Natural gas centrifugal compressor vibration fault diagnosis method
Through deep neural network and MFCC feature extraction technology, the problem of insufficient automated diagnosis capabilities in vibration fault diagnosis of natural gas centrifugal compressors is solved, and high-precision, real-time fault identification and early warning are achieved, simplifying the diagnosis process.
Patent Information
- Application Number
- CN202310642715.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-06-01
- Publication Date
- 2024-12-03
AI Technical Summary
The prior art has insufficient automated diagnostic capabilities in the diagnosis of vibration faults of natural gas centrifugal compressors, relying on manual experience and shallow machine learning algorithms, resulting in low diagnostic efficiency, low accuracy, and inability to warning in advance.
Deep neural network is used to combine MFCC feature extraction and convolutional neural network to train and deploy deep neural network models for automated diagnosis of vibration signals, real-time identification of fault types and probability.
It realizes automated fault diagnosis, improves detection accuracy and efficiency, can be early warning, reduces dependence on professional knowledge, and does not need to change the equipment structure and is easy to use.
Smart Images

Figure CN119062598A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of centrifugal compressors, and in particular to a vibration fault diagnosis method for a natural gas centrifugal compressor. Background Art
[0002] Centrifugal compressors are the core equipment of long-distance natural gas pipelines. Their safe operation can provide a stable gas source for downstream users. However, due to their complex internal structure, they are more prone to failures (rotor misalignment, rotor imbalance, surge, oil film oscillation, etc.) in actual industrial production and operation. Only by quickly discovering the failure and diagnosing the "cause" can the downstream gas supply be quickly restored to protect the national economy from losses.
[0003] Mechanical equipment will generate vibration during operation. Different operating states of the equipment are often accompanied by different forms of vibration:
[0004] 1) The periodic vibration signals of the equipment are different for normal and abnormal working conditions;
[0005] 2) For different fault types, the vibration form and signal of the equipment are different;
[0006] 3) For different degrees of the same fault, the strength of the equipment vibration signal varies greatly.
[0007] Among the various common faults of centrifugal compressors, vibration faults are the most common. According to the different analysis and calculation subjects, the existing fault analysis methods can be divided into two categories: manual analysis and machine analysis:
[0008] 1) Manual analysis method
[0009] Experts in the field can usually infer the type and cause of faults by analyzing changes in the characteristics of key vibration signals of the compressor (including vibration amplitude, phase, spectrum, trajectory, etc.). The most widely used analysis method is vibration signal characteristic analysis.
[0010] The following uses rotor imbalance fault as an example to introduce the vibration signal characteristic analysis method. Imbalance is the most common system fault of the rotor, and this fault has relatively obvious vibration characteristics: the unbalanced vibration spectrum mainly shows that the main frequency is the power frequency, the time domain waveform of the vibration is a sine wave, the rotor axis trajectory is an ellipse, the precession characteristic is synchronous positive precession, and the rotor amplitude decreases with the increase or decrease of the rotor speed. The obviousness of the spectral characteristics is an important basis for judging the rotor imbalance fault.
[0011] 2) Machine analysis
[0012] With the development of science and technology, machine learning algorithms such as expert systems, BP neural networks, Bayesian networks, and support vector machines continue to emerge. Compared with the traditional method of manually analyzing vibration characteristics, the fault diagnosis method based on machine learning is conducive to the fitting of nonlinear mapping relationships and can greatly improve the accuracy of fault identification. Based on this, Sinopec Chongqing Natural Gas Pipeline Co., Ltd. proposed an invention patent for "An Intelligent Monitoring System for Centrifugal Compressor Units in Natural Gas Long-distance Pipelines". The system consists of a data acquisition sensor, a communication unit, a data processing unit, and an intelligent fault identification unit. The fault identification unit in the scheme adopts a three-layer Bayesian network in the machine learning algorithm. The network can detect the fault type and fault probability of the compressor unit. Not only that, the scheme can also update the database of fault types through machine learning, so that it can be expanded and updated according to the state of the machine, improving the dynamic flexibility of detection.
[0013] 3) Insufficiency of existing solutions
[0014] a) The manual analysis method based on vibration characteristics requires engineers with rich experience in the compressor field. It is a post-analysis method and cannot warn of accidents in advance. It takes a day or even longer. The analysis results are subjective to a certain extent and the cost is high. It is highly dependent on the professional knowledge, field experience and equipment familiarity of the engineer.
[0015] b) Most intelligent fault analysis methods based on machine learning are shallow algorithms with weak data analysis capabilities, and their application effects depend on the ability to express features. Since there are many nonlinear factors affecting natural gas compressors, if the features cannot express the differences between signals well, the fault diagnosis and learning effects will be affected.
[0016] 4) Other methods
[0017] Compressor fault diagnosis methods also include vibration and noise detection and thermal diagnosis. Specifically:
[0018] a) The main principle of the noise monitoring method is to effectively and accurately determine whether there is wear inside the cylinder, leakage inside the valve, and other vibration conditions by combining the internal vibration detection conditions of various mechanical parts. In actual diagnosis, it is also necessary to directly install the cylinder vibration detection sensor on the entire cylinder main engine head, and use it in combination with the cylinder vibration detection signal to accurately determine whether there is any vibration fault in the entire cylinder internal structure. In addition, it is also necessary to accurately determine whether the compressor drive bearing has a related fault by combining the air pressure in the air compressor pipeline and the transmission signal characteristics of the sound wave.
[0019] b) Thermal fault diagnosis method detection is a method in which the system uses an instrument to monitor and analyze the various thermal data activities of a reciprocating compressor in real time, so as to monitor various fault causes in real time. The comprehensive monitoring and analysis of various drainage data mainly includes power supply water temperature, drainage volume, cooler drainage volume, etc. In the process of diagnosing, collecting and analyzing data, different thermal components will show large differences in the analysis data after a thermal fault occurs. Therefore, the use of traditional thermal fault diagnosis and analysis methods in the analysis and monitoring of thermal faults often lacks data accuracy. Its main research and application areas are in the fault monitoring technology of thermal operation management parameters of reciprocating compressors. Summary of the invention
[0020] The embodiments of the present invention aim to solve at least one of the technical problems existing in the prior art and provide a method for diagnosing vibration faults of a natural gas centrifugal compressor.
[0021] One aspect of an embodiment of the present invention provides a method for diagnosing vibration faults of a natural gas centrifugal compressor, which includes a training phase and an inference phase in sequence, wherein common fault characteristic signals of a natural gas centrifugal compressor include at least one of the following: slight rotor misalignment, medium rotor misalignment, and strong rotor misalignment; slight rotor imbalance, medium rotor imbalance, and strong rotor imbalance; slight oil film oscillation, medium oil film oscillation, and strong oil film oscillation; slight surge, medium surge, strong surge, and normal operating conditions;
[0022] 1) The specific implementation steps of the training phase are as follows:
[0023] Step a1: Collect at least 10,000 characteristic signals of each fault. This data is used as a data set for deep neural network training. When collecting data, label each data with one of the above faults.
[0024] Step a2: Perform MFCC transformation on each fault feature signal to extract features;
[0025] Step a3: Send the fault data and corresponding labels to the deep neural network for at least 200 rounds of training to obtain a trained deep neural network;
[0026] Step a4: deploying the trained deep neural network into an actual system;
[0027] 2) The specific steps of the application phase are:
[0028] Step b1: collecting vibration signals of natural gas compressors;
[0029] Step b2: Perform MFCC feature extraction transformation on the vibration signal;
[0030] Step b3: Send the characteristic signal obtained in step b2 to the trained deep neural network for calculation to obtain the calculated vibration fault type and its probability;
[0031] Step b4: Display the fault type and probability on the system display.
[0032] Preferably, the transformation and extraction of the fault feature signal uses the MFCC frontier analysis method, which is the most widely used in the field of speech recognition and analysis, to extract features from the collected vibration signal. MFCC is based on human auditory characteristics, firstly maps the linear spectrum to the Mel frequency nonlinear spectrum, and finally performs a cepstrum transformation.
[0033] Preferably, the Mel frequency cepstral coefficient feature extraction process is as follows: pre-emphasis, framing, windowing, fast Fourier transform, calculation of Mel filter energy, discrete cosine transform and dynamic difference parameter extraction.
[0034] Preferably, pre-emphasis uses a high-pass filter to enhance the high-frequency part of the signal, making the spectrum flatter and ensuring that the spectrum can be calculated with the same signal-to-noise ratio in the entire frequency band from low frequency to high frequency. This is the main purpose of pre-emphasis. Generally speaking, the transfer function of the high-pass filter is:
[0035] y(n)=x(n)-αx(n-1)
[0036] Where x(n) is the original signal, y(n) is the emphasized signal sequence, and α is the pre-emphasis coefficient.
[0037] Preferably, framing is to divide a long sound signal into multiple small segments. A frame covers a time of 20 to 30 ms, wherein there is an overlapping area between two adjacent frames, in order to avoid excessive changes between two adjacent frames. The present invention takes 384 points as one frame, and the length of the selected overlapping area is half a frame, that is, 192 points.
[0038] Preferably, windowing is performed to avoid signal discontinuity at both ends of each frame of signal. Commonly used window functions include square window, Hamming window and Hanning window. In contrast, the Hamming window is suitable for processing signals with complex spectral performance and multiple spectral components. Therefore, this paper selects the Hamming window to process each frame of signal.
[0039] Preferably, the fast Fourier transform is generally difficult to analyze the characteristics of the signal by observing the transformation of the signal in the time domain, so the fast Fourier transform (FFT) is usually used to perform signal time-frequency conversion to observe its energy distribution. As shown in the following formula:
[0040] X(i,k)=FFT[x i (n)]
[0041] Where X(i,k) is the desired frequency domain signal obtained by time-frequency conversion through fast Fourier transform (FFT), x i is the vibration signal of the centrifugal compressor at time i, and n is the spectrum after DCT.
[0042] Preferably, the Mel filter energy is based on the fact that the human ear has different sensitivities to signals of different frequencies, and the relationship is nonlinear, so the spectrum is divided into multiple Mel filter groups according to the different sensitivities of the human ear, and then the spectral line energy of each frame of FFT data is calculated, as shown in the following formula:
[0043] E(i,k)=[X(i,k)] 2
[0044] Where E(i,k) is the energy spectrum of the sound signal, where i represents the i-th frame and k represents the k-th spectral line. Then the spectral energy of each frame is passed through the Mel filter, that is, the energy spectrum E(i,k) of each frame is multiplied by the frequency domain of the Mel filter and then added, that is,
[0045]
[0046] Where H m (k) represents the mth filter in the Mel filter bank, M represents the number of filter banks, which is 26 here, and S(i,m) represents the energy of the i-th frame after passing through the m-th Mel filter.
[0047] Preferably, discrete cosine transform (DCT) is mainly used for compression, such as compressing data or compressing images, and it can transform spatial domain signals into frequency domain and has good decorrelation.
[0048]
[0049] In the above formula, the energy S(i,m) of the Mel filter is obtained from the previous step, where n is the spectrum line after DCT, and L is the order of the Mel frequency cepstrum coefficient, which can be 18 here.
[0050] The MFCC coefficient reflects the static characteristics of the sound signal. The dynamic characteristics of the sound signal can be further obtained by using the dynamic differential spectrum. The static plus dynamic feature analysis can more effectively improve the overall recognition performance of the system. Here, the first-order difference of MFCC is used to represent the dynamic characteristics of the sound information.
[0051] Preferably, this solution uses a convolutional neural network to train the extracted MFCC fault feature signal. The convolutional neural network used has a total of multiple resblocks, and finally passes through a fully connected layer and a softmax layer.
[0052] Compared with the prior art, the present invention has the following significant advantages:
[0053] a) Compared with the traditional fault analysis method based on vibration characteristics, this scheme can realize automatic fault diagnosis, and the diagnosis result is the fault type and the corresponding probability.
[0054] b) Compared with the fault diagnosis method based on machine learning, the trained deep neural network has already fitted a large amount of fault data, has a strong ability to express model features, stronger data analysis capabilities, strong nonlinear fitting capabilities and feature expression capabilities, and can extract discriminant information that is conducive to accurate diagnosis from complex and changeable fault data under the coupling of multiple influencing factors, with higher detection accuracy.
[0055] c) Fault diagnosis can be performed by only detecting the vibration signal of the equipment on the surface of the equipment. There is no need to change the original structure of the equipment or disassemble the machine, which is very convenient to use.
[0056] This solution uses cutting-edge big data analysis technology, neural network technology, and deep learning technology to achieve the following functions:
[0057] 1) Accurately identify and judge the type and probability of centrifugal compressor failure, and visualize the results;
[0058] 2) Real-time fault perception and early warning to ensure intrinsic safety;
[0059] 3) The optimized neural network algorithm is integrated into the chip to produce a portable dynamic equipment fault diagnosis instrument;
[0060] 4) Release high-quality human resources;
[0061] 5) The solution has a wide range of application scenarios and can be extended to vibration fault diagnosis of various mechanical equipment in the natural gas field. BRIEF DESCRIPTION OF THE DRAWINGS
[0062] Figure 1 Provides an overall framework for vibration fault diagnosis of natural gas centrifugal compressors;
[0063] Figure 2 It is the flow chart of MFCC fault feature extraction;
[0064] Figure 3 Network structure diagram for fault classification. DETAILED DESCRIPTION
[0065] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.
[0066] A natural gas centrifugal compressor vibration fault diagnosis method is provided, which includes a training stage and an inference stage in sequence. Common vibration signals of natural gas centrifugal compressors include at least one of the following: slight rotor misalignment, moderate rotor misalignment, and strong rotor misalignment; slight rotor imbalance, moderate rotor imbalance, and strong rotor imbalance; slight oil film oscillation, moderate oil film oscillation, and strong oil film oscillation; slight surge, moderate surge, and strong surge, as well as normal operating conditions.
[0067] 1) The specific implementation steps of the training phase are as follows:
[0068] Step a1: 10,000 characteristic signals of each fault are collected. This data is used as a data set for deep neural network training. During data collection, each data is labeled with one of the above faults.
[0069] Step a2: Perform MFCC transformation on each fault feature signal to extract features;
[0070] Step a3: Send the fault data and corresponding labels to the deep neural network for 200 rounds of training to obtain a trained deep neural network;
[0071] Step a4: deploying the trained neural network into an actual system;
[0072] 2) The specific steps of the application phase are:
[0073] Step b1: collecting vibration signals of natural gas compressors;
[0074] Step b2: Perform MFCC feature extraction transformation on the vibration signal;
[0075] Step b3: Send the feature signal obtained in step b2 to the trained deep neural network for calculation;
[0076] Step b4: Obtain the calculated vibration fault type and its probability;
[0077] Step b5: Display the fault type and probability on the system display.
[0078] Preferably, the transformation and extraction of the fault feature signal uses the MFCC frontier analysis method, which is the most widely used in the field of speech recognition and analysis, to extract features from the collected vibration signal. MFCC is based on human auditory characteristics, firstly maps the linear spectrum to the Mel frequency nonlinear spectrum, and finally performs a cepstrum transformation.
[0079] Preferably, the Mel frequency cepstral coefficient feature extraction process is as follows: pre-emphasis, framing, windowing, fast Fourier transform, calculation of Mel filter energy, discrete cosine transform and dynamic difference parameter extraction.
[0080] Preferably, pre-emphasis uses a high-pass filter to enhance the high-frequency part of the signal, making the spectrum flatter and ensuring that the spectrum can be calculated with the same signal-to-noise ratio in the entire frequency band from low frequency to high frequency. This is the main purpose of pre-emphasis. Generally speaking, the transfer function of the high-pass filter is:
[0081] y(n)=x(n)-αx(n-1)
[0082] Where x(n) is the original signal, y(n) is the emphasized signal sequence, and α is the pre-emphasis coefficient.
[0083] Preferably, framing is to divide a long sound signal into multiple small segments. A frame covers a time of 20 to 30 ms, wherein there is an overlapping area between two adjacent frames, in order to avoid excessive changes between two adjacent frames. The present invention takes 384 points as one frame, and the length of the selected overlapping area is half a frame, that is, 192 points.
[0084] Preferably, windowing is performed to avoid signal discontinuity at both ends of each frame of signal. Commonly used window functions include square window, Hamming window and Hanning window. In contrast, the Hamming window is suitable for processing signals with complex spectral performance and multiple spectral components. Therefore, this paper selects the Hamming window to process each frame of signal.
[0085] Preferably, the fast Fourier transform is generally difficult to analyze the characteristics of the signal by observing the transformation of the signal in the time domain, so the fast Fourier transform (FFT) is usually used to perform signal time-frequency conversion to observe its energy distribution. As shown in the following formula:
[0086] X(i,k)=FFT[x i (n)]
[0087] Preferably, the Mel filter energy is based on the fact that the human ear has different sensitivities to signals of different frequencies, and the relationship is nonlinear, so the spectrum is divided into multiple Mel filter groups according to the different sensitivities of the human ear, and then the spectral line energy of each frame of FFT data is calculated, as shown in the following formula:
[0088] E(i,k)=[X(i,k)] 2
[0089] Where E(i,k) is the energy spectrum of the sound signal, where i represents the i-th frame and k represents the k-th spectral line. Then the spectral energy of each frame is passed through the Mel filter, that is, the energy spectrum E(i,k) of each frame is multiplied by the frequency domain of the Mel filter and then added, that is,
[0090]
[0091] Where Hm (k) represents the mth filter in the Mel filter bank, M represents the number of filter banks, which is 26 here, and S(i,m) represents the energy of the i-th frame after passing through the m-th Mel filter.
[0092] Preferably, discrete cosine transform (DCT) is mainly used for compression, such as compressing data or compressing images, and it can transform spatial domain signals into frequency domain and has good decorrelation.
[0093]
[0094] In the above formula, the energy S(i,m) of the Mel filter is obtained from the previous step, where n is the spectrum line after DCT, and L is the order of the Mel frequency cepstrum coefficient, which can be 18 here.
[0095] The MFCC coefficient reflects the static characteristics of the sound signal. The dynamic characteristics of the sound signal can be further obtained by using the dynamic differential spectrum. The static plus dynamic feature analysis can more effectively improve the overall recognition performance of the system. Here, the first-order difference of MFCC is used to represent the dynamic characteristics of the sound information.
[0096] Preferably, this solution uses a convolutional neural network to train the extracted MFCC fault feature signal. The convolutional neural network used has a total of 16 resblocks, and finally passes through a fully connected layer and a softmax layer.
[0097] The specific implementation example of the application stage is as follows:
[0098] Embodiment 1:
[0099] S1: The vibration sensor obtains the vibration signal of the natural gas centrifugal compressor;
[0100] S2: Fault identification using a neural network trained with weak fault signals;
[0101] The neural network for training weak fault signals here can use a deep neural network with a large number of layers, such as VGG128. Correspondingly, since the processing speed of the neural network will slow down as the number of network layers increases, the frequency of fault detection can be reduced;
[0102] S3: Identify the type and probability of weak failure of the equipment.
[0103] Embodiment 2:
[0104] S1: The vibration sensor obtains the vibration signal of the natural gas centrifugal compressor;
[0105] S2: using the first classification algorithm to identify the level of equipment failure, where the levels include minor failure / moderate failure / serious failure;
[0106] S3: If it is a minor fault, then use the LSTM network to identify the specific fault type; if it is the other two levels of fault, then use machine learning algorithms or deep learning algorithms to identify the fault type; for minor fault signals, use the LSTM network to learn the timing characteristics of the signal, which can more effectively identify minor faults than machine learning or deep learning algorithms.
[0107] Embodiment 3:
[0108] S1: The vibration sensor obtains the vibration signal of the natural gas centrifugal compressor;
[0109] S2: using the first classification algorithm to identify the level of equipment failure, where the levels include no failure / minor failure / moderate failure / serious failure;
[0110] S3: If no fault is identified, the signal in S1 is enhanced, and the enhanced signal is sent to the second classification algorithm for fault classification to obtain the level of the equipment fault, which is the same as the fault level in S2;
[0111] The second classification algorithm is a classification algorithm obtained by training the signal after the original vibration signal is enhanced, and a general voice signal enhancement algorithm can be used;
[0112] S4: If it is a minor fault, then use the LSTM network to identify the specific fault type; if it is a fault of the other two levels, then use a machine learning algorithm or a deep learning algorithm to identify the fault type.
[0113] For minor fault signals, the LSTM network is used to learn the timing characteristics of the signals, which can more effectively identify minor faults than machine learning or deep learning algorithms.
[0114] Embodiment 4:
[0115] S1: The vibration sensor obtains the vibration signal of the natural gas centrifugal compressor;
[0116] S2: enhancing the vibration signal in S1, and then using the enhanced signal and the first classification algorithm to identify the level of equipment failure, where the levels include no failure / minor failure / moderate failure / serious failure;
[0117] S3: If no fault is identified, the signal in S1 and the signal after audio enhancement in S2 are sent to the second classification algorithm for fault classification to obtain the level of equipment fault, which is the same as the fault level in S2;
[0118] The second classification algorithm is a classification algorithm obtained by training using the original vibration signal and the enhanced vibration signal, and a general speech signal enhancement algorithm may be used;
[0119] Optionally, the audio enhancement method used here may be one or a combination of volume enhancement, pitch enhancement, denoising, style transfer speech enhancement, etc.
[0120] Beneficial effects of this step: The audio enhancement methods that can be used here include denoising, style transfer, volume enhancement, pitch enhancement, etc. Data enhancement of the enhanced speech signal can expand the training sample capacity, improve the diversity of samples, and thus improve the generalization of the model. However, for minor faults, the speech enhancement algorithm may cause the original signal to lose useful information. For example, the details of the original signal are also erased while denoising, which will cause the algorithm to perform poorly when identifying minor signals. In order to solve this problem, the data before and after the audio signal enhancement can be sent to the classification algorithm for identification, thereby improving the detection accuracy.
[0121] S4: If it is a minor fault, then use the LSTM network to identify the specific fault type; if it is a fault of the other two levels, then use a machine learning algorithm or a deep learning algorithm to identify the fault type.
[0122] For minor fault signals, the LSTM network is used to learn the timing characteristics of the signals, which can more effectively identify minor faults than machine learning or deep learning algorithms.
[0123] Optionally, for minor faults, when using the LSTM network to identify specific fault types, the signal after audio signal enhancement can be used for identification, so as to increase the diversity of training samples and thus improve the generalization of the model. If the fault type is identified, then this process ends; if the specific fault type is not identified, then the original signal in S1 and the signal after audio signal enhancement are used to identify the fault type, so as to improve the accuracy of fault type detection.
[0124] Optionally, we are not limited to using the LSTM network to identify specific fault types. We can also use other network results to identify fault types. For example, we can use Informer to identify fault types. Informer is an algorithm proposed in the past two years for processing time series signals.
[0125] It is to be understood that the above embodiments are merely exemplary embodiments used to illustrate the principles of the present invention, but the present invention is not limited thereto. For those of ordinary skill in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.
Claims
1. A method for diagnosing vibration faults of a natural gas centrifugal compressor, characterized in that: The method includes a training phase and an inference phase, wherein common fault characteristic signals of natural gas centrifugal compressors include at least one of the following: slight rotor misalignment, medium rotor misalignment, and strong rotor misalignment; Slight rotor imbalance, moderate rotor imbalance, and strong rotor imbalance; slight oil film oscillation, moderate oil film oscillation, and strong oil film oscillation; slight surge, moderate surge, strong surge, and normal operating conditions; 1) The specific implementation steps of the training phase are as follows: Step a1: Collect at least 10,000 characteristic signals of each fault. This data is used as a data set for deep neural network training. When collecting data, label each data with one of the above faults. Step a2: Perform MFCC transformation on each fault feature signal to extract features; Step a3: Send the fault data and corresponding labels to the deep neural network for at least 200 rounds of training to obtain a trained deep neural network; Step a4: deploying the trained deep neural network into an actual system; 2) The specific steps of the application phase are: Step b1: collecting vibration signals of natural gas compressors; Step b2: Perform MFCC feature extraction transformation on the vibration signal to obtain a feature signal; Step b3: sending the characteristic signal obtained in step b2 to the trained deep neural network for calculation to obtain the calculated vibration fault type and its probability; Step b4: Display the fault type and probability on the system display.
2. A natural gas centrifugal compressor vibration fault diagnosis method according to claim 1, characterized in that: In step a2, MFCC is based on human auditory characteristics. First, the linear spectrum is mapped to the Mel-frequency nonlinear spectrum, and then the cepstrum transform is performed. The Mel-frequency cepstrum coefficient feature extraction process is as follows: pre-emphasis, framing, windowing, fast Fourier transform, calculation of Mel filter energy, discrete cosine transform and dynamic difference parameter extraction.
3. A natural gas centrifugal compressor vibration fault diagnosis method according to claim 2, characterized in that: Pre-emphasis uses a high-pass filter to enhance the high-frequency part of the signal. The transfer function of the high-pass filter is: y(n)=x(n)-αx(n-1) Where x(n) is the original signal, y(n) is the emphasized signal sequence, and α is the pre-emphasis coefficient.
4. A natural gas centrifugal compressor vibration fault diagnosis method according to claim 2, characterized in that: Framing is to divide a long sound signal into multiple small segments. One frame covers a period of 20 to 30 ms. There is an overlapping area between two adjacent frames, and the selected overlapping area length is half a frame.
5. A natural gas centrifugal compressor vibration fault diagnosis method according to claim 2, characterized in that: The purpose of windowing is to avoid signal discontinuity at both ends of each frame signal. The window function selects the Hamming window to process each frame signal.
6. A natural gas centrifugal compressor vibration fault diagnosis method according to claim 3, characterized in that: Use Fast Fourier Transform (FFT) to convert the signal into time-frequency and observe its energy distribution, as shown in the following formula: X(i,k)=FFT[x i (n)] Where X(i,k) is the desired frequency domain signal obtained by time-frequency conversion through fast Fourier transform (FFT), x i is the vibration signal of the centrifugal compressor at time i, and n is the spectrum after DCT.
7. A natural gas centrifugal compressor vibration fault diagnosis method according to claim 6, characterized in that: Mel filter energy is based on the fact that the human ear has different sensitivities to signals of different frequencies, and the relationship is nonlinear. Therefore, the spectrum is divided into multiple Mel filter groups according to the different sensitivities of the human ear, and then the spectral line energy of each frame of FFT data is calculated, as shown in the following formula: E(i,k)=[X(i,k)] 2 Where E(i,k) is the energy spectrum of the sound signal, where i represents the i-th frame and k represents the k-th spectral line; then the spectral energy of each frame is passed through the Mel filter, that is, the energy spectrum E(i,k) of each frame is multiplied by the frequency domain of the Mel filter and then added, that is, Where H m (k) represents the mth filter in the Mel filter bank, M represents the number of filter banks, and S(i, m) represents the energy of the i-th frame after passing through the m-th Mel filter.
8. A natural gas centrifugal compressor vibration fault diagnosis method according to claim 7, characterized in that: Discrete cosine transform is mainly used for compression, such as compressing data or compressing images. It can convert spatial domain signals into frequency domain and has good decorrelation. In the above formula, the energy S(i,m) of the Mel filter is obtained from the previous step, where n is the spectrum line after DCT, and L is the order of the Mel frequency cepstrum coefficient.
9. A natural gas centrifugal compressor vibration fault diagnosis method according to claim 1, characterized in that: A convolutional neural network is used to train the extracted MFCC fault feature signal. The convolutional neural network used has multiple resblocks and finally passes through a fully connected layer and a softmax layer.
10. A natural gas centrifugal compressor vibration fault diagnosis method according to claim 1, characterized in that: If no fault is identified in step b3, the signal in step b1 and the signal after audio enhancement are sent to a second classification algorithm for fault classification to obtain the level of equipment fault. The second classification algorithm is a classification algorithm trained using the signal after the original vibration signal is enhanced, and a general voice signal enhancement algorithm can be used.
11. A natural gas centrifugal compressor vibration fault diagnosis method according to claim 1, characterized in that: The audio enhancement method adopted in step 10 is one or a combination of volume enhancement, pitch enhancement, denoising, and style transfer speech enhancement; the second classification algorithm is a classification algorithm trained using the original vibration signal and the enhanced vibration signal, and adopts a general speech signal enhancement algorithm.
12. A natural gas centrifugal compressor vibration fault diagnosis method according to claim 1, characterized in that: If a minor fault is identified in step b3, the LSTM network or Informer is used to identify the specific fault type, and the signal after the audio signal enhancement is used for identification. If the fault type is identified, then this process ends; If the specific fault type is not identified, the original signal in S1 and the signal after the audio signal enhancement are used to identify the fault type.
13. A natural gas centrifugal compressor vibration fault diagnosis method according to claim 1, characterized in that: If a moderate fault or a severe fault is identified in step b3, a machine learning algorithm or a deep learning algorithm is used to identify the fault type.
