A method and system for identifying compressor stall failure by sound

Abstract

The application discloses a compressor stall failure acoustic identification method and system, and relates to the field of engine state monitoring and fault diagnosis. The method comprises the following steps: collecting acoustic signals in the operation process of the compressor; performing frame division and windowing processing on the acoustic signals; performing spectrum analysis and frequency band division on each frame of windowed signals to generate a frame energy vector; evaluating the importance of each frequency band based on the random forest out-of-bag replacement method to screen out key frequency bands; constructing a triangular filter bank with each frequency in the key frequency set as a vertex, and performing neighborhood expansion when the interval between adjacent frequencies exceeds a threshold interval to obtain a reconstructed adaptive filter bank and a final filter energy feature vector; and inputting the final filter energy feature vector into a deep neural network model to obtain an identification result of whether the compressor is in a stall state. The application has the advantages of reducing equipment cost, simplifying installation process, improving stall feature targeted analysis capability, and enhancing the robustness of the identification result.

Description

Technical Field

[0001] This invention relates to the field of engine condition monitoring and fault diagnosis, specifically to an acoustic identification method and system for compressor stall faults. Background Technology

[0002] The operational stability of an axial compressor directly affects the overall performance of the engine. Rotational stall is a common and serious aerodynamic instability phenomenon during compressor operation, which can lead to a sharp decline in compressor performance, a significant increase in blade load, and in severe cases, may trigger a chain reaction such as surge, causing mechanical damage or even catastrophic consequences.

[0003] Currently, the main methods for detecting compressor stall are as follows: installing dynamic pressure sensors on the casing to monitor pressure fluctuation characteristics; using particle image velocimetry to monitor changes in flow field velocity distribution; and using laser Doppler sensors to measure blade vibration characteristics. While these methods can effectively identify stall phenomena, they generally suffer from high equipment costs, complex system structures, and limited installation locations, making it difficult to promote and apply them on a large scale in practical engineering.

[0004] In comparison, acoustic sensors offer advantages such as simple deployment, lower cost, wide frequency response range, and sensitivity to aerodynamic instability, and have been increasingly applied in the condition monitoring of rotating machinery in recent years. Existing stall identification methods based on acoustic signals suffer from two main shortcomings: first, they often rely on empirical thresholds for feature selection, lacking specific analysis of compressor stall characteristics; second, they directly apply general-purpose deep learning networks, failing to fully consider the correlation between acoustic signals and stall conditions, resulting in insufficient robustness of the identification results.

[0005] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0006] In view of this, the present invention provides an acoustic identification method and system for compressor stall faults, which has the advantages of reducing equipment costs, simplifying the installation process, improving the ability to analyze stall characteristics in a targeted manner, and enhancing the robustness of the identification results.

[0007] In a first aspect, the present invention provides an acoustic identification method for compressor stall faults, comprising the following steps:

[0008] S1. Acoustic signals during compressor operation are collected using an acoustic sensor to obtain a discrete time-domain sequence; the acoustic sensor is installed at the opening of the compressor casing.

[0009] S2. Perform framing and windowing processing on the discrete time-domain sequence to obtain the windowed signal for each frame;

[0010] S3. Perform spectral analysis on the windowed signal of each frame to calculate the spectrum corresponding to the windowed signal of each frame; and divide the amplitude spectrum of each frame into multiple frequency bands, calculate the energy of each frequency band, and generate a frame energy vector; the amplitude spectrum is obtained by extracting the amplitude information from the spectrum;

[0011] S4. The importance of each frequency band is evaluated based on the random forest bag permutation method, and key frequency bands related to the stall state are screened to obtain a set of key frequencies; the set of key frequencies is composed of the center frequencies corresponding to each key frequency band.

[0012] S5. Construct a triangular filter bank with each frequency in the key frequency set as the vertex, and expand the neighborhood when the distance between adjacent frequencies exceeds the distance threshold to obtain the reconstructed adaptive filter bank.

[0013] S6. Process each frame of signal through the reconstructed adaptive filter bank to obtain the filter energy feature vector;

[0014] S7. Input the filter energy feature vector into the deep neural network model for processing to obtain the identification result of whether the compressor is in a stall state.

[0015] In one alternative implementation, S2 includes:

[0016] S21. Calculate the total number of frames in the discrete-time sequence. K Its expression is:

[0017] (1)

[0018] in, This represents the total number of sampling points for the acoustic signal. For frame length, For frame shift;

[0019] S22. A window function is used for weighted processing to obtain the windowed signal for each frame, the expression of which is:

[0020] (2)

[0021] in, For the index of the sampling points within each frame, For window functions, For the first Windowing signal of a frame;

[0022] Window function The expression is:

[0023] (3).

[0024] In one optional implementation, S3 includes:

[0025] S31. Perform a Fourier transform on the windowed signal of each frame to obtain the spectrum corresponding to each frame signal. Its expression is:

[0026] (4)

[0027] in, The signal sampling rate is expressed in Hz. For the first A discrete frequency; Frequency index;

[0028] S32. Divide the spectrum of each frame into multiple equal-width frequency bands according to the frequency conversion ratio;

[0029] S33. Calculate the energy of each frequency band, the expression is:

[0030] (5)

[0031] in, For the first m The first frame in the spectrum j The energy value of the frequency band; For the first m The first frame in the spectrum The frequency index set contained in each frequency band;

[0032] S34. Generate a frame energy vector based on the energy of each frequency band of the windowed signal in each frame.

[0033] In one alternative implementation, S4 includes:

[0034] S41. Using frame energy vectors as features and operating condition labels as outputs, train a random forest classifier; the operating condition labels include normal operating conditions and stall operating conditions.

[0035] S42. The importance of each frequency band is evaluated using the external bag permutation method, and its expression is:

[0036] (6)

[0037] (7)

[0038] (8)

[0039] in, This represents the number of samples outside the bag. The prediction function for the classifier; The loss function is set to 0-1 loss. This refers to the error outside the bag. When calculating the importance of permutations, the first... Random permutation values ​​of the energy characteristics of each frequency band on the samples outside the bag; For the first j The sample vector obtained by randomly permuting the energy of each frequency band in the samples outside the bag; For the first j The importance score of each frequency band;

[0040] S43. Set the importance threshold, the expression of which is:

[0041] (9)

[0042] in, and The importance sets for each frequency band are respectively. The mean and standard deviation, Importance threshold;

[0043] S44. Select the frequency bands whose importance scores are greater than the importance threshold to obtain the key frequency set.

[0044] In one optional implementation, S5 includes:

[0045] S51. Construct a triangular filter bank using each frequency in the key frequency set as a vertex; where, the... k Each filter at frequency f The expression for the response is:

[0046] (10)

[0047] in, Indicates the first Each filter at frequency The response at the location; Indicates the first i The center frequencies corresponding to the key frequency bands; Indicates the first i- The center frequency corresponding to a key frequency band; Indicates the first i+ The center frequency corresponding to a key frequency band;

[0048] No. Each filter at frequency The response at a given point is defined using the principle of shared values ​​between adjacent vertices and linear interpolation.

[0049] S52. When the spacing between adjacent frequencies exceeds the spacing threshold:

[0050] (11)

[0051] in, The spacing threshold; For compressor frequency;

[0052] Then perform a neighborhood expansion within the interval. Within, from the complete set of candidate frequencies Select two neighboring frequency points , As supplementary vertices, they are as follows:

[0053] (12)

[0054] The complete set of candidate frequencies is composed of coverage intervals. It consists of all the discrete frequency points within;

[0055] The expanded set of key frequencies is as follows:

[0056] (13)

[0057] S53. Reconstruct the triangular filter bank based on the expanded key frequency set to obtain the reconstructed adaptive filter bank.

[0058] In one optional implementation, S6 includes:

[0059] S61. Pass the spectrum corresponding to the windowed signal of each frame through the reconstructed triangular filter bank and calculate the output energy of each filter; where, the first... The filter in the first... The output energy on a frame is extracted using the following formula:

[0060] (14)

[0061] in, express exist The value at;

[0062] S62. Combine the output energies of each filter to form a frame-level filter energy feature vector.

[0063] In one alternative implementation, the deep neural network model includes the following components connected in sequence:

[0064] The feature embedding layer is used to map the input filter energy vector into a high-dimensional feature representation;

[0065] The local feature extraction module is used to extract locally relevant features from the embedded features;

[0066] The global feature fusion module is used to capture long-term dependencies in feature sequences;

[0067] The classification output layer is used to output the final working condition identification result.

[0068] In one optional implementation, the local feature extraction module includes a one-dimensional convolutional layer, a batch normalization layer, a ReLU activation layer, and an average pooling layer connected in sequence.

[0069] The global feature fusion module is a Transformer encoder, which includes a multi-head self-attention layer, a feedforward neural network layer, a residual connection mechanism, and a normalization layer connected in sequence.

[0070] In one optional implementation, the classification output layer includes:

[0071] A global average pooling layer aggregates sequence features;

[0072] The first fully connected layer compresses the feature dimension to 32 dimensions;

[0073] Dropout layer, drop rate set to 0.2;

[0074] The second fully connected layer outputs a 2D feature vector.

[0075] The Softmax function calculates the final classification probability.

[0076] Secondly, the present invention also provides an acoustic identification system for compressor stall faults, comprising:

[0077] The acquisition module is used to acquire acoustic signals during the operation of the compressor through an acoustic sensor to obtain a discrete time-domain sequence; the acoustic sensor is installed at the opening of the compressor casing;

[0078] The first preprocessing module is used to perform frame division and windowing processing on the discrete time domain sequence to obtain the windowed signal of each frame;

[0079] The second preprocessing module is used to perform spectral analysis on the windowed signal of each frame, calculate the spectrum corresponding to the windowed signal of each frame, and divide the amplitude spectrum of each frame into multiple frequency bands, calculate the energy of each frequency band, and generate a frame energy vector; the amplitude spectrum is obtained by extracting the amplitude information from the spectrum.

[0080] The feature filtering module is used to evaluate the importance of each frequency band based on the random forest bag-out permutation method, and to filter out the key frequency bands related to the stall state to obtain a set of key frequencies; the set of key frequencies is composed of the center frequencies corresponding to each key frequency band.

[0081] An adaptive filtering unit is used to construct a triangular filter bank with each frequency in the key frequency set as the vertex, and to expand the neighborhood when the spacing between adjacent frequencies exceeds the spacing threshold, so as to obtain the reconstructed adaptive filter bank.

[0082] The filter energy feature vector generation module is used to process each frame of signal through the reconstructed adaptive filter bank to obtain the filter energy feature vector;

[0083] The classification module is used to input the filter energy feature vector into the deep neural network model for processing, and obtain the identification result of whether the compressor is in a stall state.

[0084] The present invention has the following beneficial effects:

[0085] This invention acquires signals using acoustic sensors and employs spectrum analysis, key frequency band screening, adaptive filter bank construction, and deep neural network processing to achieve accurate identification of compressor stall conditions. It has the advantages of reducing equipment costs, simplifying installation procedures, improving the ability to analyze stall characteristics, and enhancing the robustness of identification results. Attached Figure Description

[0086] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0087] Figure 1 This is a flowchart illustrating an acoustic identification method for compressor stall faults according to an embodiment of the present invention.

[0088] Figure 2 This is a schematic diagram of the structure of a single-stage compressor test bench according to an embodiment of the present invention;

[0089] Figure 3 The diagram shows the response results of a triangular filter bank according to an embodiment of the present invention; wherein, (a) is the response result diagram of an adaptive triangular filter bank; and (b) is the response result diagram of the extended and reconstructed triangular filter bank.

[0090] Figure 4 This is a structural framework diagram of a deep neural network model according to an embodiment of the present invention;

[0091] Figure 5 This is a distribution diagram of the importance scores of frequency band features according to an embodiment of the present invention;

[0092] Figure 6 This is a comparison diagram of the recognition results according to an embodiment of the present invention. Detailed Implementation

[0093] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0094] In existing technologies, monitoring the operating status of axial compressors mainly relies on dynamic pressure sensors, particle image velocimetry, or laser Doppler sensors. These methods require complex equipment and are subject to installation limitations, making them difficult to implement in engineering sites. While acoustic signal-based monitoring technologies offer cost advantages, traditional methods rely on manual experience to set frequency thresholds, and general neural network models struggle to capture the time-varying characteristics of stall frequency features, resulting in insufficient identification accuracy. When the compressor enters stall conditions, the energy distribution of specific frequency bands in the acoustic signal changes, but existing technologies lack targeted feature optimization mechanisms, failing to effectively distinguish the subtle differences between normal and stall states.

[0095] To address the aforementioned issues, the inventors discovered that existing technologies rely on manual experience for frequency band selection, leading to a lack of adaptability in feature extraction. By analyzing the acoustic characteristics of compressor stall, a strong correlation was found between energy changes in key frequency bands and the stall state. Therefore, they proposed using machine learning methods to automatically select sensitive frequency bands and construct a dedicated filter bank to enhance feature representation. Further considering the energy coupling effect between adjacent frequency bands, they proposed supplementing neighboring vertices when the spacing between key frequency bands is too large, ensuring continuous coverage of stall features by the filter bank. Finally, by combining the multi-scale modeling capabilities of deep neural networks, end-to-end operating condition identification is achieved.

[0096] Therefore, as Figure 1 As shown, the present invention provides an acoustic identification method for compressor stall faults, comprising the following steps:

[0097] S1. Acoustic signals during compressor operation are collected using acoustic sensors to obtain discrete time-domain sequences. The acoustic sensor is installed at the opening in the compressor casing.

[0098] For example, the experimental platform of this invention uses a single-stage compressor test bench, and the acoustic sensor is installed in the following position: Figure 2 As shown in the figure. Stall was induced by gradually closing the throttle valve, and acoustic signals of the compressor under normal operation and stall conditions at different speeds were collected in 17 groups. The sampling rate of each group was 25.6 kHz, and the duration of each group was about 30 seconds.

[0099] S2. Perform framing and windowing processing on the discrete time-domain sequence to obtain the windowed signal for each frame.

[0100] S3. Perform spectrum analysis on the windowed signal of each frame to calculate the spectrum corresponding to the windowed signal of each frame; and divide the amplitude spectrum of each frame into multiple frequency bands, calculate the energy of each frequency band, and generate the frame energy vector; the amplitude spectrum is obtained by extracting the amplitude information in the spectrum.

[0101] S4. The importance of each frequency band is evaluated based on the random forest bag permutation method, and key frequency bands related to stall state are screened out to obtain the key frequency set; the key frequency set is composed of the center frequency corresponding to each key frequency band.

[0102] S5. Construct a triangular filter bank with each frequency in the key frequency set as the vertex, and expand the neighborhood when the distance between adjacent frequencies exceeds the distance threshold to obtain the reconstructed adaptive filter bank.

[0103] S6. Process each frame of signal through the reconstructed adaptive filter bank to obtain the filter energy feature vector.

[0104] S7. Input the filter energy feature vector into the deep neural network model for processing to obtain the identification result of whether the compressor is in a stall state.

[0105] Framing and windowing processes involve dividing a continuous acoustic signal into short-segment frame sequences and applying a window function to reduce spectral leakage. Specifically, Hamming or Hanning windows can be used. This process enhances the signal's time-frequency local stationarity. Spectrum analysis involves performing a Fourier transform on the windowed signal to obtain its frequency domain energy distribution. This can be achieved using a Fast Fourier Transform algorithm, converting time-domain vibrations into frequency-domain energy features. The random forest out-of-bag permutation method evaluates feature importance using out-of-bag samples from the classifier. This is achieved by randomly permuting feature values ​​and calculating the change in classification error, enabling objective selection of key frequency bands. The triangular filter bank constructs triangular response curves with key frequencies as vertices, using adjacent vertex sharing and linear interpolation principles. This structure focuses on extracting energy from stall-sensitive frequency bands. Neighborhood expansion involves adding neighboring frequencies as new vertices when the distance between adjacent key frequencies is too large. This is achieved by selecting neighboring frequencies from the entire candidate frequency set, enhancing the filter bank's continuous coverage of characteristic frequency bands. Deep neural network models refer to multi-layer network structures that include convolutional layers and self-attention mechanisms. Specifically, they are achieved by fusing local feature extraction with global dependency fusion, which improves the robustness of feature representation.

[0106] Specifically, the raw signals acquired by the acoustic sensors are processed through frame-by-frame windowing and then converted into a short-time stationary frame sequence. Each frame signal undergoes a Fourier transform to generate a spectrum, and the energy of the frequency bands divided according to the frequency shift ratio forms an initial feature vector. A random forest classifier uses out-of-bag permutation to quantify the discriminative power of each frequency band for stall states, selecting key frequency bands whose importance exceeds a threshold. A triangular filter bank constructed with key frequencies as vertices automatically supplements neighboring frequency points when the distance between adjacent vertices exceeds a frequency shift correlation threshold, forming an adaptive filter bank that continuously covers the key frequency domain. The reconstructed filter bank performs energy remapping on the spectrum of each frame, generating feature vectors focused on stall-sensitive frequency bands. A deep neural network extracts local frequency domain correlations through convolutional layers, captures long-term dependencies using a self-attention mechanism, and finally outputs the operating condition classification probability through a fully connected layer.

[0107] Compared to existing technologies, traditional methods rely on manually setting frequency band thresholds, leading to strong subjectivity in feature selection. This solution uses machine learning to automatically select key frequency bands, eliminating interference from human experience. Existing technologies using fixed filter banks struggle to adapt to frequency band shifts under different operating conditions. This solution dynamically adjusts the filter structure through neighborhood expansion, enhancing the tracking capability of feature frequency bands. Traditional neural network models lack optimization for the frequency domain characteristics of acoustic signals. This solution combines convolution and self-attention mechanisms to achieve collaborative modeling of local features and global contextual information.

[0108] Through the above technical solution, this application can automatically identify stall-sensitive frequency bands in compressor acoustic signals, enhance feature representation through adaptive filter banks, and achieve high-precision operating condition classification by combining deep neural networks. This method avoids the high cost problem of traditional sensors, overcomes the limitations of manually setting frequency band thresholds, and improves the robustness of identification under different speed conditions.

[0109] In one optional implementation, step S2 includes:

[0110] S21. Calculate the total number of frames in the discrete-time sequence. K Its expression is:

[0111] (1)

[0112] in, This represents the total number of sampling points for the acoustic signal. For frame length, For frame shift.

[0113] S22. A window function is used for weighted processing to obtain the windowed signal for each frame, the expression of which is:

[0114] (2)

[0115] in, For the index of the sampling points within each frame, For window functions, For the first Windowing signal of a frame;

[0116] Window function The expression is:

[0117] (3).

[0118] Preferably, the window function can be any one of the following: Hamming window, Hanning window, rectangular window, Blackman window, or an equivalent window.

[0119] Compared to existing technologies, traditional methods often use fixed framing parameters or empirical values ​​to determine the number of frames, which can easily lead to incomplete or redundant signal segmentation. This scheme calculates the total number of frames precisely using mathematical formulas, ensuring the integrity of signal processing. Existing technologies using rectangular or Hamming windows suffer from problems such as excessively wide main lobes or insufficient side lobe attenuation. In contrast, the normalized cosine square window function used in this scheme is more advantageous in suppressing spectral leakage, and its parameter design balances main lobe width and side lobe attenuation.

[0120] Through the above technical solution, this application solves the signal truncation problem caused by unreasonable framing parameters and avoids energy leakage in spectrum analysis. The specific design of the window function effectively reduces spectral aliasing caused by abrupt changes in signal edges, improves the accuracy of subsequent frequency band energy calculation, and provides a more reliable preprocessing basis for fault feature extraction.

[0121] In one optional implementation, step S3 includes:

[0122] S31. Perform a Direct Fourier Transform (DFT) on the windowed signal of each frame to obtain the spectrum corresponding to the windowed signal of each frame. Its expression is:

[0123] (4)

[0124] in, The signal sampling rate is expressed in Hz. For the first A discrete frequency; Frequency index;

[0125] The above process converts the windowed signal from the time domain to the frequency domain.

[0126] S32. Divide the spectrum of each frame into multiple equal-width frequency bands according to the frequency conversion ratio;

[0127] S33. Calculate the energy of each frequency band of the windowed signal in each frame. The expression is as follows:

[0128] (5)

[0129] in, For the first m The first frame in the spectrum j The energy value of the frequency band; For the first m The first frame in the spectrum The frequency index set contained in each frequency band;

[0130] S34. Generate a frame energy vector based on the energy of each frequency band of the windowed signal in each frame; where, the first... m The frame energy vector can be represented as , J This represents the total number of frequency bands.

[0131] Preferably, the frequency band division range at least covers the low-frequency range where the stall characteristic frequency is located, and extends to the frequency position of the blade passage frequency, so that both the stall characteristic frequency and the blade passage frequency are included within the frequency band set. As an example, the amplitude spectrum is divided into 150 equal-width frequency bands according to the rotational speed ratio, covering an area of... Includes the frequency of blade passage ( In this embodiment, the compressor frequency is... ).

[0132] Fourier transform refers to the fundamental operation of converting a time-domain signal into a frequency-domain representation, which can be implemented using the Fast Fourier Transform algorithm to provide spectral distribution data for subsequent bandwidth energy calculation. Frequency ratio allocation refers to using the fundamental frequency corresponding to the actual compressor speed as the reference unit. Specifically, bandwidth boundaries can be determined according to integer or fractional multiples of the frequency, ensuring a correspondence between bandwidth allocation and the compressor's physical characteristics. Constant bandwidth means that the span of each bandwidth on the frequency axis remains constant. This can be achieved by automatically calculating the bandwidth through a preset number of bandwidths, ensuring the systematic and repeatable nature of bandwidth allocation. Bandwidth energy calculation involves calculating the sum of squares of the spectral amplitudes within a specified frequency range. Specifically, this can be done by traversing all discrete frequency points within the bandwidth and accumulating the sums to quantify the energy contribution of each frequency band.

[0133] Specifically, after signal framing and windowing, a Fast Fourier Transform (FFT) is performed on each frame's time-domain signal to obtain a complex spectrum, and amplitude spectrum data is obtained through modulo operations. The compressor's real-time rotational speed is used to calculate the rotational frequency, and the analysis frequency band is divided into equal-width sub-bands according to a preset multiple of the rotational frequency. For example, the upper limit of the analysis frequency band is set to 1.5 times the blade passage frequency, and the frequency bands are divided with a step size of 0.5 times the rotational frequency. For each equal-width frequency band, all discrete frequency points covered by the band are traversed, and the squares of the corresponding amplitude values ​​are summed to obtain the energy characterization value of that frequency band. Finally, the energy values ​​of each frequency band are arranged in frequency order to form a feature vector characterizing the frequency energy distribution of the current frame signal.

[0134] Compared to existing technologies, traditional methods often employ fixed bandwidth allocation or empirical frequency band selection, such as dividing the 0-500Hz range into equally spaced frequency bands. This allocation method is disconnected from the actual operating parameters of the compressor, making it difficult to effectively capture stall characteristic frequencies. In contrast, this method uses a frequency ratio allocation mechanism to mathematically correlate the frequency band boundaries with key parameters such as blade passage frequency and stall characteristic frequency. For example, when the compressor speed changes, the frequency band allocation range automatically and dynamically adjusts to follow the frequency change, ensuring that the characteristic frequency always remains within a specific frequency band range.

[0135] Through the above technical solutions, this application achieves a dynamic correlation between frequency band allocation and compressor physical characteristics, solving the problem of missing characteristic frequencies caused by manual experience-based allocation. The frequency rotation ratio mechanism enables the frequency band boundaries to automatically adapt to different operating speeds, enhancing the physical meaning of the frequency band energy characteristics. The equal-width frequency band allocation, combined with energy accumulation calculation, compresses high-dimensional spectral information into structured feature vectors, providing interpretable, high signal-to-noise ratio input data for subsequent machine learning-based feature selection.

[0136] In one alternative implementation, step S4 includes:

[0137] S41. Using frame energy vectors as features and operating condition labels as outputs, train a random forest classifier; the operating condition labels include normal operating conditions and stall operating conditions.

[0138] Example, using frame energy vectors The sample feature vector has the following class label: Where 0 represents normal operating conditions and 1 represents stall conditions. A random forest classifier is used for training, with the following parameters: forest size of 500 trees, Gini index as the splitting criterion, minimum number of samples per leaf node of 1, and training method of classification. 13 features are randomly selected as candidates for each split.

[0139] S42. The importance of each frequency band is evaluated using the external bag permutation method, and its expression is:

[0140] (6)

[0141] (7)

[0142] (8)

[0143] in, This represents the number of samples outside the bag. The prediction function for the classifier; The loss function is set to 0-1 loss. When calculating the importance of permutations, the first... Random permutation values ​​of the energy characteristics of each frequency band on the samples outside the bag; This refers to the error outside the bag. For the first j The sample vector obtained by randomly permuting the energy of each frequency band in the samples outside the bag; For the first j The importance score of each frequency band;

[0144] S43. Set the importance threshold, the expression of which is:

[0145] (9)

[0146] in, and These are the importance score sets for each frequency band. The mean and standard deviation, This is the importance threshold.

[0147] S44. Select the frequency bands whose importance scores are greater than the importance threshold to obtain the key frequency set.

[0148] Specifically, when At that time, the first Each frequency band was identified as a key frequency band, and its center frequency was recorded in the key frequency set. And sort them by frequency from smallest to largest. Indicates the first i The key frequency band (i.e., the first key frequency band) The center frequency corresponding to each frequency band i This represents the index of the key frequency band within the key frequency set. For example... Figure 5 The diagram shows the importance distribution of each frequency band obtained by the out-of-bag permutation method. Frequency bands exceeding the threshold are selected as critical frequency bands. In this embodiment, a total of 14 frequency bands were automatically selected as critical frequency bands.

[0149] The frame energy vector is a vector composed of the energy of each frequency band after dividing the spectrum of each frame of signal into multiple frequency bands. Specifically, it can be implemented by dividing the frequency bands according to the frequency conversion ratio after discrete Fourier transform and calculating the energy value, used to characterize the energy distribution of different frequency components. The out-of-bag permutation method utilizes out-of-bag data that was not selected as training samples during random forest training. After randomly permuting a certain feature dimension, the error change is calculated. Specifically, a 0-1 loss function can be used to quantify the difference in classification error before and after the permutation, used to evaluate the influence of a single frequency band on the classification result. The importance threshold is a selection criterion dynamically determined based on the statistical characteristics of the importance scores of all frequency bands. Specifically, a threshold can be calculated using a linear combination of the mean and standard deviation, used to adaptively distinguish between critical and non-critical frequency bands.

[0150] Specifically, a random forest classifier is trained to establish a mapping relationship between frame energy vectors and operational states. The importance of each frequency band is assessed using the changes in prediction error of out-of-bag samples. In the out-of-bag samples, after randomly permuting the energy of a certain frequency band, the difference in classification error before and after the permutation is calculated; a larger difference indicates a more significant impact of that frequency band on the classification result. A threshold is dynamically set based on the mean and standard deviation of the importance scores of all frequency bands, and frequency bands with importance scores exceeding the threshold are selected as key frequency bands. This process quantifies the impact of feature permutation on classification performance, automatically identifies sensitive frequency bands strongly correlated with stall states, and avoids subjective biases imposed by manual experience.

[0151] Compared to existing technologies, traditional methods rely on fixed thresholds or manual experience to select frequency bands, making it difficult to adapt to changes in data distribution under different operating conditions, leading to the omission of key features or interference from redundant frequency bands. This solution objectively quantifies the contribution of frequency bands through out-of-bag permutation and combines it with a dynamic threshold mechanism to achieve automated selection of key frequency bands, effectively solving the problem of insufficient recognition accuracy caused by reliance on manual experience.

[0152] Through the above technical solution, this application can accurately locate the sensitive frequency components of compressor stall faults, eliminate irrelevant noise interference, provide a high-discrimination frequency band set for subsequent feature extraction and classification, and significantly improve the accuracy and robustness of stall state identification.

[0153] In one optional implementation, step S5 includes:

[0154] S51, using key frequencies as a set Using each frequency as a vertex, construct a triangular filter bank; where the 1st frequency is the vertex. k Each filter at frequency f The expression for the response is:

[0155] (10)

[0156] in, Indicates the first Each filter at frequency The response at the location; Indicates the first i The center frequencies corresponding to the key frequency bands; Indicates the first i- The center frequency corresponding to a key frequency band; Indicates the first i+ The center frequency corresponding to a key frequency band;

[0157] It should be noted that the first Each filter at frequency The response at a given point is defined using the principle of shared values ​​between adjacent vertices and linear interpolation.

[0158] S52. When the spacing between adjacent frequencies exceeds the spacing threshold:

[0159] (11)

[0160] in, The spacing threshold; For compressor frequency;

[0161] Then in the interval Within, from the complete set of candidate frequencies Select two neighboring frequency points , As supplementary vertices, they are as follows:

[0162] (12)

[0163] The complete set of candidate frequencies is composed of coverage intervals. It consists of all the discrete frequency points within;

[0164] The expanded set of key frequencies is as follows:

[0165] (13)

[0166] S53. Reconstruct the triangular filter bank based on the expanded key frequency set to obtain the reconstructed adaptive filter bank.

[0167] The triangular filter bank refers to a set of filters with triangular frequency responses constructed using key frequencies as vertices. This can be achieved by defining the response curve of each filter using adjacent vertex sharing and linear interpolation principles, ensuring that each filter covers the transition region between adjacent key frequency bands. Neighborhood expansion refers to adding new vertices when the distance between adjacent key frequencies is too large. This can be achieved by selecting the two frequencies closest to the original vertices from the candidate frequency set within the interval, ensuring the continuity of the filter bank's coverage between key frequency bands. The candidate frequency set refers to the set of all discrete frequency points covering the target frequency range. This can be constructed using all discrete frequency points involved in acoustic signal spectral analysis, providing optional sources of supplementary vertices for neighborhood expansion.

[0168] Specifically, the construction of the triangular filter bank is based on the center frequency of the selected key frequency bands. Each filter forms a linear transition response between adjacent vertices, ensuring energy capture of the key frequency bands and their transition regions. When the distance between adjacent vertices exceeds a threshold set by the rotation factor, it indicates that the key frequency band distribution is too sparse. In this case, the two nearest frequencies in the candidate frequency set within the interval between the original vertices are selected as supplementary vertices. The expanded vertex set reconstructs a triangular filter bank with a denser distribution. This process dynamically adjusts the filter vertex positions, enabling the filter bank to adaptively match the actual distribution of stall characteristic frequency bands and avoid energy extraction omissions caused by the sparse initial selection of key frequency bands.

[0169] Through the above technical solution, this application achieves dynamic adaptation to the distribution of stall feature frequency bands, solving the problem of incomplete feature energy extraction caused by the fixed frequency band division of traditional fixed filter banks. By supplementing the transition vertices between key frequency bands through neighborhood expansion, the coverage gaps caused by the sparse distribution of the initial key frequency bands are effectively filled, ensuring that the filter bank can completely capture stall feature energy and improve the recognition accuracy of subsequent classification models.

[0170] In one optional implementation, step S6 includes:

[0171] S61. Pass the spectrum corresponding to the windowed signal of each frame through the reconstructed triangular filter bank and calculate the output energy of each filter; where, the first... The filter in the first... The output energy on a frame is extracted using the following formula:

[0172] (14)

[0173] in, express exist The value at;

[0174] S62. Combine the output energies of each filter to form a frame-level filter energy feature vector.

[0175] The candidate frequency set refers to the set of all discrete frequency points covering the spectral analysis range of the compressor's operating acoustic signal. It can be generated using equal-interval sampling or by frequency band division, and is used to provide candidate frequencies for supplementary vertices during neighborhood expansion. Supplementary vertices are two neighboring frequency points selected from the entire candidate frequency set when the distance between adjacent critical frequencies is too large. They can be determined using a nearest neighbor search algorithm, and are used to form transitional frequency points between the original critical frequencies. Neighborhood expansion refers to the process of dynamically adjusting the critical frequency set according to a distance threshold. Specifically, it can be triggered by determining whether the distance between adjacent vertices exceeds the threshold, and is used to enhance the local response capability of the filter bank to stall-sensitive frequencies.

[0176] Specifically, when the distance between adjacent key frequencies exceeds the frequency-relevance threshold, the system automatically searches the entire candidate frequency set for frequencies adjacent to the current key frequency, selecting the two nearest neighbors as supplementary vertices to add to the key frequency set. The expanded set reconstructs the triangular filter bank, ensuring that the distance between adjacent vertices meets the threshold requirement. This process dynamically adjusts the vertex distribution of the filter bank to ensure that the filter response covers the dense region where the stall characteristic frequency is located, avoiding the loss of characteristic energy information due to excessively large distances between the original key frequency bands.

[0177] Through the above technical solution, this application effectively solves the problem of discontinuous frequency response caused by excessive spacing between adjacent key frequency bands. By dynamically supplementing transition frequency points, the filter bank closely covers the stall-sensitive region, enhancing the completeness of energy feature extraction. While retaining the advantages of automatic key frequency band selection, this design further improves the filter bank's adaptability to stall characteristic frequencies under complex operating conditions, providing more discriminative feature inputs for subsequent neural network classification.

[0178] Continuing the example, an initial filter bank is constructed from the 14 key frequency bands obtained in step S4. This is then expanded to include 11 filters using a neighborhood expansion strategy to better extract stall features. Ultimately, the reconstructed filter bank contains 25 triangular filters, thus forming a 25-dimensional energy vector for each frame in subsequent classification. The adaptation of the filter bank and neighborhood expansion in this embodiment is as follows: Figure 3 As shown.

[0179] In one alternative implementation, such as Figure 4 As shown, the deep neural network model includes the following components connected in sequence:

[0180] The feature embedding layer is used to map the input filter energy vector into a high-dimensional feature representation.

[0181] Specifically, the feature embedding layer for the first Frame energy vector ( Perform a linear mapping for each component. Perform a linear transformation to obtain dimensional vector

[0182] (15)

[0183] in, , These are the weight matrix and bias vector of the feature embedding layer, respectively; For the first The filter in the first... Energy value on the frame;

[0184] Then, arranged according to the filter index order, the embedding matrix is ​​obtained. :

[0185] (16)

[0186] The local feature extraction module is used to extract locally relevant features from the embedded features.

[0187] The global feature fusion module is used to capture long-term dependencies in feature sequences;

[0188] The classification output layer is used to output the final working condition identification result.

[0189] In one alternative implementation, the local feature extraction module includes a one-dimensional convolutional layer, a batch normalization layer, a ReLU activation layer, and an average pooling layer connected in sequence.

[0190] Specifically, in the filter index direction pair Perform a one-dimensional convolution with a kernel size of 3, a stride of 1, and padding of the same length. Assume the number of output channels is... (Pick ), convolution kernel weights bias The convolution output is:

[0191] (17)

[0192] in, For the first Frame in frequency band index The multi-channel representation after linear mapping; For the neighboring area Apply as a local representation after convolution. .

[0193] right Perform batch normalization and ReLU activation:

[0194] (18)

[0195] (19)

[0196] in, To Batch normalization (BN) and nonlinear activation characteristics eliminate differences in amplitude and dimensions; To The output after ReLU activation; and These represent the learnable scaling factor and translation factor, respectively; and This indicates that the mean and variance are calculated by channel. This represents element-wise multiplication; It is a constant.

[0197] right Average pooling is performed along the filter index direction to obtain the embedding vector for this frame:

[0198] (20).

[0199] The global feature fusion module is a Transformer encoder, which includes a multi-head self-attention layer, a feedforward neural network layer, a residual connection mechanism, and a normalization layer connected in sequence.

[0200] Specifically, the frame sequence After adding positional encoding, the input is fused into a two-layer Transformer encoder for global feature fusion; each layer consists of multi-head self-attention (4 heads in this example) + feedforward network + residual and normalization. The attention calculation is as follows:

[0201] (twenty one)

[0202] in , , This is the query, key, and value matrix obtained by linear mapping from the input sequence. To scale the dimensions.

[0203] Specifically, this deep neural network achieves multi-scale feature learning through a hierarchical processing mechanism. The feature embedding layer maps the input filter energy vectors into a high-dimensional space, enhancing feature separability. The local feature extraction module performs one-dimensional convolution operations along the filter index direction, capturing energy distribution patterns between adjacent frequency bands through a sliding window to detect local spectral distortion features caused by stall. The global feature fusion module performs self-attention calculations on the frame sequence after local feature extraction, establishing correlation weights between features at different time steps to capture the temporal evolution patterns across frames during stall development. The classification output layer performs dimensionality compression and nonlinear transformation on the fused spatiotemporal features, ultimately outputting the operating condition discrimination probability. These components work collaboratively to form a comprehensive feature learning system from local details to global context.

[0204] Compared to existing technologies, traditional methods often employ single convolutional or fully connected networks to process acoustic features, making it difficult to simultaneously extract local mutation features and long-term evolution patterns. Existing convolutional neural networks are limited by their local receptive fields, failing to effectively model temporal dependencies across frames; while recurrent neural networks suffer from the vanishing gradient problem, resulting in insufficient ability to model long sequences. This proposed solution, through a combination of local convolution and global self-attention, retains sensitivity to local features while achieving long-range dependency modeling, overcoming the inherent limitations of single architectures.

[0205] Through the above technical solution, this application effectively improves the ability to capture the acoustic features of compressor stall, enabling the simultaneous extraction of local frequency band energy mutation features and global temporal evolution patterns, thus enhancing adaptability and robustness in complex operating conditions. This network structure, through a multi-scale feature fusion mechanism, solves the problem of insufficient modeling of the dynamic stall process in traditional methods, significantly improving the accuracy and stability of fault identification.

[0206] In one alternative implementation, the classification output layer includes:

[0207] A global average pooling layer aggregates sequence features;

[0208] The first fully connected layer compresses the feature dimension to 32 dimensions;

[0209] Dropout layer, drop rate set to 0.2;

[0210] The second fully connected layer outputs a 2D feature vector.

[0211] The Softmax function calculates the final classification probability.

[0212] Specifically, the output of the Transformer encoder is globally pooled and then input into the first fully connected layer. This first compresses the feature dimension from 64 to 32, then maps it to a 2D output, corresponding to the normal and stall classes. The ReLU activation function is used in the hidden layers, and a Dropout of 0.2 is introduced to improve generalization performance. Let the 2D vector output by the fully connected layer be... ,in and These represent the scores for the "normal" and "stall" categories, respectively. The final classification probability is obtained using Softmax.

[0213] (twenty two)

[0214] in, Under normal operating conditions, This is a stall condition.

[0215] This embodiment was validated on 17 sets of stall test data collected. The recognition accuracy of this method under different speed conditions reached over 95%. Figure 6 The recognition results of this method on the test set are presented.

[0216] Secondly, the present invention also provides an acoustic identification system for compressor stall faults, comprising:

[0217] The acquisition module is used to acquire acoustic signals during the operation of the compressor through an acoustic sensor to obtain a discrete time-domain sequence; the acoustic sensor is installed at the opening of the compressor casing;

[0218] The first preprocessing module is used to perform frame division and windowing processing on the discrete time domain sequence to obtain the windowed signal of each frame;

[0219] The second preprocessing module is used to perform spectral analysis on the windowed signal of each frame, calculate the spectrum corresponding to the windowed signal of each frame, and divide the amplitude spectrum of each frame into multiple frequency bands, calculate the energy of each frequency band, and generate a frame energy vector; the amplitude spectrum is obtained by extracting the amplitude information from the spectrum.

[0220] The feature filtering module is used to evaluate the importance of each frequency band based on the random forest bag-out permutation method, and to filter out the key frequency bands related to the stall state to obtain a set of key frequencies; the set of key frequencies is composed of the center frequencies corresponding to each key frequency band.

[0221] An adaptive filtering unit is used to construct a triangular filter bank with each frequency in the key frequency set as the vertex, and to expand the neighborhood when the spacing between adjacent frequencies exceeds the spacing threshold, so as to obtain the reconstructed adaptive filter bank.

[0222] The filter energy feature vector generation module is used to process each frame of signal through the reconstructed adaptive filter bank to obtain the filter energy feature vector;

[0223] The classification module is used to input the filter energy feature vector into the deep neural network model for processing, and obtain the identification result of whether the compressor is in a stall state.

[0224] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. An acoustic identification method for compressor stall faults, characterized in that, Includes the following steps: S1. Acoustic signals during compressor operation are collected using an acoustic sensor to obtain a discrete time-domain sequence; the acoustic sensor is installed at the opening of the compressor casing; S2. Perform framing and windowing processing on the discrete time-domain sequence to obtain the windowed signal for each frame; S3. Perform spectrum analysis on the windowed signal of each frame and calculate the spectrum corresponding to the windowed signal of each frame; and divide the amplitude spectrum of each frame into multiple equal-width frequency bands according to the frequency conversion ratio, calculate the energy of each frequency band, and generate a frame energy vector. The amplitude spectrum is obtained by extracting amplitude information from the spectrum; S4. The importance of each frequency band is evaluated based on the random forest bag permutation method, and key frequency bands related to the stall state are screened to obtain a set of key frequencies; the set of key frequencies is composed of the center frequencies corresponding to each key frequency band. S4 includes: S41. Using frame energy vectors as features and operating condition labels as outputs, train a random forest classifier; the operating condition labels include normal operating conditions and stall operating conditions. S42. The importance of each frequency band is evaluated using the external bag permutation method, and its expression is: （6） （7） （8） in, This represents the number of samples outside the bag. The prediction function for the classifier; The loss function is set to 0-1 loss. When calculating the importance of permutations, the first... Random permutation values ​​of the energy characteristics of each frequency band on the samples outside the bag; This is for the error outside the bag; For the first j The sample vector obtained by randomly permuting the energy of each frequency band in the samples outside the bag; For the first j The importance score of each frequency band; S43. Set the importance threshold, the expression of which is: （9） in, and The importance sets for each frequency band are respectively. The mean and standard deviation, Importance threshold; S44. Select the frequency bands whose importance scores are greater than the importance threshold to obtain the key frequency set; S5. Construct a triangular filter bank with each frequency in the key frequency set as the vertex, and expand the neighborhood when the distance between adjacent frequencies exceeds the distance threshold to obtain the reconstructed adaptive filter bank. S5 includes: S51. Construct a triangular filter bank using each frequency in the key frequency set as a vertex; where, the... k Each filter at frequency f The expression for the response is: （10） in, Indicates the first Each filter at frequency The response at the location; Indicates the first i The center frequencies corresponding to the key frequency bands; Indicates the first i- The center frequency corresponding to a key frequency band; Indicates the first i+ The center frequency corresponding to a key frequency band; No. Each filter at frequency The response at a given point is defined using the principle of shared values ​​between adjacent vertices and linear interpolation. S52. When the spacing between adjacent frequencies exceeds the spacing threshold: （11） in, The spacing threshold; For compressor frequency; Then perform a neighborhood expansion within the interval. Within, from the complete set of candidate frequencies Select two neighboring frequency points , As supplementary vertices, they are as follows: （12） The complete set of candidate frequencies is composed of coverage intervals. It consists of all the discrete frequency points within; The expanded set of key frequencies is as follows: （13） S53. Reconstruct the triangular filter bank based on the expanded key frequency set to obtain the reconstructed adaptive filter bank. S6. Process the windowed signal of each frame through the reconstructed adaptive filter bank to obtain the filter energy feature vector; S7. Input the filter energy feature vector into the deep neural network model for processing to obtain the identification result of whether the compressor is in a stall state; The deep neural network model includes the following components connected in sequence: The feature embedding layer is used to map the input filter energy vector into a high-dimensional feature representation; The local feature extraction module is used to extract locally relevant features from the embedded features; The global feature fusion module is used to capture long-term dependencies in feature sequences; The classification output layer is used to output the final working condition identification result; The local feature extraction module includes a one-dimensional convolutional layer, a batch normalization layer, a ReLU activation layer, and an average pooling layer connected in sequence. The global feature fusion module is a Transformer encoder, which includes a multi-head self-attention layer, a feedforward neural network layer, a residual connection mechanism, and a normalization layer connected in sequence.

2. The method according to claim 1, characterized in that, S2 includes: S21. Calculate the total number of frames in the discrete-time sequence. K Its expression is: （1） in, The total number of sampling points for the acoustic signal. For frame length, For frame shift; S22. A window function is used for weighted processing to obtain the windowed signal for each frame, the expression of which is: （2） in, For the index of the sampling points within each frame, For window functions, For the first Windowing signal of a frame; Window function The expression is: （3）。 3. The method according to claim 2, characterized in that, S3 includes: S31. Perform a Fourier transform on the windowed signal of each frame to obtain the spectrum corresponding to each frame signal. Its expression is: （4） in, The signal sampling rate is expressed in Hz. For the first A discrete frequency; Frequency index; S32. Divide the spectrum of each frame into multiple equal-width frequency bands according to the frequency conversion ratio; S33. Calculate the energy of each frequency band, the expression is: （5） in, For the first m The first frame in the spectrum j The energy value of the frequency band; For the first m The first frame in the spectrum The frequency index set contained in each frequency band; S34. Generate a frame energy vector based on the energy of each frequency band of the windowed signal in each frame.

4. The method according to claim 1, characterized in that, S6 includes: S61. Pass the spectrum corresponding to the windowed signal of each frame through the reconstructed triangular filter bank and calculate the output energy of each filter; where, the first... The filter in the first... The output energy on a frame is extracted using the following formula: （14） in, express exist The value at; S62. Combine the output energies of each filter to form a frame-level filter energy feature vector.

5. The method according to claim 1, characterized in that, The classification output layer includes: A global average pooling layer aggregates sequence features; The first fully connected layer compresses the feature dimension to 32 dimensions; Dropout layer, drop rate set to 0.2; The second fully connected layer outputs a 2D feature vector. The Softmax function calculates the final classification probability.

6. An acoustic identification system for compressor stall faults, used to implement the method described in any one of claims 1-5, characterized in that, include: The acquisition module is used to acquire acoustic signals during the operation of the compressor through an acoustic sensor to obtain a discrete time-domain sequence; the acoustic sensor is installed at the opening of the compressor casing; The first preprocessing module is used to perform frame division and windowing processing on the discrete time domain sequence to obtain the windowed signal of each frame; The second preprocessing module is used to perform spectral analysis on the windowed signal of each frame and calculate the spectrum corresponding to the windowed signal of each frame. Additionally, the frame spectrum is divided into multiple equal-width frequency bands according to the frequency conversion ratio, the energy of each frequency band is calculated, and a frame energy vector is generated. The amplitude spectrum is obtained by extracting amplitude information from the spectrum; The feature filtering module is used to evaluate the importance of each frequency band based on the random forest bag-out permutation method, and to filter out the key frequency bands related to the stall state to obtain a set of key frequencies; the set of key frequencies is composed of the center frequencies corresponding to each key frequency band. An adaptive filtering unit is used to construct a triangular filter bank with each frequency in the key frequency set as the vertex, and to expand the neighborhood when the spacing between adjacent frequencies exceeds the spacing threshold, so as to obtain the reconstructed adaptive filter bank. The filter energy feature vector generation module is used to process the windowed signal of each frame through the reconstructed adaptive filter bank to obtain the filter energy feature vector. The classification module is used to input the filter energy feature vector into the deep neural network model for processing, and obtain the identification result of whether the compressor is in a stall state.

Images