Combustion chamber unstable combustion state monitoring method and system based on acoustic sensor

Abstract

The invention relates to the technical field of acoustic state detection, and discloses a combustion chamber unstable combustion state monitoring method and system based on an acoustic sensor, and the method comprises the steps: collecting a pressure time domain signal in a combustion chamber through an acoustic pressure sensor, and calculating a frequency segmentation threshold value; generating a low-frequency spectrogram lower than the frequency segmentation threshold, generating a high-frequency spectrogram higher than the frequency segmentation threshold, and obtaining a fused spectrogram; generating an entropy weight fusion spectrogram; constructing a three-channel feature tensor, and inputting the three-channel feature tensor into a convolutional neural network to obtain a preliminary classification probability of a combustion state; setting a first energy variance threshold value and a second energy variance threshold value, calculating an energy variance, and when the energy variance is lower than the first energy variance threshold value, enabling the combustion state to be stable; and when the energy variance is higher than the second energy variance threshold, the system is unstable. Noise interference can be inhibited, and the accuracy of an unstable combustion state monitoring result is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of acoustic state detection, in particular to a combustion chamber unstable combustion state monitoring method and system based on an acoustic sensor. BACKGROUND

[0002] Thermal equipment such as aero-engines and gas turbines are threatened by unstable combustion in the combustion chamber during operation. The pressure oscillation generated by this phenomenon can induce severe structural vibration, leading to component fatigue damage, and even catastrophic accidents. Current monitoring methods mainly fall into two categories: one is classical signal analysis: relying on time-domain analysis of the pressure signal, such as calculating the root mean square value and peak value; or relying on frequency-domain analysis of the pressure signal, such as determining the dominant frequency through fast Fourier transform. The second is time-frequency analysis: using short-time Fourier transform (STFT) to generate a spectrogram to visually display the distribution of signal energy over time and frequency.

[0003] However, the above current monitoring methods have the following limitations: time-frequency resolution contradiction of short-time Fourier transform: short-time Fourier transform is subject to the Heisenberg uncertainty principle, resulting in a mutual constraint between time resolution and frequency resolution. The selection of a single window function cannot simultaneously consider the accurate characterization of low-frequency oscillation in the early stage of unstable combustion and high-frequency transient events such as deflagration. Insufficient reliability of deep learning: although inputting a spectrogram into a convolutional neural network (CNN) for feature extraction and pattern recognition has become a trend, it still faces the following challenges: one is feature flooding, and key weak feature information may be flooded or distorted. Two is unstable decision: existing methods excessively rely on the classification probability of the convolutional neural network or simple fixed thresholds for diagnosis, making it difficult to adapt to complex variable operating conditions, and the model output is prone to misjudgment in operating conditions not covered by the training samples. SUMMARY

[0004] The present application provides a combustion chamber unstable combustion state monitoring method and system based on an acoustic sensor to solve the problem of low accuracy of unstable combustion state monitoring results in the prior art.

[0005] In a first aspect, the present application provides a combustion chamber unstable combustion state monitoring method based on an acoustic sensor, comprising the following steps: The pressure time domain signal in the combustion chamber is collected by an acoustic pressure sensor, and is divided according to a preset analysis period, and the spectral centroid of the signal in each analysis period is calculated as the frequency segmentation threshold of the corresponding analysis period; the pressure time domain signal is respectively subjected to short-time Fourier transform by using a first window function and a second window function; a low-frequency spectrogram lower than the frequency segmentation threshold is generated based on the transform result of the first window function, and a high-frequency spectrogram higher than the frequency segmentation threshold is generated based on the transform result of the second window function, and the low-frequency spectrogram and the high-frequency spectrogram are spliced to obtain a fusion spectrogram; the spectral entropy of the fusion spectrogram on each time frame is calculated to form a spectral entropy time sequence, the amplitudes of each time frame are weighted according to the principle that the spectral entropy value is inversely proportional to the weight coefficient, and an entropy-weighted fusion spectrogram is generated; the entropy-weighted fusion spectrogram, a first-order difference graph along the time axis and a first-order difference graph along the frequency axis are constructed into a three-channel feature tensor, including: the entropy-weighted fusion spectrogram is taken as a first channel; the first-order difference graph along the time axis is generated by calculating the difference value between adjacent frames of the entropy-weighted fusion spectrogram on the time axis, and the first-order difference graph along the time axis is taken as a second channel; the first-order difference graph along the frequency axis is generated by calculating the difference value between adjacent frequency points of the entropy-weighted fusion spectrogram on the frequency axis, and the first-order difference graph along the frequency axis is taken as a third channel; the first channel, the second channel and the third channel are stacked in order to form a three-channel feature tensor, the three-channel feature tensor is input into a convolutional neural network, and a preliminary classification probability of the combustion state is obtained; the first-order difference absolute value mean of the spectral entropy time sequence is calculated as a spectral entropy change rate, a first energy variance threshold and a second energy variance threshold are set, the first energy variance threshold is inversely proportional to the spectral entropy change rate, and the second energy variance threshold is proportional to the spectral entropy change rate; the energy variance of the pressure time domain signal in the analysis period is calculated, when the energy variance is lower than the first energy variance threshold, the combustion state is stable; when the energy variance is higher than the second energy variance threshold, the combustion state is unstable; when the energy variance is between the first energy variance threshold and the second energy variance threshold, the combustion state is determined according to the maximum value of the preliminary classification probability.

[0006] Preferably, the first window function and the second window function are both Hanning windows.

[0007] Preferably, the time length of the first window function is 100 milliseconds, and the time length of the second window function is 25 milliseconds.

[0008] Preferably, the overlap rate of the first window function and the second window function is 50%.

[0009] Preferably, the amplitudes of each time frame are weighted according to the principle that the spectral entropy value is inversely proportional to the weight coefficient to generate the entropy-weighted fusion spectrogram, including: the weight coefficient is calculated by the following formula, ; wherein i represents the i th time frame of the fusion spectrogram, is a preset weighting adjustment coefficient, is the spectral entropy value of the i-th time frame; is the weight coefficient of the i-th time frame; The amplitude of each time frame is weighted by the following formula, ; wherein, is the weighted amplitude of the i-th time frame, is the amplitude before weighting of the i-th time frame, is the weight coefficient of the i-th time frame.

[0010] Preferably, the first energy variance threshold is calculated by the following formula, ; wherein, is the first energy variance threshold, is a positive real constant calibrated according to experimental data, is the spectral entropy change rate, is a small normal number to prevent the denominator from being zero, is set to 1e-6.

[0011] Preferably, the second energy variance threshold is calculated by the following formula, ; wherein, is the second energy variance threshold, is the spectral entropy change rate, , is a positive real constant calibrated according to experimental data.

[0012] Preferably, when the pressure time domain signal in the combustion chamber is collected by the acoustic pressure sensor, the preset analysis period is divided, including: installing the acoustic pressure sensor on the wall surface of the combustion chamber to collect the pressure signal, and taking 1 second as the length of an analysis period.

[0013] In the second aspect, the combustion chamber unstable combustion state monitoring system based on an acoustic sensor of the present application comprises a memory and a processor, the memory stores computer instructions, and the processor executes the computer instructions to realize the above-mentioned combustion chamber unstable combustion state monitoring method based on an acoustic sensor.

[0014] The beneficial effects of the present application are: the present application adopts a frequency segmentation strategy based on signal spectrum centroid, and combines long and short window functions to process low and high frequency components respectively, constructs a fusion spectrogram considering frequency resolution and time resolution, and can reveal key time-frequency details in the evolution process of the combustion state. The spectrogram is weighted using spectral entropy, highlighting the non-stationary characteristics and suppressing noise interference. The weighted spectrogram and its first-order difference graph are constructed as a three-channel feature tensor, providing more abundant and discriminative input information for the convolutional neural network. The preliminary judgment of the model is combined with an energy variance interval set according to the physical characteristics of the signal, and a reliable criterion is adopted for judgment, relying on model classification in ambiguous cases, and enhancing the reliability of the monitoring results. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 A flowchart of a combustion chamber unstable combustion state monitoring method based on an acoustic sensor is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0016] The embodiments of the present application will be described in detail below, examples of which are shown in the accompanying drawings. The embodiments described below by reference to the drawings are exemplary and are intended to explain the present application, and cannot be understood as a limitation of the present application.

[0017] As shown in Figure 1 , the embodiments of the combustion chamber unstable combustion state monitoring method based on an acoustic sensor provided by the present application include the following steps: S1, acquiring a pressure time-domain signal in the combustion chamber through an acoustic pressure sensor, dividing according to a preset analysis period, and calculating the spectral centroid of the signal in each analysis period as the frequency segmentation threshold of the corresponding analysis period.

[0018] Specifically, an acoustic pressure sensor is installed on the wall of the combustion chamber to collect pressure signals, and 1 second is taken as the length of an analysis period; the fast Fourier transform of the pressure time-domain signal in the analysis period is performed to obtain its frequency spectrum , the spectral centroid of the signal in the analysis period is calculated according to the formula , and the value of the spectral centroid is taken as the frequency segmentation threshold for subsequent spectrogram segmentation, wherein represents the frequency point.

[0019] S2, performing short-time Fourier transform on the pressure time-domain signal using a first window function and a second window function respectively; generating a low-frequency spectrogram below the frequency segmentation threshold based on the transformation result of the first window function, generating a high-frequency spectrogram above the frequency segmentation threshold based on the transformation result of the second window function, and splicing the low-frequency spectrogram and the high-frequency spectrogram to obtain a fusion spectrogram.

[0020] In an optional embodiment, both the first window function and the second window function are Hanning windows; the duration of the first window function is 100 milliseconds, and the duration of the second window function is 25 milliseconds; the overlap rate of both the first window function and the second window function is 50%.

[0021] The pressure signal in the combustion chamber contains steady-state and transient components. Steady-state components change slowly, such as the continuous resonance of the equipment; transient components change rapidly, such as the impact of valve opening and closing. To effectively analyze these two types of characteristics, this invention employs two window functions of different lengths to achieve a balance between time and frequency resolution. Long window functions provide high frequency resolution but low time resolution, while short window functions offer the opposite.

[0022] For example, taking a pressure signal with a sampling rate of 1000 Hz as an example, the first window function's time length is set to 100 milliseconds, i.e., 100 sampling points. After short-time Fourier transform, it can distinguish similar frequency components such as 10 Hz and 20 Hz, which is suitable for analyzing the steady-state operating frequency of the equipment. The second window function's time length is set to 25 milliseconds, i.e., 25 sampling points. It can locate the moment of signal abrupt change at a higher time, such as a pressure pulse with a duration of only 30 milliseconds. The Hanning window is used to reduce spectral leakage and ensure that energy is concentrated in the main lobe, while the 50% overlap rate ensures a smooth transition between time frames and avoids the loss of boundary information caused by windowing.

[0023] Specifically, from the transformation result of the first window function, the portion with a frequency lower than the previously calculated frequency segmentation threshold is extracted as a low-frequency spectrogram; from the transformation result of the second window function, the portion with a frequency higher than the frequency segmentation threshold is extracted as a high-frequency spectrogram; the low-frequency spectrogram and the high-frequency spectrogram are then stitched together along the frequency axis to form a fused spectrogram with high frequency resolution in the low-frequency band and high temporal resolution in the high-frequency band.

[0024] S3. Calculate the spectral entropy of the fused spectrogram at each time frame to form a spectral entropy time series. Based on the principle that the spectral entropy value is inversely proportional to the weight coefficient, weight the amplitude of each time frame to generate an entropy-weighted fused spectrogram. Construct the entropy-weighted fused spectrogram, the first-order difference map along the time axis, and the first-order difference map along the frequency axis into a three-channel feature tensor, and input it into a convolutional neural network to obtain the preliminary classification probability of the combustion state.

[0025] Specifically, for each time frame in the fused spectrogram, the amplitude spectrum is normalized to obtain the probability distribution. Then according to the formula Calculate the spectral entropy H for all time frames, and the spectral entropy of all time frames constitutes the spectral entropy time series.

[0026] In an optional embodiment, the step of weighting the amplitude of each time frame according to the principle that the spectral entropy value is inversely proportional to the weighting coefficient to generate an entropy-weighted fused spectrogram includes: the weighting coefficient is calculated using the following formula, Where i represents the i-th time frame of the fused spectrogram, The preset weighted adjustment coefficient, Let be the spectral entropy value of the i-th time frame; The weight coefficients for the i-th time frame; The amplitude values ​​of each time frame are weighted using the following formula. ;in, The weighted amplitude for the i-th time frame. The magnitude before weighting in the i-th time frame. represents the weighting coefficient for the i-th time frame.

[0027] Spectral entropy is an indicator of the uniformity of energy distribution in a signal's spectrum. When signal energy is concentrated at a few frequency points, such as pure tones or harmonics, the spectral entropy value is low, indicating a large amount of information in the signal. Conversely, when signal energy is evenly distributed across the entire frequency band, such as white noise, the spectral entropy value is high, indicating strong randomness in the signal. Utilizing this characteristic, spectral entropy is mapped to weighting coefficients using an exponential function; the higher the spectral entropy, the lower the weighting, and vice versa. By calculating the weighting coefficient for each time frame and multiplying all amplitude points within each time frame by their corresponding weighting coefficient, the entire fused spectrogram is weighted, resulting in an entropy-weighted fused spectrogram, thereby suppressing disordered components similar to white noise.

[0028] For example, suppose at a certain time i, the signal exhibits clear equipment fault characteristic frequencies, and the calculated spectral entropy H is 0.3. Then the weighting coefficient W is 0.64, and the amplitude M of this time frame is multiplied by a larger weight, thus enhancing it in the spectrogram. At another time j, the signal mainly consists of random environmental noise, and the calculated spectral entropy H is 0.9. Therefore, its weighting coefficient W is 0.26, and the amplitude M of this time frame is multiplied by a smaller weight, thus suppressing it. Entropy-weighted fusion spectrograms can automatically highlight the effective components of the signal while reducing noise interference, improving the signal-to-noise ratio of the features.

[0029] To expand the static features of the original two-dimensional spectrogram into multi-dimensional features containing change information, and to adapt to deep learning models capable of handling three-channel input, in an optional embodiment, the construction of a three-channel feature tensor from the entropy-weighted fused spectrogram, the first-order difference map along the time axis, and the first-order difference map along the frequency axis includes: using the entropy-weighted fused spectrogram as the first channel; calculating the difference between adjacent frames of the entropy-weighted fused spectrogram on the time axis to generate the first-order difference map along the time axis, which serves as the second channel; calculating the difference between adjacent frequency points of the entropy-weighted fused spectrogram on the frequency axis to generate the first-order difference map along the frequency axis, which serves as the third channel; and stacking the first, second, and third channels sequentially to form the three-channel feature tensor, the shape of which is (height, width, 3).

[0030] Entropy-weighted fusion spectrograms reflect the energy distribution at various times and frequencies, preserving core static information as the first channel. The first-order difference plot on the time axis represents the rate of change of spectral energy over time, enhancing the transient characteristics of the signal, such as sudden increases or decreases in the energy of a frequency component, which is crucial for detecting impact failure events. The first-order difference plot on the frequency axis reflects the local contour of the spectrum and the relationship between energy in adjacent frequency bands, helping to describe the harmonic structure or spectral slope, and providing important information for distinguishing the operating states of different devices. For example, a spectrogram with a height of 257 frequency points and a width of 300 time frames will generate first-order difference plots on the time axis and frequency axis of the same size, stacked into a 257*300*3 three-channel feature tensor, similar to an RGB color image. This three-channel feature tensor is then input into a pre-trained convolutional neural network containing multiple convolutional layers, pooling layers, and fully connected layers. At the output layer, the Softmax function is used to obtain preliminary classification probabilities for three states: stable, instability precursors, and unstable.

[0031] S4. Calculate the mean absolute value of the first-order difference of the spectral entropy time series as the rate of change of spectral entropy. Set a first energy variance threshold and a second energy variance threshold. The first energy variance threshold is inversely proportional to the rate of change of spectral entropy, and the second energy variance threshold is directly proportional to the rate of change of spectral entropy. Calculate the energy variance of the pressure time-domain signal within the analysis period. When the energy variance is lower than the first energy variance threshold, the combustion state is stable. When the energy variance is higher than the second energy variance threshold, the combustion state is unstable. When the energy variance is between the first energy variance threshold and the second energy variance threshold, the combustion state is determined based on the maximum value of the preliminary classification probability.

[0032] Specifically, the spectral entropy time series is first-differenced, and the average of its absolute values ​​is taken to obtain the rate of change of spectral entropy. In an optional embodiment, the first energy variance threshold is calculated using the following formula: ;in, The first energy variance threshold, These are positive real constants calibrated based on experimental data. The rate of change of spectral entropy. To prevent tiny positive numbers with a denominator of zero, Set to 1e-6.

[0033] The second energy variance threshold is calculated using the following formula. ;in, This is the second energy variance threshold. The rate of change of spectral entropy. , These are positive real constants calibrated based on experimental data.

[0034] Specifically, the rate of change of spectral entropy R reflects the degree of change in signal stationarity; the value of R increases when the signal transitions from a stationary noise segment to a non-stationary effective signal segment. Assuming calibration based on a large amount of data... It is 5. It is 20. The value is 0.5. When the signal is in a stable background noise environment, the rate of change of spectral entropy R is very small, for example, 0.05, at which point the first energy variance threshold is reached. The value is calculated to be 100, the second energy variance threshold. The calculated value is 1.5. When a stress event occurs, the signal complexity increases, and the rate of change of spectral entropy R increases to 2.0, at which point the first energy variance threshold... It will dynamically decrease to 2.5, while the second energy variance threshold This will increase to 40.5. The 1-second pressure time-domain signal is divided into 10 segments of 0.1 seconds each. The sum of squares of the signal within each segment is calculated to obtain its energy, resulting in a sequence containing 10 energy values. The variance of this energy sequence is calculated; this is the energy variance. The energy variance is then compared with a first energy variance threshold. Second energy variance threshold Comparison: If the energy variance is less than If the energy variance is greater than 1, then the combustion state is directly determined to be stable; If the combustion state is considered unstable, then the energy variance is considered unstable. and Between these states, examine the probability values ​​of the three states output by the convolutional neural network and select the state with the highest probability value as the monitoring result.

[0035] The implementation principle of the combustion chamber unstable combustion state monitoring method based on acoustic sensors in this invention is as follows: This invention successfully resolves the resolution contradiction in time-frequency analysis by employing a frequency segmentation strategy based on the centroid of the signal spectrum. A long window function is used to process low-frequency components, and a short window function is combined to process high-frequency components, ultimately constructing a fused spectrogram that balances high frequency and high time resolution, thereby revealing key time-frequency details in the combustion state evolution process more comprehensively. In terms of feature extraction, this invention uses spectral entropy to weight the spectrogram, effectively highlighting non-stationary features and suppressing noise interference. Then, the weighted spectrogram and its first-order difference graphs of the time and frequency axes are constructed into a three-channel feature tensor, providing richer and more discriminative input information for the convolutional neural network. Furthermore, by combining the energy variance interval of the signal with a dynamic threshold set by the rate of change of spectral entropy, a reliable criterion judgment logic is designed, significantly enhancing the reliability and anti-false-judgment capability of the monitoring results.

[0036] An embodiment of the combustion chamber unstable combustion state monitoring system based on acoustic sensors provided by the present invention includes a memory and a processor. The memory stores computer instructions, and when the processor executes the computer instructions, it implements the combustion chamber unstable combustion state monitoring method based on acoustic sensors in the above embodiment.

[0037] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for monitoring unstable combustion state in a combustion chamber based on acoustic sensors, characterized in that, The process includes the following steps: acquiring time-domain pressure signals within the combustion chamber using an acoustic pressure sensor, dividing the analysis into preset time periods, and calculating the spectral centroid of the signal within each time period as the frequency segmentation threshold for that time period; performing short-time Fourier transforms on the pressure time-domain signals using a first window function and a second window function, respectively; generating a low-frequency spectrogram below the frequency segmentation threshold based on the transform result of the first window function, generating a high-frequency spectrogram above the frequency segmentation threshold based on the transform result of the second window function, and stitching the low-frequency spectrogram and the high-frequency spectrogram together to obtain a fused spectrogram. The spectral entropy of the fused spectrogram at each time frame is calculated to form a spectral entropy time series. The amplitude of each time frame is weighted according to the principle that the spectral entropy value is inversely proportional to the weight coefficient to generate an entropy-weighted fused spectrogram. A three-channel feature tensor is constructed by using the entropy-weighted fused spectrogram, the first-order difference map along the time axis, and the first-order difference map along the frequency axis. This includes: using the entropy-weighted fused spectrogram as the first channel; calculating the difference between adjacent frames of the entropy-weighted fused spectrogram on the time axis to generate the first-order difference map along the time axis, which serves as the second channel; calculating the difference between adjacent frequency points of the entropy-weighted fused spectrogram on the frequency axis to generate the first-order difference map along the frequency axis, which serves as the third channel; stacking the first, second, and third channels sequentially to form the three-channel feature tensor; and inputting the three-channel feature tensor into a convolutional neural network. The initial classification probability of combustion state is obtained; the mean absolute value of the first difference of the spectral entropy time series is calculated as the spectral entropy change rate; a first energy variance threshold and a second energy variance threshold are set, the first energy variance threshold being inversely proportional to the spectral entropy change rate and the second energy variance threshold being directly proportional to the spectral entropy change rate; the energy variance of the pressure time domain signal within the analysis period is calculated, when the energy variance is lower than the first energy variance threshold, the combustion state is stable; when the energy variance is higher than the second energy variance threshold, the combustion state is unstable; when the energy variance is between the first energy variance threshold and the second energy variance threshold, the combustion state is determined according to the maximum value of the initial classification probability.

2. The method for monitoring unstable combustion state in a combustion chamber based on acoustic sensors according to claim 1, characterized in that, Both the first and second window functions are Hanning windows.

3. The method for monitoring unstable combustion state in a combustion chamber based on acoustic sensors according to claim 2, characterized in that, The duration of the first window function is 100 milliseconds, and the duration of the second window function is 25 milliseconds.

4. The method for monitoring unstable combustion state in a combustion chamber based on acoustic sensors according to claim 3, characterized in that, The overlap rate of the first window function and the second window function is both 50%.

5. The method for monitoring unstable combustion state in a combustion chamber based on an acoustic sensor according to claim 1, characterized in that, The step of weighting the amplitude of each time frame according to the principle that the spectral entropy value is inversely proportional to the weighting coefficient to generate an entropy-weighted fused spectrogram includes: The weighting coefficient is calculated using the following formula. Where i represents the i-th time frame of the fused spectrogram, The preset weighted adjustment coefficient, Let be the spectral entropy value of the i-th time frame; The weight coefficients for the i-th time frame; The amplitude values ​​of each time frame are weighted using the following formula. ;in, The weighted amplitude for the i-th time frame. The magnitude before weighting in the i-th time frame. represents the weighting coefficient for the i-th time frame.

6. The method for monitoring unstable combustion state in a combustion chamber based on an acoustic sensor according to claim 1, characterized in that, The first energy variance threshold is calculated using the following formula. ;in, The first energy variance threshold, These are positive real constants calibrated based on experimental data. The rate of change of spectral entropy. To prevent tiny positive numbers with a denominator of zero, Set to 1e-6.

7. The method for monitoring unstable combustion state in a combustion chamber based on an acoustic sensor according to claim 6, characterized in that, The second energy variance threshold is calculated using the following formula: ;in, This is the second energy variance threshold. The rate of change of spectral entropy. , These are positive real constants calibrated based on experimental data.

8. The method for monitoring unstable combustion state in a combustion chamber based on an acoustic sensor according to claim 1, characterized in that, The process of acquiring time-domain pressure signals within the combustion chamber via an acoustic pressure sensor and dividing the data into preset analysis time periods includes: An acoustic pressure sensor is installed on the combustion chamber wall to collect pressure signals, and an analysis period of 1 second is used.

9. A combustion chamber unstable combustion state monitoring system based on acoustic sensors, characterized in that, It includes a memory and a processor. The memory stores computer instructions. When the processor executes the computer instructions, it implements the method for monitoring unstable combustion state in a combustion chamber based on an acoustic sensor as described in any one of claims 1-8.

Images