A Fault Diagnosis Method for Fan Equipment in Thermal Power Plants Based on Multi-sensor Feature Fusion

Abstract

The present invention discloses a method for diagnosing faults of fan equipment in a thermal power plant based on multi-sensor feature fusion, the method comprising the following steps: S1: collecting sound signals, vibration signals and temperature signals generated when the equipment is running; S2: pre-processing the sound signals collected in step S1; S3: extracting features from the pre-processed sound signal segments, vibration signals and temperature signals, and using the signal enhanced Mel cepstral coefficient method for the sound signal; S4: fusing the features of the feature vectors obtained by multi-sensors in step S3, bringing the fused final vector into the constructed SVM model, and obtaining the fault conditions of the mechanical equipment. The present invention selects a multi-sensor feature fusion method to fuse the features of the sound signal, vibration signal and temperature signal, thereby enhancing the system's ability to capture dynamic signals, more accurately obtaining the state of the equipment, and improving the detection rate of the abnormal information detection system.

Description

Technical Field

[0001] The invention relates to the field of fault diagnosis methods for fan equipment in thermal power plants, and in particular to a fault diagnosis method for fan equipment in thermal power plants based on multi-sensor feature fusion. Background Art

[0002] Thermal power plants contain many complex mechanical equipment. As one of the important auxiliary equipment in thermal power plants, the operation of fan equipment is directly related to whether the thermal power plant can generate electricity stably in the long term. Once a certain equipment fails, the efficiency of the entire power plant may be reduced or even a safety accident may occur. Therefore, it is very important to ensure the safe and stable operation of the fan in the thermal power plant. Traditional equipment detection methods are inefficient and sudden failures may not be remedied in time. In addition, most of them lack accurate technical data for reference and reference, and the reliability of fault detection is not strong. Therefore, by using artificial intelligence to detect equipment failures, the detection efficiency is greatly improved, the safety of personnel and equipment is greatly enhanced, and the operation and maintenance costs and time costs are reduced.

[0003] When the fan equipment in a thermal power plant is running, it will vibrate and then make a sound. When a fault occurs, the motion state between the fuselage and the parts or the parts themselves changes, the vibration changes, and the sound also changes accordingly. At present, the evaluation of fan equipment fault diagnosis mostly relies on a single vibration signal, which can easily cause problems such as unstable diagnostic system and unreliable diagnostic results. Therefore, the present invention selects a multi-sensor data fusion method to perform feature fusion of sound signals, vibration signals, temperature signals, etc., so as to more accurately obtain the status of the equipment and improve the reliability of fault diagnosis. Summary of the invention

[0004] The purpose of the present invention is to provide a method for fault diagnosis of fan equipment in a thermal power plant based on multi-sensor feature fusion, which performs feature fusion on signals collected by multiple sensors to form a fault diagnosis system that combines sound signals, vibration signals and temperature signals. The system has a more comprehensive acquisition of equipment status, greatly improves the efficiency and reliability of fault diagnosis, and ensures the safety of personnel and equipment.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is:

[0006] A method for diagnosing fan equipment faults in a thermal power plant based on multi-sensor feature fusion, the method comprising the following steps:

[0007] S1: collects sound signals, vibration signals and temperature signals generated when the equipment is running;

[0008] S2: pre-processing the sound signal collected in step S1, processing the sound signal into a sound signal segment of a smooth transition minimum speech segment with a noise portion removed;

[0009] S3: Extract features from the preprocessed sound signal fragments, vibration signals, and temperature signals. The sound signal fragments are subjected to the signal enhancement Mel-frequency cepstral coefficient method. The sound signal fragments are adaptively decomposed according to the time scale using the empirical mode decomposition to fully obtain the dynamic characteristics of the signal. The characteristic parameters are obtained by the Mel-frequency triangular filter. The obtained characteristic parameters are combined with time domain features such as short-time energy and short-time zero-crossing rate to obtain the characteristic vector of the sound signal fragment. For the vibration signal and temperature signal, their time domain parameters such as root mean square, standard deviation, kurtosis, skewness, etc. are selected as feature vectors.

[0010] S4: Perform feature fusion on the feature vectors obtained by multiple sensors in step S3, and bring the final fused vector into the constructed SVM model to obtain the fault status of the mechanical equipment.

[0011] Furthermore, in step S2, the preprocessing process includes beamforming, pre-emphasis, and frame windowing, wherein the sound signal is enhanced by a beamforming algorithm, and then the signal is pre-emphasized to compensate for the loss caused by the attenuation of high-frequency components. Finally, the signal is framed and windowed, and each frame of the signal is used as the minimum unit for analysis. The signal leakage in the frequency domain is reduced through the windowing operation.

[0012] Furthermore, in step S3, the frequency domain characteristic parameters of the sound signal segment are obtained by using the signal enhanced MFCC, and the specific method is as follows:

[0013] S3a: Determine the sound signal S i The local maximum and minimum points of (n) are obtained to obtain the upper and lower envelopes z 1 (n) and z 2 (n), calculate h(n), which is defined as follows:

[0014] h 1 (n) = S (n) - m 1 (n);

[0015] Where m 1 (n) is the mean of the upper and lower envelopes;

[0016] S3b:h 1 (n) must meet two restrictions: 1. The number of zero-crossing points and extreme value points is equal or differs by at most one; 2. The mean of the upper and lower envelopes is 0. If not, h 1 (n) is re-input as the signal to be decomposed and step 1 is repeated until the constraint condition is met. The first IMF is determined and recorded as c 1 (n).

[0017] S3c: c 1 (n) is separated from S(n) to obtain r 1 (n), r1 (n) Repeat steps 1 and 2 as the input signal.

[0018]

[0019] Where r t (n) represents the trend component of the speech signal. If r in the formula t (n) is less than S d Then the iteration stops, where S d The definition of

[0020]

[0021] Where T is the signal time length;

[0022] Therefore, the original signal can be decomposed into the sum of P IMF components and trend remainders as follows:

[0023]

[0024] Perform FFT transformation on the signal and calculate the energy of each frame spectral line:

[0025] E(i,k)=(FFT[S i (n)]) 2 ;

[0026] Where i represents the i-th frame, and k represents the k-th spectral line in the frequency domain.

[0027] S3d: Set the Mel triangle filter group to filter and amplify the low-frequency and medium-low-frequency components, and filter and reduce the weight of the medium-high frequency and high frequency. These filters are expressed as

[0028]

[0029] Where: m = 1, 2, 3...24, f(m) is the center frequency,

[0030] Then calculate its logarithmic spectrum to make it more robust. Its logarithmic spectrum is

[0031]

[0032] Where: H m (k) is the filter, S(m) is the log spectrum;

[0033] Finally, discrete cosine transform is performed to obtain the MFCC feature vector, the formula is:

[0034]

[0035] Where: C(n) is the MFCC feature vector of each frame signal.

[0036] Furthermore, in step S3, the obtained characteristic parameters are combined with characteristic parameters such as short-time energy and short-time zero-crossing rate in the time domain analysis to obtain a new characteristic vector. The specific method is as follows:

[0037] S3e: Calculate short-time energy. Short-time energy refers to the energy value in the speech signal in frames. The short-time energy formula of the i-th frame audio signal is:

[0038]

[0039] where s i (n) is the audio signal after preprocessing of the i-th frame, and N is the frame length.

[0040] S3f: Calculate the short-time zero-crossing rate. The short-time zero-crossing rate indicates the number of times the signal crosses the horizontal axis (zero point) in the waveform. The formula for the short-time zero-crossing rate is as follows:

[0041]

[0042] Where sgn[*] is a sign function, which is expressed as follows:

[0043]

[0044] Furthermore, in step S3, time domain characteristic parameters such as root mean square, standard deviation, kurtosis and skewness of the vibration signal and the temperature signal are obtained, and feature fusion is performed on the feature vectors obtained by multiple sensors.

[0045] RMS formula:

[0046]

[0047] Standard deviation formula:

[0048]

[0049] Kurtosis formula:

[0050]

[0051] Skewness formula:

[0052]

[0053] Furthermore, in step S4, the features of the sound, vibration and temperature signals collected by the multiple sensors are fused and used together as feature vectors for detecting fault signals, which are input into the SVM model for fault identification.

[0054] Furthermore, in step S4, the SVM model adopts a classification model of a linear support vector machine, which can be expressed as:

[0055] W*X+b=0

[0056] Where: X is the feature vector of the data set, W is the weight, and b is the bias vector.

[0057] The objective function is constructed by Lagrange multipliers and can be expressed as:

[0058]

[0059] Constraints:

[0060] Where: N is the number of samples, λ i , j is the i-th and j-th sample X i , X j The corresponding Lagrange multiplier, y i ,y j is the fault type value corresponding to the i-th and j-th samples.

[0061] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention adopts a multi-sensor feature fusion method for fault diagnosis of fan equipment in thermal power plants, and performs feature fusion on sound signals, vibration signals and temperature signals collected by multiple types of sensors, so that the system can obtain a more comprehensive understanding of the equipment status. The feature extraction of sound signals adopts the signal enhanced MFCC method, and the system's ability to obtain dynamic information is enhanced, which greatly improves the efficiency and reliability of fault diagnosis and is of great significance in practical engineering applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 : Flow chart of the fault diagnosis method for fan equipment in thermal power plants based on multi-sensor feature fusion proposed by the present invention.

[0063] Figure 2 : Flow chart of sound signal preprocessing in the present invention.

[0064] Figure 3 : MFCC feature extraction flow chart of acoustic signal enhancement in the present invention. DETAILED DESCRIPTION

[0065] The technical solution of the invention is described in detail below in conjunction with the accompanying drawings. Obviously, the described examples are only some examples of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0066] It is worth mentioning that, unless otherwise specified, the technical terms used in the present application should have the common meanings understood by those skilled in the art to which the present invention belongs.

[0067] like Figure 1 As shown, the fault diagnosis method flow of the fan equipment in a thermal power plant based on multi-sensor feature fusion includes the following steps:

[0068] S1: Collect the sound signal, vibration signal and temperature signal generated when the equipment is running;

[0069] S2: Preprocess the sound data collected in step S1, and process the signal into a sound signal segment with a smooth transition and minimum sound segment with the noise part removed. The preprocessing process includes beamforming, pre-emphasis, frame division and windowing. The flow chart is as follows: Figure 2 shown.

[0070] S2a: The collected sound signal is enhanced using a beamforming algorithm, and a beam pointing in the direction is formed by delay compensation of the array signal, thereby removing the noise influence in other directions.

[0071] S2b: After obtaining the signal beam, the signal is pre-emphasized to emphasize the high-frequency part of the sound and increase the high-frequency resolution of the speech. Generally, the transfer function is H(z)=1-az -1 A first-order FIR high-pass digital filter is used to implement pre-emphasis.

[0072] S2c: The processed sound signal is framed and divided into several sound segments. In order to make the characteristic parameters of the sound signal change smoothly, a part of the overlap between two frames is set, i.e., frame shift. Finally, a windowing operation is performed to reduce the leakage of the signal in the frequency domain.

[0073] S3: Extract features from the collected sound signal segments, vibration signals, and temperature signals, and fuse the extracted feature vectors, such as Figure 3 As shown in Figure 1, the traditional MFCC is improved by using empirical mode decomposition.

[0074] S3a: In the frequency domain feature extraction of the sound signal segment, the signal enhanced Mel frequency cepstral coefficient method is used, and the speech signal is adaptively decomposed according to the time scale by using the empirical mode decomposition to obtain several intrinsic mode function components (IMF). The specific operation steps are as follows:

[0075] Determine the speech signal S i The local maximum and minimum points of (n) are obtained to obtain the upper and lower envelopes z 1 (n) and z 2 (n), calculate h(n), which is defined as follows:

[0076] h 1 (n) = S (n) - m 1 (n);

[0077] Where m 1 (n) is the mean of the upper and lower envelopes.

[0078] h 1 (n) must meet two restrictions: 1. The number of zero-crossing points and extreme value points is equal or differs by at most one; 2. The mean of the upper and lower envelopes is 0. If not, h 1 (n) is re-input as the signal to be decomposed and step 1 is repeated until the constraint condition is met. The first IMF is determined and recorded as c 1 (n).

[0079] c 1 (n) is separated from S(n) to obtain r 1 (n), r 1 (n) Repeat steps S3a and S3b as the input signal.

[0080] S(n)-c 1 (n) = r 1 (n)

[0081] r 1 (n)-c 2 (n) = r 2 (n)

[0082] r t-1 (n)-c t (n) = r t (n)

[0083] Where r t (n) represents the trend component of the speech signal. If r in the formula t (n) is less than S d Then the iteration stops, where S d The definition of

[0084] Where T is the signal time length.

[0085] Therefore, the original signal can be decomposed into the sum of P IMF components and trend remainders, as follows

[0086]

[0087] Perform FFT transformation on the signal and calculate the energy of each frame spectral line

[0088] E(i,k)=(FFT[S i (n)]) 2

[0089] Where i represents the i-th frame, and k represents the k-th spectral line in the frequency domain.

[0090] The Mel triangle filter group is set to filter and amplify the low-frequency and medium-low-frequency components, and filter and reduce the weight of the medium-high frequency and high frequency. These filters are expressed as

[0091]

[0092] Where: m = 1, 2, 3...24, f(m) is the center frequency,

[0093] Then calculate its logarithmic spectrum to make it more robust. Its logarithmic spectrum is

[0094]

[0095] Where: H m (k) is the filter, S(m) is the log spectrum;

[0096] Finally, discrete cosine transform is performed to obtain the MFCC feature vector, the formula is:

[0097]

[0098] Where: C(n) is the MFCC feature vector of each frame signal.

[0099] S3b: Extract the time domain features of the sound signal, select the short-time energy and short-time zero-crossing rate and add them to the feature vector. Short-time energy refers to the energy value of the sound signal in frames. The short-time energy formula of the i-th frame audio signal is

[0100]

[0101] where s i (n) is the audio signal after preprocessing of the i-th frame, and N is the frame length.

[0102] The short-time zero-crossing rate indicates the number of times the signal crosses the horizontal axis (zero point) in the waveform. The formula for the short-time zero-crossing rate is as follows:

[0103]

[0104] Where sgn[*] is a sign function, which is expressed as follows:

[0105]

[0106] S3c: Obtain the time domain characteristic parameters such as the root mean square, standard deviation, kurtosis and skewness of the vibration signal and temperature signal, and perform feature fusion on the feature vectors obtained by multiple sensors.

[0107] RMS formula:

[0108]

[0109] Standard deviation formula:

[0110]

[0111] Kurtosis formula:

[0112]

[0113] Skewness formula:

[0114]

[0115] S4: Normalize each feature vector and use the deviation standardization method to map the result value to [0,1]. The formula is as follows:

[0116]

[0117] The classification model using linear support vector machine can be expressed as

[0118] W*X+b=0

[0119] Where: X is the feature vector of the data set, W is the weight, and b is the bias vector.

[0120] The objective function is constructed by Lagrange multipliers and can be expressed as:

[0121]

[0122] Constraints:

[0123] Where: N is the number of samples, λ i , j is the i-th and j-th sample X i , X j The corresponding Lagrange multiplier, y i ,y j is the fault type value corresponding to the i-th and j-th samples.

[0124] Finally, it should be noted that the above implementation case is only a detailed explanation of the technical solution, rather than a limitation thereof; although the technical solution of the present invention has been described with reference to this specific implementation case, ordinary technicians in the field should understand that they can still modify the solution of the implementation case, or make equivalent replacements for some of them; such modification or replacement cannot essentially deviate from the scope of the technical solution proposed by the present invention, and should be included in the scope of the claims and specification of the present invention.

Claims

1. A method for fault diagnosis of fan equipment in thermal power plants based on multi-sensor feature fusion, characterized in that: The following steps are involved: S1: collects sound signals, vibration signals and temperature signals generated when the equipment is running; S2: pre-processing the sound signal collected in step S1, processing the sound signal into a sound signal segment of a smooth transition minimum speech segment with a noise portion removed; S3: Feature extraction is performed on the preprocessed sound signal fragments, vibration signals, and temperature signals. The sound signal fragments use the signal enhancement Mel cepstral coefficient method, and use empirical mode decomposition to adaptively decompose the sound signal according to the time scale to fully obtain the dynamic characteristics of the signal. The characteristic parameters are obtained through the Mel triangular filter, and the obtained characteristic parameters are combined with the short-time energy and short-time zero-crossing rate time domain features to obtain the characteristic vector of the sound signal fragment. The vibration signal and the temperature signal select their time domain parameters root mean square, standard deviation, kurtosis, and skewness as feature vectors; in step S3, in the frequency domain feature extraction of the sound signal fragment, the traditional MFCC is improved by empirical mode decomposition. The specific method is as follows: S3a: Determine the speech signal S i (n), obtain the upper and lower envelopes z1(n) and z2(n), and calculate h(n), which is defined as follows: h1(n)=S(n)-m1(n); Where m1(n) is the mean of the upper and lower envelopes; S3b: h1(n) must meet two restrictions:

1. The number of zero-crossing points and extreme value points is equal or differs by at most one; the mean of the upper and lower envelopes is 0; if not, h1(n) is re-input as the signal to be decomposed, and step S3a is repeated until the restrictions are met, and the first IMF is determined and recorded as c1(n); S3c: Separate c1(n) from S(n) to obtain r1(n), and use r1(n) as the input signal to repeat steps S3a and S3b; S(n)-c1(n)=r1(n) r1(n)-c2(n)=r2(n); r t-1 (n)-c t (n)=r t (n) Where r t (n) represents the trend component of the speech signal. If r in the formula t (n) is less than S d Then stop the iteration, where S d is defined as follows: Where T is the signal time length; Therefore, the original signal can be decomposed into the sum of P IMF components and trend remainders as follows: Perform FFT transformation on the signal and calculate the energy of each frame spectral line: E(i,k)=(FFT[S i (n)]) 2 ; Where i represents the i-th frame, and k represents the k-th spectrum line in the frequency domain; S3d: Set the Mel triangle filter group to filter and amplify the low frequency and medium-low frequency components, and filter and reduce the weight of the medium-high frequency and high frequency. These filters are expressed as Where: m = 1, 2, 3...24, f(m) is the center frequency, Then calculate its logarithmic spectrum to make it more robust. Its logarithmic spectrum is Where: H m (k) is the filter, S(m) is the log spectrum; Finally, discrete cosine transform is performed to obtain the MFCC feature vector, the formula is: Where: C(n) is the MFCC feature vector of each frame signal; S4: Perform feature fusion on the feature vectors obtained by multiple sensors in step S3, and bring the final fused vector into the constructed SVM model to obtain the fault status of the mechanical equipment.

2. The method for diagnosing fan faults in thermal power plants based on multi-sensor feature fusion according to claim 1 is characterized in that: In step S2, the preprocessing process includes beamforming, pre-emphasis, frame division and windowing. The sound signal obtained in step S1 is enhanced by a beamforming algorithm, and then the enhanced signal is pre-emphasized to compensate for the loss caused by the attenuation of high-frequency components. Finally, the signal is framed and windowed, and each frame signal is used as the minimum unit for analysis. The signal leakage in the frequency domain is reduced through the windowing operation.

3. The method for diagnosing fan faults in thermal power plants based on multi-sensor feature fusion according to claim 1 is characterized in that: In step S3, the frequency domain characteristic parameters of the sound signal segment are combined with the short-time energy and short-time zero-crossing rate characteristic parameters in the time domain analysis to obtain a new characteristic vector. The specific method is as follows: S3e: Calculate short-time energy; short-time energy refers to the energy value in the speech signal in frames. The short-time energy formula of the i-th frame audio signal is: where s i (n) is the audio signal after preprocessing of the i-th frame, and N is the frame length; S3f: Calculate the short-time zero-crossing rate; the short-time zero-crossing rate indicates the number of times the signal crosses the horizontal axis in the waveform. The formula for the short-time zero-crossing rate is as follows: Where sgn[*] is a sign function, which is expressed as follows:

4. The method for diagnosing fan faults in thermal power plants based on multi-sensor feature fusion according to claim 1 is characterized in that: Step S3 extracts features from the vibration signal and the temperature signal, and selects the root mean square value, standard deviation, kurtosis, and skewness as time domain feature parameters; S3g: Calculate the root mean square value, standard deviation, kurtosis, and skewness of the vibration signal and temperature signal; RMS formula: Standard deviation formula: Kurtosis formula: Skewness formula:

5. The method for diagnosing fan faults in thermal power plants based on multi-sensor feature fusion according to claim 1 is characterized in that: In step S4, the features of the sound, vibration and temperature signals collected by the multiple sensors are fused and used together as feature vectors for detecting fault signals, which are input into the SVM model for fault identification.

6. The method for diagnosing faults of fan equipment in a thermal power plant based on multi-sensor feature fusion according to claim 1 is characterized in that: In step S4, the SVM model adopts a linear support vector machine classification model, which can be expressed as W*X+b=0; Where: X is the feature vector of the data set, W is the weight, and b is the bias vector; The objective function is constructed by Lagrange multipliers and can be expressed as: Constraints: Where: N is the number of samples, λ i , j is the i-th and j-th sample X i , X j The corresponding Lagrange multiplier, y i ,y j is the fault type value corresponding to the i-th and j-th samples.

Images