Combustion oscillation on-line monitoring method and device based on recursive quantitative analysis

Abstract

The invention discloses a combustion oscillation on-line monitoring method and device based on recursive quantitative analysis. The method comprises the steps that dynamic signals near a combustion chamber are obtained; performing recursive analysis on the dynamic signals in a period of time window, reconstructing a phase space, calculating time delay and embedding dimensions, and constructing a recursive matrix to obtain a recursive plot; extracting features from the recursion plot to obtain recursion quantization features; taking the recursive quantization features as input variables of the neural network, and inputting the recursive quantization features into the trained neural network; and the neural network gives a judgment result of the combustion oscillation state, and triggers an early warning mechanism when an abnormality is detected. According to the method, recursive quantitative analysis and deep learning are combined, online real-time monitoring of the combustion oscillation state is achieved, the precision and real-time performance of combustion oscillation monitoring are improved, efficient recognition and early warning of combustion oscillation are achieved, operation of a combustion system is optimized, combustion stability and safety are improved, the operation and maintenance cost of equipment is reduced, and the working efficiency is improved. And the method has wide engineering application value.

Description

Technical Field

[0001] The present invention belongs to the technical field of aeroengine and gas turbine combustors, and relates to combustion instability monitoring and diagnosis technologies. In particular, it relates to a warning and monitoring method and device for combustion oscillation phenomena, which are used to obtain dynamic signals of the combustor in real time, extract combustion oscillation characteristics based on recurrence quantification analysis, improve the accuracy and adaptability of online monitoring of combustion oscillation, and ensure the stability and safety of the combustion system. Background Art

[0002] In order to make the emission level of nitrogen oxides (NOx) meet increasingly strict environmental standards, lean premixed combustion technology (LPM) is generally adopted in the design of modern aeroengines and gas turbine combustors. However, the lean premixed combustion technology adjusts the equivalence ratio to be close to the extinction boundary, which may lead to combustion oscillation phenomena. The generation mechanism of combustion oscillation is complex and usually involves the coupling effect between acoustics, fluid mechanics and chemical reactions in the combustor. When the unsteadiness of fuel supply, air flow and combustion process in the combustor interact with each other, local flame oscillation may trigger aeroacoustic resonance, thereby leading to combustion oscillation. The occurrence of combustion oscillation can induce significant pressure fluctuations, which may deteriorate the flow field conditions, and then may trigger flame extinction, flashback or an increase in emissions. These effects may seriously affect the service life or efficiency of the combustor. In extreme cases, combustion oscillation may even cause structural vibration, abnormal local temperature rise in the combustor, or damage to the structural integrity. Therefore, it is necessary to develop combustion oscillation warning technologies, combustion oscillation state recognition methods, and formulate effective combustion oscillation suppression strategies.

[0003] The current combustion oscillation warning system mainly relies on the spectrograms of dynamic pressure and vibration acceleration, and judges the combustion state according to the spectral peak or time-frequency change. However, it is difficult to determine the warning threshold of this method, and the false alarm rate and missed alarm rate are high. In addition, by monitoring various sensor signals such as dynamic pressure, exhaust temperature, vibration, etc., and fusing the warning results of multiple sensors, the accuracy of combustion oscillation warning can be improved. However, this method requires a lot of sensors to be arranged, the structure is complex, and it is difficult to determine the weight system of different sensors. In addition, there are also methods using photomultiplier tubes, flame imaging, etc. to monitor combustion oscillation phenomena. However, these methods are limited by the requirements of the equipment installation environment and cannot be applied in engineering.

[0004] Taking the combustion oscillation monitoring device and method disclosed in Chinese invention patent application CN110966100A as an example, this method collects multiple dynamic signals from the combustion chamber and its surroundings, including combustion chamber pressure, high-pressure rotor speed, etc., analyzes them in the time domain and frequency domain, and fuses the indication values ​​of multiple signals for judgment. However, this method relies on manually set thresholds and weighting coefficients, has limited adaptability, and fails to fully utilize the time-frequency characteristics of the signal, making it difficult to identify early combustion oscillations. Another example is the combustion oscillation phenomenon early warning monitoring method and device disclosed in another Chinese invention patent application CN113267291B. This method predicts the combustion chamber pressure signal through a high-order expanded state observer to achieve early warning. However, the prediction accuracy of this method is limited by the phase lag caused by the observer, and the calculation amount is large, which poses challenges in situations with high real-time requirements.

[0005] In summary, the existing combustion oscillation monitoring technology has problems such as difficulty in determining the warning threshold, high missed alarm rate and false alarm rate, complex system, difficulty in sensor layout, and limited engineering application of optical methods. Therefore, how to improve the real-time, adaptability and ease of engineering application of online monitoring of combustion oscillation while ensuring monitoring accuracy has become a technical problem that needs to be urgently solved in the field of combustion oscillation monitoring. Summary of the invention

[0006] 1. Purpose of the invention In view of the above-mentioned defects and shortcomings of the prior art and to solve at least one of the above-mentioned and other technical problems, the present invention aims to provide a method and device for online monitoring of combustion oscillations based on recursive quantitative analysis. The method obtains the dynamic signal near the combustion chamber, constructs a recursive graph by a recursive analysis method, and extracts recursive quantitative features. Then, the trained neural network model is used to perform intelligent judgment on the combustion oscillation state, thereby realizing high-precision and real-time monitoring of combustion oscillations, improving the accuracy, real-time and automation level of combustion chamber combustion oscillation monitoring, and realizing online monitoring and early warning of the combustion oscillation state during the operation of the combustion system, thereby ensuring the safe and stable operation of the combustion chamber.

[0007] (II) Technical solution In order to solve the above technical problems, the present invention proposes a combustion oscillation online monitoring method based on recursive quantitative analysis, which is suitable for monitoring the combustion process of an aircraft engine or a gas turbine combustion chamber to achieve real-time detection and identification of the combustion oscillation state. The method at least includes the following steps: SS1. Dynamic signal acquisition and preprocessing: Acquire dynamic signals near the combustion chamber in real time, and perform denoising and / or standardization processing on the acquired original signals to generate time series signals; SS2. Recursion analysis and recursion graph construction: Perform a recurrence analysis on the dynamic signal time series within a certain time window, reconstruct its phase space, convert the time series into a multi-dimensional phase space delay embedding vector, and calculate the time delay and embedding dimension required for reconstructing the phase space. Then, calculate the recurrence matrix based on the phase space delay embedding vector and form a recurrence plot, where the matrix elements are used to indicate whether there is a temporal recurrence relationship between the states at different time points; SS3. Recurrence quantification feature extraction: Extract recurrence quantification features from the recurrence plot, where the recurrence quantification features include at least one of recurrence rate, determinism, mean of diagonal lengths, longest diagonal, recurrence entropy, laminarity, trapping time, longest vertical line length, recurrence time of the first type, recurrence time of the second type, recurrence period density entropy, clustering coefficient, and transitivity; SS4. Input the recurrence quantification features into a neural network: Use the extracted recurrence quantification features as input variables and input them into a pre-trained neural network for identifying combustion oscillation states. The neural network learns the feature mapping relationship between different oscillation modes based on historical combustion oscillation data. The number of input neural nodes corresponds to the number of recurrence quantification features, and the number of output neural nodes corresponds to the number of combustion oscillation states; SS5. Determine and output the combustion oscillation state: Based on the output result of the neural network, determine the current combustion oscillation state of the combustion chamber, and trigger an early warning mechanism when an abnormal combustion oscillation state is detected.

[0008] To solve the above technical problems, another aspect of the present invention proposes an online monitoring device for combustion oscillation based on recurrence quantification analysis, including: An information acquisition module that real-time collects dynamic signals near the combustion chamber by one or more sensors, and performs denoising and / or standardization processing on the collected signals to generate time series signals; A recurrence quantification analysis module that performs recurrence analysis on the collected time series signals, reconstructs the phase space, converts the time series into a multi-dimensional phase space vector, calculates the time delay and embedding dimension required for reconstructing the phase space, and calculates a recurrence matrix based on the reconstructed phase space vector to obtain a recurrence plot to characterize the state recurrence relationship at different time points; A feature extraction module for extracting recurrence quantification features from the recurrence plot; A pre-trained neural network module for receiving the recurrence quantification features extracted by the feature extraction module and inputting them as input variables into a pre-trained neural network model. The neural network learns the feature mapping relationship between different oscillation modes based on the collected historical combustion oscillation test data, and optimizes and adjusts the neural network model parameters by labeling the historical data; The combustion oscillation state monitoring module determines the current combustion oscillation state of the combustion chamber based on the output results of the pre-trained neural network module, and triggers the warning mechanism when an abnormal combustion oscillation state is detected.

[0009] (III) Technical Effects Compared with the traditional oscillation monitoring methods, the online combustion oscillation monitoring method and device based on recurrence quantification analysis of the present invention have the following advantages: (1) By combining recurrence quantification analysis and neural network technology, the present invention can not only accurately extract the characteristics of the combustion oscillation signal, but also train the neural network based on historical data, so as to realize the online real-time monitoring of the combustion oscillation state.

[0010] (2) Through recurrence quantification analysis, the present invention can extract the deep features of the dynamic signals of the combustion chamber. Combining with the neural network model of deep learning, it can quickly and accurately output the oscillation state, trigger a warning when an abnormal combustion oscillation state is found, meet the requirements of real-time monitoring of aeroengine and gas turbine combustion chambers, contribute to the stable control of the combustion chamber operation state, avoid problems such as flame extinction, flashback, and structural damage caused by combustion oscillation, and improve the reliability and service life of the combustion system.

[0011] (3) The online monitoring method and device of the present invention are applicable to various types of dynamic signals, such as pressure, vibration, sound, etc., can be adapted to different combustion chamber structures and operating conditions, and have strong adaptability and versatility. Through the present invention, the precise monitoring of the combustion process can be realized, thereby optimizing the combustion efficiency, improving the equipment safety, reducing the operation and maintenance cost, and having a wide application prospect. Description of the Drawings

[0012] Figure 1 is the flow chart of the online combustion oscillation monitoring method based on recurrence quantification analysis of the present invention.

[0013] Figure 2 is the recurrence plot of 6 different combustion states. In the figure, (a)-(f) are stable combustion, intermittent combustion oscillation, limit cycle combustion oscillation, difference frequency combustion oscillation, flutter, and two-period combustion oscillation in sequence.

[0014] Figure 3 is the confusion matrix of the combustion oscillation monitoring results. Case1-Case6 represent stable combustion, intermittent combustion oscillation, limit cycle combustion oscillation, difference frequency combustion oscillation, flutter, and two-period combustion oscillation respectively.

[0015] Figure 4 is the schematic diagram of the online combustion oscillation monitoring device based on recurrence quantification analysis of the present invention. Detailed Embodiments

[0016] The present invention aims to provide an online monitoring method and device for combustion oscillation based on recurrence quantification analysis, which is applicable to the combustion process monitoring of aeroengines or gas turbine combustors to achieve real-time detection and identification of the combustion oscillation state. To more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be further described below in conjunction with specific implementation manners and the accompanying drawings. The embodiments described by referring to the accompanying drawings are exemplary and are intended to explain the present invention, and should not be construed as limiting the present invention.

[0017] Embodiment 1: Online monitoring method As a specific example, as Figure 1 shown, the online monitoring method for combustion oscillation based on recurrence quantification analysis provided in this Embodiment 1 mainly includes the following steps when implemented: SS1. Dynamic signal acquisition and preprocessing: Obtain the dynamic signals near the combustor, and perform denoising and / or normalization processing on the collected original signals to generate time series signals.

[0018] Optionally, the dynamic signals can be collected in real time by sensors arranged at different positions of the combustor, which can be at least one of dynamic pressure signals, vibration acceleration signals, vibration velocity signals, sound signals, etc. The original signal processing includes using a low-pass filter or a band-pass filter for denoising to remove high-frequency noise or irrelevant frequency components in the signal, and performing normalization processing on the signal to eliminate the influence of different dimensions and orders of magnitude.

[0019] SS2. Recurrence analysis and recurrence plot construction: Take the time series of dynamic signals within a time window for recurrence analysis, reconstruct its phase space, convert the time series into a multi-dimensional phase space delay embedding vector, and calculate the time delay and embedding dimension required for reconstructing the phase space. Then, calculate the recurrence matrix based on the phase space delay embedding vector to obtain a recurrence plot, and the matrix elements are used to indicate whether there is a temporal recurrence relationship between the states at different time points.

[0020] Specifically, the time series of dynamic signals within a time window is expressed as x = x 1, x 2, x 3,..., x N . Perform recurrence analysis on the time series of dynamic signals to construct a new vector, that is, the phase space delay embedding vector X ( i )= x ( i ), x ( i + τ ),x ( i +2 τ ),…, x ( i +( d -1) τ )] maps the one-dimensional time series into the d -dimensional phase space, where τ represents the time delay, d is the embedding dimension, and i ranges from 1 to N -( d -1) τ , N is the length of the time series. The optimal time delay τ required for reconstructing the phase space is calculated using the autocorrelation function method or the average mutual information method, and the embedding dimension d required for reconstructing the phase space is calculated using the false nearest neighbor method. The dimension increase stops when the false nearest neighbor ratio is less than the preset threshold.

[0021] The recurrence plot presents as a N × N matrix structure, and its matrix element R i,j is used to indicate whether there is a recurrence relationship between the states of time i and time j . Specifically, if R i,j equals 1, it indicates that the states of these two time points are close to each other and are presented as corresponding marked points in the graph; conversely, if R i,j equals 0, it means that the states of these two time points are not close enough, so they are not marked in the graph.

[0022] The calculation method of the matrix element R i,j is: R i,j ( ε ) = Θ( ε - || X i - X j ||), where Θ is the Heaviside step function, ε is the threshold parameter and is adaptively adjusted based on the standard deviation of the time series signal, || X i - X j || is the Euclidean distance between the vectors X i and X j .

[0023] The recurrence plots of six different combustion states are as Figure 2 shown, Figure 2 where (a) - (f) are stable combustion, intermittent combustion oscillation, limit cycle combustion oscillation, difference frequency combustion oscillation, beat vibration, and two - period combustion oscillation in sequence. From the structural characteristics of the recurrence plots, it can be observed that different combustion oscillation states present significantly different spatio - temporal structure patterns. Among them, stable combustion shows a random dot - like distribution, while other oscillation states exhibit their own unique periodicity and diagonal structure characteristics.

[0024] SS3. Recurrence Quantification Feature Extraction: Extract features from the recurrence plot to obtain recurrence quantification features. Optionally, the recurrence quantification features include: recurrence rate, determinism, mean diagonal length, longest diagonal, recurrence entropy, laminarity, trapping time, longest vertical line length, recurrence time of the first type, recurrence time of the second type, recurrence period density entropy, clustering coefficient, and transitivity.

[0025] Specifically, among the above recurrence quantification features, the recurrence rate represents the proportion of recurrence points and is used to measure global repeatability; determinism reflects the predictability of the system state; the mean diagonal length measures state persistence; the longest diagonal evaluates maximum stability; recurrence entropy describes complexity; laminarity characterizes the hierarchical structure; trapping time measures the state residence time; the longest vertical line length measures system stability; recurrence time evaluates the state repetition interval; recurrence period density entropy analyzes periodic characteristics; the clustering coefficient and transitivity reflect the correlation between states.

[0026] Recurrence rate RR It is obtained by calculating the ratio of the total number of recurrence points in the recurrence matrix to the total number of elements in the matrix and is used to measure the global repeatability of the time series. Its calculation formula is: where, is the recurrence matrix, is the length of the time series.

[0027] Determinism DET It is obtained by calculating the ratio of the total diagonal length in the recurrence plot to the total number of recurrence points. This index can be used to measure the predictability of the system state. Its calculation formula is: where, P ( l ) is the number of diagonals with length l , l min is the minimum diagonal length considered.

[0028] Diagonal lengthL ave The mean value is obtained by calculating the average length of all valid diagonals, which can reflect the average duration of the system state. Its calculation formula is: The longest diagonal L max It is obtained by traversing all diagonal lengths and selecting the maximum value. This feature can be used to evaluate the maximum stability of the system state. Its calculation formula is:

[0029] Recurrence entropy ENTR It is obtained by calculating the probability distribution of the lengths of each diagonal in the recurrence plot and is used to measure the complexity of the recurrence pattern. Its calculation formula is: Among them, is the proportion of diagonals with length l in all diagonals.

[0030] Layered degree LAM It is obtained by calculating the ratio of the total length of vertical lines to the total number of recurrence points in the recurrence plot and can be used to characterize the hierarchical structure of the system state. Its calculation formula is: Among them, is the number of vertical lines with length in the recurrence plot.

[0031] Capture time TT It is obtained by calculating the weighted average of the lengths of vertical lines and is used to characterize the persistence of the system state. Its calculation formula is: The length of the longest vertical line P max It is obtained by finding the longest continuous vertical line structure and is used to measure the stability and residence time of the system state. Its calculation formula is: The recurrence time of the first type It is obtained by calculating the time interval for the system state to repeat in the same recurrence region. Its calculation formula is: The recurrence time of the second type It is used to measure the time interval from repetition to non-repetition of the state: Recurrence period density entropy RPDEObtained by calculating the probability distribution of recurrence periods in the recurrence plot, it can be used to analyze the nonlinear periodic characteristics of the system, and its calculation formula is: Clustering coefficient C 1 is used to characterize the local clustering of the system state, and its calculation formula is: Transitivity C 2 is used to characterize the nonlinear transfer relationship between states in the system, and its calculation formula is: SS4. Input the recurrence quantification features into the neural network: Take the extracted recurrence quantification features as the input variables of the neural network and input them into the trained neural network. This neural network learns the feature mapping relationship between different oscillation modes based on historical combustion oscillation data. The number of input neural nodes corresponds to the number of recurrence quantification features, and the number of output neural nodes corresponds to the number of combustion oscillation states.

[0032] Optionally, the pre-trained neural network adopts a deep belief network structure, and uses the recurrence quantification features and the corresponding combustion states as the input and output variables of the neural network for neural network training. The recurrence quantification features for the neural network input variables can be one or more of recurrence rate, determinism, mean of diagonal lengths, longest diagonal, recurrence entropy, laminarity, capture time, longest vertical line length, recurrence time of the first type, recurrence time of the second type, recurrence period density entropy, clustering coefficient, and transitivity. The number of input neural nodes is determined according to the number of extracted quantified recurrence features. In this embodiment, the number of input neural nodes is equal to 13. The number of output neural nodes is equal to the number of combustion states. In this embodiment, the number of output neural nodes is equal to 6. The number of hidden layers is set to 3, the number of nodes in each layer is set to 12, the learning rate is set to 0.02, and the backpropagation algorithm is used to optimize the weights and structure parameters during the training process.

[0033] SS5. Combustion oscillation state determination and output: Based on the output result of the neural network, determine the current combustion oscillation state of the combustion chamber, and trigger an early warning mechanism when an abnormal combustion oscillation state is detected.

[0034] Optionally, the determined combustion oscillation states include at least one of different oscillation states such as stable combustion, intermittent combustion oscillation, limit cycle combustion oscillation, difference frequency combustion oscillation, flutter, and two-period combustion oscillation. Each combustion oscillation state is marked by historical test data, and the recognition accuracy of the neural network is evaluated through a confusion matrix to ensure the accuracy of the monitoring results; and when the output result of the neural network is one of the above combustion oscillation states, a hierarchical early warning mechanism is triggered.

[0035] Specifically, as Figure 3 shown in the confusion matrix of the combustion oscillation monitoring results, Case1-Case6 represent stable combustion, intermittent combustion oscillation, limit cycle combustion oscillation, difference frequency combustion oscillation, flutter, and two-period combustion oscillation respectively. For the 6 combustion states, there are 581 sample data for each combustion state, and the number of correctly classified samples is 581, 580, 580, 580, 579, and 580 respectively. The overall monitoring accuracy of the combustion oscillation state is 99.83%. This result indicates that the combustion oscillation monitoring method based on recurrence quantification analysis and neural network of the present invention has extremely high accuracy and reliability in identifying different combustion oscillation states, and can effectively meet the real-time monitoring and early warning requirements of the combustion system.

[0036] Embodiment 2: On-line monitoring device Based on the combustion oscillation on-line monitoring method shown in the above Embodiment 1, in order to further improve its practicability and realize engineering application, as another preferred example, Figure 4 shows the combustion oscillation on-line monitoring device based on recurrence quantification analysis of the present invention, including: An information acquisition module 10, which collects signals such as dynamic pressure, vibration, and sound near the combustion chamber by one or more sensors, and performs denoising and / or standardization processing on the signals. The denoising processing can adopt methods such as low-pass filtering, band-pass filtering, wavelet transform denoising, or median filtering to remove high-frequency noise or irrelevant frequency components and improve the signal-to-noise ratio of the signal. The standardization processing is used to eliminate the dimension difference of different measurement signals, ensure the stability of the input data, provide input data with stronger consistency for subsequent recurrence quantification analysis, and finally generate a time series signal; A recurrence quantification analysis module 20, which performs recurrence analysis on the collected dynamic signal time series, reconstructs the phase space, converts the time series into a multi-dimensional phase space vector, calculates the time delay and embedding dimension required for reconstructing the phase space, calculates a recurrence matrix based on the reconstructed phase space vector, obtains a recurrence plot, and performs quantization processing on the recurrence plot to characterize the state recurrence relationship at different time points. During the phase space reconstruction process, this module calculates the time delay and embedding dimension. The time delay can be determined by the autocorrelation function method or the average mutual information method, and the embedding dimension can be determined by the false nearest neighbor method. Subsequently, this module calculates a recurrence matrix based on the reconstructed phase space vector and generates a recurrence plot accordingly, where each element of the recurrence matrix is used to indicate whether there is a temporal recurrence relationship between the states at different time points. In order to improve the accuracy of data analysis, this module performs quantization processing on the recurrence plot and optimizes its threshold parameters to enable it to more accurately characterize the time evolution pattern of combustion oscillation.

[0037] The feature extraction module 30 is used to extract recurrence quantification features from the recurrence plot, including at least one of recurrence rate, determinism, mean of diagonal length, longest diagonal, recurrence entropy, laminarity, capture time, longest vertical line length, recurrence time of the first type, recurrence time of the second type, recurrence period density entropy, clustering coefficient, and transitivity.

[0038] Among the recurrence quantification features, the recurrence rate is used to measure the repeatability of states in the time series; determinism reflects the predictability of the signal; the mean of diagonal length and the longest diagonal are used to evaluate the stability of the system; recurrence entropy is used to measure the dynamic complexity of the system; laminarity and capture time can analyze the hierarchy and stability of the combustion process; and indicators such as recurrence time, recurrence period density entropy, clustering coefficient, and transitivity are used to deeply analyze the periodicity, nonlinear characteristics, and interaction relationships of the oscillatory signal.

[0039] The pre-trained neural network module 40 is used to receive the recurrence quantification features extracted by the feature extraction module and input them as input variables into the pre-trained neural network model. The neural network learns the feature mapping relationship between different oscillation modes based on the collected historical combustion oscillation test data, and optimizes and adjusts the neural network model parameters by labeling the historical data.

[0040] During the training process, first, the historical data is labeled to ensure that different combustion states correspond to different categories; then, it is trained through deep learning models such as deep belief networks or convolutional neural networks to optimize the structural parameters and weights of the neural network. The number of neural nodes in the input layer of the neural network depends on the number of recurrence quantification features extracted, and the number of neural nodes in the output layer corresponds to the number of categories of combustion oscillation states. In this embodiment, the number of input neural nodes is equal to 13, and the number of output neural nodes is equal to 6, corresponding to 6 combustion oscillation states. The training of the neural network uses the backpropagation algorithm to optimize the weights and sets hyperparameters such as learning rate, regularization parameter, number of hidden layers, and number of neurons to improve the generalization ability and classification accuracy of the model.

[0041] The combustion oscillation state monitoring module 50 uses the output result of the trained neural network model to give the combustion oscillation state and trigger the warning mechanism when an abnormal combustion oscillation state is detected. The determination results of the combustion oscillation state include at least one of different combustion states such as stable combustion, intermittent combustion oscillation, limit cycle combustion oscillation, difference frequency combustion oscillation, flutter, and two-period combustion oscillation. This module further combines historical test data with the confusion matrix to evaluate the recognition accuracy of the neural network and adopts a hierarchical warning mechanism to set different levels of alarm thresholds according to the degree of combustion oscillation. When an abnormal combustion oscillation is detected, this module can issue a warning through various feedback settings to ensure the safe and stable operation of the combustion system, improve the combustion efficiency, and reduce the equipment maintenance cost.

[0042] In summary, the online combustion oscillation monitoring device provided in this Embodiment 2 realizes high-precision and real-time monitoring and early warning of the combustion oscillation state by integrating information acquisition, recurrence quantification analysis, feature extraction, neural network recognition, and oscillation state monitoring modules.

[0043] Through the above embodiments, the object of the present invention is fully and effectively achieved. Those skilled in the art can understand that the present invention includes but is not limited to the content described in the drawings and the above specific embodiments. Although the present invention has been described with reference to the currently considered most practical and preferred embodiments, it should be understood that the present invention is not limited to the disclosed embodiments, and any modifications that do not deviate from the functional and structural principles of the present invention will be included within the scope of the claims.

Claims

1. A combustion oscillation online monitoring method based on recursive quantitative analysis, characterized in that: The online monitoring method comprises at least the following steps: SS1. Real-time acquisition of dynamic signals near the combustion chamber, and denoising and / or standardization of the acquired raw signals to generate time series signals; SS2. Take the dynamic signal time series within a time window for recursive analysis, reconstruct its phase space, convert the time series into a multidimensional phase space delay embedding vector, and calculate the time delay and embedding dimension required to reconstruct the phase space. Then, calculate the recursive matrix based on the delay embedding vector and form a recursive graph. The matrix elements are used to indicate whether there is a temporal recursive relationship between the states at different time points; SS3. Extract recursive quantitative features from recursive graphs; SS4. The extracted recursive quantization features are used as input variables and input into a pre-trained neural network for identifying combustion oscillation states, wherein the neural network learns feature mapping relationships between different oscillation modes based on historical combustion oscillation data, wherein the number of input neural nodes corresponds to the number of recursive quantization features, and the number of output neural nodes corresponds to the number of combustion oscillation states; SS5. Based on the output of the neural network, the current combustion oscillation state of the combustion chamber is determined, and when an abnormal combustion oscillation state is detected, an early warning mechanism is triggered.

2. The method according to claim 1, characterized in that In the above step SS1, dynamic signals are acquired by sensors arranged at different positions in the combustion chamber, and the acquired dynamic signals include at least one of dynamic pressure signals, vibration acceleration signals, vibration velocity signals, and sound signals; the original signal processing includes denoising using a low-pass filter or a band-pass filter to remove high-frequency noise or irrelevant frequency components in the signal, and standardizing the signal.

3. The method according to claim 1, characterized in that In the above step SS2, the recursive analysis and recursive graph construction include the following sub-steps: SS21. Dynamic signal time series based on the intercepted time window x =[ x 1, x 2, x 3,..., x N ], construct the phase space delay embedding vector X ( i )=[ x ( i ), x ( i + τ ), x ( i +2 τ ),…, x ( i +( d -1) τ )], mapping the one-dimensional time series to d dimensional phase space, where τ is the time delay parameter, d is the phase space embedding dimension, i The value range is 1 to N -( d -1) τ , N is the length of the time series; SS22. Delay embedding vector construction based on phase space N × N -dimensional recursive matrix, matrix elements R i,j pass R i,j ( ε )=Θ( ε -|| X i - X j ||) calculation, where i , j =1,2,..., N , Θ is the Heaviside step function, ε is the threshold parameter, || X i - X j || represents a vector X i and X j If the Euclidean distance between R i,j =1, indicating that the two states are close to each other and are marked. R i,j =0, indicating that the states are not close enough and are not marked; A recursion graph is constructed based on the recursion matrix, and the recursion graph is used to show the state recursion relationship at different time points.

4. The method according to claim 3, characterized in that In the above step SS2, the time delay parameter τ The autocorrelation function method or the average mutual information method is used for calculation, and the embedding dimension d Determined by the false neighbor method, when the false neighbor ratio is less than the preset threshold, the dimension is stopped from being increased; the threshold parameter ε Adaptive adjustment is performed based on the standard deviation of the time series signal.

5. The method according to claim 1, characterized in that In the above step SS3, the recursive quantization features include at least one of recursion rate, certainty, mean of diagonal length, longest diagonal line, recursion entropy, layeredness, capture time, longest vertical line length, first type recurrence time, second type recurrence time, recursive period density entropy, clustering coefficient, and transitivity.

6. The method according to claim 5, characterized in that The recursion rate is obtained by calculating the ratio of the total number of recursion points in the recursion matrix to the total number of matrix elements; the certainty is obtained by calculating the ratio of the total length of the diagonal structure in the recursion graph to the total number of recursion points; the mean diagonal length is obtained by calculating the average length of all valid diagonals; the recursion entropy is obtained by calculating the probability distribution of the length of each diagonal line in the recursion graph; the layered degree is obtained by calculating the ratio of the total length of the vertical lines in the recursion graph to the total number of recursion points.

7. The method according to claim 5, characterized in that The capture time is obtained by calculating the weighted average of the vertical line lengths; the longest vertical line length is obtained by finding the longest continuous vertical line structure; the longest diagonal line is obtained by traversing all diagonal line lengths and selecting the maximum value; the first type and second type recurrence times are obtained by calculating the first and conditional recurrence times of the midpoint in the state space, respectively; the recursive cycle density entropy is obtained by calculating the probability distribution of the recursive cycle in the recursive graph; the clustering coefficient and transitivity are obtained by analyzing the local density and connectivity of the recursive points.

8. The method according to claim 1, characterized in that In the above step SS4, the neural network adopts a deep belief network structure, and uses recursive quantization features and corresponding combustion states as input and output variables of the neural network for neural network training. Its structure includes an input layer, 3 hidden layers and an output layer. The number of nodes in each hidden layer is set to 12, and the learning rate is set to 0.

02. The back propagation algorithm is used to optimize the weights and structural parameters during the training process.

9. The method according to claim 1, characterized in that: In the above step SS5, the determined combustion oscillation state includes at least one of stable combustion, intermittent combustion oscillation, limit cycle combustion oscillation, difference frequency combustion oscillation, beat oscillation and two-cycle combustion oscillation, wherein each combustion oscillation state is marked by historical test data, and the recognition accuracy of the neural network is evaluated by a confusion matrix to ensure the accuracy of the monitoring results; and when the output result of the neural network is one of the above combustion oscillation states, a graded early warning mechanism is triggered.

10. A combustion oscillation online monitoring device based on recursive quantitative analysis, characterized in that: The online monitoring device at least comprises: An information acquisition module, which uses a single or multiple sensors to collect dynamic signals near the combustion chamber in real time, and performs denoising and / or standardization processing on the collected signals to generate time series signals; A recursive quantitative analysis module performs recursive analysis on the collected time series signals, reconstructs the phase space, converts the time series into a multidimensional phase space vector, and calculates the time delay and embedding dimension required for the reconstructed phase space. The recursive matrix is ​​calculated based on the reconstructed phase space vector to obtain a recursive graph to characterize the state recursive relationship at different time points. A feature extraction module, used to extract recursive quantitative features from the recursive graph; A pre-trained neural network module is used to receive the recursive quantization features extracted by the feature extraction module and input them as input variables into a pre-trained neural network model, wherein the neural network learns the feature mapping relationship between different oscillation modes based on the collected historical combustion oscillation test data, and optimizes and adjusts the parameters of the neural network model by marking the historical data; The combustion oscillation state monitoring module determines the current combustion oscillation state of the combustion chamber based on the output results of the pre-trained neural network module, and triggers the early warning mechanism when an abnormal combustion oscillation state is detected.

Images