Compressor instability surge data analysis method based on moment function

Abstract

The present application belongs to the technical field of aero-engine compressor test data processing and state discrimination, and relates to a compressor instability surge data analysis method based on moment functions. The present application takes the compressor dynamic pressure time series as the processing object, obtains the analysis sequence under the unified time scale through the finite impulse response filtering and downsampling, and can reduce the influence of high-frequency sampling data redundancy and local noise on the subsequent discrimination; the present application uses the sliding window surge index to automatically calibrate the surge starting reference time, and forms the steady-state sample interval and the risk sample interval accordingly, thereby reducing the subjective difference caused by manual visual calibration; the present application constructs the statistical moment characteristic table containing the mean, variance, skewness, kurtosis, variance growth rate, skewness offset and excess kurtosis, and can more completely represent the changes of pressure signal fluctuation intensity, distribution offset and sharpness compared with single amplitude discrimination.

Description

Technical Field

[0001] This invention belongs to the technical field of test data processing and state discrimination of aero-engine compressors. Specifically, it relates to a method for analyzing instability and surge data of compressor dynamic pressure time series, and particularly to a method for analyzing compressor instability and surge data by using finite impulse response filtering downsampling, automatic calibration of surge time, statistical moment characteristic table, logarithmic scaling mapping, and support vector machine risk index. Background Technology

[0002] The compressor is a crucial component of an aero-engine, and its aerodynamic stability directly affects engine operational safety. During compressor testing or operation, as operating conditions gradually approach the stability boundary, dynamic pressure signals exhibit changes in amplitude, fluctuation intensity, and distribution pattern. Failure to analyze these changes promptly and objectively may lead to inconsistent judgments regarding the surge initiation time, thereby impacting test data verification, sample segmentation, and subsequent state identification.

[0003] Current compressor surge data analysis typically relies on manual visual interpretation or fixed amplitude thresholds. Manual interpretation requires examining long time sequences one by one, which is inefficient, and different personnel may have different judgments on the surge onset time. While the fixed threshold method is easy to implement, it is greatly affected by the measurement point location, sampling frequency, signal dimensions, noise level, and operating conditions, making it difficult to establish a unified discrimination standard among multiple batches of test data.

[0004] Furthermore, compressor dynamic pressure data typically features high sampling rates, large data volumes, and a large number of measurement points. Directly performing statistical analysis and discrimination on the raw high-frequency sequences is not only computationally intensive but also susceptible to interference from local pulses or measurement noise. Therefore, it is necessary to first transform the raw sequences into analytical sequences at a unified time scale, and then form statistical features that can be batch-calculated, verified, and input into discrimination models.

[0005] Therefore, it is necessary to establish a data analysis method for compressor instability surge that can perform anti-aliasing downsampling on high-frequency dynamic pressure sequences, automatically calibrate the surge reference time, construct a statistical moment feature table, unify the numerical scale of different features, and use support vector machines to form an interpretable risk index curve, thereby improving the consistency and automation of multi-batch test data processing. Summary of the Invention

[0006] To address the problems of low efficiency in manual calibration, insufficient adaptability of fixed thresholds across operating conditions, inconsistent statistical feature scales, and unstable risk segment identification in existing compressor surge test data analysis, this invention provides a compressor instability surge data analysis method based on moment functions.

[0007] To achieve the above objectives, the present invention adopts the following technical solution.

[0008] A method for analyzing compressor instability and surge data based on moment functions includes the following steps: Step 1: Acquire the compressor dynamic pressure time series and perform finite impulse response (FIR) filtering downsampling. Pressure time series are acquired using dynamic pressure sensors positioned at key sections of the compressor. The original time series is then subjected to FIR filtering for anti-aliasing, and an analysis sequence is obtained according to a preset downsampling factor.

[0009] Step two: Automatically determine the surge onset time based on the sliding window surge index. A sliding window is set for the analysis sequence, the amplitude fluctuation index within each window is calculated, and the surge onset reference time is determined according to the continuous exceedance rule. The surge onset reference time is used to divide the steady-state sample interval and the risk sample interval.

[0010] Step 3: Construct a sliding window statistical moment characteristic table. Perform sliding statistics on the analysis sequence according to the preset window length and step size, calculate the first moment mean, second central moment variance, third standardized moment skewness, and fourth standardized moment kurtosis for each window, and further generate derived statistical indicators such as variance growth rate, skewness offset, and kurtosis excess to form a statistical moment characteristic table.

[0011] Step four involves logarithmic scaling and standardization of the statistical moment features. For positive or non-negative statistical features such as mean magnitude, variance, and kurtosis excess, logarithmic compression mapping with shift is used. For statistical features that can be positive or negative, such as skewness, variance growth rate, and skewness offset, standardization is used to bring statistical features of different magnitudes into a unified and comparable numerical range.

[0012] Step 5: Generate the risk index based on radial basis function (RBF) kernel support vector machine. Train the RBF kernel support vector machine discriminant model using steady-state samples and risk samples to obtain the discriminant function values ​​of each sliding window relative to the discriminant boundary; normalize and smooth the discriminant function values ​​to obtain the risk index curve, and output the surge risk segment according to the rule that the risk index of multiple consecutive windows reaches or exceeds 0.8.

[0013] Optionally, the original pressure segments before and after the risk index abrupt change can be reviewed in the frequency domain or time-frequency domain to confirm that the risk segment corresponds to the low-frequency oscillation of surge or broadband energy change. This review does not change the risk index generation process.

[0014] Compared with the prior art, the method proposed in this invention has the following advantages: First, the present invention automatically calibrates the surge initiation time by using a sliding window surge index, reducing the differences in manual visual calibration and providing a unified benchmark for multiple batches of test data.

[0015] Second, this invention transforms high-frequency dynamic pressure data into a statistical moment feature table, which can describe the average level, fluctuation intensity, skewness and sharpness of pressure signals within different time windows in tabular form, facilitating batch processing and verification.

[0016] Third, this invention reduces the impact of a single large-scale feature on the discrimination model by compressing the magnitude differences of different statistical features through logarithmic scaling and standardization.

[0017] Fourth, this invention uses radial basis kernel support vector machine discriminant function values ​​to construct risk index, avoiding the discriminant result being expressed only as a fixed amplitude threshold alarm, thus making the risk segment output continuous and traceable.

[0018] Fifth, this invention does not rely on a specific single measuring point or manually given surge time, and can be used for automated analysis of dynamic pressure test data of multiple batches of compressors. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of the compressor instability surge data analysis method based on statistical moment features and support vector machine risk index described in this invention; Figure 2 This is a schematic diagram of the finite impulse response filtering downsampling and surge moment automatic calibration module described in this invention; Figure 3 This is a schematic diagram illustrating the construction of the sliding window statistical moment feature table according to the present invention; Figure 4 This is a schematic diagram of the logarithmic scaling and standardization process described in this invention; Figure 5 This is a schematic diagram of the support vector machine risk index curve and continuous over-limit criterion described in this invention; Figure 6 This is a schematic diagram of the optional derived statistical state space of the present invention, wherein the three-dimensional coordinates are the variance growth rate, skewness offset and kurtosis excess, respectively. Detailed Implementation

[0020] The present invention will be further described below with reference to specific embodiments. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. Without departing from the concept of the present invention, those skilled in the art can adjust the window length, step size, threshold, and support vector machine parameters according to the specific test bench, measurement point arrangement, and sampling conditions.

[0021] In this embodiment, the time series data collected by the compressor dynamic pressure sensor is used as input, and the surge risk index curve is generated and the risk segment is output using the method of this invention.

[0022] Step 1: Data acquisition and finite impulse response filtering downsampling to obtain the raw time series collected by the compressor dynamic pressure sensor, denoted as: ; in, This represents the original number of sampling points. The original time series can be derived from dynamic pressure measurement points at the compressor inlet section, outlet section, or other key sections.

[0023] Because the original dynamic pressure sequence has a high sampling rate, directly calculating the sliding statistical characteristics would increase the computational load. To obtain an analysis sequence with a uniform time scale, this embodiment first performs finite impulse response filtering anti-aliasing processing on the original time series, and then applies a preset downsampling factor. Downsampling was performed to obtain the analysis sequence: ; in, This represents the number of sampling points after downsampling. Downsampling factor. It can be determined based on the original sampling frequency and the target analysis frequency band.

[0024] Step 2: Automatic calibration of surge onset time for the analysis sequence. Set the length to Step size is A sliding window. For the first There are 3 windows, and the sequence within each window is denoted as . Calculate the surge index for the window: ; in, and These are the maximum and minimum values ​​within the window, respectively. The root mean square value of the window. To prevent positive numbers with excessively small denominators.

[0025] When continuous Each window satisfies At that time, the center time of the first window that satisfies the continuous over-limit condition is marked as the surge initiation reference time. .in, The surge index threshold, This represents the number of consecutive windows exceeding the limit.

[0026] according to Divide the sample into intervals. For example, divide the sample into intervals far from the sample. The time period is divided into steady-state sample intervals, and The abnormal evolution period within a pre-set time range is divided into risk sample intervals. The length of the sample interval can be determined based on the specific experimental process and the amount of data.

[0027] Step 3: Construct the sliding window statistical moment characteristic table for the analysis sequence. Set the length to Step size is The feature window. For the first Using feature windows, calculate the following basic statistical moment features: Mean of first moment: ; Second-order central moment variance: ; Third-order normalized moment skewness: ; Fourth-order normalized moment kurtosis: ; Based on the basic statistical moment characteristics, further derived statistical indicators are constructed, including but not limited to: Variance growth rate: ; Skewness offset: ; Kuroshi excess: ; in, For steady-state sample intervals, skewness sequence This represents the baseline kurtosis value within the steady-state sample interval. Through the above calculations, a statistical moment characteristic table arranged by time window is obtained. .

[0028] Step 4: Logarithmic scaling and standardization. Since the numerical scales of features such as mean magnitude, variance, and kurtosis excess may vary significantly, this embodiment uses logarithmic scaling with a shift for positive or non-negative features: ; in, This represents the new eigenvalues ​​after logarithmic scaling with translation. Features to be mapped To ensure that the logarithmic input is a positive translation amount.

[0029] Features such as skewness, variance growth rate, and skewness offset are standardized. ; in, and These are the mean and standard deviation of the corresponding features in the training samples, respectively.

[0030] After scaling, the input features for the support vector machine are obtained: ; in, The input feature dimension.

[0031] Step 5: Generate a risk index based on support vector machine. Assign class labels to the scaled feature vectors based on the steady-state sample interval and risk sample interval obtained in Step 2. Label steady-state samples as negative classes and risk samples as positive classes, and train a radial basis function kernel support vector machine discriminant model.

[0032] The radial basis kernel function is: ; in, These are the kernel function parameters. The penalty parameters for the support vector machine. and kernel function parameters This can be determined through cross-validation or preset parameters. and This represents the feature vector of two data points input into the radial basis kernel function for similarity calculation.

[0033] For any sliding window feature The support vector machine outputs the discriminant function value. .when The closer the window is to the risk sample side, the closer it is to a surge risk state. For ease of display and threshold determination, the discriminant function value is normalized to obtain the risk index: ; in, The value range can be set to When continuous Each window satisfies When this time period is reached, the corresponding time segment is output as the surge risk segment. Among them, The risk index threshold is set to 0.8 in this embodiment. This represents the number of consecutive windows exceeding the limit.

[0034] This risk index curve can show the process of the compressor dynamic pressure signal developing from stable fluctuations to abnormal fluctuations, and can be compared and verified with the surge initiation reference time obtained by automatic calibration.

[0035] Optional Step 6: Physical Consistency Verification. After obtaining the risk segment, the original pressure segments before and after the risk segment can be verified in the frequency domain or time-frequency domain to determine whether the risk segment corresponds to enhanced low-frequency oscillations, broadband energy changes, or other pressure fluctuation characteristics related to surge. This verification step is used to assist in interpreting the risk segment and is not a necessary condition for generating the risk index.

[0036] Visualization of derived statistical state space To facilitate observation of changes in statistical states, this embodiment can also construct a derived statistical state space. This space is represented by the variance growth rate. skewness offset and kurtosis excess Using the mean, variance, and skewness as coordinate axes, each sliding window is mapped to a spatial point. Compared to directly using the mean, variance, and skewness as coordinates, derived statistical indicators more directly reflect changes in the intensity of pressure fluctuations, distribution shifts, and increases in sharpness, which helps in the review and visualization of risk segments.

[0037] Example 2: Reference Figure 1 This embodiment uses the dynamic pressure signal at measuring point 39 during a compressor surge test on March 25, 2024, as the object to illustrate the method described in this invention.

[0038] This embodiment includes steps 1 to 5: Step 1: Obtain the original dynamic pressure time series with a sampling rate of 4096 Hz, and select 176.0 s to 188.0 s as the analysis interval; perform FIR low-pass anti-aliasing filtering on the original signal, and downsample it by 4 times to obtain the analysis sequence with a sampling rate of 1024 Hz; Step 2: Calculate the sliding window surge index based on the downsampled analysis sequence and automatically calibrate the surge initiation reference time; in this embodiment, the surge index threshold is 8.5359, and the automatically calibrated surge initiation reference time is 184.7412 s; Step 3: Construct a statistical moment feature table based on the above analysis sequence. The statistical moment feature table includes mean, variance, skewness, kurtosis, and derived statistical indicators such as variance growth rate, skewness offset, and kurtosis excess. In this embodiment, a total of 116 sets of window feature samples are formed. Step 4: Perform logarithmic scaling and standardization on the feature quantities in the statistical moment feature table to convert statistical indicators with different dimensions and different ranges of change into feature vectors suitable for support vector machine input. Step 5: Input the feature vector obtained in Step 4 into the radial basis kernel support vector machine to obtain the discriminant function value corresponding to each window, and normalize and smooth the discriminant function value to generate the risk index curve. In this embodiment, the risk threshold is set to 0.8. When the risk index reaches or exceeds 0.8 and forms a continuous over-limit, the corresponding time period is determined to be the surge risk segment.

[0039] As per the instruction manual Figure 5As shown, within the analysis interval of 176.0 s to 188.0 s, the risk index forms a clear continuous over-limit range around 184.5176 s to 184.8164 s, with the risk index reaching 1.0000 at 184.7168 s. This moment roughly corresponds to the automatically calibrated surge initiation reference time of 184.7412 s, and is slightly earlier than this reference time, indicating that the present invention can provide a concentrated and stable response to the risk state before and after the surge occurs. Meanwhile, although short-term increases in the risk index occur around 181.4297 s, 181.5293 s, and 183.8203 s, they do not form the same continuous change as the main risk segment. The impact of random fluctuations on the judgment results can be reduced through the continuous over-limit rule.

[0040] As per the instruction manual Figure 6 As shown, in the derived statistical state space composed of variance growth rate, skewness offset, and kurtosis excess, sample points migrate from the steady-state region to the high-risk region when surge risk increases. Taking the 184.7168 s window where the risk index reaches its peak as an example, its variance growth rate is 0.3674, skewness offset is 0.2469, and kurtosis excess is 0.1885, indicating that this window not only exhibits changes in energy intensity but also is accompanied by increased waveform asymmetry and impact. This embodiment illustrates that the present invention does not rely on a single amplitude threshold for judgment but achieves automatic calibration and quantitative identification of compressor surge risk segments through a combination of statistical moment characteristics, logarithmic scaling mapping, and SVM risk index, thereby improving the consistency and anti-interference capability of batch test data analysis.

[0041] The above embodiments show that the present invention realizes a complete analysis process from high-frequency dynamic pressure data to surge risk segment output through finite impulse response filtering downsampling, automatic surge moment calibration, construction of statistical moment feature table, logarithmic scaling mapping and support vector machine risk index generation.

Claims

1. A method for analyzing compressor instability and surge data based on moment functions, characterized in that, Includes the following steps: Step 1: Obtain the dynamic pressure time series of the compressor and perform finite impulse response filtering downsampling; The pressure time series is collected by dynamic pressure sensors arranged at key sections of the compressor, the original time series is processed by finite impulse response filtering to prevent aliasing, and the analysis sequence is obtained according to the preset downsampling factor. Step 2: Automatically calibrate the surge initiation time based on the sliding window surge index; A sliding window is set for the analysis sequence, the amplitude fluctuation index within each window is calculated, and the surge initiation reference time is determined according to the continuous over-limit rule; the surge initiation reference time is used to divide the steady-state sample interval and the risk sample interval. Step 3: Construct a sliding window statistical moment feature table; Perform sliding statistics on the analysis sequence according to the preset window length and step size, calculate the first moment mean, second central moment variance, third standardized moment skewness, and fourth standardized moment kurtosis of each window, and further generate variance growth rate, skewness offset, and kurtosis excess derived statistical indicators to form a statistical moment feature table. Step four: Perform logarithmic scaling and standardization on the statistical moment features; for positive or non-negative statistical features such as mean magnitude, variance, and kurtosis excess, use logarithmic compression mapping with shift; for statistical features such as skewness, variance growth rate, and skewness offset that can be positive or negative, use standardization to bring statistical features of different magnitudes into a unified and comparable numerical range. Step 5: Generate a risk index based on a radial basis function (RBF) kernel support vector machine; train the RBF kernel support vector machine discriminant model using steady-state samples and risk samples to obtain the discriminant function value of each sliding window relative to the discriminant boundary; normalize and smooth the discriminant function value to obtain the risk index curve, and output the surge risk segment according to the rule that the risk index of multiple consecutive windows reaches or exceeds 0.

8.

2. The method for analyzing compressor instability and surge data based on moment functions as described in claim 1, characterized in that, The specific steps for step two are as follows: Set the window length for the analysis sequence to be [value]. Step size is The sliding window; note the first The sequence within each sliding window is The dimensionless surge index is defined as: ， in, and These are the maximum and minimum values ​​within the window, respectively. The root mean square value of the window. To prevent positive numbers with excessively small denominators; When continuous Each window satisfies At that time, the center time of the first window that satisfies the continuous over-limit condition is marked as the surge initiation reference time. ;in, The surge index threshold, The number of consecutive windows exceeding the limit, and is an integer greater than or equal to 2.

3. The method for analyzing compressor instability and surge data based on moment functions as described in claim 1, characterized in that, The specific steps of step one are as follows: The original time series is subjected to finite impulse response filtering for anti-aliasing, and then downsampled according to a preset factor. Downsampling was performed to obtain the analysis sequence. ， in, This represents the number of sampling points after downsampling.

4. The method for analyzing compressor instability and surge data based on moment functions as described in claim 1, characterized in that, The specific steps for step three are as follows: Using window length Step size is A sliding window maps the analysis sequence from a time-domain signal to a statistical moment feature table, which includes at least the following basic statistical moment features: The first-order moment mean is used to reflect the average level of the pressure signal within the current window; The second-order central moment variance is used to reflect the fluctuation intensity of the pressure signal within the current window; The third-order normalized moment skewness is used to describe the degree of deviation of the pressure signal distribution relative to the steady-state reference. Fourth-order normalized moment kurtosis is used to describe the sharpness and impact of pressure signal distribution.

5. The method for analyzing compressor instability and surge data based on moment functions as described in claim 4, characterized in that, The statistical moment feature table also includes derived statistical indicators, which include at least one of variance growth rate, skewness offset, and kurtosis excess. The variance growth rate is determined by the ratio of the difference between the variances of adjacent sliding windows to the absolute value of the variance of the previous sliding window. The skewness offset is determined by the degree of deviation between the skewness of the current sliding window and the skewness benchmark value of the steady-state sample interval. The kurtosis excess is determined by the portion of the kurtosis of the current sliding window that exceeds the kurtosis benchmark value of the steady-state sample interval.

6. The method for analyzing compressor instability and surge data based on moment functions as described in claim 1, characterized in that, The specific steps for step four are as follows: For positive or non-negative features, apply a logarithmic scaling mapping with a shift: ， in, This represents the new eigenvalues ​​after logarithmic scaling with translation. Features to be mapped To ensure that the logarithmic input is a positive shift amount; Standardization was applied to the skewness, variance growth rate, and skewness offset features: ， in, and These are the mean and standard deviation of the corresponding features in the training samples, respectively; After scaling, the input features for the support vector machine are obtained: ， in, The input feature dimension.

7. The method for analyzing compressor instability and surge data based on moment functions as described in claim 1, characterized in that, The specific steps for step five are as follows: Steady-state samples are labeled as negative classes, and risk samples are labeled as positive classes. A radial basis function kernel support vector machine discriminant model is then trained. The radial basis kernel function is: ， in, Kernel function parameters; penalty parameters for support vector machines. and kernel function parameters This can be determined through cross-validation or preset parameters. and This represents the feature vector of two data points input into the radial basis kernel function for similarity calculation; For any sliding window feature The support vector machine outputs the discriminant function value. ; Normalizing the discriminant function value yields the risk index: ， in, The value range can be set to When continuous Each window satisfies When the time period is specified, the corresponding time segment is the surge risk segment; among which, This is the threshold for the risk index.

8. The method for analyzing compressor instability and surge data based on moment functions as described in claim 1, characterized in that, The risk index is obtained by normalizing the discriminant function value output by the radial basis function support vector machine, and the value range of the risk index is from 0 to 1. When the risk index of multiple consecutive sliding windows reaches or exceeds the risk index threshold... When the value is 0.8, the corresponding time period is determined to be the surge risk period.

9. The method for analyzing compressor instability and surge data based on moment functions as described in claim 1, characterized in that, It also includes step six, a visualization and verification step of the derived statistical state space; the derived statistical state space uses the variance growth rate, skewness offset and kurtosis excess as coordinates to observe the migration process of the sliding window statistical state over time, and to verify the surge risk segment.

Images