A Machine Learning-Based Fault Classification Method for Air Conditioner Compressors

Abstract

This invention relates to the field of data processing technology, specifically to a machine learning-based method for classifying faults in air conditioning compressors. The method includes: collecting vibration signals, rotational speed data, and noise vibration signals; calculating noise intensity; calculating the theoretical fault frequency; extracting the main noise frequency and calculating the overlap between the theoretical fault frequency and the main noise frequency; determining the noise type based on frequency domain energy entropy; determining the wavelet basis and decomposition level to decompose the vibration signal and obtain fault-sensitive feature signals; generating enhanced samples and using these enhanced samples to train a random forest model to obtain a noise-resistant fault classification model; inputting the fault-sensitive feature signals corresponding to the vibration signals of the air conditioning compressor to be analyzed into the noise-resistant fault classification model, and outputting the fault analysis results for the air conditioning compressor to be analyzed through the noise-resistant fault classification model. This invention can improve the accuracy of fault classification for air conditioning compressors.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and specifically to a machine learning-based method for classifying faults in air conditioning compressors. Background Technology

[0002] The air conditioning compressor is the core power component of the refrigeration system, and its operating status directly determines the reliability and safety of the entire system. In actual operation, common compressor failures include bearing wear, rotor eccentricity, valve damage, and refrigerant leakage. If these failures are not detected and addressed in a timely manner, they may lead to unplanned unit shutdowns, refrigerant leaks causing environmental pollution, and even equipment damage or safety accidents.

[0003] Currently, the mainstream method for compressor fault classification is based on vibration or acoustic signal analysis techniques. The basic process is as follows: vibration or acoustic signals during compressor operation are collected by sensors, and time-domain features (such as peak value, root mean square, kurtosis, etc.) and frequency-domain features (such as amplitude at characteristic frequencies, sideband energy, etc.) are extracted. Then, the extracted features are input into machine learning models such as support vector machines, neural networks, or random forests to achieve automatic classification of fault types.

[0004] However, complex environmental noise interference is prevalent in real-world applications. Taking a refrigeration room as an example, multiple noise sources often exist around the compressor: aerodynamic noise from the fan, resonance noise from the cooling water pipes due to fluid excitation, vibration noise transmitted through the foundation between multiple parallel units, and radiated noise from external traffic or industrial equipment. These environmental noises couple with the compressor's own operating signals, significantly reducing the signal-to-noise ratio of the collected acoustic and vibration signals, and submerging fault characteristics in the strong noise background. The environmental noise characteristics vary greatly depending on the installation scenario and the compressor model; no single set of filtering parameters or feature thresholds is applicable to all operating conditions. Specifically, if the filtering parameters are set too strictly, some weak early fault characteristics will be filtered out along with the noise, leading to missed fault detection; if the filtering parameters are set too loosely, a large amount of noise will be retained, causing misjudgments by the classification model. Currently, most engineering practices use manual calibration, setting feature thresholds individually for each compressor. This method has inherent drawbacks such as high calibration costs, poor adaptability to changes in the installation environment, and inability to dynamically respond to noise evolution, resulting in generally low fault classification accuracy and making it difficult to meet the actual needs of intelligent operation and maintenance. Summary of the Invention

[0005] This invention provides a machine learning-based method for classifying faults in air conditioner compressors to address existing problems.

[0006] The present invention provides a machine learning-based method for classifying air conditioner compressor faults, which employs the following technical solution: One embodiment of the present invention provides a machine learning-based method for classifying air conditioner compressor faults, the method comprising the following steps: The system collects vibration signals and speed data of the air conditioner compressor during operation, and also collects noise and vibration signals. Calculate noise intensity based on noise vibration signals; The theoretical failure frequency of the air conditioner compressor is calculated based on its factory parameters and speed data. Extract the main noise frequency from the noise vibration signal and calculate the overlap between the theoretical fault frequency and the main noise frequency; Perform Fourier transform on the vibration signal and divide it into frequency bands, calculate the frequency domain energy entropy, and determine the noise type based on the frequency domain energy entropy; Based on the noise type and overlap, the wavelet basis and the number of decomposition layers are determined respectively, and the vibration signal is decomposed to obtain the fault-sensitive characteristic signal. Simulated noise samples are generated based on noise type and noise intensity. These simulated noise samples are then superimposed onto fault-sensitive feature signals to generate enhanced samples. Finally, the enhanced samples are used to train a random forest model to obtain a noise-resistant fault classification model. The sensitive feature signal of the fault to be analyzed corresponding to the vibration signal of the air conditioner compressor to be analyzed is input into the noise-resistant fault classification model, and the fault analysis result of the air conditioner compressor to be analyzed is output through the noise-resistant fault classification model.

[0007] Furthermore, the noise intensity is calculated based on the noise vibration signal, specifically including: The noise intensity is obtained by squaring each vibration amplitude in the noise vibration signal and summing the results, then taking the square root of the ratio of the sum to the number of data points in the noise vibration signal.

[0008] Furthermore, the theoretical failure frequency includes the theoretical failure frequency of the compressor inner ring, the theoretical failure frequency of the compressor outer ring, and the theoretical failure frequency of the compressor rolling elements. The calculation of the theoretical failure frequency based on the air conditioning compressor's factory parameters and speed data specifically includes: Obtain the factory parameters of the air conditioner compressor, which include the number of rolling elements, pitch circle diameter, rolling element diameter, and contact angle of the compressor bearing. Half the number of rolling elements is defined as the number of half rolling elements, and the rotational speed data per second is used as the unit rotational speed; Calculate the ratio of the rolling element diameter to the pitch circle diameter, and multiply this ratio by the cosine of the contact angle to determine the contact proportionality coefficient; The theoretical failure frequency of the compressor inner ring is determined by multiplying the product of the number of semi-rolling elements and the unit speed, and then multiplying it by a factor plus the contact ratio coefficient. The theoretical failure frequency of the compressor outer ring is determined by multiplying the product of the number of semi-rolling elements and the unit speed, and then multiplying by the difference between the contact ratio coefficient and the product of the product of the number of semi-rolling elements and the unit speed. The theoretical failure frequency of the compressor rolling element is determined by multiplying the ratio of the pitch circle diameter to the rolling element diameter by the unit speed, and then multiplying the product by the square of the contact ratio coefficient.

[0009] Furthermore, the main noise frequency in the noise vibration signal is extracted, and the overlap between the theoretical fault frequency and the main noise frequency is calculated, specifically including: Perform a Fourier transform on the noise vibration signal and determine the frequency with the highest amplitude ranking as the main noise frequency. For each theoretical fault frequency, calculate the frequency difference between the theoretical fault frequency and each main noise frequency, and determine the theoretical fault frequencies with frequency differences less than a preset error threshold as overlapping fault frequencies. The ratio of the number of overlapping fault frequencies to the total number of theoretical fault frequencies is determined as the overlap between the theoretical fault frequencies and the main noise frequencies.

[0010] Furthermore, the vibration signal is subjected to Fourier transform and divided into frequency bands, the frequency domain energy entropy is calculated, and the noise type is determined based on the frequency domain energy entropy, specifically including: Perform a Fourier transform on the vibration signal to obtain a frequency domain signal, and divide the frequency range of the frequency domain signal into at least three equal-width frequency bands; For each equal-width frequency band, calculate the proportion of energy within that band to the total energy of the entire frequency domain signal. Multiply the logarithm of this proportion (base 2) by the proportion itself to determine the energy entropy contribution value of that band. The frequency domain energy entropy is obtained by summing the energy entropy contributions of all equal-width frequency bands and taking the negative of the sum. The frequency domain energy entropy is compared with a preset entropy threshold. If the frequency domain energy entropy is less than the preset entropy threshold, the noise type is determined to be steady-state noise; otherwise, the noise type is determined to be transient noise.

[0011] Furthermore, based on the noise type and overlap, the wavelet basis and decomposition level are determined respectively, and the vibration signal is decomposed to obtain fault-sensitive characteristic signals, specifically including: Perform a Fourier transform on the vibration signal to obtain the frequency domain signal, and calculate the total energy across the entire frequency range in the frequency domain signal; Determine the total fault band energy corresponding to the fault characteristic band for each theoretical fault frequency; The ratio of the total energy of the fault frequency band corresponding to each theoretical fault frequency to the total energy of the vibration signal is determined as the energy proportion of the fault characteristic frequency band. When the noise type is steady-state noise, the db4 wavelet is selected as the first wavelet basis; when the noise type is transient noise, the sym5 wavelet is selected as the second wavelet basis. When the overlap is less than the preset overlap threshold, the first decomposition layer is selected; otherwise, the second decomposition layer is selected, wherein the first decomposition layer is greater than the second decomposition layer. The vibration signal was decomposed using a defined wavelet basis and a number of decomposition levels to obtain wavelet coefficients in different frequency ranges. The wavelet coefficients of the main noise frequency are set to zero, while other wavelet coefficients are retained. The reconstructed signal is obtained through wavelet reconstruction. Calculate the energy percentage of each fault characteristic frequency band in the reconstructed signal. If the energy percentage after reconstruction is lower than the energy percentage before reconstruction, adjust the number of decomposition layers and re-decompose and reconstruct until the energy percentage after reconstruction is not lower than the energy percentage before reconstruction, and obtain the fault sensitive characteristic signal.

[0012] Furthermore, the total fault band energy corresponding to the fault characteristic band for each theoretical fault frequency is determined, specifically including: For each theoretical fault frequency, the difference between the theoretical fault frequency and the preset offset is determined as the lower limit of the frequency band, the sum of the theoretical fault frequency and the preset offset is determined as the upper limit of the frequency band, and the frequency band interval formed by the lower limit and the upper limit of the frequency band is determined as the fault characteristic frequency band corresponding to the theoretical fault frequency. Extract the components of all frequencies within the fault characteristic frequency band and sum them up to obtain the total fault frequency band energy corresponding to the theoretical fault frequency.

[0013] Furthermore, the simulated noise samples include steady-state simulated noise samples and transient simulated noise samples. The generation of simulated noise samples based on noise type and noise intensity specifically includes: When the noise type is steady-state noise, Gaussian white noise is generated, and a sine wave corresponding to the main noise frequency is superimposed on the Gaussian white noise. The intensity of the superimposed noise is adjusted to match the noise intensity to obtain a steady-state simulated noise sample. When the noise type is transient noise, an impulse pulse signal with an amplitude matching the noise intensity and a randomly set pulse interval is generated to obtain a transient simulated noise sample.

[0014] Furthermore, simulated noise samples are superimposed onto the fault-sensitive feature signal to generate enhanced samples, specifically including: The simulated noise samples are superimposed on the fault-sensitive feature signals at at least two preset intensity multiples to obtain the initial enhanced samples; Each initial enhanced sample is labeled with its corresponding fault type to obtain the enhanced sample.

[0015] This invention proposes a machine learning-based air conditioner compressor fault classification system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of a machine learning-based air conditioner compressor fault classification method.

[0016] The beneficial effects of the technical solution of the present invention are: In this embodiment of the invention, by first analyzing the environmental noise characteristics and then adaptively refining the fault signal, and generating simulated noise samples based on the noise characteristics for data augmentation training of the noise-resistant model, the interference of complex environmental noise on fault characteristics can be effectively reduced, the filtering out of early weak fault signals can be avoided, the generalization ability and classification accuracy of the model under different noise scenarios can be improved, and at the same time, there is no need to manually calibrate the feature threshold for a single compressor, which significantly reduces the calibration cost and achieves highly reliable and adaptive compressor fault classification. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a machine learning-based fault classification method for air conditioner compressors, as provided in one embodiment of the present invention; Figure 2 This is a structural diagram of a machine learning-based fault classification system for air conditioner compressors, provided as an embodiment of the present invention. Detailed Implementation

[0019] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a machine learning-based air conditioner compressor fault classification method proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0021] The following description, in conjunction with the accompanying drawings, details a specific scheme for a machine learning-based fault classification method for air conditioner compressors provided by this invention.

[0022] This invention provides a machine learning-based method for classifying air conditioner compressor faults. Please refer to [link / reference]. Figure 1 The diagram illustrates a flowchart of a machine learning-based fault classification method for air conditioner compressors according to an embodiment of the present invention. The method includes the following steps: S101. Collect vibration signals and speed data of the air conditioner compressor during operation, and collect noise vibration signals at the same time.

[0023] In this embodiment, vibration signal acquisition is achieved using a piezoelectric vibration sensor. For example, a piezoelectric vibration sensor of model PCB 352C33 is used, and it is bolted to a high-fault area on the compressor housing. High-fault areas include locations near the bearings and near the crankshaft. During installation, the resonance points of the compressor housing must be avoided. The resonance points are determined by tapping the housing to test the resonance frequency, and selecting a location outside the resonance frequency as the sensor installation point.

[0024] The sampling frequency of the vibration signal is set to 10 kHz. This value is selected based on the Nyquist sampling theorem: the characteristic frequencies of compressor faults are usually distributed in the range of 0 to 5 kHz, and the sampling frequency needs to be higher than twice the highest fault characteristic frequency, that is, higher than 10 kHz, to ensure that frequency aliasing does not occur during signal acquisition.

[0025] Each batch collects 10 seconds of continuous vibration data and records the collection timestamp simultaneously. The collected vibration data format is as follows: 10,000 vibration amplitudes are collected per second, with the unit expressed as gravitational acceleration g, and the data is presented as an amplitude sequence in the form of 0.02, 0.03, 0.01, ..., 0.04.

[0026] Rotational speed data is acquired using a Hall effect speed sensor. For example, a Hall effect speed sensor, model HC-020K, is used and installed near the compressor shaft to acquire real-time rotational speed data in revolutions per minute (rpm).

[0027] Ambient temperature data is acquired using thermocouple temperature sensors. For example, a type K thermocouple is used, attached to the middle of the compressor casing, to collect ambient temperature data in degrees Celsius. This ambient temperature data can be used to determine if the compressor's operating conditions are abnormal and to assist in confirming the type of fault.

[0028] The noise and vibration signals were acquired as follows: with the compressor off, one second of vibration data was collected as a pure noise segment. This pure noise segment only includes ambient background noise and does not contain any operating vibration signals from the compressor itself, thus it can be used as a benchmark for subsequent noise analysis.

[0029] It should be noted that the vibration signal, rotational speed data, and ambient temperature data were time-stamped during acquisition to ensure that the three sets of data were synchronized in time, providing a basis for subsequent correlation analysis.

[0030] S102. Calculate noise intensity based on noise vibration signal.

[0031] In this embodiment, the calculation of noise intensity based on noise vibration signals specifically includes: The noise intensity is obtained by squaring each vibration amplitude in the noise vibration signal and summing the results, then taking the square root of the ratio of the sum to the number of data points in the noise vibration signal.

[0032] For example, noise intensity is calculated based on noise vibration signals in order to quantify the average energy level of environmental background noise and provide an intensity benchmark for the subsequent generation of simulated noise samples.

[0033] Specifically, the noise and vibration signals are acquired. These signals are pure noise segments acquired when the compressor is not running, with a length of 1 second and containing N vibration amplitude data points.

[0034] Because noise vibration signals exhibit alternating positive and negative oscillations, directly calculating the arithmetic mean of the amplitude will cause the positive and negative values ​​to cancel each other out, resulting in a calculation result close to zero and failing to accurately reflect the noise's energy intensity. Therefore, this embodiment uses the root mean square (RMS) value to describe noise intensity. This physical quantity effectively characterizes the average energy level of the signal. Its calculation logic is as follows: first, each data point is squared to eliminate the influence of the positive and negative signs; then, the mean of the squared values ​​is calculated; finally, the square root is taken to restore the dimension to the original amplitude unit.

[0035] The calculation formula can be: ; in, Indicates noise intensity. Indicating the first noise vibration signal One vibration amplitude, This indicates the number of data points in the noise vibration signal.

[0036] The physical meaning of this formula is as follows: First, calculate the arithmetic mean of the squares of each vibration amplitude (reflecting the average power of the signal), then take the square root (converting the power into an equivalent amplitude), and finally obtain the root mean square value of the noise signal. The magnitude of the root mean square value directly reflects the average energy strength of the noise signal. The larger the value, the stronger the background noise. The smaller the value, the quieter the environment.

[0037] S103. Calculate the theoretical failure frequency of the air conditioner compressor based on its factory parameters and speed data.

[0038] In this embodiment, the theoretical failure frequency includes the theoretical failure frequency of the compressor inner ring, the theoretical failure frequency of the compressor outer ring, and the theoretical failure frequency of the compressor rolling elements. The theoretical failure frequency is calculated based on the air conditioner compressor's factory parameters and speed data, specifically including: Obtain the factory parameters of the air conditioner compressor, which include the number of rolling elements, pitch circle diameter, rolling element diameter, and contact angle of the compressor bearing. Half the number of rolling elements is defined as the number of half rolling elements, and the rotational speed data per second is used as the unit rotational speed; Calculate the ratio of the rolling element diameter to the pitch circle diameter, and multiply this ratio by the cosine of the contact angle to determine the contact proportionality coefficient; The theoretical failure frequency of the compressor inner ring is determined by multiplying the product of the number of semi-rolling elements and the unit speed, and then multiplying it by a factor plus the contact ratio coefficient. The theoretical failure frequency of the compressor outer ring is determined by multiplying the product of the number of semi-rolling elements and the unit speed, and then multiplying by the difference between the contact ratio coefficient and the product of the product of the number of semi-rolling elements and the unit speed. The theoretical failure frequency of the compressor rolling element is determined by multiplying the ratio of the pitch circle diameter to the rolling element diameter by the unit speed, and then multiplying the product by the square of the contact ratio coefficient.

[0039] Determine the total fault band energy corresponding to the fault characteristic band for each theoretical fault frequency, specifically including: For each theoretical fault frequency, the difference between the theoretical fault frequency and the preset offset is determined as the lower limit of the frequency band, the sum of the theoretical fault frequency and the preset offset is determined as the upper limit of the frequency band, and the frequency band interval formed by the lower limit and the upper limit of the frequency band is determined as the fault characteristic frequency band corresponding to the theoretical fault frequency. Extract the components of all frequencies within the fault characteristic frequency band and sum them up to obtain the total fault frequency band energy corresponding to the theoretical fault frequency.

[0040] For example, theoretical fault frequencies include the theoretical fault frequencies of the compressor inner ring, the compressor outer ring, and the compressor rolling elements. The purpose of calculating theoretical fault frequencies is to determine the vibration frequencies that should theoretically occur when different types of faults occur in the compressor, providing a frequency reference for subsequent frequency band localization and fault feature extraction.

[0041] The theoretical failure frequency of the air conditioner compressor is calculated based on its factory parameters and speed data. The specific process is as follows: First, obtain the factory specifications of the air conditioner compressor. These specifications include the number of rolling elements in the compressor bearings. Pitch circle diameter (Unit: mm) Rolling element diameter (Unit: mm) and contact angle (Unit: degrees). Simultaneously, acquire real-time rotational speed data. (Unit: revolutions per minute).

[0042] Then, half the number of rolling elements, i.e. The number of rolling elements is determined as half the number of rolling elements. This parameter represents half the number of rolling elements simultaneously bearing load on the rolling element distribution circle, and is a common factor for calculating the failure frequency of the inner and outer races.

[0043] The rotational speed data per second, i.e. The value is determined to be the unit rotational speed. This parameter converts the rotational speed from "revolutions per minute" to "revolutions per second," making the unit of the fault frequency calculation hertz (Hz), which facilitates subsequent frequency domain analysis.

[0044] Calculate the ratio of the rolling element diameter to the pitch circle diameter, i.e. And this ratio is compared with the cosine of the contact angle. The product of, i.e. This is determined to be the contact ratio coefficient. This coefficient reflects the geometric characteristics of the bearing and is a key parameter for distinguishing between inner ring failure, outer ring failure, and rolling element failure.

[0045] Based on the above intermediate parameters, the three theoretical fault frequencies are calculated respectively: Multiply the product of the number of semi-rolling bodies and the unit speed by a factor, and then multiply by the sum of the contact ratio factor, i.e. The theoretical failure frequency of the compressor inner ring was determined. The physical meaning of this formula is: when the inner ring fails, the frequency at which the rolling elements pass through the inner ring defect is affected by both the bearing geometry and the rotational speed. The "one plus" term in the formula reflects the increasing effect of the inner ring rotation on the failure frequency.

[0046] Multiply the product of the number of semi-rolling bodies and the unit speed by a factor minus the contact ratio coefficient, i.e. The theoretical failure frequency of the compressor outer ring was determined. The physical meaning of this formula is: with the outer ring fixed, the frequency of the rolling elements passing through the defects on the outer ring is affected by the bearing geometry. The "minus one" term in the formula reflects the frequency characteristics of the outer ring when it is stationary.

[0047] Multiply the ratio of the pitch circle diameter to the rolling element diameter by the unit speed, and then multiply the product by a factor minus the square of the contact proportionality coefficient, i.e. The theoretical failure frequency of the compressor rolling elements was determined. The physical meaning of this formula is that when a rolling element fails, its rotation frequency and revolution frequency work together to affect the vibration signal. The square term in the formula reflects the complex modulation relationship of the rolling element failure.

[0048] The calculation parameters in the above three formulas can all be obtained from the bearing manufacturer's instruction manual. The theoretical failure frequency changes linearly with the real-time rotational speed. The higher the rotational speed, the higher the failure frequency; the lower the rotational speed, the lower the failure frequency.

[0049] S104. Extract the main noise frequency from the noise vibration signal and calculate the overlap between the theoretical fault frequency and the main noise frequency.

[0050] In this embodiment, the main noise frequency in the noise vibration signal is extracted, and the overlap between the theoretical fault frequency and the main noise frequency is calculated, specifically including: Perform a Fourier transform on the noise vibration signal and determine the frequency with the highest amplitude ranking as the main noise frequency. For each theoretical fault frequency, calculate the frequency difference between the theoretical fault frequency and each main noise frequency, and determine the theoretical fault frequencies with frequency differences less than a preset error threshold as overlapping fault frequencies. The ratio of the number of overlapping fault frequencies to the total number of theoretical fault frequencies is determined as the overlap between the theoretical fault frequencies and the main noise frequencies.

[0051] For example, the main noise frequency in the noise vibration signal is extracted, and the overlap between the theoretical fault frequency and the main noise frequency is calculated. The purpose is to identify the frequency components with the strongest energy in the environmental noise and quantify the frequency proximity between these noise frequencies and the theoretical fault frequency of the compressor, so as to provide a basis for the subsequent adaptive selection of the wavelet decomposition layer.

[0052] First, a Fourier transform is performed on the noise vibration signal to convert the time-domain noise vibration signal into a frequency-domain signal. After the Fourier transform, the horizontal axis of the frequency-domain signal represents frequency (unit: Hz), and the vertical axis represents amplitude (unit: g). The magnitude of the amplitude reflects the energy strength of the frequency component in the noise signal.

[0053] It should be noted that the preset ranking K is a pre-defined positive integer used to determine the number of main noise frequencies. This parameter is set based on the following considerations: environmental noise typically contains multiple frequency components, but only a few of the strongest frequencies have the most significant masking effect on fault characteristics. If too few noise frequencies are selected, important strong noise sources may be missed; if too many are selected, a large number of weak noise frequencies will be introduced, increasing unnecessary computational overhead and potentially reducing the accuracy of overlap determination.

[0054] For example, in this embodiment, the preset rank K is set to 5. This value is based on the following: In a real refrigeration room environment, the main noise sources typically do not exceed five (such as fan noise, pipe resonance, parallel unit vibration, traffic noise, etc.), and selecting the top five main noise frequencies is sufficient to cover most strong interference sources. It should be noted that this value is only an example; in practical applications, it can be adjusted according to the complexity of the noise environment. For instance, in scenarios with many noise sources, the K value can be appropriately increased, while in scenarios with fewer noise sources, the K value can be appropriately decreased.

[0055] After obtaining the main noise frequencies, it is necessary to calculate the degree of overlap between the theoretical fault frequencies and these noise frequencies. The specific steps are as follows: First, for each theoretical failure frequency (including the inner race theoretical failure frequency, the outer race theoretical failure frequency, and the rolling element theoretical failure frequency), the frequency difference between this theoretical failure frequency and each main body noise frequency is calculated. The frequency difference is calculated by subtracting the main body noise frequency from the theoretical failure frequency and taking the absolute value. ,in, Represents the theoretical fault frequency (i.e. and ), Indicates the frequency of the main noise.

[0056] Then, the frequency difference is less than the preset error threshold. The theoretical fault frequency is determined as the overlapping fault frequency. The overlapping fault frequency indicates that the theoretical fault frequency is too close to a certain strong noise frequency in the frequency domain, and the noise will mask or interfere with the identification of the fault characteristics.

[0057] For example, a preset error threshold The value is set to 5Hz. Assume the theoretical fault frequency of the inner ring. =250Hz, the main noise frequency includes 248Hz, and the frequency difference between the two is 2Hz, which is less than 5Hz. Therefore, the theoretical fault frequency of the inner ring is determined to be an overlapping fault frequency. If the other theoretical fault frequency is 500Hz, and all main noise frequencies are far from 500Hz (the difference is greater than 5Hz), then this frequency is not determined to be an overlapping fault frequency.

[0058] It should be noted that the above-mentioned preset error threshold This is a pre-defined frequency tolerance parameter used to determine whether the theoretical fault frequency overlaps with the main noise frequency. The setting of this parameter is based on the following considerations: In actual operation, due to factors such as speed fluctuations, measurement errors, and Fourier transform frequency resolution, there is a certain deviation between the theoretical fault frequency and the actual fault frequency. If the error threshold is set too small, it may miss actual overlapping frequency pairs; if it is set too large, it will excessively expand the overlap range, misjudging frequencies that do not overlap as overlapping, leading to overly conservative selection of subsequent filtering parameters.

[0059] For example, in this embodiment, a preset error threshold is used. The value is set at 5Hz. This value is based on the following: compressor speed fluctuations are typically controlled within ±5%. Taking 3000 rpm as an example, the fault frequency deviation is approximately ±10Hz; considering the frequency resolution of the Fourier transform (usually 1-2Hz), the total deviation generally does not exceed ±15Hz. Therefore, a 5Hz error threshold ensures that overlapping frequency differences within a small range are effectively identified. This value is for illustrative purposes only and can be adjusted accordingly in practical applications based on factors such as compressor speed stability and measurement accuracy.

[0060] Finally, the number M of overlapping fault frequencies is compared with the total number of theoretical fault frequencies. The ratio, i.e. The overlap between the theoretical fault frequency and the main noise frequency was determined. Among them, the total number of theoretical fault frequencies In this embodiment, the value is fixed at 3 (corresponding to three fault types: inner ring, outer ring, and rolling element). Overlap degree The value range is from 0 to 1.

[0061] S105. Perform Fourier transform on the vibration signal and divide it into frequency bands, calculate the frequency domain energy entropy, and determine the noise type based on the frequency domain energy entropy.

[0062] In this embodiment, the vibration signal is subjected to Fourier transform and divided into frequency bands, the frequency domain energy entropy is calculated, and the noise type is determined based on the frequency domain energy entropy. Specifically, this includes: Perform a Fourier transform on the vibration signal to obtain a frequency domain signal, and divide the frequency range of the frequency domain signal into at least three equal-width frequency bands; For each equal-width frequency band, calculate the proportion of energy within that band to the total energy of the entire frequency domain signal. Multiply the logarithm of this proportion (base 2) by the proportion itself to determine the energy entropy contribution value of that band. The frequency domain energy entropy is obtained by summing the energy entropy contributions of all equal-width frequency bands and taking the negative of the sum. The frequency domain energy entropy is compared with a preset entropy threshold. If the frequency domain energy entropy is less than the preset entropy threshold, the noise type is determined to be steady-state noise; otherwise, the noise type is determined to be transient noise.

[0063] For example, the vibration signal is subjected to Fourier transform and divided into frequency bands, the frequency domain energy entropy is calculated, and the noise type is determined based on the frequency domain energy entropy. The purpose is to quantify the uniformity of the distribution of the vibration signal spectrum energy, thereby distinguishing whether the current environmental noise belongs to steady-state noise or transient noise, and providing a basis for subsequent adaptive selection of wavelet basis.

[0064] The specific process is as follows: First, a Fourier transform is performed on the vibration signal to convert it from a time-domain signal to a frequency-domain signal. After the Fourier transform, the horizontal axis of the frequency-domain signal represents frequency (in Hz), and the vertical axis represents amplitude (in g). The amplitude is then squared to obtain the energy value of each frequency component.

[0065] Then, the frequency range of the frequency domain signal is divided into at least three equal-width frequency bands. Equal-width frequency bands mean that each band has the same frequency width, i.e., bandwidth = maximum frequency range / number of bands. The purpose of dividing the spectrum into equal-width bands is to discretize the continuous spectrum into several frequency band units, facilitating subsequent calculations of the energy distribution ratio across each band.

[0066] For example, in this embodiment, the frequency range (0Hz to 5000Hz) of the frequency domain signal is divided into 50 equal-width frequency bands, each with a width of 100Hz. Specifically, the first frequency band is 0-100Hz, the second frequency band is 100-200Hz, and so on, with the 50th frequency band being 4900-5000Hz.

[0067] For each equal-width frequency band, calculate the proportion of energy within that band to the total energy of the entire frequency domain signal. The specific calculation method is as follows: First, sum the energy of all frequency components within the band to obtain the band energy. Then, the frequency band energy of all frequency bands is accumulated to obtain the total energy of the entire frequency domain signal; finally, the frequency band energy is... Divide by the total energy to obtain the energy percentage of that frequency band. .

[0068] The calculation of frequency domain energy entropy is based on the concept of information entropy. Information entropy is used to measure the degree of disorder or uncertainty of a system: the entropy value is the largest when energy is evenly distributed across all frequency bands; the entropy value is the smallest when energy is concentrated in a few frequency bands.

[0069] Regarding the first Each equal-width frequency band, the energy proportion of that frequency band Logarithm base 2 With this energy ratio The product of, i.e. This is determined as the energy entropy contribution value for that bandwidth. This product is usually negative because... exist When the time is negative.

[0070] Then, the energy entropy contributions of all equal-width frequency bands are summed to obtain... The sum is either negative or zero. Finally, the opposite of the sum is taken, i.e., the sum is subtracted, to obtain the frequency domain energy entropy H.

[0071] Frequency domain energy entropy The calculation formula can be: ; Where B represents the number of equal-width frequency bands; Indicates the first Energy percentage of each equal bandwidth frequency band.

[0072] The frequency domain energy entropy H ranges from 0 to... The smaller the H value, the more concentrated the energy distribution and the simpler the spectral structure; the larger the H value, the more dispersed the energy distribution and the more complex the spectral structure.

[0073] After calculating the frequency domain energy entropy H, the frequency domain energy entropy H is compared with a preset entropy threshold, and the noise type corresponding to the current vibration signal is determined based on the comparison result.

[0074] The judgment rules are as follows: If the frequency domain energy entropy H is less than a preset entropy threshold, the noise type is determined to be steady-state noise. The characteristics of steady-state noise are: spectral energy concentrated in a few fixed frequency bands, relatively uniform energy distribution, and a stable spectral structure. Typical steady-state noises include: broadband noise generated by continuous fan operation, electromagnetic noise from motors, and steady-state flow noise in pipelines.

[0075] If the frequency domain energy entropy H is greater than or equal to a preset entropy threshold, the noise type is determined to be transient noise. Transient noise is characterized by its spectral energy being dispersed across multiple frequency bands, exhibiting a chaotic energy distribution, complex spectral structure, and impactfulness. Typical transient noises include: impact noise generated during equipment start-up and shutdown, pulse noise generated by water hammer in pipelines, and transient vibrations caused by external impacts.

[0076] S106. Based on the noise type and overlap, determine the wavelet basis and the number of decomposition layers, decompose the vibration signal, and obtain the fault-sensitive characteristic signal.

[0077] In this embodiment, the wavelet basis and decomposition level are determined based on the noise type and overlap, respectively, and the vibration signal is decomposed to obtain fault-sensitive feature signals, specifically including: Perform a Fourier transform on the vibration signal to obtain the frequency domain signal, and calculate the total energy across the entire frequency range in the frequency domain signal; Determine the total fault band energy corresponding to the fault characteristic band for each theoretical fault frequency; The ratio of the total energy of the fault frequency band corresponding to each theoretical fault frequency to the total energy of the vibration signal is determined as the energy proportion of the fault characteristic frequency band. When the noise type is steady-state noise, the db4 wavelet is selected as the first wavelet basis; when the noise type is transient noise, the sym5 wavelet is selected as the second wavelet basis. When the overlap is less than the preset overlap threshold, the first decomposition layer is selected; otherwise, the second decomposition layer is selected, wherein the first decomposition layer is greater than the second decomposition layer. The vibration signal was decomposed using a defined wavelet basis and a number of decomposition levels to obtain wavelet coefficients in different frequency ranges. The wavelet coefficients of the main noise frequency are set to zero, while other wavelet coefficients are retained. The reconstructed signal is obtained through wavelet reconstruction. Calculate the energy percentage of each fault characteristic frequency band in the reconstructed signal. If the energy percentage after reconstruction is lower than the energy percentage before reconstruction, adjust the number of decomposition layers and re-decompose and reconstruct until the energy percentage after reconstruction is not lower than the energy percentage before reconstruction, and obtain the fault sensitive characteristic signal.

[0078] Determine the total fault band energy corresponding to the fault characteristic band for each theoretical fault frequency, specifically including: For each theoretical fault frequency, the difference between the theoretical fault frequency and the preset offset is determined as the lower limit of the frequency band, the sum of the theoretical fault frequency and the preset offset is determined as the upper limit of the frequency band, and the frequency band interval formed by the lower limit and the upper limit of the frequency band is determined as the fault characteristic frequency band corresponding to the theoretical fault frequency. Extract the components of all frequencies within the fault characteristic frequency band and sum them up to obtain the total fault frequency band energy corresponding to the theoretical fault frequency.

[0079] For example, the wavelet basis and decomposition level are determined based on the noise type and overlap, respectively, and the vibration signal is decomposed to obtain fault-sensitive feature signals. The purpose of this step is to adaptively select the optimal wavelet basis and decomposition level according to the main noise frequency, noise type, and overlap, and to perform wavelet decomposition and reconstruction on the original vibration signal. This removes environmental noise while preserving complete fault features, including early weak faults, providing high-quality fault-sensitive feature signals for subsequent model training.

[0080] The specific process is as follows: Before performing wavelet decomposition, it is necessary to first calculate the total energy of the vibration signal and the energy proportion of each fault characteristic frequency band, so as to compare the energy after subsequent filtering and determine whether the filtering is excessive.

[0081] Perform a Fourier transform on the vibration signal to obtain the frequency domain signal. Then, sum the energies of all frequency components in the frequency domain signal to obtain the total energy of the entire vibration signal across the entire frequency range. .

[0082] After obtaining the theoretical fault frequencies, it is necessary to determine the fault characteristic frequency band corresponding to each theoretical fault frequency and calculate the total energy within that frequency band. The specific process is as follows: First, for each theoretical failure frequency (including the theoretical failure frequency of the inner race, the theoretical failure frequency of the outer race, and the theoretical failure frequency of the rolling element), determine its corresponding fault characteristic frequency band range: Theoretical fault frequency and preset offset The difference is determined as the lower limit of the frequency band; the theoretical fault frequency is compared with the preset offset. The sum of these values ​​is determined as the upper limit of the frequency band; the closed interval formed by the lower limit and the upper limit of the frequency band is... This was determined to be the fault characteristic frequency band corresponding to the theoretical fault frequency.

[0083] For example, preset offset The value is set to 50 Hz. If the theoretical fault frequency of the inner ring... Therefore, its fault characteristic frequency band is 200Hz to 300Hz. The rationale for setting this frequency band range is that in actual vibration signals, the fault characteristic frequency is not absolutely stable and will drift within a certain range due to factors such as rotational speed fluctuations and measurement errors. Setting an offset of ±50Hz can ensure that the fault energy is completely captured.

[0084] It should be noted that the above preset offset This is a pre-set frequency tolerance parameter. The setting of this parameter is based on the following considerations: In actual operation, the compressor speed is not absolutely constant and exhibits slight fluctuations; simultaneously, the measurement accuracy of vibration sensors is limited, and Fourier transforms also have limitations in frequency resolution. These factors lead to a certain deviation between the actual fault characteristic frequency and the theoretically calculated value. If the fault characteristic frequency band is set too narrowly, the true fault energy may be missed due to frequency drift; if it is set too wide, too many irrelevant frequency components will be introduced, reducing the signal-to-noise ratio.

[0085] For example, in this embodiment, a preset offset is used. The value is set at 50 Hz. This value is based on the following: compressor speed fluctuations are typically controlled within ±5%. Taking 3000 rpm as an example, the fault frequency deviation caused by speed fluctuations is approximately ±10 Hz; adding measurement errors and frequency resolution errors (usually ±20 Hz), the total deviation generally does not exceed ±50 Hz. Therefore, This ensures that fault energy is completely captured while avoiding the introduction of excessive noise components. It should be noted that this value is for illustrative purposes only and can be adjusted according to the specific compressor model and operating conditions in actual applications.

[0086] Then, the vibration signal is transformed in the frequency domain to obtain the frequency domain signal. In the frequency domain signal, the energy values ​​of all frequency components within the aforementioned fault characteristic frequency band are extracted. These energy values ​​are then summed to obtain the total fault frequency band energy corresponding to the theoretical fault frequency. .

[0087] by =250Hz Taking 50Hz as an example, the fault characteristic frequency band is 200Hz to 300Hz. Extract the energy values ​​corresponding to all frequency components within this frequency band (e.g., 205Hz, 210Hz, ..., 295Hz, etc.), and sum them to obtain the total fault frequency band energy. .

[0088] The total energy of the fault band corresponding to each theoretical fault frequency Total energy of vibration signal To perform ratio calculations, that is... Divide by The calculation results are determined as the energy proportion of the fault characteristic frequency band. Energy percentage This reflects the significance of fault characteristics in the overall vibration signal: The larger the value, the more obvious the fault characteristics; The smaller the value, the weaker the fault characteristics.

[0089] Adaptive selection of wavelet basis based on noise type: When the noise type is steady-state noise, the db4 wavelet is selected as the wavelet basis. The db4 wavelet (Daubechies 4th order wavelet) has good regularity and tight support properties, making it suitable for analyzing stationary signals and capable of effectively handling fault feature extraction under steady-state noise background.

[0090] When the noise type is transient noise, the sym5 wavelet is selected as the wavelet basis. The sym5 wavelet (Symlets 5th order wavelet) has approximately symmetric properties, is suitable for analyzing non-stationary signals, and can effectively handle fault feature extraction in the context of transient impulse noise.

[0091] The wavelet decomposition level is adaptively selected based on the overlap C: Compare the overlap C with the preset overlap threshold A comparison is made. When the overlap C is less than the preset overlap threshold... When the overlap between the theoretical fault frequency and the main noise frequency is low, the noise masks the fault characteristics less. In this case, choosing a smaller number of decomposition layers (the first decomposition layer) can meet the noise reduction requirements while taking into account computational efficiency.

[0092] When the overlap C is greater than or equal to the preset overlap threshold This indicates a high degree of overlap between the theoretical fault frequency and the main noise frequency. Some fault frequencies and noise frequencies are too close in the frequency domain, requiring a larger number of decomposition layers (the second decomposition layer) for more precise frequency separation to effectively separate the fault signal from the noise signal. The value of the second decomposition layer is greater than the value of the first decomposition layer.

[0093] Using the wavelet basis and decomposition level determined above, wavelet decomposition is performed on the original vibration signal. Wavelet decomposition separates the vibration signal layer by layer according to frequency from high to low, obtaining wavelet coefficients for different frequency ranges. Each wavelet coefficient corresponds to a signal component within a specific frequency range.

[0094] The wavelet coefficients corresponding to the extracted main noise frequencies are set to zero. This means that the frequency components with the strongest noise energy are directly removed from the signal, thus achieving targeted denoising. Wavelet coefficients in other frequency ranges (including coefficients corresponding to fault characteristic frequencies) are fully preserved.

[0095] When the frequency overlap is high (i.e., the theoretical fault frequency is very close to the main noise frequency), a selective attenuation strategy can be adopted to avoid filtering out weak fault signals that overlap with the noise frequency: only 80% of the wavelet coefficients of the main noise frequency are set to zero, while 20% of the wavelet coefficients are retained. This soft processing method can achieve a balance between noise reduction and fidelity preservation.

[0096] All processed wavelet coefficients (coefficients after noise bands are set to zero or attenuated and coefficients retained in other frequency bands) are reconstructed using wavelet reassembly, and the signals of each frequency band are resynthesized to obtain a preliminarily purified reconstructed signal.

[0097] Calculate the energy percentage of each fault characteristic frequency band in the reconstructed signal. The calculation method is the same as the first step: perform a Fourier transform on the reconstructed signal, extract the energy within the fault characteristic frequency band, divide it by the total energy of the reconstructed signal, and obtain the energy ratio after reconstruction.

[0098] The reconstructed energy proportion Compared with the original energy ratio before reconstruction Comparison: If the energy percentage after reconstruction is not lower than the energy percentage before reconstruction, it indicates that the filtering process has not caused any additional loss of fault characteristic energy, the filtering parameters are set reasonably, and the current reconstructed signal is the final fault-sensitive characteristic signal.

[0099] If the energy percentage after reconstruction is lower than that before reconstruction, it indicates that the filtering process has caused attenuation of the fault characteristics, possibly due to over-filtering. In this case, the decomposition parameters need to be adjusted before re-decomposition and reconstruction. Specific adjustment methods include: reducing the proportion of noise figures set to zero (e.g., from 100% to 70%), reducing the number of decomposition levels, or changing the wavelet basis. After adjustment, repeat the above steps of wavelet decomposition, zeroing coefficients, reconstruction, and energy comparison until the energy percentage after reconstruction is not lower than that before reconstruction, thus obtaining the final fault-sensitive characteristic signal.

[0100] S107. Generate simulated noise samples based on noise type and noise intensity, superimpose the simulated noise samples onto the fault-sensitive feature signal to generate enhanced samples, and use the enhanced samples to train the random forest model to obtain a noise-resistant fault classification model.

[0101] In this embodiment, the simulated noise samples include steady-state simulated noise samples and transient simulated noise samples. The simulated noise samples are generated based on the noise type and noise intensity, specifically including: When the noise type is steady-state noise, Gaussian white noise is generated, and a sine wave corresponding to the main noise frequency is superimposed on the Gaussian white noise. The intensity of the superimposed noise is adjusted to match the noise intensity to obtain a steady-state simulated noise sample. When the noise type is transient noise, an impulse pulse signal with an amplitude matching the noise intensity and a randomly set pulse interval is generated to obtain a transient simulated noise sample.

[0102] Simulated noise samples are superimposed onto fault-sensitive feature signals to generate enhanced samples, specifically including: The simulated noise samples are superimposed on the fault-sensitive feature signals at at least two preset intensity multiples to obtain the initial enhanced samples; Each initial enhanced sample is labeled with its corresponding fault type to obtain the enhanced sample.

[0103] For example, simulated noise samples are generated based on noise type and noise intensity. These simulated noise samples are then superimposed on fault-sensitive feature signals to generate enhanced samples. The enhanced samples are then used to train a random forest model, resulting in a noise-resistant fault classification model. The purpose of this step is to use data augmentation techniques to expose the model to various possible noise scenarios during the training phase, thereby improving the model's anti-interference ability and generalization performance, enabling it to adapt to the complex and varied environmental noise in real-world applications.

[0104] The simulated noise samples include two types: steady-state simulated noise samples and transient simulated noise samples, which correspond to two different noise types.

[0105] When the noise type is steady-state noise, a steady-state simulated noise sample is generated. The specific generation method is as follows: First, Gaussian white noise is generated. Gaussian white noise is a type of random noise with a power spectral density uniformly distributed across the entire frequency range. Its time-domain amplitude follows a Gaussian distribution (normal distribution), with a mean of zero and a variance related to the noise intensity. Gaussian white noise is used to simulate broadband continuous noise, such as the aerodynamic noise generated by the continuous operation of a fan or the electromagnetic noise of a motor.

[0106] Then, a sine wave corresponding to the main noise frequency is superimposed onto the Gaussian white noise. The top K frequencies with the highest amplitudes in the main noise frequency range represent the most energetic frequency components in the environmental noise. The purpose of superimposing the sine wave is to simulate narrowband periodic noise, such as single-frequency noise generated by pipe resonance or harmonic noise generated by unbalanced rotation of equipment. The frequency of the sine wave is taken as the main noise frequency, and the initial amplitude is taken as unit amplitude.

[0107] Finally, the intensity of the superimposed analog noise signal is adjusted to match the noise intensity Z. Intensity adjustment means calculating the current root mean square (RMS) value of the superimposed signal, and then multiplying the entire signal by a scaling factor to make its RMS value equal to Z. After this processing, a steady-state analog noise sample is obtained.

[0108] When the noise type is transient noise, a transient simulated noise sample is generated. The specific generation method is as follows: An impact pulse signal is generated, where the amplitude of each pulse is matched to the noise intensity Z, and the pulse interval is randomly set. The impact pulse signal is used to simulate transient impact noise, such as the impact generated by equipment start-up and shutdown, the pulse generated by water hammer effect in pipelines, and transient vibrations generated by external impacts.

[0109] Specifically, a zero signal with a duration equal to the sampling duration (e.g., 10 seconds) is first generated. Then, pulse signals with an amplitude of Z are inserted at random time points (e.g., 1.2 seconds, 3.5 seconds, 5.8 seconds, 7.1 seconds, and 9.4 seconds). The duration of each pulse is set to a small fixed value (e.g., 0.01 seconds), and the pulse shape can be a rectangular pulse or an exponentially decaying pulse. The pulse interval is randomly set to a random number between 1 and 5 seconds to simulate irregular transient impacts in a real environment. After the above processing, a transient simulated noise sample is obtained.

[0110] After generating simulated noise samples, they need to be superimposed on the fault-sensitive feature signal to generate enhanced samples. Specifically: Simulated noise samples are superimposed on fault-sensitive feature signals at at least two preset intensity multiples. Superposition means multiplying the simulated noise sample by an intensity multiple and then adding it to the fault-sensitive feature signal to obtain a noisy fault feature signal. The purpose of setting the preset intensity multiples is to cover scenarios with different noise intensities, enabling the model to adapt to dynamic changes in noise intensity during the training phase.

[0111] In this embodiment, the preset intensity multipliers include 0.5x, 1x, and 1.5x. For each original fault-sensitive feature signal, the following superposition operation is performed: Superimposed simulated noise at 0.5 times the intensity: The simulated noise sample is multiplied by 0.5 and added to the fault-sensitive feature signal to obtain an enhanced sample with low noise intensity, simulating a relatively quiet operating environment.

[0112] Superimposed simulated noise of 1x intensity: The simulated noise sample is multiplied by 1 and then added to the fault-sensitive feature signal to obtain an enhanced sample with medium noise intensity, simulating the current actual noise level of the operating scenario.

[0113] Superimposed simulated noise of 1.5 times intensity: The simulated noise sample is multiplied by 1.5 and added to the fault-sensitive feature signal to obtain an enhanced sample with high noise intensity, simulating extreme scenarios with severe noise.

[0114] Furthermore, hybrid noise samples can be generated: simultaneously superimposing steady-state and transient simulated noise to cover more complex real-world noise scenarios. For example, by superimposing transient impacts on steady-state noise, a complex environment can be simulated where occasional pipe resonance impacts occur within the continuous noise of a wind turbine motor.

[0115] Each initial augmented sample is labeled with its corresponding fault type, resulting in augmented samples that can be used for supervised learning. The labeling means assigning a label to each sample, indicating the fault category it corresponds to.

[0116] Fault types include, but are not limited to: normal operation, inner ring wear, outer ring wear, and rolling element wear. Other possible fault types include rotor eccentricity, valve plate damage, and refrigerant leakage, which can be expanded according to the actual application scenario.

[0117] The enhanced samples generated above are used to train a random forest model, resulting in a noise-resistant fault classification model. Random forest is an ensemble learning algorithm that classifies faults by constructing multiple decision trees and combining their voting results.

[0118] The augmented samples are divided into training and validation sets according to a preset ratio. The training set is used to train the model parameters, and the validation set is used to evaluate the model's classification performance.

[0119] Key parameters of the random forest model include the number of decision trees, maximum depth, and random seed.

[0120] The number of decision trees represents the total number of decision trees in the random forest. Too few decision trees will result in insufficient model accuracy, while too many decision trees will significantly increase training time and lead to diminishing marginal returns. For example, in this embodiment, the number of decision trees is set to 100.

[0121] The maximum depth represents the maximum number of layers in each decision tree. Excessive depth can lead to overfitting (the model remembers noise from the training set instead of learning patterns), while insufficient depth results in inadequate model expressive power. For example, in this embodiment, the maximum depth is set to 10.

[0122] The random seed is used to control the generation of random numbers during model training, ensuring that the training results are reproducible. For example, in this embodiment, the random seed is set to 42.

[0123] The model is trained using the sklearn.ensemble.RandomForestClassifier utility library in Python, with augmented samples from the training set. During training, the model automatically learns the mapping relationship between different fault types and fault-sensitive features.

[0124] The classification performance of the trained random forest model is evaluated using a validation set. Evaluation metrics include: classification accuracy (the proportion of correctly classified samples out of the total number of samples), precision, recall, and F1 score.

[0125] Specifically, it is necessary to record the accuracy reduction rate under different noise intensities. The specific method is to: separately calculate the classification accuracy under no noise, low noise (0.5 times the intensity), medium noise (1 times the intensity), and high noise (1.5 times the intensity), and calculate the percentage decrease in accuracy for each noise scenario relative to the no noise scenario.

[0126] After training, a noise-resistant fault classification model is obtained. This model can adapt to fault identification tasks in different noise scenarios and maintain a high classification accuracy even in complex environmental noise backgrounds.

[0127] S108. Input the sensitive feature signal of the fault to be analyzed corresponding to the vibration signal of the air conditioner compressor to be analyzed into the noise-resistant fault classification model, and output the fault analysis result of the air conditioner compressor to be analyzed through the noise-resistant fault classification model.

[0128] In this embodiment, the vibration signal of the air conditioner compressor to be analyzed, processed in the same way to obtain the sensitive feature signal of the fault to be analyzed, is input into the trained noise-resistant fault classification model. The noise-resistant fault classification model then outputs the fault analysis result corresponding to the air conditioner compressor to be analyzed. The purpose of this step is to apply the trained model to a real-world scenario to diagnose faults in the real-time collected compressor operating data, while combining a confidence assessment mechanism to ensure the reliability of the classification results and avoid misjudgments or omissions caused by noise interference.

[0129] The specific process is as follows: First, following the same method, the vibration signal to be analyzed, the speed data to be analyzed, and the ambient temperature data to be analyzed during the operation of the air conditioner compressor are collected, along with the noise vibration signal to be analyzed. The acquisition parameters are consistent with those of S101: a piezoelectric vibration sensor is used to collect vibration signals at a sampling frequency of 10kHz, with 10 seconds of data collected per batch; a Hall effect speed sensor is used to collect real-time speed data; and a thermocouple temperature sensor is used to collect ambient temperature data.

[0130] Then, the collected data to be analyzed is processed in the same way as S102 to S106 to obtain the sensitive feature signal of the fault to be analyzed, which can be directly input into the noise-resistant fault classification model for identification.

[0131] The fault-sensitive feature signals obtained above are input into the noise-resistant fault classification model that has been trained.

[0132] The noise-resistant fault classification model consists of 100 decision trees, each of which outputs a fault type prediction based on the input feature signal. The model then votes on the predictions from all decision trees, and the fault type with the most votes is output as the final classification result.

[0133] The fault classification results output by the model include, but are not limited to: normal, inner ring wear, outer ring wear, rolling element wear, rotor eccentricity, valve plate damage, refrigerant leakage, etc. The specific category of the output result depends on the fault type labeled in the training samples.

[0134] In practical applications, if the noise intensity of the current signal differs significantly from that during training, the model's classification accuracy will decrease accordingly, thus reducing the reliability of the classification results. To ensure the reliability of the output results, this embodiment introduces a confidence assessment and resampling mechanism.

[0135] First, calculate the signal-to-noise ratio of the vibration signal to be analyzed. The calculation formula can be: ; in, This indicates that the vibration data of the compressor being analyzed for 1 second when it is not running is taken as the power corresponding to the pure noise segment being analyzed. This indicates that the vibration data segment in the vibration signal to be analyzed before the current time is taken as the power corresponding to the fault signal to be analyzed. This represents the logarithmic function with base 10, that is... .

[0136] Signal-to-noise ratio The larger the value, the cleaner the signal; the smaller the value, the more severe the noise pollution.

[0137] During the model evaluation phase, the accuracy degradation rate (Dr) of the model is recorded under different signal-to-noise ratio (SNR) intervals. For example: When the signal-to-noise ratio (SNR) S ≥ 20 dB, the accuracy decrease rate Dr = 0%; when the SNR is 15 dB ≤ S < 20 dB, the accuracy decrease rate Dr = 1%; when the SNR is 10 dB ≤ S < 15 dB, the accuracy decrease rate Dr = 3%; when the SNR is 5 dB ≤ S < 10 dB, the accuracy decrease rate Dr = 8%; when the SNR is 0 dB ≤ S < 5 dB, the accuracy decrease rate Dr = 15%; and when the SNR is S < 0 dB, the accuracy decrease rate Dr = 25%.

[0138] Based on the currently calculated signal-to-noise ratio Find the corresponding accuracy decline rate. .

[0139] Calculate confidence level The calculation formula is: ; Set confidence threshold This is used to determine whether the current classification result is reliable. For example, in this embodiment, the confidence threshold... The value is 0.9, which means that the reliability of the classification result is required to be no less than 90%.

[0140] The calculated confidence level (Conf) is compared with the confidence threshold: If Conf≥ (i.e., Conf≥0.9) indicates that the confidence level of the current classification result meets the requirements, the result is reliable, and the fault classification result is directly output.

[0141] If Conf< If Conf < 0.9, it indicates that the current signal noise is too high, the reliability of the classification result is insufficient, and the resampling mechanism is triggered. The specific operation of the resampling mechanism is as follows: the acquisition time is extended to 20 seconds (originally 10 seconds), the vibration signal to be analyzed is reacquired, and the amount of data is increased to improve the analysis accuracy. After reacquisition, all the steps of feature purification, classification, and confidence assessment are repeated until the confidence level meets the requirements and the classification result is output.

[0142] After the above confidence assessment and resampling mechanism, the final output is the fault analysis result corresponding to the air conditioner compressor to be analyzed. The output results include: fault type (such as "inner ring wear", "outer ring wear" and "normal", etc.), confidence value (such as 0.97), and suggested handling measures (such as "suggest arranging maintenance in the near future", "suggesting emergency shutdown", etc., which can be configured according to actual needs).

[0143] In summary, this invention purifies vibration signals by quantifying environmental noise intensity, frequency distribution, and spectral energy entropy, adaptively selecting wavelet bases and decomposition levels, and introducing a feedback adjustment mechanism based on energy proportion. This effectively avoids the loss of weak fault features caused by traditional fixed-parameter filtering, especially preserving early weak fault features to reduce the false negative rate. Simultaneously, it generates simulated noise samples based on noise characteristics for data augmentation, significantly improving the generalization ability and classification accuracy of the random forest model under different noise environments. Furthermore, it introduces a confidence assessment and resampling mechanism based on signal-to-noise ratio in the real-time classification stage to ensure the reliability of the output results.

[0144] This invention also proposes a machine learning-based fault classification system for air conditioner compressors. Please refer to [link / reference]. Figure 2 The diagram shows a structural diagram of a machine learning-based air conditioner compressor fault classification system provided in an embodiment of the present invention. The system includes: a data acquisition module 101, a data analysis module 102, and a fault output module 103.

[0145] Data acquisition module 101 is used to acquire vibration signals and speed data of the air conditioner compressor during operation, and at the same time acquire noise vibration signals; Data analysis module 102 is used to calculate noise intensity based on noise vibration signals; The theoretical failure frequency of the air conditioner compressor is calculated based on its factory parameters and speed data. Extract the main noise frequency from the noise vibration signal and calculate the overlap between the theoretical fault frequency and the main noise frequency; Perform Fourier transform on the vibration signal and divide it into frequency bands, calculate the frequency domain energy entropy, and determine the noise type based on the frequency domain energy entropy; Based on the noise type and overlap, the wavelet basis and the number of decomposition layers are determined respectively, and the vibration signal is decomposed to obtain the fault-sensitive characteristic signal. Simulated noise samples are generated based on noise type and noise intensity. These simulated noise samples are then superimposed onto fault-sensitive feature signals to generate enhanced samples. Finally, the enhanced samples are used to train a random forest model to obtain a noise-resistant fault classification model. The fault output module 103 is used to input the sensitive feature signal of the fault to be analyzed corresponding to the vibration signal of the air conditioner compressor to be analyzed into the noise-resistant fault classification model, and output the fault analysis result of the air conditioner compressor to be analyzed through the noise-resistant fault classification model.

[0146] It should be noted that the system provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the computer device can be divided into different functional modules to complete all or part of the functions described above. In addition, the air conditioner compressor fault classification system based on machine learning and the air conditioner compressor fault classification method based on machine learning provided in the above embodiments belong to the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0147] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0148] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0149] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A machine learning-based method for classifying faults in air conditioner compressors, characterized in that, include: The system collects vibration signals and speed data of the air conditioner compressor during operation, and also collects noise and vibration signals. Calculate noise intensity based on noise vibration signals; The theoretical failure frequency of the air conditioner compressor is calculated based on its factory parameters and speed data. Extract the main noise frequency from the noise vibration signal and calculate the overlap between the theoretical fault frequency and the main noise frequency; Perform Fourier transform on the vibration signal and divide it into frequency bands, calculate the frequency domain energy entropy, and determine the noise type based on the frequency domain energy entropy; Based on the noise type and overlap, the wavelet basis and the number of decomposition layers are determined respectively, and the vibration signal is decomposed to obtain the fault-sensitive characteristic signal. Simulated noise samples are generated based on noise type and noise intensity. These simulated noise samples are then superimposed onto fault-sensitive feature signals to generate enhanced samples. Finally, the enhanced samples are used to train a random forest model to obtain a noise-resistant fault classification model. The sensitive feature signal of the fault to be analyzed corresponding to the vibration signal of the air conditioner compressor to be analyzed is input into the noise-resistant fault classification model, and the fault analysis result of the air conditioner compressor to be analyzed is output through the noise-resistant fault classification model.

2. The air conditioner compressor fault classification method based on machine learning according to claim 1, characterized in that, The calculation of noise intensity based on noise vibration signals specifically includes: The noise intensity is obtained by squaring each vibration amplitude in the noise vibration signal and summing the results, then taking the square root of the ratio of the sum to the number of data points in the noise vibration signal.

3. The air conditioner compressor fault classification method based on machine learning according to claim 1, characterized in that, The theoretical failure frequency includes the theoretical failure frequency of the compressor inner ring, the theoretical failure frequency of the compressor outer ring, and the theoretical failure frequency of the compressor rolling elements. The calculation of the theoretical failure frequency based on the air conditioner compressor's factory parameters and speed data specifically includes: Obtain the factory parameters of the air conditioner compressor, which include the number of rolling elements, pitch circle diameter, rolling element diameter, and contact angle of the compressor bearing. Half the number of rolling elements is defined as the number of half rolling elements, and the rotational speed data per second is used as the unit rotational speed; Calculate the ratio of the rolling element diameter to the pitch circle diameter, and multiply this ratio by the cosine of the contact angle to determine the contact proportionality coefficient; The theoretical failure frequency of the compressor inner ring is determined by multiplying the product of the number of semi-rolling elements and the unit speed, and then multiplying it by a factor plus the contact ratio coefficient. The theoretical failure frequency of the compressor outer ring is determined by multiplying the product of the number of semi-rolling elements and the unit speed, and then multiplying by the difference between the contact ratio coefficient and the product of the product of the number of semi-rolling elements and the unit speed. The theoretical failure frequency of the compressor rolling element is determined by multiplying the ratio of the pitch circle diameter to the rolling element diameter by the unit speed, and then multiplying the product by the square of the contact ratio coefficient.

4. The air conditioner compressor fault classification method based on machine learning according to claim 1, characterized in that, The extraction of the main noise frequency from the noise vibration signal and the calculation of the overlap between the theoretical fault frequency and the main noise frequency specifically include: Perform a Fourier transform on the noise vibration signal and determine the frequency with the highest amplitude ranking as the main noise frequency. For each theoretical fault frequency, calculate the frequency difference between the theoretical fault frequency and each main noise frequency, and determine the theoretical fault frequencies with frequency differences less than a preset error threshold as overlapping fault frequencies. The ratio of the number of overlapping fault frequencies to the total number of theoretical fault frequencies is determined as the overlap between the theoretical fault frequencies and the main noise frequencies.

5. The air conditioner compressor fault classification method based on machine learning according to claim 1, characterized in that, The process of performing Fourier transform on the vibration signal, dividing it into frequency bands, calculating the frequency domain energy entropy, and determining the noise type based on the frequency domain energy entropy specifically includes: Perform a Fourier transform on the vibration signal to obtain a frequency domain signal, and divide the frequency range of the frequency domain signal into at least three equal-width frequency bands; For each equal-width frequency band, calculate the proportion of energy within that band to the total energy of the entire frequency domain signal. Multiply the logarithm of this proportion (base 2) by the proportion itself to determine the energy entropy contribution value of that band. The frequency domain energy entropy is obtained by summing the energy entropy contributions of all equal-width frequency bands and taking the negative of the sum. The frequency domain energy entropy is compared with a preset entropy threshold. If the frequency domain energy entropy is less than the preset entropy threshold, the noise type is determined to be steady-state noise; otherwise, the noise type is determined to be transient noise.

6. The air conditioner compressor fault classification method based on machine learning according to claim 1, characterized in that, The process of determining the wavelet basis and decomposition level based on noise type and overlap, respectively, to decompose the vibration signal and obtain fault-sensitive feature signals specifically includes: Perform a Fourier transform on the vibration signal to obtain the frequency domain signal, and calculate the total energy across the entire frequency range in the frequency domain signal; Determine the total fault band energy corresponding to the fault characteristic band for each theoretical fault frequency; The ratio of the total energy of the fault frequency band corresponding to each theoretical fault frequency to the total energy of the vibration signal is determined as the energy proportion of the fault characteristic frequency band. When the noise type is steady-state noise, the db4 wavelet is selected as the first wavelet basis; when the noise type is transient noise, the sym5 wavelet is selected as the second wavelet basis. When the overlap is less than the preset overlap threshold, the first decomposition layer is selected; otherwise, the second decomposition layer is selected, wherein the first decomposition layer is greater than the second decomposition layer. The vibration signal was decomposed using a defined wavelet basis and a number of decomposition levels to obtain wavelet coefficients in different frequency ranges. The wavelet coefficients of the main noise frequency are set to zero, while other wavelet coefficients are retained. The reconstructed signal is obtained through wavelet reconstruction. Calculate the energy percentage of each fault characteristic frequency band in the reconstructed signal. If the energy percentage after reconstruction is lower than the energy percentage before reconstruction, adjust the number of decomposition layers and re-decompose and reconstruct until the energy percentage after reconstruction is not lower than the energy percentage before reconstruction, and obtain the fault sensitive characteristic signal.

7. The air conditioner compressor fault classification method based on machine learning according to claim 6, characterized in that, The determination of the total fault band energy corresponding to the fault characteristic band for each theoretical fault frequency specifically includes: For each theoretical fault frequency, the difference between the theoretical fault frequency and the preset offset is determined as the lower limit of the frequency band, the sum of the theoretical fault frequency and the preset offset is determined as the upper limit of the frequency band, and the frequency band interval formed by the lower limit and the upper limit of the frequency band is determined as the fault characteristic frequency band corresponding to the theoretical fault frequency. Extract the components of all frequencies within the fault characteristic frequency band and sum them up to obtain the total fault frequency band energy corresponding to the theoretical fault frequency.

8. The air conditioner compressor fault classification method based on machine learning according to claim 1, characterized in that, The simulated noise samples include steady-state simulated noise samples and transient simulated noise samples. The generation of simulated noise samples based on noise type and noise intensity specifically includes: When the noise type is steady-state noise, Gaussian white noise is generated, and a sine wave corresponding to the main noise frequency is superimposed on the Gaussian white noise. The intensity of the superimposed noise is adjusted to match the noise intensity to obtain a steady-state simulated noise sample. When the noise type is transient noise, an impulse pulse signal with an amplitude matching the noise intensity and a randomly set pulse interval is generated to obtain a transient simulated noise sample.

9. The air conditioner compressor fault classification method based on machine learning according to claim 1, characterized in that, The step of superimposing simulated noise samples onto the fault-sensitive feature signal to generate enhanced samples specifically includes: The simulated noise samples are superimposed on the fault-sensitive feature signals at at least two preset intensity multiples to obtain the initial enhanced samples; Each initial enhanced sample is labeled with its corresponding fault type to obtain the enhanced sample.

10. A machine learning-based fault classification system for air conditioner compressors, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed by the processor, it implements the steps of the machine learning-based air conditioning compressor fault classification method as described in any one of claims 1-9.

Images