Turbine high throttle fault detection method based on improved attention mechanism

Abstract

The steam turbine high throttle fault detection method based on the improved attention mechanism, the mathematical simulation model of the steam turbine control system and various fault models of the high throttle are established, and the fault data set for training is generated; short-term data with high sampling frequency and long-term data with low sampling frequency are spliced into a mixed sequence, multi-scale attention mechanism is adopted to extract short-term and long-term data features, and system operation features under different time scales are fused; the context representation of each time point is obtained through the multi-head attention mechanism, and the fault type prediction result of the time point is output by using the full connection layer and the softmax function. Through the above method, the abnormal state detection of the steam turbine high throttle is realized, and a new technical means is provided for the fault detection and prevention of the power system.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment fault diagnosis technology, and provides a method for detecting faults in the high-pressure control valve of a steam turbine based on an improved attention mechanism. Background Technology

[0002] In modern power systems, steam turbines, as key equipment for converting thermal energy into mechanical energy, directly affect the stability and reliability of the entire power system. The high-pressure regulating valve (or simply high-pressure regulating valve) of the steam turbine, as a crucial component, undertakes the critical task of precisely controlling steam flow and regulating unit power. With the large-scale development of new energy power systems, the grid connection and consumption of highly stochastic power sources such as wind power and photovoltaics pose a severe challenge to grid stability. To mitigate the volatility of new energy sources, high-power thermal power units frequently participate in deep peak-shaving and frequency regulation operations, causing their operating conditions to deviate significantly from design values, placing even higher demands on the reliability and safety of the steam turbine's high-pressure regulating valve.

[0003] The normal operation of high-pressure control valves plays an irreplaceable role in maintaining stable unit operation, improving energy efficiency, and ensuring continuous power supply. However, under complex operating environments such as frequent operations and significant load changes, not only are hardware aging problems such as valve stem wear and seal failure exacerbated, but faults such as valve characteristic deviation, actuator jamming, and abnormal position feedback can also occur. If these faults are not detected in time, they can lead to minor issues like unit load oscillations and efficiency decline, or even major accidents such as valve jamming and unit loss of control, seriously threatening the safe operation of the power system. Therefore, developing efficient and reliable high-pressure control valve fault detection technology is of great significance for realizing unit status early warning and optimizing maintenance strategies.

[0004] Currently, fault detection technologies for high-speed control valves in steam turbines are mainly divided into three categories: threshold alarms based on mechanistic models, classification methods based on traditional machine learning, and end-to-end detection methods based on deep learning. However, these methods all have significant shortcomings in practical applications.

[0005] Threshold alarm methods based on mechanistic models establish a high-frequency physical model and combine it with expert experience to set alarm thresholds for key parameters. An alarm is triggered when the monitored value exceeds the preset threshold. While this type of method played an important role in the early stages of industrialization, it has the following limitations: First, threshold setting relies on historical data and expert experience, lacking adaptability to new types of faults; second, parameters fluctuate frequently under dynamic operating conditions, and fixed thresholds are prone to misjudgment; finally, an alarm is only triggered when the fault accumulates to a significant deviation from the normal range, failing to achieve early warning.

[0006] With the popularization of industrial big data technology, traditional machine learning methods, such as support vector machines and random forests, have been introduced into the field of fault diagnosis. These methods extract time-domain, frequency-domain, and time-frequency-domain indicators through feature engineering, and combine them with classifiers to achieve fault identification. Although they are improvements over traditional methods, they still have problems such as complex feature engineering, insufficient generalization ability, and weak time-series modeling ability.

[0007] First, traditional machine learning methods require manually designed features, making it difficult to capture nonlinear relationships in high-dimensional time-series data. For example, early features of jamming failures in high-pressure door actuators may be hidden in millisecond-level pressure pulsations, which traditional methods cannot effectively extract. Second, traditional machine learning models are sensitive to data distribution, and their classification performance drops sharply when operating conditions change abruptly. Finally, traditional machine learning models struggle to model the gradual evolution of faults. For example, mismatches in flow characteristic curves typically exhibit periodic intermittent fluctuations, requiring long-term trend analysis for accurate identification.

[0008] In recent years, deep learning, with its automatic feature extraction and end-to-end learning capabilities, has seen the application of deep learning models such as Long Short-Term Memory Networks and Convolutional Neural Networks in fault detection. Despite the potential of these methods, they still face significant challenges in high-gated detection.

[0009] First, there's the issue of dynamically focusing key fault features. High-gated fault signals often manifest as transient pulses or local distortions, but traditional deep learning models treat all time steps equally, causing key features to be submerged in long-sequence noise. Furthermore, multi-sensor data contains a large amount of redundant information, and the model lacks a dynamic feature selection mechanism, affecting classification accuracy. Second, there's the contradiction between long-distance dependencies and computational efficiency. Early signs of high-gated faults have long-term correlations with subsequent faults, requiring models to have long-sequence modeling capabilities. Traditional RNNs struggle to capture dependencies exceeding 100 steps due to the vanishing gradient problem, while Transformers, although addressing this issue through self-attention, have computational complexity that increases quadratically with sequence length. For high-gated data with a sampling frequency of 1kHz, directly applying Transformers would lead to GPU memory overflow, hindering practical application. Finally, there's insufficient generalization performance in small-sample scenarios. Real-world fault samples are scarce, and fault types are diverse, making supervised learning models prone to overfitting. Existing solutions often rely on data augmentation, but the distribution of generated data differs significantly from real faults, limiting model generalization improvement. Summary of the Invention

[0010] The purpose of this invention is to address the problems of scarce actual fault data, significant detection lag, and insufficient long-distance dependency modeling in the fault detection of high-pressure control valves in steam turbines. It proposes a deep learning method based on multi-scale attention mechanisms and simulation modeling. By constructing a high-fidelity simulation model to generate diverse fault data, and combining dynamic feature focusing and real-time point-by-point prediction techniques, early warning, accurate classification, and evolution tracking of high-pressure control valve faults can be achieved.

[0011] The technical solution adopted in this invention is as follows: A method for detecting faults in high-pressure control valves of steam turbines based on an improved attention mechanism, the steps of which are:

[0012] Step 1: By establishing a mathematical simulation model of the steam turbine control system and various fault models of the high-voltage gate, the final simulation model of the steam turbine control system is formed, and a fault dataset for training is generated.

[0013] For a DEH control system without a primary frequency regulation loop, its forward loop includes power control, pressure control, software cam characteristics, servo amplification characteristics, electro-hydraulic conversion characteristics, hydraulic actuator characteristics, turbine characteristics, and generator characteristics, with transfer functions G1-G8 respectively. The feedback loop includes stroke feedback, pressure feedback, and power feedback, with transfer functions H1-H3 respectively. The characteristic modeling of each part is as follows:

[0014] Power control and pressure control use PI controllers:

[0015]

[0016] Among them, K P1 K P2 The proportional gain for power control and pressure control are respectively, T I1 T I2 The integral time constant;

[0017] The software cam characteristic is the turbine output characteristic f t inverse function:

[0018]

[0019] Among them, f t This is the output characteristic function of the steam turbine;

[0020] Servo amplification is considered a linear process, while electro-hydraulic conversion and hydraulic actuators are approximated as first-order hysteresis processes.

[0021] G4 = K4

[0022]

[0023] Wherein, K4, K5, and K6 are the proportional gains of the servo amplification stage, electro-hydraulic conversion, and hydraulic motor, respectively, and T5 and T6 are the integral time constants;

[0024] The pressure in the steam chamber is approximately a first-order hysteresis loop.

[0025]

[0026] Where K7 is the proportional gain and T7 is the integral time constant;

[0027] The generator is approximately a combination of a first-order lag element and the output characteristics of a steam turbine.

[0028]

[0029] Where K8 is the proportional gain, T8 is the integral time constant, and f t (u) represents the output characteristics of the steam turbine.

[0030] The travel feedback, pressure feedback, and power feedback stages all use unit feedback.

[0031] H1 = H2 = H3 = 1

[0032] By modifying the parameters or structure of the simulation model, three typical faults are simulated: changes in the characteristics of the high-speed gate, jamming of the actuator, and abnormal position feedback.

[0033] Modify the output characteristics f of the steam turbine t The slope range of (u) causes a non-linear relationship between the total valve position command and the actual flow rate. When the slope decreases in a certain range, the command needs to be adjusted significantly to change the flow rate, causing load oscillations. When the slope increases, even a small change in the command can lead to a sudden change in flow rate, resulting in load instability. The fault modeling is expressed as:

[0034] f′ t (u)=f t (u) α

[0035] Where α is the nonlinear distortion factor.

[0036] The motion is restricted within the jamming interval (x1, x2) of the displacement x of the hydraulic actuator. The model is expressed as follows:

[0037]

[0038] Among them, Indicates the direction of motion of the hydraulic motor.

[0039] The jamming that occurs with the piston itself can be simulated using a rounding function:

[0040]

[0041] Where m is the discrete resolution factor, and the smaller the value, the greater the impact of jamming. This model can simulate different degrees of failure: from mild jamming with increased resistance to movement within the interval, to normal movement outside the interval, and complete jamming where movement is impossible within the interval;

[0042] By modifying the output of the linear displacement differential sensor LVDT, three common fault modes are simulated: feedback failure, feedback hold, and feedback offset.

[0043] Feedback failure anomaly refers to a situation where the sensor output suddenly drops to zero due to a disconnection in electrical connection or a power outage, i.e.:

[0044] h′1=0

[0045] Feedback persistence anomaly refers to a situation where the sensor continues to output historical values ​​due to a stuck signal acquisition card or outdated cached data.

[0046] h′1=h c

[0047] Where h c Output historical values ​​for LVDT;

[0048] Feedback offset anomaly refers to a constant deviation in the output caused by core position misalignment or temperature drift, i.e.:

[0049] h′1=h1+Δh.

[0050] Where Δh represents a constant deviation.

[0051] Step 2, Multi-scale data sampling and splicing: Splice short-term data with high sampling frequency and long-term data with low sampling frequency into a mixed sequence;

[0052] Using a high sampling frequency T S1 =Captures real-time dynamic features in 1 second, using a low sampling frequency T 12 = Extract historical trends and global patterns in 1 minute, align them by time, and then concatenate them into a mixed sequence.

[0053] The input sequence is divided into two parts of length n1 and n2, i.e. X1 represents short-term data with a high sampling frequency, and X2 represents long-term data with a low sampling frequency, with continuous and non-overlapping sampling ranges; concatenating X1 and X2 yields X, i.e.:

[0054]

[0055] Where T s1 and T s2 These are the sampling periods for the two parts respectively.

[0056] Step 3: Improved multi-scale attention mechanism: Employ a multi-scale attention mechanism to extract short-term and long-term data features and fuse system operation features at different time scales;

[0057] An improved multi-scale attention mechanism: mapping the input sequence X to the embedding space X E =XW E +P, where the position code P is generated using a sine / cosine function; then, a query matrix Q, a key matrix K, and a value matrix V are generated through a linear transformation, and dynamic weights are calculated.

[0058] The input X is mapped to d through a linear transformation. m In the embedding space, sinusoidal positional encoding is used to embed temporal information as input to the self-attention mechanism, enhancing the model's perception of temporal relationships:

[0059] X E =XW E +P

[0060] p t,2i =sin tθ i

[0061] p t,2i+1 =cos tθ i

[0062]

[0063] in It is a linear mapping matrix. For the position encoding matrix, θ i It is the frequency scaling factor, where i is the dimension index and the maximum index is 1.

[0064] X E Generate query Q, key K, and value V using a learnable weight matrix:

[0065] Q = X E W Q K = X E W K V = X E W V

[0066] Among them W Q , All are linear transformation weight matrices;

[0067] Calculate the attention weight matrix A:

[0068]

[0069] Where d kis the dimension of the key vector.

[0070] Step 4: Point-by-point fault classification and real-time alarm: Obtain the context representation of each time point through a multi-head attention mechanism, and output the fault type prediction result of that time point using a fully connected layer and a softmax function.

[0071] The self-attention output vector a i After concatenation, the data is input into a fully connected layer, and the failure probability is calculated using the softmax function. When the failure probability exceeds the threshold, an alarm signal is immediately triggered and the failure type and timestamp are recorded.

[0072] Self-attention It is considered that the output at each time point has a dimension of d. v vector a i This vector is used for point-by-point fault type prediction; for multiple attention heads, the corresponding a needs to be... i The data is concatenated, and then the classification result at that time point is obtained through a fully connected layer and a softmax function.

[0073]

[0074] in, Let represent the representation vector output by the j-th attention head at time i. concat concatenates the output vectors of multiple attention heads along the feature dimension. W and b are the weight matrix and bias vector of the fully connected layer.

[0075] Cross-entropy loss is used as the loss function:

[0076]

[0077] Where C is the number of fault categories, u is the time point, and y is the time point. ij For the i-th time point, this represents the actual value of the j-th type of fault. Let be the predicted probability of the j-th type of fault at the i-th time point.

[0078] By analyzing the prediction results at each time point, the evolution of the fault can be tracked to determine the specific time and cause of the fault.

[0079] The device employed in this invention comprises four parts: a simulation data generation system, a multi-sensor data acquisition module, an edge computing processing unit, and a human-computer interaction terminal. The simulation data generation system generates simulation data covering changes in valve characteristics, actuator jamming, and abnormal position feedback, addressing the problem of insufficient actual fault samples. The multi-sensor data acquisition module captures multi-dimensional signals of the high-voltage valve's operating state in real time and preprocesses the raw signals to improve data quality and reduce subsequent analysis errors. The edge computing processing unit performs multi-scale stitching of the acquired real-time data, dynamically allocates feature weights through a self-attention mechanism, and ultimately outputs the probability of the fault type. The human-computer interaction terminal provides intuitive fault location and evolution tracking, supports historical data backtracking and model performance analysis, and optimizes maintenance strategies.

[0080] The beneficial effects of this invention are as follows:

[0081] 1. This invention generates simulation data covering changes in valve characteristics, actuator jamming, and abnormal position feedback by modeling transfer functions and fault mechanisms, thereby expanding the diversity of training samples.

[0082] 2. This invention solves the problem of long sequence dependencies by integrating a multi-scale attention mechanism to simultaneously capture short-term abnormal fluctuations and long-term fault evolution trends.

[0083] 3. The present invention designs a point-by-point fault prediction mechanism, which outputs the fault probability at each time point, reducing the processing delay of the traditional window sliding method. Attached Figure Description

[0084] Figure 1 This is a structural diagram of a steam turbine DEH control system without a primary frequency regulation loop;

[0085] Figure 2 For the output characteristics f of the steam turbine t (u) Schematic diagram;

[0086] Figure 3 This is a schematic diagram illustrating the changes in the loss function during model training.

[0087] Figure 4 A confusion matrix diagram of the model's classification results on the validation set;

[0088] Figure 5 A comparison chart of the confusion matrix of ablation experiment classification results;

[0089] Figure 6 This is a comparison chart of indicators from the ablation experimental model. Detailed Implementation

[0090] The technical solutions in the embodiments of the present invention will be described in detail below with reference to the accompanying drawings. The following embodiments are applicable to the present invention, but are not intended to limit the scope of the invention.

[0091] A method for detecting high-voltage valve faults in steam turbines based on an improved attention mechanism is proposed. The method includes establishing a simulation model of the steam turbine control system, constructing a fault state dataset, multi-scale data sampling and stitching, an improved multi-scale attention mechanism, point-by-point fault classification and real-time alarm, and result analysis.

[0092] The computer used to build the simulation model of the steam turbine control system and construct the fault state dataset was an AMD Ryzen 7 7700X with an NVIDIA RTX 4070 GPU, running Windows 11. The simulation model was built using MATLAB and Simulink.

[0093] Multi-scale data sampling and stitching, an improved multi-scale attention mechanism, point-by-point fault classification, and real-time alarms were implemented on an AMD Ryzen 7 7700X computer. The attention model was built using Python 3.10 and PyTorch.

[0094] A transfer function model of the turbine DEH control system is constructed to simulate changes in high-voltage valve characteristics, actuator jamming, and position feedback anomalies, generating diverse fault data. A multi-scale attention mechanism is designed to integrate mixed sequences with high and low sampling frequencies and capture local and global features through self-attention dynamic weighting. A point-by-point prediction method is proposed to output the fault probability at each time point in real time, realizing detection and evolution tracking.

[0095] The first step is to establish a simulation model of the turbine control system: This involves constructing a transfer function model of the DEH control system, including modules for power control, pressure control, software cams, servo amplifiers, electro-hydraulic conversion, and hydraulic actuators. All feedback loops use unity feedback to simplify the model's complexity.

[0096] Step 2: Constructing a fault state dataset: This is done by modifying the turbine output characteristic function f. t (u) simulates load oscillation and load change faults by adjusting the slope of the flow curve in a specific range; simulates valve actuator jamming faults by limiting the change of hydraulic actuator displacement x in the range (x1,x2) through a piecewise function; and simulates valve position feedback abnormal faults by modifying the output of the linear displacement differential sensor (LVDT).

[0097] The specific methods for the first and second steps are as follows:

[0098] Based on the physical mechanism of the digital electro-hydraulic control (DEH) system for steam turbines, a transfer function model is constructed that includes power control, pressure control, servo amplification, electro-hydraulic conversion, and hydraulic actuators, such as... Figure 1 As shown, the mathematical model of the key components is as follows:

[0099] Power control and pressure control use PI controllers:

[0100]

[0101] Among them, K P1 K P2 The proportional gain for power control and pressure control are respectively, T I1 T I2 is the integration time constant.

[0102] The software cam characteristic is the turbine output characteristic f t inverse function:

[0103]

[0104] Among them, f t This is the output characteristic function of the steam turbine.

[0105] The servo amplifier can be considered a linear circuit, while the electro-hydraulic converter and the hydraulic actuator can be approximated as first-order hysteresis circuits.

[0106] G4 = K4#(4)

[0107]

[0108] Where K4, K5, and K6 are the proportional gains of the servo amplification stage, electro-hydraulic conversion stage, and hydraulic motor, respectively, and T5 and T6 are the integral time constants.

[0109] The pressure in the steam chamber is approximately a first-order hysteresis loop.

[0110]

[0111] Where K7 is the proportional gain and T7 is the integral time constant.

[0112] The generator is approximately a combination of a first-order lag element and the output characteristics of a steam turbine.

[0113]

[0114] Where K8 is the proportional gain, T8 is the integral time constant, and f t (u) represents the output characteristics of the steam turbine.

[0115] Steam turbine output characteristics f t (u) is achieved through piecewise linear functions.

[0116] Each feedback stage involves feedback from the relevant unit:

[0117] H1=H2=H3=1#(9)

[0118] This patent considers the following types of high-adjustment gate failures: changes in gate characteristics, jamming of the gate actuator, and abnormal gate position feedback.

[0119] A change in the control valve characteristics refers to a shift in the flow characteristic curve of the turbine's high-pressure control valve, altering the relationship between the unit's flow characteristics—specifically, the total valve position command and the actual flow output. Ideally, the total valve position command and actual flow output have a linear relationship; however, after a change in control valve characteristics, this relationship becomes non-linear, with the slope potentially increasing or decreasing in certain intervals. This change in the unit's flow characteristics can lead to instability in the closed-loop control system, resulting in load oscillations and sudden load changes. When the slope of the unit's flow characteristic curve decreases in a certain interval, a significant change in the total valve position command is required to adjust the actual flow, causing oscillations in the command, control valve, and steam pressure, leading to load oscillations. Conversely, when the slope of the unit's flow characteristic curve increases in a certain interval, a small change in the total valve position command can cause a large change in the actual flow, resulting in sudden load changes. Control valve characteristic change faults can be addressed by modifying f... t (u) is implemented, and the fault modeling is as follows:

[0120] f′ t (u)=f t (u) α #(10)

[0121] Where α is the nonlinear distortion factor.

[0122] Jamming of the control valve actuator refers to the jamming of the valve stem or its connected mechanism, which can lead to complete jamming in severe cases. The control valve actuator typically consists of a valve stem, a hydraulic actuator, and corresponding servo valves and amplifiers. Since the control valve actuator usually forms a closed-loop control with the control valve position feedback, mild jamming will not significantly affect unit operation. However, once the jamming force increases above the actuator's maximum driving force, it can cause the control valve to jam, potentially leading to unit malfunction and accidents. Therefore, it is necessary to detect control valve actuator jamming abnormalities promptly to prevent control valve jamming. The control valve actuator jamming fault is implemented by modifying the hydraulic actuator model. Let its displacement be x. For jamming within a fixed interval, let the interval be (x1, x2), then:

[0123]

[0124] in, Indicates the direction of motion of the hydraulic motor.

[0125] The jamming that occurs in the piston itself can be simulated using a rounding function:

[0126]

[0127] Where m is the discrete resolution factor, the smaller the value, the greater the impact of jamming. This model can simulate different degrees of failure: from mild jamming with increased resistance to movement within the interval, to normal movement outside the interval, and complete jamming where movement is impossible within the interval.

[0128] As a core component of the turbine control closed loop, the valve position feedback system directly impacts the stability and dynamic response characteristics of power regulation. Linear displacement differential sensors (LVDTs) are widely used for high-voltage valve position detection due to their non-contact measurement, high resolution, and strong anti-interference capabilities. However, operating under harsh conditions of high temperature, high-frequency vibration, and steam erosion, LVDTs are prone to faults such as winding insulation aging, core wear, or signal circuit failure, causing the position feedback value to deviate from the actual physical displacement. If such anomalies are not identified in time, they can affect the closed-loop control of the valve position, leading to malfunctions in control commands and actuators, and in severe cases, causing sudden load changes or valve oscillations, threatening the safe operation of the unit.

[0129] This patent establishes a valve position feedback anomaly model based on the LVDT sensor status, starting from the fault mechanism, including feedback failure anomaly, feedback holding anomaly, and feedback offset anomaly.

[0130] Feedback failure anomaly refers to a situation where the sensor output suddenly drops to zero due to a disconnection in electrical connection or a power outage, i.e.:

[0131] h′1=0#(13)

[0132] Feedback persistence anomaly refers to a situation where the sensor continues to output historical values ​​due to a stuck signal acquisition card or outdated cached data.

[0133] h′1=h c #(14)

[0134] Where h c Output historical values ​​for LVDT.

[0135] Feedback offset anomaly refers to a constant deviation in the output caused by core position misalignment or temperature drift, i.e.:

[0136] h′1=h1+Δh#(15)

[0137] Where Δh represents a constant deviation.

[0138] Step 3: Multi-scale data sampling and stitching: Using a high sampling frequency T S1 =Captures real-time dynamic features in 1 second, using a low sampling frequency T S2 = Extract historical trends and global patterns in 1 minute, align them by time, and then concatenate them into a mixed sequence.

[0139] Step 4: Improved multi-scale attention mechanism: Mapping the input sequence X to the embedding space X E =XW E +P, where the position code P is generated using a sine / cosine function. Then, a linear transformation is used to generate the query matrix Q, key matrix K, and value matrix V, and dynamic weights are calculated.

[0140] The specific methods for steps three and four are as follows:

[0141] To ensure the model's real-time performance, the input sequence is sampled at a high frequency. However, due to computational constraints, the input sequence length *n* is finite, preventing the model from obtaining long-term historical data. This patent addresses this issue by employing a multi-scale attention mechanism. The multi-scale attention mechanism is a deep learning technique that combines multi-scale feature extraction and attention mechanisms. Its core idea is to extract key information from features at multiple scales (such as semantic information at different levels, receptive fields of different sizes, or contexts of different granularities) and dynamically allocate weights using an attention mechanism, thereby enhancing the model's ability to focus on important features.

[0142] The input sequence is divided into two parts of length n1 and n2, i.e. X1 represents short-term data with a high sampling frequency, and X2 represents long-term data with a low sampling frequency, with continuous and non-overlapping sampling ranges. Then, X1 and X2 are concatenated to obtain X, i.e.:

[0143]

[0144] Where T s1 and T s2 These are the sampling periods for the two parts respectively.

[0145] Through a multi-scale attention mechanism, the model can balance the timeliness of short-term data with the greater information provided by long-term data, all within a given computational limit. Combined with dynamic weighting of the attention mechanism, the model can more comprehensively capture local details and global features, fully utilizing data over a longer time span and improving accuracy. Table 1 shows a comparison of the model's immediacy and time range at different sampling frequencies.

[0146] Table 1 Comparison of different sampling frequencies

[0147]

[0148] The input X is mapped to d through a linear transformation. m In the embedding space, sinusoidal positional encoding is used to embed temporal information as input to the self-attention mechanism:

[0149] X E =XW E+P#(17)

[0150] p t,2i =sin tθ i #(18)

[0151] p t,2i+1 =cos tθ i #(19)

[0152]

[0153] in It is a linear mapping matrix. For the position encoding matrix, θ i It is the frequency scaling factor, where i is the dimension index and the maximum index is 1.

[0154] Then, through linear transformation, we obtain the query matrix Q, the key matrix K, and the value matrix V:

[0155] Q = X E W Q K = X E W K V = X E W V #(twenty one)

[0156] Among them W Q , All are linear transformation weight matrices;

[0157] Calculate the attention weight matrix A:

[0158]

[0159] Where d k is the dimension of the key vector.

[0160] Step 5: Point-by-point fault classification and real-time alarm. The self-attention output vector a... i After concatenation, the data is input into a fully connected layer, where the fault probability is calculated using the softmax function. When the detected anomaly probability exceeds a threshold, an alarm signal is immediately triggered, and the fault type and timestamp are recorded.

[0161] Step 6: Result Analysis: Based on the alarm information displayed on the human-machine terminal and the visualization analysis, guide operation and maintenance decisions.

[0162] Self-attention This can be viewed as outputting a dimension of d at each time point. v vector a i Therefore, this vector can be used for point-by-point fault type prediction. For multiple attention heads, the corresponding a needs to be... iThe data is concatenated, and then the classification result at that time point is obtained through a fully connected layer and a softmax function.

[0163]

[0164] in, Let represent the representation vector output by the j-th attention head at time i. concat concatenates the output vectors of multiple attention heads along the feature dimension. W and b are the weight matrix and bias vector of the fully connected layer.

[0165] Cross-entropy loss is used as the loss function:

[0166]

[0167] Where C is the number of fault categories, i is the time point, and y is the time point. ij For the i-th time point, this represents the actual value of the j-th type of fault. Let be the predicted probability of the j-th type of fault at the i-th time point.

[0168] By making point-by-point predictions, the model can capture changes in system state in real time, reducing lag and identifying potential problems in the early stages before a failure occurs. Point-by-point predictions fully utilize the continuity of time series data, capturing trends and patterns over time, thus providing a more comprehensive understanding of the system and improving the accuracy of fault detection. By analyzing the prediction results at each time point, the evolution of the fault can be tracked, determining the specific time of occurrence and possible causes, which helps to more accurately locate problems and optimize maintenance strategies.

[0169] This invention addresses the problems of scarce fault data, detection lag, and insufficient long-distance dependency modeling in traditional turbine high-pressure control valve fault detection methods. It proposes a turbine high-pressure control valve fault detection method based on an improved attention mechanism. First, a transfer function model of the turbine DEH control system is constructed, including a power control module, a pressure control module, and a software cam module, etc., such as... Figure 1 As shown. Secondly, fault states are incorporated into the model through mechanistic analysis. The turbine output characteristic function f is modified. t (u), such as Figure 2As shown, the slope of the flow curve is adjusted within a specific interval to simulate load oscillation and sudden load changes; the displacement x of the hydraulic actuator within the interval (x1, x2) is limited by a piecewise function to simulate valve actuator jamming; and the output of the linear displacement differential sensor is modified to simulate valve position feedback anomaly. Dynamic parameter adjustment generates diverse training data, effectively compensating for the lack of actual fault samples. Next, a multi-scale attention network is designed, concatenating high-sampling-frequency short-term data (1-second signals) with low-sampling-frequency long-term data (1-minute signals) into a mixed sequence. Weights are dynamically allocated through a self-attention mechanism to capture local abnormal fluctuations and global evolution patterns. Based on a multi-head attention layer, the contextual relationships of multi-dimensional sensor information are extracted in parallel, significantly improving the sensitivity to early fault signs. Finally, a point-by-point fault prediction method is proposed, outputting the fault type probability distribution at each time point to achieve real-time fault detection and evolution path tracking.

[0170] In the experiment Figure 3 A schematic diagram of the loss function changes during the model training process is given. It can be seen that the training and validation losses gradually decrease and tend to level off as the training progresses. The convergence trend of the training loss and the validation loss is consistent, and no significant overfitting phenomenon is observed. Figure 4 The confusion matrix of the model's classification results on the validation set is presented, demonstrating that the model achieves high accuracy across all categories. To verify the effectiveness of the key components of the proposed model, ablation experiments were conducted to evaluate the contributions of the pointwise prediction strategy and the multi-scale attention mechanism to the detection performance. The classification confusion matrix is ​​shown below. Figure 5 As shown, the model metrics are as follows Figure 6 As shown in the figure, experiments demonstrate that the point-by-point prediction strategy, through fine-grained modeling in the time dimension, effectively reduces fault confusion; the multi-scale attention mechanism, through cross-time domain feature fusion, significantly improves the model's ability to capture fault evolution patterns. The combined effect of these two mechanisms enables the complete model to achieve high-precision classification of complex fault modes while maintaining high real-time performance. This invention, through the synergistic innovation of simulation modeling, multi-scale feature fusion, and real-time point-by-point prediction technology, provides a reliable technical means for anomaly detection of high-voltage control valves in steam turbines, and also provides important assurance for the safe and stable operation of power systems. Furthermore, the framework of this method can be extended to fault detection tasks of other key equipment, showing broad application prospects.

Claims

1. A method for detecting faults in high-pressure control valves of steam turbines based on an improved attention mechanism, characterized in that, The steps are as follows: Step 1: By establishing a mathematical simulation model of the steam turbine control system and various fault models of the high-voltage gate, the final simulation model of the steam turbine control system is formed, and a fault dataset for training is generated. Step 2, Multi-scale data sampling and splicing: Splice short-term data with high sampling frequency and long-term data with low sampling frequency into a mixed sequence; Step 3: Improved multi-scale attention mechanism: Employ a multi-scale attention mechanism to extract short-term and long-term data features and fuse system operation features at different time scales; Improved multi-scale attention mechanism: mixing sequences Mapping to Embedded Space Position encoding Generates the query matrix Q, key matrix K, and value matrix V through sine / cosine functions; then generates them through linear transformations, and calculates the dynamic weights. ; The method is as follows: The input is transformed by linear transformation Mapped to In the embedding space, sinusoidal positional encoding is used to embed temporal information as input to the self-attention mechanism, enhancing the model's perception of temporal relationships: in It is a linear mapping matrix. For the position encoding matrix, It is the frequency scaling factor. It is a dimensional index and the maximum index is ; Generate query Q, key K, and value V using a learnable weight matrix: in All are linear transformation weight matrices; Calculate the attention weight matrix : in The dimension of the key vector; Step 4: Point-by-point fault classification and real-time alarm: Obtain the context representation of each time point through a multi-head attention mechanism, and output the fault type prediction result at that time point using a fully connected layer and a softmax function; Self-attention It is considered that the output at each time point has a dimension of vector This vector is used for point-by-point fault type prediction; for multiple attention heads, the corresponding... Concatenate, then through full connection and The function obtains the classification result at that time point: in, Indicates the first Attention is focused on the point in time. The output representation vector, The output vectors of multiple attention heads are concatenated along the feature dimension. and These are the weight matrix and bias vector of the fully connected layer; Cross-entropy loss is used as the loss function: in, Number of fault categories For a point in time, For the first The time point for the first The actual value of the fault type, For the first The time point for the first Predicted probability of type of failure; By analyzing the prediction results at each time point, the evolution of the fault can be tracked to determine the specific time and cause of the fault.

2. The turbine high-pressure control valve fault detection method based on an improved attention mechanism according to claim 1, characterized in that: The specific method in step one is as follows: For a DEH control system without a primary frequency regulation loop, its forward loop includes power control, pressure control, software cam characteristics, servo amplification characteristics, electro-hydraulic conversion characteristics, hydraulic actuator characteristics, turbine characteristics, and generator characteristics. The transfer functions of each part are as follows: - The feedback loop includes travel feedback, pressure feedback, and power feedback, with transfer functions as follows: - The characteristics of each part are modeled as follows: Power and pressure control use PI controllers: in, These are the proportional gains for power control and pressure control, respectively. The integral time constant; Software cam characteristics are turbine output characteristics inverse function: in, This is the output characteristic function of the steam turbine; Servo amplification is considered a linear process, while electro-hydraulic conversion and hydraulic actuators are approximated as first-order hysteresis processes. in, These are the proportional gains of the servo amplification stage, electro-hydraulic conversion, and hydraulic motor, respectively. The integral time constant; The pressure in the steam chamber is approximately a first-order hysteresis loop. in, For proportional gain, The integral time constant; The generator is approximately a combination of a first-order lag element and the output characteristics of a steam turbine. in, For proportional gain, The integral time constant is... For the output characteristics of the steam turbine; The travel feedback, pressure feedback, and power feedback stages all use unit feedback. By modifying the parameters or structure of the simulation model, three typical faults are simulated: changes in the characteristics of the high-speed gate, jamming of the actuator, and abnormal position feedback. Modify turbine output characteristics The slope range causes a non-linear relationship between the total valve position command and the actual flow rate. When the slope decreases in a certain range, the command needs to be adjusted significantly to change the flow rate, causing load oscillations; when the slope increases, a small change in the command leads to a sudden change in flow rate, causing load instability. The fault modeling is expressed as: in This is the nonlinear distortion factor; Displacement of the hydraulic motor The range of sluggishness Internally restricted motion, modeled as: in, Indicates the direction of motion of the hydraulic motor; The jamming that occurs with the piston itself is simulated using a rounding function: in, The smaller the value, the greater the impact of jamming. This model can simulate different degrees of failure: from mild jamming with increased resistance to movement within the range, to normal movement outside the range, and complete jamming where movement is impossible within the range. By modifying the output of the linear displacement differential sensor LVDT, three common fault modes are simulated: feedback failure, feedback hold, and feedback offset. Feedback failure anomaly refers to a situation where the sensor output suddenly drops to zero due to a disconnection in electrical connection or a power outage, i.e.: Feedback persistence anomaly refers to a situation where the sensor continues to output historical values ​​due to a stuck signal acquisition card or outdated cached data. in Output historical values ​​for LVDT; Feedback offset anomaly refers to a constant deviation in the output caused by core position misalignment or temperature drift, i.e.: in This represents a constant deviation that exists.

3. The turbine high-pressure control valve fault detection method based on an improved attention mechanism according to claim 1, characterized in that: In step two, a high sampling frequency is used. Capture real-time dynamic features using a low sampling frequency Historical trends and global patterns are extracted, aligned by time, and then concatenated into a mixed sequence. .

4. The turbine high-pressure control valve fault detection method based on an improved attention mechanism according to claim 3, characterized in that: The specific method in step two is as follows: The mixed sequence is divided into lengths of... and The two parts, namely , , For short-term data with a high sampling frequency, For long-term data with low sampling frequency, and with continuous and non-overlapping sampling ranges; and spliced ​​together ,Right now: in and These are the sampling periods for the two parts respectively.

5. The turbine high-pressure control valve fault detection method based on an improved attention mechanism according to claim 1, characterized in that: In step four: the self-attention output vector After concatenation, input into a fully connected layer, and then... The function calculates the probability of failure; when the probability of an anomaly exceeds the threshold, an alarm signal is immediately triggered and the fault type and timestamp are recorded.

Images