its goal is to learn the patterns of normal data and identify anomalies, a large amount of data covering various normal operating conditions is mainly used. Data enhancement methods such as noise injection and time distortion can be applied to these normal training data, with the aim of improving the model's learning ability and robustness for the diversity of normal patterns, so that the model can more accurately identify those patterns that have never been seen and are truly abnormal. In the training of this stage, fault labels are not required. For the training of the lightweight fault classifier in stage 2, a data set with clear fault type labels is required. This data set can be real fault cases accumulated from actual operation, or it can contain a large amount of synthetic fault data generated by simulation or rules. For the original signal window X_win in this labeled data set, data enhancement techniques such as fault injection and pattern mixing similar to those in Example 2 can be applied. The key step is to convert these augmented, labeled raw data windows X_win and their corresponding fault labels into a series of feature-label pairs (f_clf, fault_label) using the trained Stage 1 model (or at least the portion of its logic used to calculate the reconstruction error) and the complete processing logic of the feature construction unit 55. These converted feature-label pairs are then used to train the lightweight fault classifier 56.

[0057] The training process of the model in this embodiment is also divided into two main stages. The first is stage 1: training the improved Informer anomaly detection model 54. In terms of data preparation, only a large amount of exhaust temperature sensor data and related operating parameter data covering various normal operating conditions are used. The aforementioned signal preprocessing and feature windowing operations are performed on these data. In terms of model initialization, an improved Informer model with the encoder-decoder structure described above is constructed. In terms of the choice of loss function, since it is a reconstruction task, the mean square error (MSE) or mean absolute error (MAE) is usually used to quantify the difference between the original input window X_win and the output window X_rec reconstructed by the model. The setting of the optimizer and learning rate can refer to the description in Example 2. The goal of the training cycle is to minimize the reconstruction error on normal training data. After the model training is completed, an important step is to determine the anomaly threshold Th_anomaly. This is usually accomplished by running the trained model on an independent validation set that also contains only normal data. Record the reconstruction error values ​​generated by all data windows on the validation set, and then finally determine the value of Th_anomaly based on the statistical distribution of these error values ​​(for example, the 99.5th percentile of the distribution can be taken, or set based on domain knowledge and tolerance for false positives).

[0058] Next comes Phase 2: Training a lightweight fault classifier 56. Data preparation requires a dataset containing labels for various known fault types. For each fault sample window X_win_fault in this dataset, the corresponding feature vector f_clf_fault, used for classification, is extracted through the processing logic of the feature construction unit 55 (this may require first inputting X_win_fault into the trained Phase 1 model to obtain necessary information such as its reconstruction error). This method converts the original fault dataset into a series of feature-label pairs (f_clf_fault, fault_label). For model selection and training, an appropriate lightweight classifier model (such as SVM or random forest) is selected based on actual needs. Standard supervised classification model training is then performed using the generated feature-label pairs. During training, methods such as cross-validation may be required to select and optimize the classifier's hyperparameters.

[0059] Through the detailed description of the three aforementioned embodiments, including a general system framework and two specific diagnostic module implementations based on the core "Improved Informer Model" but with different application strategies, the present invention demonstrates how to build an advanced and powerful gas turbine exhaust temperature sensor fault diagnosis system. It should be emphasized that these embodiments are not mutually exclusive or isolated; the module designs and technical concepts contained within them can be flexibly combined, adjusted, and optimized based on actual engineering needs and application scenarios. For example, the direct classification model described in Example 2 can also obtain additional input features from the physical model residual calculation module 40 to enhance its performance. Alternatively, in certain scenarios where the accuracy of fault type identification is not required to be high, the anomaly detection results in Example 3 can be directly output as the final alarm signal, eliminating the need for subsequent, more detailed fault classification steps. The ultimate goal of all these variations and applications is to fully utilize the excellent capabilities of the "Improved Informer Model" proposed in this invention in processing complex time series data, thereby achieving the goal of efficient, accurate, and robust diagnosis of gas turbine exhaust temperature sensor faults.

[0060] Compared with the prior art, the gas turbine exhaust temperature sensor fault diagnosis system and method based on the improved Informer model disclosed in the present invention, through the technical solutions described in the aforementioned embodiments, can achieve at least one or more of the following beneficial effects: 1. Significantly improve the accuracy and reliability of fault diagnosis, and effectively reduce the false alarm rate and missed alarm rate: The "improved Informer model" adopted at the core of the present invention (whether it is used for direct classification in Example 2 or for anomaly detection in Example 3), its internally integrated ProbSparse self-attention mechanism can efficiently capture long-range dependencies in time series data, while the separable convolution module can effectively extract local context features. This combination enables the model to more deeply understand the complex dynamic behavior and subtle differences of sensor signals under normal and various fault modes. Compared with traditional statistical methods or some shallow machine learning models, the present invention can learn more discriminative feature representations from high-dimensional, nonlinear sensor and operating condition data, thereby showing higher accuracy in distinguishing normal, early minor faults and different fault types. In addition, through the data enhancement strategies described in the embodiments (such as Gaussian noise injection, time warping, SMOTE oversampling, synthetic specific fault mode injection, etc.) and the use of weighted loss functions in model training, the problems of scarce fault samples and class imbalance in industrial practice are effectively alleviated, further improving the model's ability to recognize minority class faults and overall generalization performance, thereby reducing false positives and missed detections caused by model bias or overfitting.

[0061] 2. Achieving early warning and accurate identification of early, subtle sensor faults: Many faults in gas turbine exhaust temperature sensors (such as early drift, slight noise increase, and intermittent, small fluctuations) exhibit very subtle characteristics in their initial stages and are easily masked by normal operating fluctuations. The deep learning model employed in this invention, particularly the improved Informer model, can construct a refined baseline model of the sensor under various normal operating conditions by learning from a large amount of historical data. When a sensor exhibits an early, subtle fault, causing its output signal to deviate from this normal baseline, even the slightest deviation, the model's high sensitivity allows it to capture these abnormal patterns. For example, in Example 3, the anomaly detection mechanism based on reconstruction error is highly sensitive to patterns that deviate from normal behavior. This ability to detect early faults enables operations and maintenance personnel to receive early warnings before the fault develops sufficiently to affect measurement accuracy or cause system misjudgments, buying valuable time for implementing preventive maintenance and avoiding unplanned downtime.

[0062] 3. Enhanced coverage and identification accuracy for multiple complex fault modes: Traditional fault diagnosis methods are often designed for a few known, well-characterized fault types, limiting their ability to identify new or complex faults. The data-driven diagnosis module 50 of the present invention, particularly the direct classification model described in Example 2, can learn and distinguish multiple predefined fault types, such as linear drift (over / under), signal stuckness, open circuits, short circuits, and increased noise, by including as comprehensive and clearly labeled known fault samples as possible in the training data (including real faults and high-quality synthetic faults). The improved Informer model's multi-layer nonlinear mapping capabilities enable it to learn highly complex decision boundaries between different fault modes. In contrast, the two-stage approach described in Example 3, with its first-stage unsupervised anomaly detection, has the potential to detect any "unknown" anomalies that deviate from the normal pattern. The second-stage lightweight classifier can perform more detailed feature extraction and classification for identified anomalies. This combination balances the discovery of unknown anomalies with the identification of known faults.

[0063] 4. Improved adaptability and robustness of the diagnostic system to variable gas turbine operating conditions: Gas turbine operating conditions (such as load, ambient temperature, humidity, fuel characteristics, etc.) are complex and variable. These operating condition changes will directly affect the normal output range and dynamic characteristics of the exhaust temperature sensor, posing a significant challenge to fault diagnosis. When constructing features, the present invention uses relevant gas turbine operating condition parameters and sensor signals as model inputs (as described in module 30), allowing the improved Informer model to learn the inherent correlation between operating condition changes and exhaust temperature signals. This effectively removes the impact of normal operating condition changes on the exhaust temperature signal during diagnosis and more accurately identifies anomalies caused by sensor faults. In addition, the optional physical model residual calculation module 40 in Example 1 predicts normal exhaust temperature fluctuations by introducing prior knowledge based on physical mechanisms. The calculated residual signal can further enhance the model's ability to distinguish between true faults and operating condition disturbances. At the same time, by regularly retraining or incrementally learning the model using the latest operating data, the diagnostic system can continuously adapt to the gradual changes in gas turbine equipment characteristics over time (caused by normal aging or component replacement) and maintain stable diagnostic performance.

[0064] 5. Improved safety, economy, and operation and maintenance efficiency of gas turbines: Through early and accurate diagnosis of exhaust temperature sensor failures, the present invention can avoid misjudgment of the exhaust temperature control system, over-temperature operation of the gas turbine, or unnecessary protective shutdowns caused by sensor failures, thereby directly ensuring the safe and stable operation of the gas turbine unit. Accurate fault location and type identification enable maintenance personnel to quickly lock on to the problem sensor and take targeted repair or replacement measures, significantly shortening the troubleshooting and repair time, and reducing the huge economic losses caused by unplanned downtime. At the same time, based on early warning information, unplanned emergency repairs can be transformed into planned preventive maintenance, optimizing spare parts management and allocation of maintenance resources, and reducing overall operation and maintenance costs. For example, it avoids unnecessary replacement of normal sensors due to misjudgment, or the higher repair costs caused by minor faults turning into major accidents due to missed judgments.

[0065] 6. The system design offers flexibility and scalability: The proposed system framework (Example 1) features a modular design. For example, the physical model residual calculation module 40 is optional, and the data-driven diagnosis module 50 can select either the direct classification scheme of Example 2 or the two-stage scheme of Example 3 based on specific needs. This flexibility allows the present invention to be customized and optimized for different gas turbine models, varying field data conditions, and varying diagnostic accuracy requirements. Furthermore, the design of the human-machine interface module 80 not only displays diagnostic results but also has the potential to enhance model interpretability by visualizing the model's internal state (such as attention weights), helping operators understand and trust the diagnostic results. Furthermore, the system is easily upgradeable and expandable as more fault data accumulates or more advanced model algorithms emerge.

[0066] In summary, the present invention introduces an advanced improved Informer deep learning model and combines comprehensive signal processing, feature engineering, data enhancement and optional physical model information fusion strategy to construct a system and method that can significantly improve the fault diagnosis performance of gas turbine exhaust temperature sensors. It has important practical application value and promotion prospects.

[0067] 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 exhaust temperature sensor fault diagnosis system, characterized in that: The system comprises: A sensor signal acquisition module is used to establish an interface with the exhaust temperature sensor array of the gas turbine and the gas turbine control system or data acquisition system, and to collect the original temperature signal of the exhaust temperature sensor in real time or periodically, and to synchronously obtain relevant gas turbine operating parameters; A signal preprocessing module is used to obtain the original temperature signal and gas turbine operating parameters from the sensor signal acquisition module and perform preprocessing operations; The feature extraction and windowing module is used to obtain the data after the preprocessing operation of the signal preprocessing module, cut the data into a series of multidimensional time series segments according to the preset window length L_win and sliding step size S_step, record the multidimensional time series segments as windows, and calculate the statistical features or frequency domain features within each window; The data-driven diagnosis module integrates one or more pre-trained improved informer models. This module receives windowed multidimensional time series features from the feature extraction and windowing module. Using the deep learning capabilities of the improved informer model, it assesses the health status of the exhaust temperature sensor within the current time window and outputs preliminary diagnostic results. The decision fusion module is used to obtain the preliminary diagnostic results from the data-driven diagnosis module and the signals of other preset independent diagnostic channels, and to perform a comprehensive analysis of the multi-source information obtained, and to make a final fault diagnosis judgment based on the preset fusion logic. The output of the decision fusion module is the final fault diagnosis result, which includes the fault type, the time when the fault occurred, and the confidence level of the diagnosis. The preset other independent diagnostic channels include an alarm system based on threshold comparison and a judgment system based on an expert rule base.

2. The gas turbine exhaust temperature sensor fault diagnosis system according to claim 1, characterized in that: Gas turbine operating parameters include turbine load, compressor outlet pressure, ambient temperature and fuel flow.

3. The gas turbine exhaust temperature sensor fault diagnosis system according to claim 1, characterized in that: The preprocessing operations specifically include: performing filtering operations to remove high-frequency noise and power frequency interference; performing normalization or standardization operations; performing resampling operations to unify data with different sampling rates to a fixed time interval according to actual needs; and performing data cleaning operations to interpolate short-term missing data or remove bad pixels.

4. The gas turbine exhaust temperature sensor fault diagnosis system according to claim 1, characterized in that: The system further comprises: The physical model residual calculation module is used to predict the theoretical output value of the exhaust temperature sensor at the current moment based on the current gas turbine operating parameters and / or historical exhaust temperature sensor data from the signal preprocessing module or the feature extraction and windowing module. Then, the physical model residual calculation module calculates the residual between the actually measured exhaust temperature sensor data from the signal preprocessing module and the predicted theoretical output value, and inputs the residual as an additional feature into the data-driven diagnosis module, or transmits the statistical characteristics of the residual to the decision fusion module.

5. The gas turbine exhaust temperature sensor fault diagnosis system according to claim 1, characterized in that: The system further comprises: The alarm output module is used to obtain the fault diagnosis results from the decision fusion module and, when a confirmed fault is detected, generate a corresponding alarm signal and send it to the upper system through a predetermined interface, or directly notify the relevant operation and maintenance personnel via SMS or email; The human-machine interface module is used to display all relevant information of the system, including real-time data, diagnostic results and alarm information on the system.

6. The gas turbine exhaust temperature sensor fault diagnosis system according to claim 1, characterized in that: The improved Informer model includes: Input embedding layer, which consists of a fully connected layer or a one-dimensional convolutional layer with superimposed position encoding; Improved Informer encoder, which is composed of multiple identical encoder sublayers, each of which contains a ProbSparse multi-head self-attention module; Classification head,The classification head includes a pooling layer and a fully connected classification layer.

7. A method for diagnosing a gas turbine exhaust temperature sensor fault, characterized in that: Applied to the system according to any one of claims 1 to 6, the method comprises the following steps: The data-driven diagnosis module receives a windowed multidimensional time series feature X_win from the feature extraction and windowing module; The input embedding layer converts the received X_win into the embedded representation X_emb inside the model; The improved Informer encoder performs complex spatiotemporal feature extraction on X_emb and outputs a deep feature representation H_enc; Convert H_enc to a fixed-length vector h_final; The classification head maps h_final to the probability P_fault of each predefined fault category; The calculated probability vector P_fault is output to the decision fusion module for subsequent comprehensive judgment.

8. A method for diagnosing a gas turbine exhaust temperature sensor fault, characterized in that: Applied to the system according to any one of claims 1 to 6, the method comprises the following steps: S301: The data-driven diagnosis module receives a windowed multidimensional time series feature X_win from the feature extraction and windowing module; S302: The improved Informer anomaly detection model reconstructs or predicts the input feature sequence X_win to obtain an output sequence X_rec; S303: The anomaly scoring unit inside the module calculates the reconstruction error between X_win and X_rec to obtain AnomalyScore; S304: The system compares AnomalyScore with the preset anomaly threshold Th_anomaly. If AnomalyScore is less than or equal to Th_anomaly, it is determined to be normal and the process proceeds to step S308. However, if AnomalyScore is greater than Th_anomaly, it is determined to be abnormal and the process proceeds to step S305. S305: The system first outputs a preliminary abnormal signal. At the same time, the feature construction unit is activated to construct a feature vector f_clf for subsequent classification for the window X_win currently determined to be abnormal. S306: The pre-trained lightweight fault classifier receives the feature vector f_clf, identifies the specific fault type, and outputs the probability P_fault_specific of each fault category; S307: The probability vector P_fault_specific containing the specific fault category information is output to the decision fusion module for it to make the final comprehensive decision; S308: Output the "normal" state or an abnormality score less than the preset value, and the processing flow of the current window ends.

Images