A method and device for suppressing surge of a centrifugal air compressor

Abstract

The present invention relates to the technical field of air compressors, and discloses a method and device for suppressing surging of a centrifugal air compressor. The method includes: extracting pressure pulsation characteristics from the pressure signal of the centrifugal air compressor to obtain pressure pulsation characteristics; extracting flow statistical characteristics and band energy characteristics of the flow signal to obtain a surging early warning feature space; performing feature decomposition on the bearing vibration signal, extracting feature frequencies, and obtaining a surging early warning feature set; inputting the surging early warning feature set into a three-channel neural network for surging early warning prediction to obtain a surging early warning level; generating control parameters for a guide vane angle controller, an anti-surge valve opening controller, and a reflux valve opening controller according to the surging early warning level, and synchronously adjusting the guide vane angle, the anti-surge valve opening, and the reflux valve opening, and calculating a surging margin value. The present invention realizes precise control of the guide vane angle, the anti-surge valve opening, and the reflux valve opening, calculates the surging margin value, and ensures real-time optimization and adjustment of the control strategy.

Description

Technical Field

[0001] The present invention relates to the technical field of air compressors, and in particular to a method and device for suppressing surge of a centrifugal air compressor. Background Art

[0002] As a key equipment in industrial production, the operation stability of centrifugal air compressors directly affects production efficiency and safety. In actual operation, due to the frequent changes in operating conditions, the internal flow field structure of the air compressor is complex and changeable, and surge is prone to occur. At present, most surge warning methods mainly rely on a single pressure or flow signal for analysis, which makes it difficult to accurately capture the weak signs before surge occurs, resulting in insufficient warning accuracy and affecting the surge suppression effect.

[0003] Traditional surge suppression methods usually use a fixed threshold control strategy, which cannot be dynamically adjusted according to the real-time operating status of the air compressor. This control method is prone to response lag or over-adjustment when facing rapid changes in parameters such as speed and pressure ratio. At the same time, due to the lack of collaborative analysis capabilities for multi-source signals, existing methods are difficult to accurately determine the critical state of surge occurrence, resulting in insufficient optimization of control parameters and affecting the operating efficiency of the air compressor. Summary of the invention

[0004] The present invention provides a method and device for suppressing surge of a centrifugal air compressor, which realizes accurate control of the guide vane angle, the opening of an anti-surge valve and the opening of a return valve, calculates the surge margin value, and ensures real-time optimization and adjustment of the control strategy.

[0005] In a first aspect, the present invention provides a method for suppressing surge of a centrifugal air compressor, the method comprising:

[0006] Extract the pressure pulsation characteristics of the pressure signal of the centrifugal air compressor to obtain the pressure pulsation characteristics;

[0007] Extracting flow statistics characteristics and frequency band energy characteristics of the flow signal in the centrifugal air compressor, and combining them with the pressure pulsation characteristics to obtain a surge warning feature space;

[0008] Performing feature decomposition on the bearing vibration signal of the centrifugal air compressor, extracting the characteristic frequency, and combining it with the surge warning feature space to obtain a surge warning feature set;

[0009] Inputting the surge warning feature set into a three-channel neural network to perform surge warning prediction to obtain a surge warning level;

[0010] The control parameters of the guide vane angle controller, the anti-surge valve opening controller and the return valve opening controller are generated according to the surge warning level, and the guide vane angle, the anti-surge valve opening and the return valve opening are adjusted synchronously to calculate the surge margin value.

[0011] In a second aspect, the present invention provides a surge suppression device for a centrifugal air compressor, and the surge suppression device for the centrifugal air compressor includes:

[0012] A feature extraction module, configured to perform pressure pulsation feature extraction on the pressure signal of the centrifugal air compressor to obtain pressure pulsation features;

[0013] A feature combination module, configured to extract the flow statistical features and band energy features of the flow signal in the centrifugal air compressor, and combine them with the pressure pulsation features to obtain a surge warning feature space;

[0014] A feature decomposition module, configured to perform feature decomposition on the bearing vibration signal of the centrifugal air compressor, extract feature frequencies, and combine them with the surge warning feature space to obtain a surge warning feature set;

[0015] A warning prediction module, configured to input the surge warning feature set into a three-channel neural network for surge warning prediction to obtain a surge warning level;

[0016] A calculation module, configured to generate control parameters of a guide vane angle controller, an anti-surge valve opening controller, and a reflux valve opening controller according to the surge warning level, and synchronously adjust the guide vane angle, the anti-surge valve opening, and the reflux valve opening, and calculate a surge margin value.

[0017] In the technical solution provided by the present invention, through multi-scale analysis of wavelet transform and Fourier transform on the pressure signal, combined with dimensionality reduction processing by the principal component analysis method, the time-frequency features of pressure pulsation are effectively extracted, redundant information is reduced, and the efficiency of feature extraction is improved. The flow signal is processed by a method combining Butterworth filtering and wavelet packet transform. By extracting statistical features and band energy features, the dynamic characteristics of flow fluctuations are accurately captured, and the reliability of surge warning is enhanced. Empirical mode decomposition and Hilbert envelope spectrum analysis are introduced to process the bearing vibration signal. Combined with the envelope demodulation technology, the vibration characteristics of each bearing component are accurately extracted, and the information dimension of surge warning is expanded. A three-channel parallel deep neural network structure is designed. Different types of features are processed by convolutional layers, fully connected layers, and LSTM networks respectively, and an attention mechanism is used for feature fusion to construct an accurate surge warning model. A multi-parameter collaborative optimization method based on fuzzy rules and genetic algorithms is proposed. Through dynamic weight update and objective function optimization, precise control of the guide vane angle, the anti-surge valve opening, and the reflux valve opening is achieved. A complete surge suppression effect evaluation system is established. By monitoring the pressure ratio change rate, the flow fluctuation amplitude, and the bearing vibration intensity, the surge margin value is calculated to ensure the real-time optimization and adjustment of the control strategy. Description of the Drawings

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0019] Figure 1 It is a schematic diagram of the steps of the centrifugal air compressor surge suppression method in the embodiments of the present invention;

[0020] Figure 2 It is a schematic diagram of the structure of the centrifugal air compressor surge suppression device in the embodiments of the present invention. Specific embodiments

[0021] The embodiments of the present invention provide a centrifugal air compressor surge suppression method and device. Terms such as "first", "second", "third", "fourth", etc. (if any) in the specification, claims and above-mentioned drawings of the present invention are used to distinguish similar objects, and do not have to be used to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments described here can be implemented in an order different from that shown or described here. In addition, the terms "include" or "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0022] For ease of understanding, the following describes the specific process of the embodiments of the present invention. Please refer to Figure 1 , an embodiment of the centrifugal air compressor surge suppression method in the embodiments of the present invention includes:

[0023] Step S1: Extract the pressure pulsation characteristics of the pressure signal of the centrifugal air compressor to obtain the pressure pulsation characteristics;

[0024] It can be understood that the execution subject of the present invention can be a centrifugal air compressor surge suppression device, or a terminal or a server. Specifically, it is not limited here. The embodiments of the present invention take the server as the execution subject as an example for illustration.

[0025] Specifically, the inlet pressure signal, outlet pressure signal, and intermediate stage pressure signal of the centrifugal air compressor are collected to construct the original pressure signal sequence. These pressure signals are from real-time measurements of sensors, obtained at a high sampling rate, and due to the complexity of fluid dynamics, these signals contain rich transient and periodic components, and are affected by turbulent pulsation, pipeline resonance, and changes in operating conditions. The original pressure signal sequence is input into the db4 wavelet basis function for 4-layer wavelet decomposition to analyze the local characteristics of the signal at multiple scales. Wavelet transform has good time-frequency localization characteristics and can analyze the time-domain and frequency-domain information of the signal simultaneously. The high-frequency detail components of the signal are extracted at the first layer, which contain high-frequency pulse noise or rapidly changing information, while the medium-low frequency components of the signal are gradually extracted from the second layer to the fourth layer to capture the fluid instability characteristics before surge occurs. The multi-scale characteristics of the pressure signal are obtained through decomposition. The discrete Fourier transform is performed on the multi-scale characteristics to obtain the frequency-domain characteristic sequence of the signal. The discrete Fourier transform converts the time-series signal into the expression form of frequency components and identifies the characteristic frequencies related to surge. During the calculation process, the Fourier transform is performed on the wavelet components of different scales respectively. Statistical analysis is performed on the original pressure signal sequence to extract important parameters for evaluating the dynamic characteristics of the signal, including indicators such as root mean square value, peak factor, waveform factor, and impulse factor. The root mean square value measures the overall change in signal energy, the peak factor is used to evaluate the extreme value characteristics of the signal, the waveform factor reflects the complexity of the signal waveform, and the impulse factor characterizes the mutation phenomenon in the signal. These statistical characteristics are combined with the frequency-domain characteristics to form the pressure pulsation feature vector. Principal component analysis is performed on the pressure pulsation feature vector to reduce the data dimension and remove redundant information. By constructing a new orthogonal coordinate system to maximize the data variance, the computational complexity is reduced while key information is retained. Specifically, the covariance matrix of the eigenvectors is calculated, and its eigenvalues and eigenvectors are solved, and then the principal components with the largest variance information content are selected, and finally the reduced-dimensional pressure pulsation characteristics are obtained.

[0026] Step S2: Extract the flow statistical characteristics and frequency band energy characteristics of the flow signal in the centrifugal air compressor, and combine them with the pressure pulsation characteristics to obtain the surge warning feature space;

[0027] Specifically, the flow signals collected by the inlet flow sensor and the outlet flow sensor are sampled to obtain the original flow signals. The original flow signals are denoised. Methods such as wavelet denoising, adaptive filtering, or empirical mode decomposition are used to process the original flow signals to remove high-frequency noise while retaining the key surge characteristic information. After filtering, a relatively stable flow signal is obtained. Statistical analysis is performed on the filtered flow signal to extract the key statistical features that characterize the flow variation characteristics. The statistical analysis mainly involves the mean, standard deviation, skewness, and kurtosis. The mean reflects the overall level of the flow, the standard deviation measures the amplitude of the flow fluctuation, the skewness describes the symmetry of the flow distribution, and the kurtosis is used to identify the sharpness of the flow signal. Usually, when a surge is about to occur, the standard deviation of the flow signal will increase abnormally, and the changes in skewness and kurtosis will also show specific trends. Therefore, these statistical features play an important role in surge warning. Wavelet packet decomposition is performed on the filtered flow signal to obtain its band coefficient matrix. Wavelet packet decomposition is an extended version of wavelet transform that can decompose the entire frequency band of a signal and provides finer frequency resolution compared to traditional wavelet transform. Wavelet packet decomposition decomposes the flow signal layer by layer into sub-signals of different frequency bands, and each sub-signal corresponds to the information in a different frequency band, forming a band coefficient matrix. Energy eigenvalues are calculated for the band coefficient matrix to extract the energy distribution of each frequency band. The calculation method of the energy feature is based on the square integral of the signal, that is, the signals of each band component are squared and then summed to obtain the energy of that band. By calculating the energy eigenvalues of all sub-bands, a band energy feature vector is obtained, which reflects the energy distribution of the flow signal in different frequency ranges. The flow statistical features, band energy features, and pressure pulsation features are stitched together to construct a surge warning feature space.

[0028] Step S3: Perform feature decomposition on the bearing vibration signal of the centrifugal air compressor, extract the characteristic frequencies, and combine them with the surge warning feature space to obtain a surge warning feature set;

[0029] Specifically, high-precision acceleration sensors are installed at the inlet and outlet bearings of the centrifugal air compressor, and bearing vibration signals are collected at a sufficiently high sampling rate to obtain the original vibration signal sequence. The original vibration signal is subjected to empirical mode decomposition to decompose the non-stationary signal into a series of intrinsic mode functions, and these modal components reflect the vibration characteristics at different scales, obtaining the component matrix of the vibration signal. Hilbert transform is applied to each intrinsic mode function component of the vibration signal component matrix to calculate its instantaneous frequency and envelope spectrum, obtaining the spectral feature sequence. Feature frequency extraction is performed on the spectral feature sequence to calculate the key frequency parameters of the bearing, including the inner race passing frequency, outer race passing frequency, rolling element passing frequency, and cage rotation frequency. These characteristic frequencies are determined by the structural parameters of the bearing, are related to the damage state inside the bearing, and when surging occurs, the force state of the bearing changes, resulting in abnormal energy distribution of these characteristic frequencies. By calculating these bearing characteristic frequencies, the abnormal vibration of the bearing caused by surging can be effectively identified. Envelope demodulation analysis is performed on the bearing characteristic frequencies to extract the vibration characteristics of each component of the bearing. Envelope demodulation uses band-pass filtering and Hilbert transform to extract low-frequency modulation information from high-frequency signals, and this method is suitable for detecting the fault characteristics of bearings. By demodulating the envelope signal of the bearing characteristic frequencies, the vibration characteristics of each component are obtained, such as the vibration modes of the inner race, outer race, rolling elements, and cage, and their time-domain and frequency-domain characteristics are calculated to form a bearing vibration feature vector to reflect the health state of the bearing. When surging occurs, the instability of the air flow will cause changes in the bearing load, resulting in the enhancement or attenuation of specific vibration frequencies. The bearing vibration feature vector is combined with the surging warning feature space to fuse the multi-modal information of the pressure signal, flow signal, and bearing vibration signal to form a high-dimensional feature set. The feature combination method uses feature splicing or data fusion technology to arrange the features from different signal sources in a certain order, and optimizes the data structure through dimensionality reduction or feature selection methods to ensure the effectiveness of the features and the calculation efficiency, obtaining the surging warning feature set.

[0030] Step S4: Input the surging warning feature set into a three-channel neural network for surging warning prediction to obtain the surging warning level;

[0031] Specifically, the surge warning feature set is input into a three-channel neural network, and corresponding features are extracted in three different channels respectively, so as to make full use of the characteristics of different signals and improve the accuracy and robustness of surge prediction. In the entire network architecture, the pressure pulsation features in the surge warning feature set are input into the first channel of the three-channel neural network, and layer-by-layer feature extraction is performed through four convolutional layers in this channel to obtain the deep pattern information of the pressure signal. During the convolution operation, multiple different convolutional kernels are applied in each layer to extract local features, and the ReLU activation function is introduced to bring in the non-linear mapping ability, enabling the model to learn more complex feature representations. A max-pooling operation is added after each convolutional layer to reduce the size of the feature map, improve the computational efficiency, and enhance the translational invariance of the model. After these four convolutional processes, the pressure pulsation features are converted into a high-dimensional pressure feature map, reflecting the pattern changes of pressure pulsation at different time scales and effectively capturing the abnormal trends leading to surge. At the same time, the flow statistical features and band energy features in the surge warning feature set are input into the second channel of the three-channel neural network, and feature transformation is sequentially performed through three fully connected layers in this channel to obtain a more compact and distinguishable flow feature representation. In the first fully connected layer, the number of neurons is relatively large to ensure sufficient capture of the distribution information of the original flow data. After passing through the ReLU activation, a non-linear transformation is introduced to enhance the modeling ability for complex flow patterns. The second fully connected layer performs dimensionality reduction on the features, enabling the model to maintain the key information of the flow signal in a lower-dimensional space and eliminating redundant features. In the third fully connected layer, the network further transforms the dimensionality-reduced flow features to ensure the final flow feature representation is obtained, which effectively depicts the temporal changes of the flow and the abnormal pattern of the flow energy distribution during surge. At the same time, the bearing vibration feature vector in the surge warning feature set is input into the third channel of the three-channel neural network, and temporal feature extraction is performed through two layers of LSTM (Long Short-Term Memory network) in this channel to obtain the temporal features of the bearing vibration signal. In the first LSTM network layer, the model learns the short-term dependence features of the vibration signal and captures the local change trends, while in the second LSTM network layer, the model integrates the long-term dependence information to extract the temporal evolution law of the bearing vibration features before surge. After two layers of LSTM processing, the obtained frequency feature sequence fully characterizes the bearing vibration pattern changes during surge. Attention weights are calculated for the pressure feature map, flow feature representation, and frequency feature sequence respectively, and a feature weight matrix is generated. When calculating the attention weights, each feature is adaptively weighted, that is, through a learnable parameter matrix, the contribution degree of each feature is dynamically adjusted, so that features with greater predictive value obtain higher weights, while features with more noise or higher redundancy are weakened.According to the feature weight matrix, weighted fusion is performed on the pressure feature mapping, flow rate feature representation, and frequency feature sequence to form a fused feature vector. During the fusion process, the attention mechanism assigns different weights to different features to ensure that key features occupy a larger proportion in the final prediction, while avoiding the degradation of model performance caused by information redundancy. The fused feature vector is input into the output layer of the three-channel neural network, and the probability distribution of different surge warning levels is calculated through a softmax classifier. The softmax classifier maps the fused feature vector to three different surge levels and calculates the corresponding probability values for each level. Finally, the level corresponding to the maximum probability value is selected as the final surge warning level.

[0032] Step S5: Generate the control parameters of the guide vane angle controller, anti-surge valve opening controller, and reflux valve opening controller according to the surge warning level, synchronously adjust the guide vane angle, anti-surge valve opening, and reflux valve opening, and calculate the surge margin value.

[0033] Specifically, discretize the control ranges of the guide vane angle controller, the anti-surge valve opening controller, and the return valve opening controller so that optimization calculations can be performed on a finite parameter set. During the discretization process, the range of the controller is divided into multiple discrete levels to form a control parameter space. Construct a fuzzy rule base based on the surge warning level for fuzzy inference operations according to different surge warning states. Fuzzy control uses linguistic rules to describe the relationship between the input and the output. Therefore, a series of fuzzy rules are set according to empirical data and surge characteristics. For example, "If the surge level is low, moderately adjust the guide vane angle, while the anti-surge valve and the return valve maintain the minimum adjustment amount", "If the surge level is high, quickly increase the opening of the anti-surge valve, and moderately adjust the guide vane angle and the return valve opening", etc. Input the discretized controller parameter space into the fuzzy rule base and use the fuzzy inference method to calculate and generate a fuzzy control decision sequence. This sequence adjusts the control parameters according to different surge levels to achieve dynamic regulation of surges. Set an optimization objective function according to the fuzzy control decision sequence so that the control strategy can achieve a reasonable balance among surge suppression, energy loss, and adjustment time. In the optimization objective function, set the surge suppression effect as the first weight to ensure that the primary goal of the control strategy is to effectively prevent surges from occurring. At the same time, set the energy loss as the second weight to avoid excessive adjustment of control parameters resulting in increased system energy consumption, and set the adjustment time as the third weight to ensure that the response speed of the control strategy is fast enough to complete effective adjustment before surges occur. During the construction of the objective function, different weights are adjusted according to the operating conditions. For example, in the case of a higher surge risk, increase the weight of the surge suppression effect, while in the stable operating condition, appropriately increase the weight of energy loss optimization to improve system efficiency. Based on the optimization objective function, construct a genetic algorithm population to intelligently optimize the control parameters. Generate chromosome encoding for the control parameters, where the discretized parameters of the guide vane angle, the anti-surge valve opening, and the return valve opening are converted into binary sequences, and an initial population is formed during the binary coding mapping process. Each individual represents a possible combination of control parameters, and the genetic algorithm continuously evolves these individuals to find the optimal control strategy. During the genetic calculation process, perform a selection operation to select better individuals according to the fitness value of the objective function to enter the next generation; then perform a crossover operation to generate new candidate solutions by randomly crossing the gene segments of different individuals; then perform a mutation operation to introduce more search possibilities by randomly changing some gene loci of the individuals, thus avoiding the algorithm falling into a local optimum. After several generations of evolutionary calculations, obtain the optimal chromosome, which corresponds to a set of optimal control parameters for the guide vane angle, the anti-surge valve opening, and the return valve opening. Perform decoding operations on the optimal chromosome to convert the binary coding into the corresponding physical control parameters to obtain the control parameters of the guide vane angle, the anti-surge valve opening, and the return valve opening.The adjustment of the guide vane angle is used to control the air flow entering the air compressor, thereby affecting the hydrodynamic characteristics of the system. The adjustment of the anti-surge valve opening is used to quickly release excessive pressure to prevent surging. The adjustment of the reflux valve opening realizes the reflux control of the flow rate, thereby optimizing the surge suppression effect. Based on the control parameters, the guide vane angle, the anti-surge valve opening, and the reflux valve opening are synchronously adjusted, the changes in the pressure pulsation characteristics, the flow rate statistical characteristics, and the characteristic frequencies are monitored, and the surge margin value is calculated. The surge margin is a key indicator for measuring the surge risk of the system, defined as the relative distance between the surge limit and the current operating state. By calculating the surge margin value, it is judged whether the current control strategy is effective and whether the adjustment parameters need to be further optimized. If the surge margin is high, it indicates that the current working condition is in a stable state, and the control parameters are appropriately adjusted back to reduce unnecessary energy losses. If the surge margin is low, it indicates that the system is still in the surge risk area, and the control strategy needs to be continuously optimized to ensure that surging is completely suppressed.

[0034] Input the control parameters into the servo motor drive system and the electro-hydraulic servo system to synchronously adjust the guide vane angle, the opening degree of the anti-surge valve, and the opening degree of the reflux valve. The control system receives the optimized control parameters and drives each actuator, enabling components such as the guide vane and valves to perform corresponding operations according to the set adjustment range, forming an actuator action sequence. Detect the response time of the actuator action sequence to evaluate the adjustment speed and dynamic response characteristics of the guide vane angle, the opening degree of the anti-surge valve, and the opening degree of the reflux valve. During this process, record the response time of the guide vane angle adjustment, the response time of the anti-surge valve opening degree adjustment, and the response time of the reflux valve opening degree adjustment respectively, store these data in the database of the control system, and perform dynamic feedback adjustment to optimize the control strategy of the actuator and improve the response ability of the overall system. Calculate the pressure ratio change rate, the flow rate fluctuation amplitude, and the bearing vibration intensity based on the response time data to quantitatively evaluate the anti-surge suppression effect and form a performance parameter matrix. The pressure ratio change rate is used to characterize the change trend of the compression ratio during the adjustment process of the system. If the pressure ratio change rate is too large, it indicates that the system is still in an unstable state and there is a risk of surge; the flow rate fluctuation amplitude is an important parameter to measure the flow rate adjustment effect. If the flow rate fluctuates too much, it means that the surge has not been fully suppressed, or the opening degree adjustment of the anti-surge valve and the reflux valve has not reached the optimal state; the bearing vibration intensity reflects the impact of anti-surge suppression on mechanical vibration. If the bearing vibration is still large after adjustment, the control strategy needs to be further optimized. By calculating these key parameters and integrating them into a performance parameter matrix, evaluate the dynamic behavior of the system during the anti-surge suppression process. Conduct a stability analysis on the performance parameter matrix, and calculate the overshoot, the settling time, and the steady-state error. Among them, the overshoot is used to measure whether there is an overshoot phenomenon in the control strategy during the adjustment process. An excessive overshoot causes the anti-surge suppression strategy to have a reverse effect and makes the system enter the unstable region. Therefore, it is necessary to optimize the control strategy as much as possible to reduce the overshoot; the settling time represents the time required for the system to enter the stable state from the start of adjustment. A short settling time means that the adjustment speed of the system is fast, which helps to improve the real-time performance of anti-surge suppression; while the steady-state error measures whether the final anti-surge suppression effect meets the accuracy requirements. A large steady-state error will lead to insufficient surge margin, thus affecting the long-term stable operation of the air compressor. During the stability analysis process, ensure the effectiveness of the anti-surge suppression strategy by calculating and optimizing these operating state indicators. Based on the operating state indicators obtained from the stability analysis, calculate the surge margin to quantitatively measure the current surge safety margin of the system. The calculation method of the surge margin adopts a weighted summation model. Among them, multiply the pressure ratio change rate by the first target value, the flow rate fluctuation amplitude by the second target value, and the bearing vibration intensity by the third target value, and then sum up these weighted results to obtain the final surge margin value.

[0035] In the embodiments of the present invention, through the multi-scale analysis of wavelet transform and Fourier transform on the pressure signal, combined with the dimensionality reduction processing of the principal component analysis method, the time-frequency characteristics of the pressure pulsation are effectively extracted, redundant information is reduced, and the efficiency of feature extraction is improved. The method of combining Butterworth filtering and wavelet packet transform is used to process the flow signal. By extracting statistical features and band energy features, the dynamic characteristics of the flow fluctuation are accurately captured, and the reliability of the surge warning is enhanced. The empirical mode decomposition and Hilbert envelope spectrum analysis are introduced to process the bearing vibration signal. Combined with the envelope demodulation technology, the vibration characteristics of each component of the bearing are accurately extracted, and the information dimension of the surge warning is expanded. A three-channel parallel deep neural network structure is designed. Different types of features are processed through the convolutional layer, fully connected layer and LSTM network respectively, and the attention mechanism is used for feature fusion to construct an accurate surge warning model. A multi-parameter collaborative optimization method based on fuzzy rules and genetic algorithms is proposed. Through dynamic weight update and objective function optimization, the accurate control of the guide vane angle, anti-surge valve opening and reflux valve opening is realized. A complete evaluation system for surge suppression effect is established. By monitoring the pressure ratio change rate, flow fluctuation amplitude and bearing vibration intensity, the surge margin value is calculated to ensure the real-time optimization and adjustment of the control strategy.

[0036] In a specific embodiment, the process of executing step S1 may specifically include the following steps:

[0037] Collect the inlet pressure signal, outlet pressure signal and intermediate stage pressure signal of the centrifugal air compressor to obtain the original pressure signal sequence;

[0038] Input the original pressure signal sequence into the db4 wavelet basis function for 4-layer wavelet decomposition to obtain the multi-scale features of the pressure signal;

[0039] Perform discrete Fourier transform on the multi-scale features of the pressure signal to obtain the frequency domain feature sequence;

[0040] Perform statistical analysis on the original pressure signal sequence, calculate the root mean square value, peak factor, waveform factor and pulse factor, and combine them with the frequency domain feature sequence to obtain the pressure pulsation feature vector;

[0041] Perform principal component analysis on the pressure pulsation feature vector to obtain the pressure pulsation feature.

[0042] Specifically, during the operation of the centrifugal air compressor, the pressure state of the internal air flow will be affected by factors such as impeller rotation, air flow disturbance and surge. Pressure sensors are arranged at different positions to collect the pressure signal, where the inlet pressure signal 、outlet pressure signal and intermediate stage pressure signal Obtain a complete time series through high-frequency sampling and combine these signals into an original pressure signal sequence:

[0043]

[0044] Among them, represents the pressure signal sequence varying with time, is the time variable, , and represent the pressure signals of the inlet, outlet, and intermediate stage respectively. Input the original pressure signal sequence into the db4 wavelet basis function for four-layer wavelet decomposition. By using four-layer wavelet decomposition, multi-scale analysis of the signal is carried out to obtain information in different frequency bands. The mathematical expression form of wavelet transform is as follows:

[0045]

[0046] Among them, is the wavelet coefficient, representing the component of the signal at different scales and time shift , is the wavelet basis function. For the four-layer decomposition process of the db4 wavelet basis, it is expressed as:

[0047]

[0048] Among them, represents the low-frequency component of the 4th layer, represent the high-frequency detail components of each layer respectively. The multi-scale feature can effectively separate the low-frequency trend and high-frequency perturbation in the signal. Perform discrete Fourier transform on the multi-scale features obtained by wavelet decomposition to obtain the frequency-domain feature sequence. The basic expression of Fourier transform is:

[0049]

[0050] Among them, represents the th frequency component in the frequency domain, is the signal length, is the imaginary unit, is the th sampling point of the original signal. Through Fourier transform, the amplitude spectrum of the signal is obtained:

[0051]

[0052] Among them, and are the real part and imaginary part of the Fourier transform result respectively. The amplitude spectrum Reflects the energy distribution of the signal at different frequencies, thus revealing the main oscillation modes of the pressure signal. While obtaining the frequency-domain features, statistical analysis is performed on the original pressure signal to calculate the root mean square value (RMS), peak factor, waveform factor, and impulse factor, which are used to characterize the overall characteristics of the pressure signal. The root mean square value is defined as:

[0053]

[0054] where, reflects the energy level of the signal, is the th sampling value of the signal, is the total number of samplings. The peak factor is used to measure the ratio of the maximum amplitude of the signal to the root mean square value, and its definition is:

[0055]

[0056] where, is the maximum peak value of the signal. The waveform factor measures the waveform complexity of the signal, and the calculation formula is as follows:

[0057]

[0058] The impulse factor reflects the degree of spikes in the signal and is defined as:

[0059]

[0060] By calculating the above statistical features and combining them with the frequency-domain features obtained by Fourier transform, a pressure pulsation feature vector is constructed:

[0061]

[0062] where, As the final high-dimensional feature vector, it contains key information in the time domain and frequency domain. Principal component analysis is performed on the pressure pulsation feature vector to reduce the dimension and extract the main feature components. Principal component analysis calculates the eigenvalues and eigenvectors through the covariance matrix and selects the principal components with the largest variance contribution. Calculate the covariance matrix:

[0063]

[0064] where, is the mean of the eigenvectors. Solve the eigenvalues and eigenvectors , and select the principal components corresponding to the first largest eigenvalues:

[0065]

[0066] where, is the pressure pulsation feature after dimensionality reduction.

[0067] In a specific embodiment, the process of performing step S2 may specifically include the following steps:

[0068] Sample the flow signals collected by the inlet flow sensor and the outlet flow sensor to obtain the original flow signals, and perform signal denoising on the original flow signals to obtain the filtered flow signals;

[0069] Perform statistical analysis on the filtered flow signals, calculate the mean, standard deviation, skewness, and kurtosis to obtain the flow statistical features, and perform wavelet packet decomposition on the filtered flow signals to obtain the band coefficient matrix;

[0070] Calculate the energy eigenvalues of the band coefficient matrix, extract the energy distribution of each band to obtain the band energy features;

[0071] Perform feature splicing on the flow statistical features, band energy features, and pressure pulsation features to obtain the surge warning feature space.

[0072] Specifically, obtain the inlet flow signal and the outlet flow signal through a high-precision flow sensor, and form these flow signals into an original flow signal sequence:

[0073]

[0074] wherein, represents the signal sequence of the flow changing with time, is the time variable, and and respectively represent the instantaneous flow values at the inlet and outlet. Perform signal denoising on the original flow signals to obtain a smoother and more effective flow signal. The methods of signal denoising include wavelet threshold denoising, empirical mode decomposition denoising, and adaptive filtering, etc. Among them, wavelet threshold denoising can better retain the mutation characteristics of the flow signal while removing high-frequency noise. Perform wavelet transform on the signal to decompose it into detail components at different scales and low-frequency components

[0075]

[0076] wherein, represents the wavelet low-frequency component of the fourth layer, and are the wavelet detail components of each layer respectively. Perform soft threshold processing on the detail components to remove high-frequency noise and reconstruct the signal to obtain the filtered flow signal Perform statistical analysis on the filtered flow signal to calculate the mean, standard deviation, skewness, and kurtosis of the flow rate, and form the flow rate statistical characteristics. The mean of the flow signal is used to measure the average flow rate of the system, and the calculation formula is:

[0077]

[0078] where is the total number of flow data points, represents the th data point of the filtered flow signal. The standard deviation reflects the degree of fluctuation of the flow signal, and the calculation formula is as follows:

[0079]

[0080] The skewness reflects the symmetry of the flow rate distribution, and is defined as follows:

[0081]

[0082] If , it means that the flow signal is skewed to the right, that is, there are more relatively high flow rate values. If , it means that the flow is skewed to the left, that is, there are more low flow rate values. The kurtosis measures the sharpness of the flow signal, and the calculation formula is:

[0083]

[0084] If is much greater than 3, it means that the flow signal has a spike feature and there are mutations. After obtaining the flow rate statistical characteristics, perform wavelet packet decomposition on the flow signal to extract the multi-band information of the flow signal. Wavelet packet decomposition not only decomposes the low-frequency part, but also decomposes the high-frequency part to obtain a finer time-frequency representation. For the filtered flow signal , its wavelet packet decomposition process is expressed as:

[0085]

[0086] where are the wavelet packet coefficients, representing the energy distribution of the signal at scale and time shift , is the wavelet packet basis function. After 4-layer wavelet packet decomposition, a frequency band coefficient matrix is obtained, and the coefficients of each frequency band reflect the characteristics of the flow signal at different scales. Calculate the frequency band energy characteristics, that is, perform energy calculation on the wavelet packet coefficient matrix, and the formula is as follows:

[0087]

[0088] Among them, represents the energy of the layer decomposed signal, is the corresponding decomposition coefficient. The energy values of all frequency bands constitute the frequency band energy feature vector:

[0089]

[0090] Among them, is the total number of decomposition layers. The flow rate statistical features, frequency band energy features and pressure pulsation features are concatenated to obtain the surge warning feature space:

[0091]

[0092] Among them, represents the pressure pulsation feature, is the final surge warning feature vector.

[0093] In a specific embodiment, the process of executing step S3 may specifically include the following steps:

[0094] Sample the bearing vibration signal collected by the acceleration sensors at the inlet and outlet bearings of the centrifugal air compressor to obtain the original vibration signal;

[0095] Perform empirical mode decomposition on the original vibration signal to obtain the vibration signal component matrix, and perform Hilbert transform on the vibration signal component matrix to obtain the spectral feature sequence;

[0096] Extract the characteristic frequencies from the spectral feature sequence, calculate the inner race passing frequency, outer race passing frequency, rolling element passing frequency and cage rotation frequency to obtain the bearing characteristic frequencies;

[0097] Perform envelope demodulation analysis on the bearing characteristic frequencies, extract the vibration characteristics of each bearing component to obtain the bearing vibration feature vector, and combine the bearing vibration feature vector with the surge warning feature space to obtain the surge warning feature set.

[0098] Specifically, high-precision acceleration sensors are arranged at the inlet and outlet bearings to collect the original vibration signal of the bearings:

[0099]

[0100] Among them, represents the acceleration signal sequence varying with time, and respectively represent the vibration signals of the inlet and outlet bearings, while is a time variable. Empirical Mode Decomposition (EMD) is used to decompose the signal and extract the vibration components of different modes. EMD decomposes the original signal through iterative decomposition into multiple Intrinsic Mode Functions (IMFs), and each IMF represents different frequency components of the signal. The decomposition process is expressed as:

[0101]

[0102] where represents the th intrinsic mode component, is the total number of decomposition layers, is the residual signal. To ensure that the decomposed IMF components have physical meanings, the first few IMF components are selected as the main vibration feature components, and the influence of the residual component is ignored. By this method, the original vibration signal is decomposed into multiple frequency components. The Hilbert transform is performed on the IMF components to obtain the instantaneous frequency characteristics of the vibration signal. The definition of the Hilbert transform is as follows:

[0103]

[0104] By calculating the instantaneous amplitude and instantaneous frequency, the time-frequency representation of the signal is obtained, where the instantaneous amplitude is calculated by the following formula:

[0105]

[0106] while the instantaneous frequency is calculated as follows:

[0107]

[0108] where represents the instantaneous envelope of the vibration signal, represents the instantaneous frequency. These characteristics help to identify the vibration modes of the bearing and analyze its dynamic changes during surging. Feature frequencies are extracted from the spectral feature sequence, and the Inner Race Passing Frequency (BPFI), Outer Race Passing Frequency (BPFO), Ball Spin Frequency (BSF), and Cage Rotation Frequency (FTF) of the bearing are calculated. The calculation formulas for these frequencies are as follows:

[0109]

[0110]

[0111]

[0112]

[0113] where is the number of balls, is the inner diameter of the bearing, is the diameter of the rolling element, is the contact angle, is the rotational frequency. By calculating these characteristic frequencies, the health state of the bearing is identified, and the abnormal vibration mode of the bearing caused by surge is detected. After obtaining the characteristic frequencies, envelope demodulation analysis is performed to extract the vibration characteristics of each component of the bearing. Envelope demodulation uses band-pass filtering and Hilbert transform to extract the low-frequency modulation information of the bearing vibration signal. The vibration signal is band-pass filtered, and the frequency band range is selected as , and then the Hilbert transform is performed on the filtered signal to obtain the envelope signal:

[0114]

[0115] where, is the filtered signal, is the envelope signal. By calculating the spectral characteristics of the envelope signal, the main vibration modes of the bearing are extracted, and the bearing vibration feature vector is formed:

[0116]

[0117] where, represents the vibration energy of the envelope signal at different frequencies. In order to construct the surge warning feature set, the bearing vibration feature vector is concatenated with the surge warning feature space to form the warning feature vector:

[0118]

[0119] where, represents the pressure pulsation feature, represents the flow statistical feature and the frequency band energy feature, is the bearing vibration feature vector, is the final surge warning feature vector.

[0120] In a specific embodiment, the process of executing step S4 may specifically include the following steps:

[0121] Input the pressure pulsation feature in the surge warning feature set into the first channel of the three-channel neural network, and perform feature extraction through four convolutional layers in the first channel to obtain the pressure feature map;

[0122] Input the flow statistical feature and the frequency band energy feature in the surge warning feature set into the second channel of the three-channel neural network, and perform feature transformation through three fully connected layers in the second channel in sequence to obtain the flow feature representation;

[0123] Input the bearing vibration feature vector in the surge warning feature set into the third channel of the three-channel neural network, and perform time series feature extraction through the two-layer LSTM network in the third channel to obtain the frequency feature sequence;

[0124] Calculate the attention weights for the pressure feature mapping, flow feature representation, and frequency feature sequence respectively to generate the feature weight matrix;

[0125] Perform weighted fusion on the pressure feature mapping, flow feature representation, and frequency feature sequence according to the feature weight matrix to obtain the fusion feature vector;

[0126] Input the fusion feature vector into the output layer of the three-channel neural network, calculate the probability distribution of the three levels through the softmax classifier, and take the level corresponding to the maximum probability value as the surge warning level.

[0127] Specifically, split the surge warning feature set into three parts, where the pressure pulsation feature is used as the input of the first channel, the flow statistical feature and the band energy feature are used as the input of the second channel, and the bearing vibration feature vector is used as the input of the third channel. Let the input feature set be , then it is expressed as:

[0128]

[0129] Among them, represents the pressure pulsation feature, represents the flow statistical feature and the band energy feature, represents the bearing vibration feature vector. In order to fully extract the pressure pulsation feature, is input into the first channel, and this channel uses a four-layer convolutional neural network to extract spatial features and enhance the local pattern representation of the pressure signal. Assume that the convolutional kernel is , and the bias term is , and the output of the th layer of convolution is expressed as:

[0130]

[0131] Among them, * represents the convolution operation, represents the activation function (ReLU), is the initial input feature. After four-layer convolution operations, the pressure feature mapping is obtained:

[0132]

[0133] At the same time, in order to extract the flow statistical feature and the band energy feature, is input into the second channel, and this channel uses a three-layer fully connected neural network for feature transformation and dimensionality reduction. Let the weight matrix be , the bias term is , then the calculation formula of the fully connected layer is as follows:

[0134]

[0135] Among them, is the initial input, and after being transformed by a three-layer fully connected network, the traffic feature representation is obtained:

[0136]

[0137] In order to extract the temporal features of bearing vibration, is input into the third channel, and this channel uses a two-layer long short-term memory network (LSTM) to learn the time dependence of bearing vibration. Let the input of the LSTM cell be , and its calculation method is:

[0138]

[0139] Among them, and are the weight matrices of the LSTM network, is the bias term, represents the hidden state at the -th time step. After two-layer LSTM calculation, the frequency feature sequence is obtained:

[0140]

[0141] In order to improve the prediction ability of the model, attention calculations are performed on the pressure feature map , the traffic feature representation and the frequency feature sequence to generate a feature weight matrix. Assuming that the weight parameter of the attention calculation is , then the feature attention score is calculated by the following formula:

[0142]

[0143] Among them, represents different feature channels. By calculating the attention weight matrix , the final weighted feature vector is obtained:

[0144]

[0145] The fused feature vector Input to the output layer of the three-channel neural network, and calculate the probability distribution of the surge warning level through the softmax classifier. The calculation formula of the softmax classifier is as follows:

[0146] ;

[0147] Among them, and are the weights and biases of the output layer, represents the probability belonging to the -level surge warning. Select the level corresponding to the maximum probability value as the surge warning level:

[0148] .

[0149] In a specific embodiment, the process of executing step S5 may specifically include the following steps:

[0150] Discretize the control ranges of the guide vane angle controller, the anti-surge valve opening controller, and the return valve opening controller to obtain the controller parameter space;

[0151] Construct a fuzzy rule base based on the surge warning level, input the controller parameter space into the fuzzy rule base for reasoning operations, and obtain a fuzzy control decision sequence;

[0152] Set an optimization objective function according to the fuzzy control decision sequence, set the surge suppression effect as the first weight, the energy loss as the second weight, and the adjustment time as the third weight to obtain the parameter optimization objective;

[0153] Construct a genetic algorithm population according to the parameter optimization objective, generate chromosome coding, and perform binary coding mapping on the controller parameter space to obtain the initial population;

[0154] Perform genetic evolution calculations on the initial population to obtain the optimal chromosome, and perform decoding operations on the optimal chromosome to convert the binary coding into the control parameters of the guide vane angle, the anti-surge valve opening, and the return valve opening;

[0155] Synchronously adjust the guide vane angle, the anti-surge valve opening, and the return valve opening based on the control parameters, monitor the changes in the pressure pulsation characteristics, the flow statistical characteristics, and the characteristic frequencies, and calculate the surge margin value.

[0156] Specifically, discretize the control parameters to construct the controller parameter space. In the surge control system of the centrifugal air compressor, the guide vane angle controls the angle range of the flow entering the compressor, the anti-surge valve opening controls the exhaust recirculation amount, and the return valve opening Control the redistribution amount of air flow in the system. Discretize the three control variables respectively to form a searchable control parameter space. Let the discretization levels be , and , then the controller parameter space is expressed as:

[0157]

[0158] where represents the discrete set of guide vane angles, represents the discrete set of anti-surge valve openings, represents the discrete set of reflux valve openings. The controller parameter space is composed of all possible combinations. Build a fuzzy rule base based on the surge warning level to select a reasonable control strategy through the fuzzy control method. The surge warning level is set to three states, namely low risk (L1), medium risk (L2), and high risk (L3), and the control rules are defined using fuzzy logic. After inputting the controller parameter space into the fuzzy rule base for inference calculation, a fuzzy control decision sequence is obtained, that is, multiple possible combinations. Set an optimization objective function according to the fuzzy control decision sequence to ensure that the control strategy takes into account surge suppression, energy loss, and adjustment response time. Let the optimization objective function consist of three parts: the surge suppression effect , which measures the improvement degree of surge:

[0159]

[0160] where is the pressure change amplitude, is the adjustment time, and the larger it is, the better the surge suppression effect. The energy loss , which measures the energy loss of the anti-surge valve and the reflux valve:

[0161]

[0162] where is the weight coefficient, indicating that a larger opening will result in greater energy loss. The adjustment time , which measures the response speed of the adjustment strategy:

[0163]

[0164] where represents the sum of the response times of the guide vane, anti-surge valve, and reflux valve. The optimization objective function is:

[0165]

[0166] Among them, is the weight coefficient, which is used to balance the influence of surge suppression, energy loss, and regulation speed. To optimize the control parameters, a genetic algorithm population is constructed based on the objective function. The binary coding method is adopted to represent the control parameters. Let the chromosome be:

[0167]

[0168] Among them, is the binary coding of the guide vane angle, anti-surge valve opening, and return valve opening. Let the coding length be , then the mapping relationship of the controller parameters is:

[0169]

[0170]

[0171]

[0172] Perform genetic evolution calculation on the initial population, optimize the population through selection, crossover, and mutation operations, and finally obtain the optimal chromosome , and then perform decoding operation to obtain the optimal control parameters . Synchronously adjust the guide vane angle, anti-surge valve opening, and return valve opening based on the optimal control parameters, and monitor the pressure pulsation characteristics , flow statistical characteristics , and bearing vibration characteristics , and calculate the surge margin :

[0173]

[0174] Among them, is the characteristic weight, which is used to comprehensively evaluate the surge margin.

[0175] In a specific embodiment, the process of performing the steps to synchronously adjust the guide vane angle, anti-surge valve opening, and return valve opening based on the control parameters, monitor the changes in pressure pulsation characteristics, flow statistical characteristics, and characteristic frequencies, and calculate the surge margin value may specifically include the following steps:

[0176] Input the control parameters into the servo motor drive system and the electro-hydraulic servo system to synchronously adjust the guide vane angle, anti-surge valve opening, and return valve opening, and obtain the actuator action sequence;

[0177] Detect the response time of the actuator action sequence, record the response time of the guide vane angle adjustment, the response time of the anti-surge valve opening adjustment, and the response time of the reflux valve opening adjustment, and obtain the response time data;

[0178] Calculate the pressure ratio change rate, the flow rate fluctuation amplitude, and the bearing vibration intensity based on the response time data, evaluate the anti-surge suppression effect, and obtain the performance parameter matrix;

[0179] Conduct a stability analysis on the performance parameter matrix, calculate the overshoot, the settling time, and the steady-state error, and obtain the operating state index;

[0180] Calculate the anti-surge margin based on the operating state index, multiply the pressure ratio change rate by the first target value, the flow rate fluctuation amplitude by the second target value, and the bearing vibration intensity by the third target value for weighted summation, and obtain the anti-surge margin value.

[0181] Specifically, input the optimal control parameters into the servo motor drive system and the electro-hydraulic servo system to drive the synchronous adjustment of the guide vane angle, the anti-surge valve opening, and the reflux valve opening. Assume that the optimized control parameters are the guide vane angle , the anti-surge valve opening , and the reflux valve opening , then the control command is expressed as:

[0182]

[0183] Among them, As the control input, it is transmitted to the servo motor drive system and the electro-hydraulic servo system respectively to ensure the coordinated adjustment of the guide vane angle, the anti-surge valve, and the reflux valve. The servo motor drive system is used to adjust the guide vane angle, and its dynamic response model is expressed as:

[0184]

[0185] Among them, Is the gain of the guide vane angle system, Is the guide vane angle control input, Is the time constant of the servo motor. Similarly, the dynamic responses of the electro-hydraulic servo systems of the anti-surge valve and the reflux valve are respectively:

[0186]

[0187]

[0188] Among them, And Are the gains of the anti-surge valve and the reflux valve respectively, And Are the corresponding control inputs, And are their respective time constants. Through these dynamic models, the action responses of the actuators are obtained to form an actuator action sequence:

[0189] ;

[0190] After the actuator action is completed, the response time of the actuator action sequence is detected, and the guide vane angle adjustment response time is recorded , the anti-surge valve opening adjustment response time and the reflux valve opening adjustment response time . Let the response time of the valve and the guide vane be defined as the time required to reach 95% of the set value, that is:

[0191]

[0192]

[0193]

[0194] According to the response time data, the pressure ratio change rate , the flow fluctuation amplitude and the bearing vibration intensity are calculated to evaluate the anti-surge suppression effect. The pressure ratio change rate is defined as the change amount of the pressure ratio per unit time, that is:

[0195]

[0196] where and are the pressures at the outlet and the inlet respectively, is the maximum response time for controlling the actuator. Similarly, the flow fluctuation amplitude is calculated as:

[0197]

[0198] where is the flow signal, is the average value of the flow. And the bearing vibration intensity is calculated as the root mean square value of the vibration acceleration:

[0199]

[0200] where is the value of the bearing vibration signal at the th sampling point, is the total number of samplings. The stability of the anti-surge suppression effect is analyzed, and the overshoot , the settling time and the steady-state error The overshoot is defined as the maximum deviation degree of the system response, and the calculation formula is:

[0201] ;

[0202] where, is the actuator value in the final stable state. The settling time is defined as the time required for the system response to reach and remain within a 5% error range:

[0203] ;

[0204] The steady-state error is calculated as:

[0205] ;

[0206] where, is the final stable state of the system, is the set target value. The surge margin value is calculated based on the operating state indicators , and is calculated by weighted summing the pressure ratio change rate, the flow fluctuation amplitude, and the bearing vibration intensity:

[0207] ;

[0208] where, , and are the target weights used to adjust the influence of different parameters on the surge margin.

[0209] The above describes the surge suppression method for a centrifugal air compressor in an embodiment of the present invention. Next, the surge suppression device for a centrifugal air compressor in an embodiment of the present invention will be described. Please refer to Figure 2 , an embodiment of the surge suppression device for a centrifugal air compressor in an embodiment of the present invention includes:

[0210] A feature extraction module for extracting pressure pulsation features from the pressure signal of the centrifugal air compressor to obtain pressure pulsation features;

[0211] A feature combination module for extracting the flow statistical features and frequency band energy features of the flow signal in the centrifugal air compressor and combining them with the pressure pulsation features to obtain a surge warning feature space;

[0212] A feature decomposition module for decomposing the bearing vibration signal of the centrifugal air compressor, extracting the characteristic frequencies, and combining them with the surge warning feature space to obtain a surge warning feature set;

[0213] A warning prediction module for inputting the surge warning feature set into a three-channel neural network for surge warning prediction to obtain a surge warning level;

[0214] A calculation module is used to generate control parameters for a guide vane angle controller, an anti-surge valve opening controller, and a reflux valve opening controller according to the surge warning level, synchronously adjust the guide vane angle, the anti-surge valve opening, and the reflux valve opening, and calculate the surge margin value.

[0215] Through the collaborative cooperation of the above-mentioned various components, through the multi-scale analysis of wavelet transform and Fourier transform of the pressure signal, combined with the dimensionality reduction processing of the principal component analysis method, the time-frequency characteristics of the pressure pulsation are effectively extracted, redundant information is reduced, and the efficiency of feature extraction is improved. The method of combining Butterworth filtering and wavelet packet transform is used to process the flow signal. By extracting statistical features and band energy features, the dynamic characteristics of the flow fluctuation are accurately captured, and the reliability of the surge warning is enhanced. The empirical mode decomposition and Hilbert envelope spectrum analysis are introduced to process the bearing vibration signal. Combined with the envelope demodulation technology, the vibration characteristics of each component of the bearing are accurately extracted, and the information dimension of the surge warning is expanded. A three-channel parallel deep neural network structure is designed. Different types of features are processed through the convolutional layer, the fully connected layer, and the LSTM network respectively, and the attention mechanism is used for feature fusion to construct an accurate surge warning model. A multi-parameter collaborative optimization method based on fuzzy rules and genetic algorithms is proposed. Through dynamic weight update and objective function optimization, the accurate control of the guide vane angle, the anti-surge valve opening, and the reflux valve opening is realized. A complete surge suppression effect evaluation system is established. By monitoring the pressure ratio change rate, the flow fluctuation amplitude, and the bearing vibration intensity, the surge margin value is calculated to ensure the real-time optimization and adjustment of the control strategy.

[0216] Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the above-described system, system, and unit can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated herein.

[0217] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0218] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for suppressing surge of a centrifugal air compressor, characterized in that: The method comprises: Extract the pressure pulsation characteristics of the pressure signal of the centrifugal air compressor to obtain the pressure pulsation characteristics; Extracting flow statistics characteristics and frequency band energy characteristics of the flow signal in the centrifugal air compressor, and combining them with the pressure pulsation characteristics to obtain a surge warning feature space; Performing feature decomposition on the bearing vibration signal of the centrifugal air compressor, extracting the characteristic frequency, and combining it with the surge warning feature space to obtain a surge warning feature set; Inputting the surge warning feature set into a three-channel neural network to perform surge warning prediction to obtain a surge warning level; Generate control parameters of the guide vane angle controller, the anti-surge valve opening controller and the return valve opening controller according to the surge warning level, and synchronously adjust the guide vane angle, the anti-surge valve opening and the return valve opening to calculate the surge margin value; specifically include: discretizing the control range of the guide vane angle controller, the control range of the anti-surge valve opening controller and the control range of the return valve opening controller to obtain the controller parameter space; construct a fuzzy rule base based on the surge warning level, input the controller parameter space into the fuzzy rule base for reasoning operation, and obtain a fuzzy control decision sequence; set the optimization objective function according to the fuzzy control decision sequence, and set the surge suppression effect to the first weight, The energy loss is set as the second weight, and the adjustment time is set as the third weight to obtain a parameter optimization target; a genetic algorithm population is constructed according to the parameter optimization target, a chromosome code is generated, and the controller parameter space is mapped to a binary code to obtain an initial population; a genetic evolution calculation is performed on the initial population to obtain an optimal chromosome, and a decoding operation is performed on the optimal chromosome to convert the binary code into control parameters of the guide vane angle, the anti-surge valve opening, and the return valve opening; the guide vane angle, the anti-surge valve opening, and the return valve opening are synchronously adjusted based on the control parameters, the pressure pulsation characteristics, the flow statistical characteristics, and the changes in the characteristic frequency are monitored, and the surge margin value is calculated.

2. The method for suppressing surge of a centrifugal air compressor according to claim 1, characterized in that: The pressure pulsation feature extraction of the pressure signal of the centrifugal air compressor to obtain the pressure pulsation feature includes: The inlet pressure signal, outlet pressure signal and intermediate pressure signal of the centrifugal air compressor are collected to obtain an original pressure signal sequence; Input the original pressure signal sequence into the db4 wavelet basis function to perform 4-layer wavelet decomposition to obtain multi-scale features of the pressure signal; Performing discrete Fourier transform on the multi-scale features of the pressure signal to obtain a frequency domain feature sequence; Performing statistical analysis on the original pressure signal sequence, calculating the root mean square value, peak factor, waveform factor and pulse factor, and combining them with the frequency domain feature sequence to obtain a pressure pulsation feature vector; A principal component analysis is performed on the pressure pulsation feature vector to obtain the pressure pulsation feature.

3. The method for suppressing surge of a centrifugal air compressor according to claim 1, characterized in that: The method extracts the flow statistical characteristics and the frequency band energy characteristics of the flow signal in the centrifugal air compressor and combines them with the pressure pulsation characteristics to obtain the surge warning feature space, including: Sampling the flow signals collected by the inlet flow sensor and the outlet flow sensor to obtain an original flow signal, and performing signal noise reduction on the original flow signal to obtain a filtered flow signal; Performing statistical analysis on the filtered flow signal, calculating the mean, standard deviation, skewness and kurtosis to obtain flow statistical characteristics, and performing wavelet packet decomposition on the filtered flow signal to obtain a frequency band coefficient matrix; Calculating energy eigenvalues ​​for the frequency band coefficient matrix, extracting energy distribution of each frequency band, and obtaining frequency band energy characteristics; The flow statistical feature, the frequency band energy feature and the pressure pulsation feature are feature spliced ​​to obtain a surge warning feature space.

4. The method for suppressing surge of a centrifugal air compressor according to claim 1, characterized in that: The feature decomposition of the bearing vibration signal of the centrifugal air compressor is performed to extract the feature frequency, and the feature frequency is combined with the surge warning feature space to obtain a surge warning feature set, including: Sampling the bearing vibration signal collected by the acceleration sensor at the inlet and outlet bearings of the centrifugal air compressor to obtain an original vibration signal; Performing empirical mode decomposition on the original vibration signal to obtain a vibration signal component matrix, and performing Hilbert transform on the vibration signal component matrix to obtain a frequency spectrum feature sequence; Extracting characteristic frequencies from the frequency spectrum characteristic sequence, calculating inner ring passing frequency, outer ring passing frequency, rolling element passing frequency and cage rotation frequency, and obtaining bearing characteristic frequencies; The bearing characteristic frequency is subjected to envelope demodulation analysis to extract vibration characteristics of various bearing components to obtain a bearing vibration characteristic vector, and the bearing vibration characteristic vector is combined with the surge warning feature space to obtain a surge warning feature set.

5. The method for suppressing surge of a centrifugal air compressor according to claim 1, characterized in that: The step of inputting the surge warning feature set into a three-channel neural network to perform surge warning prediction and obtain a surge warning level includes: Inputting the pressure pulsation feature of the surge warning feature set into the first channel of the three-channel neural network, performing feature extraction through four convolutional layers in the first channel to obtain a pressure feature map; Inputting the flow statistics features and the frequency band energy features in the surge warning feature set into the second channel of the three-channel neural network, and sequentially performing feature transformation through the three fully connected layers in the second channel to obtain a flow feature representation; Inputting the bearing vibration feature vector in the surge warning feature set into the third channel of the three-channel neural network, performing time series feature extraction through the two-layer LSTM network in the third channel, and obtaining a frequency feature sequence; Calculating attention weights for the pressure feature map, the flow feature representation, and the frequency feature sequence respectively to generate a feature weight matrix; Performing weighted fusion on the pressure feature map, the flow feature representation and the frequency feature sequence according to the feature weight matrix to obtain a fused feature vector; The fused feature vector is input into the output layer of the three-channel neural network, the probability distribution of three levels is calculated by the softmax classifier, and the level corresponding to the maximum probability value is taken as the surge warning level.

6. The method for suppressing surge of a centrifugal air compressor according to claim 1, characterized in that: The synchronously adjusting the guide vane angle, the anti-surge valve opening and the return valve opening based on the control parameters, monitoring the changes of the pressure pulsation characteristics, the flow statistical characteristics and the characteristic frequency, and calculating the surge margin value includes: The control parameters are input into the servo motor drive system and the electric hydraulic servo system to synchronously adjust the guide vane angle, the anti-surge valve opening and the return valve opening to obtain the actuator action sequence; Perform response time detection on the actuator action sequence, record the guide vane angle adjustment response time, the anti-surge valve opening adjustment response time and the return valve opening adjustment response time, and obtain response time data; Calculate the pressure ratio change rate, flow fluctuation amplitude and bearing vibration intensity according to the response time data, evaluate the surge suppression effect, and obtain a performance parameter matrix; Perform stability analysis on the performance parameter matrix, calculate overshoot, stabilization time and steady-state error, and obtain operating status indicators; The surge margin is calculated based on the operating state index, and the pressure ratio change rate is multiplied by the first target value, the flow fluctuation amplitude is multiplied by the second target value, and the bearing vibration intensity is multiplied by the third target value for weighted summation to obtain a surge margin value.

7. A centrifugal air compressor surge suppression device, characterized in that: Used to perform the centrifugal air compressor surge suppression method according to any one of claims 1 to 6, the centrifugal air compressor surge suppression device comprises: A feature extraction module is used to extract pressure pulsation features from the pressure signal of the centrifugal air compressor to obtain pressure pulsation features; A feature combination module, used to extract the flow statistical features and frequency band energy features of the flow signal in the centrifugal air compressor, and combine them with the pressure pulsation features to obtain a surge warning feature space; A feature decomposition module, used for performing feature decomposition on the bearing vibration signal of the centrifugal air compressor, extracting the feature frequency, and combining it with the surge warning feature space to obtain a surge warning feature set; An early warning prediction module is used to input the surge early warning feature set into a three-channel neural network to perform surge early warning prediction and obtain a surge early warning level; The calculation module is used to generate control parameters of the guide vane angle controller, the anti-surge valve opening controller and the return valve opening controller according to the surge warning level, and synchronously adjust the guide vane angle, the anti-surge valve opening and the return valve opening to calculate the surge margin value.

Images