Steam turbine blade water erosion defect monitoring and identification method

Abstract

The invention discloses a turbine blade water erosion defect monitoring and identification method, and relates to the technical field of turbine state monitoring and fault diagnosis. The method comprises a state sensing module, a feature processing module, an intelligent identification module and an early warning correction module. The state sensing module collects macroscopic feature data such as vibration signals and operation parameters; the feature processing module performs multi-dimensional extraction on the data and constructs a fusion feature vector; the intelligent identification module outputs a water erosion defect level through a dual-mode autonomous learning model; and the early warning correction module executes graded early warning and performs closed-loop correction on the model based on an actual result. Early recognition of water erosion defects is achieved, new and old units are adapted, and monitoring reliability and economical efficiency are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of steam turbine state monitoring and fault diagnosis, in particular to a steam turbine blade water erosion defect monitoring and identification method. BACKGROUND

[0002] As the core equipment of energy conversion and power output, the blade of steam turbine is prone to water erosion defects in the running process due to the wet steam environment. When the steam expands in the last stage and the penultimate stage of the low-pressure cylinder, the pressure and temperature decrease to form wet steam, and the liquid water droplets in the wet steam continuously impact the blade surface, causing damage such as pitting, pits, and cracks on the blade, which not only reduces the operating efficiency of the steam turbine, but also may cause serious safety accidents such as blade rupture.

[0003] The existing blade water erosion monitoring technology has obvious limitations: offline detection (such as endoscopy and ultrasonic flaw detection) relies on shutdown operation and cannot grasp the defect development trend in real time, making it difficult to achieve early warning; online monitoring relies on microsensors (such as distributed optical fibers and microwave radars), which require modification of the steam turbine body, installation is complex and costly, and is not suitable for retrofit of old units. At the same time, traditional early warning methods based on a single parameter (such as vibration amplitude and temperature) have low recognition accuracy, high false alarm rate, and lack of self-learning ability, making it difficult to adapt to the individualized operating characteristics of different units and meet the needs of monitoring reliability and adaptability in industrial scenarios. SUMMARY

[0004] (I) Technical problems solved In view of the deficiencies of the prior art, the present application provides a steam turbine blade water erosion defect monitoring and identification method, which solves the problem of insufficient timeliness of traditional offline detection.

[0005] (II) Technical solutions To achieve the above purpose, the present application is realized by the following technical solutions: A steam turbine blade water erosion defect monitoring and identification method, comprising: A state perception module for collecting macro feature data during the operation of the steam turbine, the macro feature data including vibration signals, operating parameters, acoustic signals, axial displacement, and differential expansion data; A feature processing module connected to the state perception module for multi-dimensional feature extraction and construction of a fusion feature vector from the macro feature data; An intelligent identification module connected to the feature processing module for analyzing the fusion feature vector through a dual-mode self-learning model and outputting a water erosion defect level; The early warning correction module is connected with the intelligent identification module, performs graded early warning according to the water erosion defect grade, and performs closed-loop correction on the dual-mode autonomous learning model based on actual detection results.

[0006] Preferably, the state perception module includes the following acquisition modes: vibration amplitude and frequency spectrum in horizontal and vertical directions are acquired by vibration sensors installed on the bearing seat and the low-pressure cylinder body, with a sampling frequency of 10 kHz; low-pressure cylinder inlet steam pressure, temperature, flow, steam turbine heat rate, low-pressure cylinder exhaust temperature, unit output power and speed fluctuation data are acquired by pressure sensors, temperature sensors and flow sensors, with a sampling interval of 1 minute; sound pressure level in the range of 80-120 dB and noise spectrum in the frequency band of 500-2000 Hz are acquired by waterproof microphones installed on the outer wall of the low-pressure cylinder; rotor axial displacement deviation and cylinder and rotor expansion difference are acquired by displacement sensors; historical maintenance data, including blade appearance inspection records and repair records, are input.

[0007] Preferably, the multi-dimensional feature extraction process of the feature processing module includes the following steps: wavelet packet decomposition is used to extract the vibration energy proportion in the frequency band of 100-300 Hz, and the kurtosis coefficient of the vibration signal is calculated, and when the kurtosis value is greater than 3, the abnormal vibration feature is marked; a 24-hour sliding window is used to calculate the cumulative deviation rate of the heat rate, i.e. (actual value-design value) / design value, and when the deviation rate is greater than 0.5% for three consecutive days, the abnormal parameter feature is marked; the impact noise feature peak near 1500 Hz is extracted by the mel-frequency cepstral coefficient to generate a water erosion soundprint feature vector; after the above features are normalized, a 128-dimensional fusion feature vector is formed.

[0008] Preferably, the dual-mode autonomous learning model of the intelligent identification module includes a basic model architecture and a learning mode, and the basic model architecture is a hybrid model of a convolutional neural network (CNN) and a long short-term memory network (LSTM), wherein: The CNN layer is used to process the frequency spectrum graph space features of the vibration and acoustic signals, including a first layer of 16 3x3 convolution kernels and a second layer of 32 3x3 convolution kernels, with 2x2 maximum pooling, a ReLU activation function, and a Dropout probability of 0.2 to prevent overfitting; the LSTM layer is used to analyze the time sequence features of the parameter trend, including 64 hidden units, a time step of 24, a dropout of 0.1, and a recurrent_dropout of 0.1; the output layer is used to predict the water erosion defect grade, which is divided into 0-4 levels, 0 level is no damage, 1 level is slight pitting, 2 level is local pit, 3 level is crack initiation, and 4 level is serious damage requiring emergency shutdown.

[0009] Preferably, the learning mode includes an online federated learning mode, specifically: in a networked state, the multi-unit edge node uploads the model parameters of the "fused feature vector-erosion defect level" to the cloud server in an encrypted manner, and the cloud server aggregates the parameters to update the global model and then issues it to each edge node.

[0010] Preferably, the learning mode includes an offline incremental learning mode, which is realized based on a knowledge distillation technology, and the specific steps are as follows: The online global model is used as a teacher model, the local model to be updated is used as a student model, and the erosion level probability distribution output by the teacher model is used as historical knowledge carrier; the distillation temperature parameter T is set to 7, the output distribution of the teacher model and the student model is calculated through a softening probability formula, and the softening probability formula is as follows: Teacher model softening probability: ; Student model softening probability: ; Wherein, , logits of the teacher model and the student model, respectively; A mixed loss function is constructed: , and an Adam optimizer is used to train for 20-30 epochs until the historical knowledge retention rate of the verification set is greater than 90%; the student model after training is used to replace the original local model, and offline updating is completed.

[0011] Preferably, the grading early warning mechanism of the early warning correction module includes: when the erosion defect level is level 1, level 1 early warning is triggered: the data acquisition frequency is increased to 1 minute / time, and the middle and upper key inspection areas of the blade steam exit edge are pushed; When the erosion defect level is level 2, level 2 early warning is triggered: an online adjustment scheme is generated, and the scheme includes increasing the steam superheat degree by 0.5 DEG C, and simultaneously reserving a shutdown maintenance plan; When the erosion defect level is level 3 or above, level 3 early warning is triggered: a sound-light alarm is started, the unit load is forcibly limited to 30%, and an emergency repair plan is pushed.

[0012] Preferably, the closed-loop correction process of the early warning correction module is: after the unit is shut down for maintenance, the actual erosion level detected by the artificial detection is input into the system; the intelligent identification module calculates the prediction error, and the feature weight is corrected through a gradient descent algorithm; when the prediction confidence is less than 80%, it is marked as a to-be-verified sample, and after networking, the cloud model is requested for collaborative verification, and the local model parameters are updated.

[0013] (Three) beneficial effects The present application provides a steam turbine blade erosion defect monitoring and identification method. It has the following beneficial effects: Through multi-dimensional macro feature fusion analysis and intelligent learning model, the water erosion defect of the blade can be accurately identified at the germination stage, potential risks can be found in advance, and damage and shutdown loss of the equipment caused by the development of the defect to a serious stage can be avoided. Through early warning and accurate identification, unnecessary shutdown maintenance is reduced, the service life of the blade is prolonged, the safe and stable operation of the steam turbine is ensured, and the energy conversion efficiency and the economic efficiency of the equipment operation are improved.

[0014] Adopting a dual-mode autonomous learning mechanism, the group recognition ability can be improved through online collaborative learning, and the model can be updated autonomously in an offline state to adapt to the differences in operating characteristics and changes in working conditions of different units and maintain long-term monitoring accuracy.

[0015] Through modular design combined with a closed-loop correction mechanism, the accuracy of the monitoring results is ensured; at the same time, a hierarchical early warning strategy provides clear decision-making basis for operation and maintenance personnel, facilitating operation and execution in actual engineering applications. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The flowchart of the present application. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0018] Please refer to Figure 1 The present application provides a technical solution: a steam turbine blade water erosion defect monitoring and identification method, comprising: A state perception module is used to collect macro feature data in the operation process of the steam turbine, and the macro feature data includes vibration signals, operating parameters, acoustic signals, axial displacement and differential expansion data; A feature processing module is connected with the state perception module, performs multi-dimensional feature extraction on the macro feature data, and constructs a fusion feature vector; An intelligent identification module is connected with the feature processing module, analyzes the fusion feature vector through a dual-mode autonomous learning model, and outputs a water erosion defect grade; A warning correction module is connected with the intelligent identification module, performs hierarchical early warning according to the water erosion defect grade, and performs closed-loop correction on the dual-mode autonomous learning model based on actual detection results.

[0019] The acquisition mode of the state perception module includes: collecting the vibration amplitude and frequency spectrum in the horizontal and vertical directions through vibration sensors installed on the bearing seat and the low-pressure cylinder body, with a sampling frequency of 10 kHz; collecting the low-pressure cylinder inlet steam pressure, temperature, flow, turbine heat rate, low-pressure cylinder exhaust temperature, unit output power and speed fluctuation data through pressure sensors, temperature sensors and flow sensors, with a sampling interval of 1 minute; collecting the sound pressure level in the range of 80-120 dB and the noise spectrum in the frequency band of 500-2000 Hz through a waterproof microphone installed on the outer wall of the low-pressure cylinder; obtaining the rotor axial displacement deviation and the expansion difference change of the cylinder and the rotor through displacement sensors; and inputting historical maintenance data, including blade appearance inspection records and repair records.

[0020] The multi-dimensional feature extraction process of the feature processing module includes: extracting the vibration energy proportion in the frequency band of 100-300 Hz by wavelet packet decomposition, and calculating the kurtosis coefficient of the vibration signal, which is marked as an abnormal vibration feature when the kurtosis value is greater than 3; calculating the cumulative deviation rate of the heat rate by a 24-hour sliding window, i.e. (actual value-design value) / design value, which is marked as an abnormal parameter feature when the deviation rate is greater than 0.5% for 3 consecutive days; extracting the impact noise feature peak near 1500 Hz by Mel cepstrum coefficient to generate a water erosion soundprint feature vector; and forming a 128-dimensional fusion feature vector after normalizing the above features.

[0021] The dual-mode autonomous learning model of the intelligent identification module includes a basic model architecture and a learning mode, and the basic model architecture is a hybrid model of convolutional neural network (CNN) and long short-term memory network (LSTM), wherein: The CNN layer is used to process the frequency spectrum space features of vibration and acoustic signals, including the first layer: 16 3x3 convolution kernels, the second layer: 32 3x3 convolution kernels, using 2x2 maximum pooling, the activation function is ReLU, and the Dropout probability for preventing overfitting is 0.2; the LSTM layer is used to analyze the time series features of parameter trends, including 64 hidden units, a time step of 24, dropout=0.1, and recurrent_dropout=0.1; the output layer is used to predict the water erosion defect level, which is divided into 0-4 levels, 0 level is no damage, 1 level is slight pitting, 2 level is local pit, 3 level is crack initiation, and 4 level is serious damage requiring emergency shutdown.

[0022] The learning mode includes an online federated learning mode, specifically: in a networked state, the multi-unit edge node uploads the model parameters of "fusion feature vector-water erosion defect level" to the cloud server after encryption, and the cloud server aggregates the parameters to update the global model and then distributes it to each edge node.

[0023] The learning mode includes an offline incremental learning mode, which is realized based on the knowledge distillation technology, and the specific steps are: Using the online global model as the teacher model and the local model to be updated as the student model, the probability distribution of water erosion levels output by the teacher model is used as a carrier of historical knowledge. The distillation temperature parameter T=7 is set, and the output distributions of the teacher and student models are calculated using the softening probability formula: Teacher model softening probability: ; Student model softening probability: ; in, , The output logits are for the teacher model and the student model, respectively. Constructing a hybrid loss function: The Adam optimizer is used to train the model for 20-30 epochs until the historical knowledge retention rate on the validation set is >90%. The trained student model is then used to replace the original local model to complete the offline update.

[0024] The graded early warning mechanism of the early warning and correction module includes: when the water erosion defect level is 1, a level 1 early warning is triggered: the data acquisition frequency is increased to 1 minute / time, and the key inspection area in the upper middle part of the steam outlet edge of the blade is pushed. When the water erosion defect level is 2, a level 2 warning is triggered: an online adjustment plan is generated, which includes increasing the steam superheat by 0.5°C and scheduling a shutdown maintenance. When the water erosion defect level is 3 or above, a level 3 early warning is triggered: an audible and visual alarm is activated, the unit load is forcibly limited to 30%, and an emergency repair plan is pushed out.

[0025] The closed-loop correction process of the early warning correction module is as follows: after the unit is shut down for maintenance, the actual water erosion level detected manually is entered into the system; the intelligent identification module calculates the prediction error and corrects the feature weights through the gradient descent algorithm; when the prediction confidence is <80%, it is marked as a sample to be verified, and after connecting to the network, it requests the cloud model for collaborative verification and updates the local model parameters.

[0026] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0027] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring and identifying water erosion defects in steam turbine blades, characterized in that, include: The state perception module is used to collect macroscopic feature data during the operation of the steam turbine. The macroscopic feature data includes vibration signals, operating parameters, acoustic signals, axial displacement and expansion difference data. The feature processing module, connected to the state perception module, performs multi-dimensional feature extraction on the macroscopic feature data and constructs a fused feature vector; The intelligent identification module is connected to the feature processing module and analyzes the fused feature vector through a dual-mode autonomous learning model to output the water erosion defect level. The early warning and correction module is connected to the intelligent recognition module. It performs graded early warning according to the water erosion defect level and performs closed-loop correction on the dual-mode autonomous learning model based on the actual detection results.

2. The method for monitoring and identifying water erosion defects in steam turbine blades according to claim 1, characterized in that, The state awareness module collects data in the following ways: Vibration sensors installed on the bearing housing and low-pressure cylinder body collect horizontal and vertical vibration amplitude and frequency spectrum data at a sampling frequency of 10kHz. Pressure, temperature and flow sensors collect low-pressure cylinder inlet pressure, temperature and flow rate, turbine heat rate, low-pressure cylinder exhaust temperature, unit output power and speed fluctuation data at a sampling interval of 1 minute. A waterproof microphone installed on the outer wall of the low-pressure cylinder collects sound pressure level in the range of 80-120dB and noise spectrum in the frequency band of 500-2000Hz. Displacement sensors acquire rotor axial displacement deviation and expansion difference between the cylinder body and rotor. Historical maintenance data, including blade appearance inspection records and repair records, are entered.

3. The method for monitoring and identifying water erosion defects in steam turbine blades according to claim 1, characterized in that, The multi-dimensional feature extraction process of the feature processing module includes: Wavelet packet decomposition was used to extract the proportion of vibration energy in the 100-300Hz frequency band, and the kurtosis coefficient of the vibration signal was calculated. When the kurtosis value was >3, it was marked as an abnormal vibration feature. A 24-hour sliding window was used to calculate the cumulative deviation rate of heat dissipation rate, i.e. (actual value - design value) / design value. When the deviation rate was >0.5% for 3 consecutive days, it was marked as an abnormal parameter feature. The impact noise feature peak near 1500Hz was extracted by Mel-frequency cepstral coefficients to generate a water erosion acoustic feature vector. After normalizing the above features, a 128-dimensional fused feature vector was formed.

4. The method for monitoring and identifying water erosion defects in steam turbine blades according to claim 1, characterized in that, The dual-mode autonomous learning model of the intelligent recognition module includes a basic model architecture and a learning mode. The basic model architecture is a hybrid model of Convolutional Neural Network (CNN) and Long Short-Term Memory Network (LSTM), wherein: The CNN layer is used to process the spatial features of the spectrum of vibration and acoustic signals, including layer 1: 16 3×3 convolutional kernels and layer 2: 32 3×3 convolutional kernels, using 2×2 max pooling, with ReLU activation function and a Dropout probability of 0.2 to prevent overfitting; the LSTM layer is used to analyze the time series features of parameter trends, including 64 hidden units, a time step of 24, dropout=0.1, and recurrent_dropout=0.1; the output layer is used to predict the level of water erosion defects, which is divided into levels 0-4, where level 0 is no damage, level 1 is slight pitting, level 2 is local pitting, level 3 is crack initiation, and level 4 is severe damage requiring emergency shutdown.

5. The method for monitoring and identifying water erosion defects in steam turbine blades according to claim 4, characterized in that, The learning mode includes an online federated learning mode, which is as follows: when connected to the network, the edge nodes of multiple units encrypt and upload the model parameters of "fusion feature vector - water erosion defect level" to the cloud server. The cloud server aggregates the parameters, updates the global model, and then distributes it to each edge node.

6. The method for monitoring and identifying water erosion defects in steam turbine blades according to claim 4, characterized in that, The learning mode includes an offline incremental learning mode, implemented based on knowledge distillation technology, with the following specific steps: Using the online global model as the teacher model and the local model to be updated as the student model, the probability distribution of water erosion levels output by the teacher model is used as a carrier of historical knowledge. The distillation temperature parameter T=7 is set, and the output distributions of the teacher and student models are calculated using the softening probability formula, which is: Teacher model softening probability: ; Student model softening probability: ; in, , The output logits are for the teacher model and the student model, respectively. Constructing a hybrid loss function: The Adam optimizer is used to train the model for 20-30 epochs until the historical knowledge retention rate on the validation set is >90%. The trained student model is then used to replace the original local model to complete the offline update.

7. The method for monitoring and identifying water erosion defects in steam turbine blades according to claim 1, characterized in that, The tiered early warning mechanism of the early warning correction module includes: When the water erosion defect level is 1, a level 1 warning is triggered: the data acquisition frequency is increased to 1 minute / time, and the key inspection area in the upper middle part of the steam outlet edge of the blade is pushed. When the water erosion defect level is 2, a level 2 warning is triggered: an online adjustment plan is generated, which includes increasing the steam superheat by 0.5°C and scheduling a shutdown maintenance plan. When the water erosion defect level is 3 or above, a level 3 early warning is triggered: an audible and visual alarm is activated, the unit load is forcibly limited to 30%, and an emergency repair plan is pushed out.

8. The method for monitoring and identifying water erosion defects in steam turbine blades according to claim 1, characterized in that, The closed-loop correction process of the early warning correction module is as follows: after the unit is shut down for maintenance, the actual water erosion level detected manually is entered into the system; the intelligent identification module calculates the prediction error and corrects the feature weights through the gradient descent algorithm. When the prediction confidence level is less than 80%, it is marked as a sample to be verified. After connecting to the network, it requests the cloud model for collaborative verification and updates the local model parameters.

Images