Turbine high control valve fault detection method based on improved attention mechanism

Abstract

The invention discloses a steam turbine high-pressure control valve fault detection method based on an improved attention mechanism, and the method comprises the steps: generating a fault data set for training through building a mathematical simulation model of a steam turbine control system and various fault models of a high-pressure control valve; the short-term data with the high sampling frequency and the long-term data with the low sampling frequency are spliced into a mixed sequence, short-term and long-term data features are extracted by adopting a multi-scale attention mechanism, and system operation features under different time scales are fused; and acquiring context representation of each time point through a multi-head attention mechanism, and outputting a fault type prediction result of the time point by using a full connection layer and a softmax function. Through the method, the abnormal state of the high control valve of the steam turbine is detected, and a new technical means is provided for fault detection and prevention of a power system.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial equipment fault diagnosis, and provides a method for detecting a high-pressure regulating valve fault of a steam turbine based on an improved attention mechanism. Background Art

[0002] In modern power systems, steam turbines are key equipment for converting thermal energy into mechanical energy, and their operating status is directly related to the stability and reliability of the entire power system. The steam turbine high-pressure regulating valve (HVV), as an important component of the steam turbine, undertakes the critical task of accurately controlling steam flow and regulating unit power. With the large-scale development of new energy power systems, the grid connection and absorption of highly random power sources such as wind power and photovoltaics poses a severe challenge to the stability of the power grid. In order to smooth out the volatility of new energy, high-power thermal power units frequently participate in deep peak-shaving and frequency-regulating operations, causing their operating conditions to deviate significantly from the design values, and placing higher requirements on the reliability and safety of the steam turbine HVV.

[0003] The normal operation of high-pressure regulating valves plays an irreplaceable role in maintaining the stable operating conditions of the unit, improving energy utilization efficiency, and ensuring the continuity of power supply. However, in complex operating environments such as frequent operation and large-scale load changes, not only will hardware aging problems such as valve stem wear and seal failure be exacerbated, but it will also lead to faults such as valve characteristic deviation, actuator jamming, and position feedback anomalies. If such faults are not detected in a timely manner, at the very least, they can lead to unit load oscillation and reduced efficiency, and at worst, they can cause major accidents such as regulating valve jamming and unit loss of control, seriously threatening the safe operation of the power system. Therefore, the development of efficient and reliable high-pressure regulating valve fault detection technology is of great significance for achieving unit status early warning and optimizing maintenance strategies.

[0004] Currently, fault detection technologies for steam turbine high-pressure regulating valves are mainly divided into three categories: threshold alarms based on mechanism 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 mechanism models establish a physical model of a high-pressure valve and, incorporating expert experience, set alarm thresholds for key parameters. Alarms are triggered when monitored values ​​exceed the preset thresholds. Although these methods played an important role in the early stages of industrialization, they suffer from the following limitations. First, threshold setting relies on historical data and expert experience, making it less adaptable to new types of faults. Second, under dynamic operating conditions, parameters fluctuate frequently, and fixed thresholds can easily lead to misjudgments. Finally, faults must accumulate to a point where they significantly deviate from the normal range before an alarm is triggered, making early warning impossible.

[0006] With the widespread adoption of industrial big data technologies, 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 metrics through feature engineering and combine them with classifiers to identify faults. While these methods offer some improvements over traditional methods, they still suffer from complex feature engineering, insufficient generalization capabilities, and weak time series modeling capabilities.

[0007] First, traditional machine learning methods require manual feature design, making it difficult to capture nonlinear relationships in high-dimensional time series data. For example, the early signs of a high-pressure valve actuator jam 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 suddenly change. Finally, traditional machine learning models struggle to model the gradual evolution of faults. For example, flow characteristic curve mismatches often exhibit periodic intermittent fluctuations, requiring long-term trend analysis for accurate identification.

[0008] In recent years, deep learning models such as long short-term memory networks and convolutional neural networks have been used for fault detection due to their automatic feature extraction and end-to-end learning capabilities. Although these methods have shown great potential, they still face significant challenges in high-profile fault detection.

[0009] First, there's the issue of dynamically focusing on key fault features. High-frequency gate fault signals often manifest as transient pulses or localized distortions, but traditional deep learning models treat all time steps equally, causing key features to be buried in long sequences of noise. Furthermore, multi-sensor data contains a large amount of redundant information, and the model lacks a dynamic feature selection mechanism, impacting classification accuracy. Second, there's a conflict between long-range dependencies and computational efficiency. Early signs of high-frequency gate faults have long-term correlations with subsequent failures, requiring the model to be able to model long sequences. Traditional RNNs struggle to capture dependencies exceeding 100 steps due to the vanishing gradient problem. While the Transformer addresses this issue through its self-attention mechanism, its computational complexity increases quadratically with the sequence length. For high-frequency gate data sampled at 1kHz, directly applying the Transformer would cause GPU memory overflow, making it difficult to implement in engineering applications. Finally, generalization performance is insufficient in small sample size scenarios. Actual fault samples are scarce and their 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, resulting in limited improvement in model generalization. Summary of the Invention

[0010] This paper aims to address the challenges of steam turbine high-pressure valve fault detection, such as a scarcity of actual fault data, significant detection lag, and inadequate modeling of long-range dependencies. By proposing a deep learning approach based on a multi-scale attention mechanism and simulation modeling, this approach generates diverse fault data through the construction of a high-fidelity simulation model. This approach, combined with dynamic feature focusing and real-time point-by-point prediction, enables early warning, accurate classification, and evolution tracking of high-pressure valve faults.

[0011] The technical solution adopted by the present invention is as follows: a method for detecting a high-pressure regulating valve fault of a steam turbine based on an improved attention mechanism, the steps of which are as follows:

[0012] Step 1: By establishing a mathematical simulation model of the steam turbine control system and various fault models of the high-pressure valve, a final steam turbine control system simulation model is formed, and a fault data set for training is generated;

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

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

[0015]

[0016] Among them, K P1 , K P2 are the proportional gains of power control and pressure control, T I1 、T I2 is the integration time constant;

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

[0018]

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

[0020] The servo amplifier is considered as a linear link, and the electro-hydraulic conversion and the oil motor are approximately first-order lag links:

[0021] G4=K4

[0022]

[0023] Among them, K4, K5, and K6 are the proportional gains of the servo amplifier link, electro-hydraulic conversion, and oil motor respectively, and T5 and T6 are the integral time constants;

[0024] The steam chamber pressure is approximately a first-order lag link:

[0025]

[0026] Among them, K7 is the proportional gain and T7 is the integral time constant;

[0027] The generator is approximately a combination of a first-order lag link and a steam turbine output characteristic:

[0028]

[0029] Among them, K8 is the proportional gain, T8 is the integral time constant, f t (u) is the turbine output characteristic.

[0030] Stroke feedback, pressure feedback, and power feedback are all unit feedback:

[0031] H1=H2=H3=1

[0032] By modifying the simulation model parameters or structure, three typical faults, namely, high-pressure valve characteristic change, actuator jamming, and position feedback abnormality, can be simulated;

[0033] Modify the turbine output characteristic f t The slope range of (u) makes the total valve position command and the actual flow present a nonlinear relationship. When the slope of a certain interval decreases, the command needs to be adjusted significantly to change the flow, causing load oscillation. When the slope increases, a small change in the command leads to a sudden change in the flow, causing load instability. The fault modeling expression is:

[0034] f′ t (u) = f t (u) α

[0035] Where α is the nonlinear distortion factor.

[0036] The movement is restricted within the stuck range (x1, x2) of the hydraulic motor displacement x. The model is expressed as:

[0037]

[0038] in, Indicates the direction of movement of the oil motor.

[0039] The jamming of the piston itself is simulated by the rounding function:

[0040]

[0041] Where m is the discrete resolution factor, and smaller values ​​indicate a greater impact of the jam. The model can simulate different fault levels: from mild jamming with increased resistance to movement within the interval, to normal movement outside the interval, and complete jamming with no movement within the interval.

[0042] By modifying the output of the linear displacement differential sensor LVDT, three common fault modes, namely feedback failure abnormality, feedback retention abnormality, and feedback offset abnormality, are simulated;

[0043] Feedback failure abnormality means that the sensor output suddenly drops to zero due to electrical disconnection or power interruption, that is:

[0044] h′1=0

[0045] Feedback retention abnormality means that the sensor continues to output historical values ​​due to the signal acquisition card being stuck or the cache data not being updated, that is:

[0046] h′1=h c

[0047] where h c Output of LVDT historical value;

[0048] Feedback offset anomaly refers to a constant deviation in the output caused by core position offset or temperature drift, that is:

[0049] h′1=h1+Δh.

[0050] Where Δh is the constant deviation that exists.

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

[0052] Using high sampling frequency T S1 = 1s to capture real-time dynamic features, using a low sampling frequency T 12 =1min Extract historical trends and global patterns, align them in time and then splice them into a mixed sequence

[0053] The input sequence is divided into two parts of length n1 and n2, namely X1 is short-term data with a higher sampling frequency, and X2 is long-term data with a lower sampling frequency. The sampling ranges are continuous and non-overlapping. X1 and X2 are concatenated to obtain X, that is:

[0054]

[0055] Where T s1 and T s2 There are two sampling periods respectively.

[0056] Step 3: Improved multi-scale attention mechanism: A multi-scale attention mechanism is used to extract short-term and long-term data features and integrate system operation characteristics at different time scales;

[0057] 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 by the sine / cosine function; then the query matrix Q, key matrix K, and value matrix V are generated through linear transformation to calculate the dynamic weight

[0058] Map the input X to d through a linear transformation m In the embedding space of , sinusoidal position encoding is used to embed time information as the input of the self-attention mechanism to enhance 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 is the linear mapping matrix, is the position encoding matrix, θ i is the frequency scaling factor, i is the dimension index and the maximum index is

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

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

[0066] Where W Q , They are all 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: The multi-head attention mechanism is used to obtain the context representation of each time point, and the fully connected layer and softmax function are used to output the fault type prediction result at that time point.

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

[0072] Attention Considered as outputting a dimension of d at each time point v Vector a i , use this vector to predict the fault type point by point; for multiple attention heads, the corresponding a i Splicing, and then using full connection and softmax function to get the classification result at that time point:

[0073]

[0074] in, represents the representation vector output by the jth attention head at time point i. concat concatenates the output vectors of multiple attention heads in 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] Among them, C is the number of fault categories, u is the time point, y ij is the true value of the j-th type of fault at the i-th time point, is the predicted probability of the jth type of failure 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 point and cause of the fault.

[0079] The device used in the present invention consists of 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 valve characteristic changes, actuator jamming, and position feedback anomalies, solving the problem of insufficient actual fault samples. The multi-sensor data acquisition module captures multi-dimensional signals of the high-pressure valve operating status in real time and pre-processes the original signals to improve data quality and reduce subsequent analysis errors. The edge computing processing unit performs multi-scale splicing on the collected real-time data, dynamically assigns 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 the present invention are:

[0081] 1. The present invention generates simulation data covering valve characteristic changes, actuator jamming and position feedback abnormalities through transfer function and fault mechanism modeling, thereby expanding the diversity of training samples.

[0082] 2. This invention solves the problem of long sequence dependency 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, outputting the fault probability at each time point, thus reducing the processing delay of the traditional window sliding method. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 This is the structure diagram of the steam turbine DEH control system without the primary frequency modulation circuit;

[0085] Figure 2 is the turbine output characteristic f t (u) Schematic diagram;

[0086] Figure 3 Schematic diagram of the change of loss function during model training;

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

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

[0089] Figure 6 A comparison chart of ablation experiment model indicators. DETAILED DESCRIPTION

[0090] The technical solutions in the embodiments of the present invention are 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 present invention.

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

[0092] The computer used to establish the steam turbine control system simulation model and construct the fault state data set is AMD Ryzen 7 7700X, the GPU is NVIDIA RTX 4070 computer, and the WINDOW 11 operating system. The simulation model is built using MATLAB and Simulink.

[0093] Multi-scale data sampling and splicing, improved multi-scale attention mechanism, point-by-point fault classification and real-time alarm run on an AMD Ryzen7 7700X computer, and the attention model is built using Python 3.10 and PyTorch.

[0094] A transfer function model of the steam turbine DEH control system is constructed to simulate changes in high-pressure valve characteristics, actuator jamming, and position feedback anomalies to generate diverse fault data. A multi-scale attention mechanism is designed to integrate mixed sequences of high and low sampling frequencies, and capture local and global features through dynamic weighting of self-attention. 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 steam turbine control system simulation model: Build a transfer function model of the DEH control system, including the power control module, pressure control module, software cam module, servo amplifier module, electro-hydraulic conversion module, and hydraulic motor module. All feedback links use unit feedback to simplify the model complexity.

[0096] Step 2: Construct the fault status data set: By modifying the turbine output characteristic function f t (u), adjust the slope of the flow curve in a specific interval to simulate load oscillation and load mutation faults; limit the change of the oil motor displacement x in the interval (x1, x2) through a piecewise function to simulate the jamming fault of the throttle actuator; and simulate the abnormal throttle position feedback fault by modifying the output of the linear displacement differential sensor (LVDT).

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

[0098] Based on the physical mechanism of the steam turbine digital electro-hydraulic control (DEH) system, a transfer function model including power control, pressure control, servo amplification, electro-hydraulic conversion and oil motor is constructed, such as Figure 1 As shown, the mathematical models of key components are as follows:

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

[0100]

[0101] Among them, K P1 , K P2 are the proportional gains of power control and pressure control, T I1 、T I2 is the integration time constant.

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

[0103]

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

[0105] The servo amplifier is considered as a linear link, and the electro-hydraulic conversion and oil motor can be approximated as first-order lag links:

[0106] G4=K4#(4)

[0107]

[0108] Among them, K4, K5, and K6 are the proportional gains of the servo amplifier link, electro-hydraulic conversion, and oil motor respectively, and T5 and T6 are the integral time constants.

[0109] The steam chamber pressure is approximately a first-order lag link:

[0110]

[0111] Among them, K7 is the proportional gain and T7 is the integral time constant.

[0112] The generator is approximately a combination of a first-order lag link and a steam turbine output characteristic:

[0113]

[0114] Among them, K8 is the proportional gain, T8 is the integral time constant, f t (u) is the turbine output characteristic.

[0115] Steam turbine output characteristic f t (u) is implemented by a piecewise linear function.

[0116] Each feedback link is unit feedback:

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

[0118] This patent considers that the high-pressure valve failure types include changes in valve characteristics, jamming of the valve actuator, and abnormal valve position feedback.

[0119] The change of the regulating valve characteristic refers to the change of the flow characteristic curve of the high regulating valve of the steam turbine, which causes the flow characteristic of the unit, that is, the relationship between the total valve position command and the actual flow output to change. Ideally, the total valve position command and the actual flow output are linearly related; when the regulating valve characteristic changes, the two have a nonlinear relationship, and the slope of some intervals may increase or decrease. After the flow characteristic of the unit changes, the closed-loop control system may no longer be stable, and load oscillation and load mutation problems may occur. When the slope of the unit's flow characteristic curve decreases in a certain interval, the total valve position command needs to change significantly to adjust the actual flow, causing oscillations in the command, regulating valve, and steam pressure, and load oscillation problems; when the slope of the unit's flow characteristic curve increases in a certain interval, a small change in the total valve position command causes a large change in the actual flow, causing load mutation problems. The regulating valve characteristic change fault can be corrected by modifying f t (u) Implementation, fault modeling is:

[0120] f′ t (u) = f t (u) α #(10)

[0121] Where α is the nonlinear distortion factor.

[0122] The jam of the throttle actuator refers to the jamming of the valve stem of the throttle or the mechanism connected to it, and in severe cases, it may become stuck. The throttle actuator is usually composed of a valve stem, an oil motor, a corresponding servo valve and an amplifier. Since the throttle actuator usually forms a closed-loop control with the throttle position feedback, mild jamming will not significantly affect the operation of the unit. However, once the jamming force increases to above the maximum driving force of the actuator, it will cause the throttle to jam, which may cause the unit to lose control and cause an accident. Therefore, it is necessary to detect the jamming abnormality of the throttle actuator in time to prevent the throttle from getting stuck. The jamming fault of the throttle actuator is realized by modifying the oil motor model. Let its displacement be x. For the jamming in a fixed interval, let the interval be (x1, x2), and we have:

[0123]

[0124] in, Indicates the direction of movement of the oil motor.

[0125] The jamming of the piston itself can be simulated by the rounding function:

[0126]

[0127] Here, m is the discrete resolution factor; smaller values ​​indicate a greater impact of the jam. The model can simulate different fault severity levels: from mild jamming, where resistance to movement increases within the range, to normal movement outside the range, to complete jamming, where movement is impossible within the range.

[0128] As the core link in the steam turbine control closed-loop, the valve position feedback system's measurement accuracy directly impacts the stability and dynamic response characteristics of power regulation. Linear displacement differential sensors (LVDTs) are widely used for high-pressure valve position detection due to their non-contact measurement, high resolution, and strong anti-interference capabilities. However, due to long-term operation in harsh operating conditions such as high temperature, high-frequency vibration, and steam erosion, LVDTs are prone to failures 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 a timely manner, they will affect the closed-loop control of the valve position, causing control commands and actuators to malfunction. In severe cases, they can cause sudden load changes or valve oscillation, threatening the safe operation of the unit.

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

[0130] Feedback failure abnormality means that the sensor output suddenly drops to zero due to electrical disconnection or power interruption, that is:

[0131] h′1=0#(13)

[0132] Feedback retention abnormality means that the sensor continues to output historical values ​​due to the signal acquisition card being stuck or the cache data not being updated, that is:

[0133] h′1=h c #(14)

[0134] where h c Output of LVDT historical value.

[0135] Feedback offset anomaly refers to a constant deviation in the output caused by core position offset or temperature drift, that is:

[0136] h′1=h1+Δh#(15)

[0137] Where Δh is the constant deviation that exists.

[0138] Step 3: Multi-scale data sampling and splicing: using high sampling frequency T S1 = 1s to capture real-time dynamic features, using a low sampling frequency T S2 =1min Extract historical trends and global patterns, align them in time and then splice them into a mixed sequence

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

[0140] The specific methods for the third and fourth steps are:

[0141] In order to ensure the immediacy of the model, the sampling frequency of the input sequence is high. Due to the consideration of computational complexity, the length n of the input sequence is limited, and the model cannot obtain long-term historical data. This patent adopts a multi-scale attention mechanism to solve this problem. The multi-scale attention mechanism is a deep learning technology that combines multi-scale feature extraction and attention mechanism. Its core idea is to extract key information through features of multiple scales (such as semantic information at different levels, receptive fields of different sizes, or contexts of different granularity), and use the attention mechanism to dynamically assign weights, 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, namely X1 is short-term data with a higher sampling frequency, and X2 is long-term data with a lower sampling frequency, and the sampling ranges are continuous and non-overlapping. Then X1 and X2 are concatenated to get X, that is:

[0143]

[0144] Where T s1 and T s2 There are two sampling periods respectively.

[0145] Through the multi-scale attention mechanism, the model is able to balance the timeliness provided by short-term data with the more information provided by long-term data, while maintaining a limited computational footprint. Combined with the dynamic weighting of the attention mechanism, the model can more comprehensively capture local details and global features, fully utilizing data with a longer time span, thereby improving accuracy. Table 1 shows a comparison of the model's immediacy and time horizon at different sampling frequencies.

[0146] Table 1 Comparison of different sampling frequencies

[0147]

[0148] Map the input X to d through a linear transformation m In the embedding space of , sinusoidal position encoding is used to embed time information as the input of 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 is the linear mapping matrix, is the position encoding matrix, θ i is the frequency scaling factor, i is the dimension index and the maximum index is

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

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

[0156] Where W Q , They are all 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 fed into the fully connected layer, where the fault probability is calculated using the softmax function. When the detected abnormal probability exceeds the threshold, an alarm signal is immediately triggered and the fault type and timestamp are recorded.

[0161] Step 6. Result analysis: Guide operation and maintenance decisions based on the alarm information and visual analysis displayed on the human-machine terminal.

[0162] Self-attention It can be regarded as outputting a dimension of d at each time point v Vector a i , so the fault type can be predicted point by point through this vector. For multiple attention heads, the corresponding a iSplicing, and then using full connection and softmax function to get the classification result at that time point:

[0163]

[0164] in, represents the representation vector output by the jth attention head at time point i. concat concatenates the output vectors of multiple attention heads in 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, y ij is the true value of the j-th type of fault at the i-th time point, is the predicted probability of the jth type of failure at the i-th time point.

[0168] Through point-by-point prediction, the model can instantly capture changes in system status, reduce lag, and identify potential issues at an early stage before a failure occurs. Point-by-point prediction fully leverages the continuity of time series data to capture trends and patterns across time, enabling a more comprehensive understanding of the system and improving fault detection accuracy. By analyzing the prediction results at each point in time, it is also possible to track the evolution of faults, determine the specific time of occurrence, and identify possible causes, helping to more accurately locate problems and optimize maintenance strategies.

[0169] This paper addresses the problems of scarcity of fault data, detection lag, and insufficient long-distance dependency modeling in traditional steam turbine high-pressure valve fault detection methods. A steam turbine high-pressure valve fault detection method based on an improved attention mechanism is proposed. First, a transfer function model of the steam turbine DEH control system is constructed, including a power control module, a pressure control module, a software cam module, etc. Figure 1 As shown. Secondly, the fault state is added to the model through mechanism analysis. Modify the turbine output characteristic function f 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 change faults; a piecewise function is used to constrain the displacement x of the hydraulic motor within the interval (x1, x2) to simulate a throttle actuator jam; and the output of the linear displacement differential sensor is modified to simulate an abnormal throttle position feedback fault. Dynamic parameter adjustment generates diverse training data, effectively compensating for the lack of actual fault samples. Next, a multi-scale attention network is designed to concatenate short-term data (1-second signals) with long-term data (1-minute signals) at a low sampling frequency into a mixed sequence. A self-attention mechanism dynamically assigns weights to capture both local abnormal fluctuations and global evolution patterns. Based on a multi-head attention layer, the contextual associations of multi-dimensional sensor information are extracted in parallel, significantly improving sensitivity to early fault signs. Finally, a point-by-point fault prediction method is proposed, outputting a probability distribution of fault types at each time point, enabling instant fault detection and evolution path tracking.

[0170] During the experiment Figure 3 A schematic diagram of the change in the loss function during model training is given. It can be seen that the training and validation losses gradually decrease and tend to be flat as the training progresses. The convergence trend of training loss and validation loss is consistent, and there is no significant overfitting phenomenon. Figure 4 The confusion matrix of the classification results of the model on the validation set is given, which shows that the model has a high accuracy rate in all categories. In order to verify the effectiveness of the key parts of the model proposed in the patent, the contribution of the point-by-point prediction strategy and the multi-scale attention mechanism to the detection performance is evaluated through ablation experiments. The classification confusion matrix is ​​as follows: Figure 5 As shown, the model indicators are as follows Figure 6 As shown. Experiments show that the point-by-point prediction strategy effectively reduces the confusion of faults through fine-grained modeling in the time dimension; the multi-scale attention mechanism significantly improves the model's ability to capture the laws of fault evolution through cross-time domain feature fusion. The combined effect of the two enables the complete model to achieve high-precision classification of complex fault modes while maintaining high real-time performance. Through the collaborative innovation of simulation modeling, multi-scale feature fusion and real-time point-by-point prediction technology, the present invention provides a reliable technical means for the abnormal detection of steam turbine high-pressure valves, and also provides an important guarantee for the safe and stable operation of the power system. In addition, the framework of this method can be extended to fault detection tasks of other key equipment, and has broad application prospects.

Claims

1. A steam turbine high-pressure valve fault detection method based on an improved attention mechanism is characterized by: The steps are: Step 1: By establishing a mathematical simulation model of the steam turbine control system and various fault models of the high-pressure valve, a final steam turbine control system simulation model is formed, and a fault data set for training is generated; Step 2: Multi-scale data sampling and splicing: Splice the short-term data with high sampling frequency and the long-term data with low sampling frequency into a mixed sequence; Step 3: Improved multi-scale attention mechanism: A multi-scale attention mechanism is used to extract short-term and long-term data features and integrate system operation characteristics at different time scales; Step 4: Point-by-point fault classification and real-time alarm: The multi-head attention mechanism is used to obtain the context representation of each time point, and the fully connected layer and softmax function are used to output the fault type prediction result at that time point.

2. The method for detecting high-pressure regulating valve faults in a steam turbine based on an improved attention mechanism according to claim 1, characterized in that: In the step 1, the specific method is: For a DEH control system without a primary frequency modulation loop, its forward loop includes power control, pressure control, software cam characteristics, servo amplification characteristics, electro-hydraulic conversion characteristics, oil motor characteristics, steam turbine and generator characteristics. The transfer functions of each part are G1-G8. The feedback loop includes stroke feedback, pressure feedback, and power feedback. The transfer functions are H1-H3. The characteristic modeling of each part is as follows: Power control and pressure control use PI controller: Among them, K P1 , K P2 are the proportional gains of power control and pressure control, T I1 、T I2 is the integration time constant; The software cam characteristic is the turbine output characteristic f t The inverse function of : G3=f t -1 Among them, f t is the turbine output characteristic function; The servo amplifier is considered as a linear link, and the electro-hydraulic conversion and the oil motor are approximately first-order lag links: G4=K4 Among them, K4, K5, and K6 are the proportional gains of the servo amplifier link, electro-hydraulic conversion, and oil motor respectively, and T5 and T6 are the integral time constants; The steam chamber pressure is approximately a first-order lag link: Among them, K7 is the proportional gain and T7 is the integral time constant; The generator is approximately a combination of a first-order lag link and a steam turbine output characteristic: Among them, K8 is the proportional gain, T8 is the integral time constant, f t (u) is the turbine output characteristic. Stroke feedback, pressure feedback, and power feedback are all unit feedback: H1=H2=H3=1 By modifying the simulation model parameters or structure, three typical faults, namely, high-pressure valve characteristic change, actuator jamming, and position feedback abnormality, can be simulated; Modify the turbine output characteristic f t The slope range of (u) makes the total valve position command and the actual flow present a nonlinear relationship. When the slope of a certain interval decreases, the command needs to be adjusted significantly to change the flow, causing load oscillation. When the slope increases, a small change in the command leads to a sudden change in the flow, causing load instability. The fault modeling expression is: f t ′(u)=f t (u) α Where α is the nonlinear distortion factor. The movement is restricted within the stuck range (x1, x2) of the hydraulic motor displacement x. The model is expressed as: in, Indicates the direction of movement of the oil motor. The piston itself is stuck and is simulated by the rounding function: Where m is the discrete resolution factor, and smaller values ​​indicate a greater impact of the jam. The model can simulate different fault levels: from mild jamming with increased resistance to movement within the interval, to normal movement outside the interval, and complete jamming with no movement within the interval. By modifying the output of the linear displacement differential sensor LVDT, three common fault modes, namely feedback failure abnormality, feedback retention abnormality, and feedback offset abnormality, are simulated; Feedback failure abnormality means that the sensor output suddenly drops to zero due to electrical disconnection or power interruption, that is: h′1＝0 Feedback retention abnormality means that the sensor continues to output historical values ​​due to the signal acquisition card being stuck or the cache data not being updated, that is: h′1=h c where h c Output of LVDT historical value; Feedback offset anomaly refers to a constant deviation in the output caused by core position offset or temperature drift, that is: h′1=h1+Δh. Where Δh is the constant deviation that exists.

3. The method for detecting high-pressure regulating valve faults in a steam turbine based on an improved attention mechanism according to claim 1, characterized in that: In the step 2, a high sampling frequency T is used. S1 = 1s to capture real-time dynamic features, using a low sampling frequency T S2 =1min Extract historical trends and global patterns, align them in time and then splice them into a mixed sequence 4. The method for detecting high-pressure regulating valve faults in a steam turbine based on an improved attention mechanism according to claim 4, characterized in that: The specific method in the step 2 is: The input sequence is divided into two parts of length n1 and n2, that is X1 is short-term data with a higher sampling frequency, and X2 is long-term data with a lower sampling frequency. The sampling ranges are continuous and non-overlapping. X1 and X2 are concatenated to obtain X, that is: Where T s1 and T s2 There are two sampling periods respectively.

5. The method for detecting high-pressure regulating valve faults in a steam turbine based on an improved attention mechanism according to claim 1, characterized in that: In step 3, the improved multi-scale attention mechanism: maps the input sequence X to the embedding space X E =XW E +P, where the position code P is generated by the sine / cosine function; then the query matrix Q, key matrix K, and value matrix V are generated through linear transformation to calculate the dynamic weight 6. The method for detecting high-pressure regulating valve faults in a steam turbine based on an improved attention mechanism according to claim 5, characterized in that: The specific method in step three is: Map the input X to d through a linear transformation m In the embedding space of , sinusoidal position encoding is used to embed time information as the input of the self-attention mechanism to enhance the model's perception of temporal relationships: X E =XW E +P p t,2i =sintθ i p t,2i+1 =costθ i in is the linear mapping matrix, is the position encoding matrix, θ i is the frequency scaling factor, i is the dimension index and the maximum index is X E Generate query Q, key K, value V through learnable weight matrix: Q=X E W Q ,K=X E W K ,V=X E W V in They are all linear transformation weight matrices; Calculate the attention weight matrix A: where d k is the dimension of the key vector.

7. The method for detecting high-pressure regulating valve faults in a steam turbine based on an improved attention mechanism according to claim 1, wherein: In the above step 4: the self-attention output vector a i After splicing, the data is input into the fully connected layer and the fault probability is calculated by the softmax function. When the abnormal probability exceeds the threshold, an alarm signal is immediately triggered and the fault type and timestamp are recorded.

8. The method for detecting high-pressure regulating valve faults in a steam turbine based on an improved attention mechanism according to claim 7, characterized in that: The specific method in step 4 is: Attention Considered as outputting a dimension of d at each time point v Vector a i , use this vector to predict the fault type point by point; for multiple attention heads, the corresponding a i Splicing, and then using full connection and softmax function to get the classification result at that time point: in, represents the representation vector output by the jth attention head at time point i. concat concatenates the output vectors of multiple attention heads in the feature dimension. W and b are the weight matrix and bias vector of the fully connected layer. Cross entropy loss is used as the loss function: Where C is the number of fault categories, i is the time point, y ij is the true value of the j-th type of fault at the i-th time point, is the predicted probability of the jth type of failure at the i-th time point. By analyzing the prediction results at each time point, the evolution of the fault can be tracked to determine the specific time point and cause of the fault.

Images