Zero sample speed reducer composite fault diagnosis method based on fast spectral kurtosis

Abstract

The invention relates to the technical field of fault diagnosis, in particular to a zero-sample speed reducer composite fault diagnosis method based on fast spectral kurtosis, which comprises the following steps: arranging multiple types of sensors at the output and input ends of a fault speed reducer, and collecting data according to a preset sampling frequency; the collected real-time signals are processed into one-dimensional time domain signals, the one-dimensional time domain signals are converted into an actual spectral kurtosis graph after accuracy and integrity analysis and adjustment, a single fault feature set is constructed by combining real-time temperature values and pressure values, and single fault categories are obtained through classification by using a clustering algorithm; dividing samples under a single fault category to construct a diagnosis model, performing combined simulation on the single fault category to construct a composite fault sample set and labeling, performing preliminary diagnosis by the diagnosis model to obtain an initial result, finally determining a final diagnosis result according to the initial result and a composite fault inference rule, evaluating a classification recognition effect, and adjusting a preset sampling frequency according to the final diagnosis result. Therefore, the accuracy of a fault diagnosis result is improved.

Description

Technical Field

[0001] The invention relates to the technical field of fault diagnosis, and in particular to a zero-sample reducer composite fault diagnosis method based on fast spectral kurtosis. Background Art

[0002] Gearboxes are widely used in various chemical machinery production fields. They are indispensable in automobiles, aviation, machinery, energy, etc., and play an important role in mechanical equipment. Gearboxes are the main components in mechanical transmission systems. In actual environments, there are often too few samples of compound faults and multiple fault sources in compound faults have fault characteristics that interfere with or mask each other, making it difficult to detect compound faults in gearboxes. Therefore, it is necessary to find an accurate and effective method for diagnosing compound faults of gearboxes when compound fault samples are scarce.

[0003] The patent document with Chinese patent publication number CN117949209A discloses a zero-sample rolling bearing composite fault diagnosis system and method based on envelope spectrum, which belongs to the technical field of rolling bearing fault diagnosis. First, the vibration signal of the rolling bearing is processed to obtain the envelope spectrum, and then a feature extractor based on a convolutional neural network is used to extract fault-related features from the one-dimensional signal. Subsequently, a fault semantic embedding module based on prior knowledge is used to embed the semantic vector of the composite fault into the semantic space as the center of each category. The Euclidean distance is used to compare the similarity between the composite fault feature and the center of each category, and the unknown composite fault is identified by calculating the similarity.

[0004] The prior art only relies on the envelope spectrum to process the rolling bearing vibration signal. For some early weak faults or fault characteristics that are not obvious in the envelope spectrum, important information may be missed, resulting in a decrease in the accuracy of fault diagnosis. Summary of the invention

[0005] To this end, the present invention provides a zero-sample reducer composite fault diagnosis method based on fast spectral kurtosis, which can solve the problem of inaccurate composite fault classification.

[0006] To achieve the above object, the present invention provides a zero-sample reducer composite fault diagnosis method based on fast spectral kurtosis, the method comprising: Step S1, respectively setting a plurality of accelerometers, a plurality of temperature sensors and a plurality of pressure sensors at the output end and the input end of a plurality of faulty reduction gearboxes, collecting data based on a preset sampling frequency, and obtaining a plurality of real-time signals, a plurality of real-time temperature values ​​and a plurality of real-time pressure values; Step S2, processing the plurality of real-time signals to obtain a plurality of one-dimensional time domain signals, analyzing the accuracy and integrity of the plurality of one-dimensional time domain signals, adjusting the one-dimensional time domain signals according to the analysis results, and converting them into actual spectral kurtosis diagrams; Step S3, constructing a single fault feature set based on a number of one-dimensional time domain signals, a number of actual spectral kurtosis diagrams, a number of real-time temperature values, and a number of real-time pressure values, and classifying the number of single fault feature sets based on a clustering algorithm to obtain a number of single fault categories; Step S4, based on any single fault category, dividing the corresponding fault samples into a training set and a test set to construct a single fault diagnosis model; Step S5, combining and simulating several single fault categories to construct a composite fault sample set, setting several labels corresponding to several composite fault samples one by one, and performing preliminary diagnosis on the composite fault sample set according to the single fault diagnosis model to obtain the initial diagnosis result of the corresponding single fault feature sample in any composite fault sample; Step S6, determining the final diagnosis result of the composite fault according to the initial diagnosis result of each single fault feature sample in the composite fault sample and the preset composite fault inference rule, evaluating the composite fault classification and identification result, and adjusting the preset sampling frequency based on the evaluation result.

[0007] Furthermore, the step of processing the plurality of real-time signals to obtain a one-dimensional time domain signal comprises: Determine the frequency characteristics corresponding to any of the real-time signals, and determine the wavelet basis function and the number of wavelet decomposition layers based on the frequency characteristics; Performing wavelet decomposition processing on any of the real-time signals based on wavelet basis functions and wavelet decomposition layer numbers to obtain a number of wavelet coefficients, and processing the number of wavelet coefficients based on a preset threshold value; The processed wavelet coefficients are reconstructed to obtain a one-dimensional time domain signal.

[0008] Further, the steps of analyzing the accuracy and integrity of a plurality of one-dimensional time domain signals include: Comparing the one-dimensional time domain signals with a preset time domain signal range, determining the abnormal proportion of the abnormal one-dimensional time domain signals based on the comparison result, and determining the accuracy of the one-dimensional time domain signals based on the comparison result of the abnormal proportion and the preset proportion; The waveform integrity of a signal waveform diagram determined based on a plurality of one-dimensional time domain signals is determined, and the integrity of the plurality of one-dimensional time domain signals is determined based on the waveform integrity.

[0009] Furthermore, the step of constructing a single fault feature set based on a plurality of one-dimensional time domain signals, a plurality of actual spectral kurtosis diagrams, a plurality of real-time temperature values, and a plurality of real-time pressure values ​​includes: Calculating time domain characteristic parameters of a plurality of said one-dimensional time domain signals; Extracting key features in any of the actual spectral kurtosis graphs; Calculating actual temperature change rates corresponding to a number of the real-time temperature values, and performing statistics on actual temperature intervals corresponding to a number of the real-time temperature values; Calculating an actual pressure average and an actual pressure fluctuation amplitude corresponding to a plurality of the real-time pressure values; The various features calculated above are normalized and combined into a single fault feature vector.

[0010] Furthermore, the step of extracting key features from any of the actual spectral kurtosis graphs comprises: Setting a spectral kurtosis threshold, and taking an area in the actual spectral kurtosis graph that is larger than the spectral kurtosis threshold as the spectral kurtosis optimal frequency band; The actual spectral kurtosis mean, actual spectral kurtosis variance and actual frequency band width within the spectral kurtosis optimal frequency band are calculated as key features.

[0011] Furthermore, the step of classifying a plurality of single fault feature sets based on a clustering algorithm includes: Calculate the actual data density of any single fault feature sample and the remaining single fault feature samples based on the clustering algorithm; A single fault feature sample whose actual data density is greater than a set density threshold is used as a cluster consisting of a cluster core point and its neighborhood; For any boundary point, if it is in the neighborhood of a cluster core point, it will be divided into the corresponding cluster, otherwise, it will be marked as a noise point.

[0012] Furthermore, the steps of constructing a single fault diagnosis model include: Build a neural network model based on a deep learning framework; Train the neural network model based on the training set to obtain an initial model; The model performance of the initial model is tested based on the test set to determine a single fault diagnosis model based on the test results.

[0013] Furthermore, the steps of performing preliminary diagnosis on the composite fault sample set according to the single fault diagnosis model include: The single fault feature vector corresponding to each sample in the composite fault sample set is sequentially input into the trained single fault diagnosis model; The model calculates each input feature vector and outputs the actual probability value corresponding to each single fault category; The actual probability value is compared with the preset probability value to determine an initial diagnosis result based on the comparison result.

[0014] Further, the step of determining the final diagnosis result of the composite fault includes: Extract the association pattern between single fault feature samples and compound fault types based on several historical fault data, and determine the compound fault inference rule based on the association pattern; Match any single fault feature sample with the composite fault inference rule one by one to obtain a matching result; The final diagnosis result is determined based on the matching result and the initial diagnosis result.

[0015] Furthermore, the step of adjusting the preset sampling frequency based on the evaluation result includes: Identify the correct classification results in the compound fault classification recognition results, calculate the proportion of correct classification results, and obtain the classification accuracy; The classification accuracy is compared with the preset accuracy, and the preset sampling frequency is adjusted based on the comparison result.

[0016] Compared with the prior art, the beneficial effect of the present invention lies in that by arranging accelerometers, temperature sensors and pressure sensors at the output and input ends of the faulty reducer, the operating information of the reducer can be obtained from multiple physical dimensions. The accelerometer can capture vibration-related data to reflect the motion state and potential faults of mechanical parts; the temperature sensor monitors the operating temperature, which is helpful to detect faults caused by overheating, such as bearing wear, poor lubrication, etc.; the pressure sensor measures pressure changes, which provides a key basis for detecting sealing problems, abnormal oil pressure and other faults, and synchronously collects data based on a preset sampling frequency to ensure the integrity and consistency of data in the time series, providing a rich data foundation for comprehensive and accurate diagnosis of reducer faults, avoiding misdiagnosis or missed diagnosis due to the limitations of a single sensor data, converting the collected real-time signal into a one-dimensional time domain signal, and analyzing its accuracy and integrity, effectively removing noise interference, outliers and data missing in the signal, improving data quality, and converting it into an actual spectral kurtosis diagram can further Revealing the characteristics of the signal and quickly locating the fault frequency components contained in the signal are of great significance for detecting frequency information related to specific faults such as abnormal gear meshing frequency and bearing fault frequency, laying the foundation for subsequent accurate judgment of the fault type. A single fault feature set is constructed by integrating one-dimensional time domain signals, actual spectral kurtosis diagrams, real-time temperature values ​​and real-time pressure values. The fault feature information in multi-source data is fully integrated to make the feature representation more comprehensive and discriminative, providing an accurate classification basis for the subsequent construction of targeted single fault diagnosis models. Corresponding diagnosis models are constructed for different single fault categories to achieve accurate identification of various single faults, providing strong support for the judgment of single fault components in composite fault diagnosis, and comprehensively covering various composite fault situations that may occur in actual operation, so that the diagnosis system can have a deeper understanding and response capabilities for composite faults, comprehensively consider the correlation and mutual influence between multiple fault factors, avoid one-sided judgment based on single fault information, and further improve the accuracy and reliability of diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A schematic flow chart of a zero-sample reducer composite fault diagnosis method based on fast spectral kurtosis provided in an embodiment of the present invention; Figure 2 A simplified structural diagram of a reduction gearbox in a zero-sample reduction gearbox composite fault diagnosis method based on fast spectral kurtosis provided in an embodiment of the present invention; Figure 3 A schematic flow chart of analyzing the accuracy and integrity of a plurality of one-dimensional time domain signals in a zero-sample reducer composite fault diagnosis method based on fast spectral kurtosis is provided for an embodiment of the present invention; Figure 4 A schematic diagram of a process for constructing a single fault feature set in a zero-sample reducer composite fault diagnosis method based on fast spectral kurtosis provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to make the objects and advantages of the present invention more clearly understood, the present invention is further described below in conjunction with embodiments; it should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0019] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these embodiments are only used to explain the technical principles of the present invention and are not intended to limit the protection scope of the present invention.

[0020] It should be noted that, in the description of the present invention, terms such as "up", "down", "left", "right", "inside" and "outside" indicating directions or positional relationships are based on the directions or positional relationships shown in the drawings. This is merely for the convenience of description and does not indicate or imply that the device or element must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it should not be understood as a limitation on the present invention.

[0021] In addition, it should be noted that in the description of the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected" and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, or it can be the internal communication of two components. For those skilled in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0022] See also Figure 1 As shown, an embodiment of the present invention provides a zero-sample reducer composite fault diagnosis method based on fast spectral kurtosis, the method comprising: Step S1, respectively setting a plurality of accelerometers, a plurality of temperature sensors and a plurality of pressure sensors at the output end and the input end of a plurality of faulty reduction gearboxes, collecting data based on a preset sampling frequency, and obtaining a plurality of real-time signals, a plurality of real-time temperature values ​​and a plurality of real-time pressure values; Step S2, processing the plurality of real-time signals to obtain a plurality of one-dimensional time domain signals, analyzing the accuracy and integrity of the plurality of one-dimensional time domain signals, adjusting the one-dimensional time domain signals according to the analysis results, and converting them into actual spectral kurtosis diagrams; Step S3, constructing a single fault feature set based on a number of one-dimensional time domain signals, a number of actual spectral kurtosis diagrams, a number of real-time temperature values, and a number of real-time pressure values, and classifying the number of single fault feature sets based on a clustering algorithm to obtain a number of single fault categories; Step S4, based on any single fault category, dividing the corresponding fault samples into a training set and a test set to construct a single fault diagnosis model; Step S5, combining and simulating several single fault categories to construct a composite fault sample set, setting several labels corresponding to several composite fault samples one by one, and performing preliminary diagnosis on the composite fault sample set according to the single fault diagnosis model to obtain the initial diagnosis result of the corresponding single fault feature sample in any composite fault sample; Step S6, determining the final diagnosis result of the composite fault according to the initial diagnosis result of each single fault feature sample in the composite fault sample and the preset composite fault inference rule, evaluating the composite fault classification and identification result, and adjusting the preset sampling frequency based on the evaluation result.

[0023] Specifically, the embodiment of the present invention can obtain the operating information of the reducer from multiple physical dimensions by arranging accelerometers, temperature sensors and pressure sensors at the output and input ends of the faulty reducer. The accelerometer can capture vibration-related data to reflect the motion state and potential faults of mechanical parts; the temperature sensor monitors the operating temperature to help discover faults caused by overheating, such as bearing wear, poor lubrication, etc.; the pressure sensor measures pressure changes, providing a key basis for detecting sealing problems, abnormal oil pressure and other faults, and synchronously collects data based on a preset sampling frequency to ensure the integrity and consistency of data in the time series, providing a rich data foundation for comprehensive and accurate diagnosis of reducer faults, avoiding misdiagnosis or missed diagnosis due to the limitations of a single sensor data, converting the collected real-time signal into a one-dimensional time domain signal, and analyzing its accuracy and completeness, effectively removing noise interference, outliers, data missing and other problems in the signal, improving data quality, and converting it into an actual spectral kurtosis diagram can further reveal the signal in the frequency domain. The characteristics of the signal can quickly locate the fault frequency components contained in the signal, which is of great significance for detecting frequency information related to specific faults such as abnormal gear meshing frequency and bearing fault frequency, laying the foundation for the subsequent accurate judgment of the fault type. A single fault feature set is constructed by integrating the one-dimensional time domain signal, the actual spectral kurtosis diagram, the real-time temperature value and the real-time pressure value, which fully integrates the fault feature information in the multi-source data, making the feature representation more comprehensive and discriminative, providing an accurate classification basis for the subsequent construction of a targeted single fault diagnosis model, and building corresponding diagnosis models for different single fault categories to achieve accurate identification of various single faults, providing strong support for the judgment of single fault components in composite fault diagnosis, and comprehensively covering various composite fault situations that may occur in actual operation, so that the diagnosis system can have a deeper understanding and response capabilities for composite faults, comprehensively considering the correlation and mutual influence between multiple fault factors, avoiding one-sided judgment based on single fault information, and further improving the accuracy and reliability of diagnosis.

[0024] See also Figure 2 , which is a simplified structural diagram of a reduction box according to an embodiment of the present invention, in which BinA, BinB, BoutA, and BoutB are bearings of the reduction box, Gin and Gout are gears of the reduction box, and sensor A and sensor B are two accelerometers respectively arranged at the input end and output end of the reduction box body; It can be understood that the embodiments of the present invention are directed to this type of reducer, and all single fault types thereof include: Single bearing failure: The outer ring (BinA_OF), inner ring (BinA_IF) and roller (BinA_RF) of the bearing (BinA) at the input end of the input shaft fail, the outer ring (BinB_OF), inner ring (BinB_IF) and roller (BinB_RF) of the bearing (BinB) at the output end of the input shaft fail, and the outer ring (BoutA_OF), inner ring (BoutA_IF) and roller (BoutA_RF) of the bearing (BoutA) at the output end of the output shaft fail, and the outer ring (BoutB_OF), inner ring (BoutB_IF) and roller (BoutB_RF) of the bearing (BoutB) at the output end of the output shaft fail; Single gear fault: input shaft gear broken tooth fault (Gin_BT) and output shaft gear broken tooth fault (Gout_BT); The types of compound faults may include: Composite failure of bearings: For example, the outer ring (BinA_OF), inner ring (BinA_IF) and roller (BinA_RF) failure of the bearing (BinA) at the input end of the input shaft and the outer ring (BinB_OF), inner ring (BinB_IF) and roller (BinB_RF) failure of the bearing (BinB) at the output end of the input shaft form a composite failure.

[0025] Composite failure of bearing and gear combination: For example, a single bearing failure, such as the outer ring (BinA_OF), inner ring (BinA_IF) and roller (BinA_RF) failure of the input shaft input end bearing (BinA) combined with the input shaft gear failure; Composite failure of the same bearing: For example, a composite failure consisting of two failures in the outer ring, inner ring and roller of a bearing; The above-mentioned compound fault is a common compound fault mode of the reducer, and may also include other types of compound faults.

[0026] Specifically, the step of processing the real-time signals to obtain a one-dimensional time domain signal comprises: Determine the frequency characteristics corresponding to any of the real-time signals, and determine the wavelet basis function and the number of wavelet decomposition layers based on the frequency characteristics; Performing wavelet decomposition processing on any of the real-time signals based on wavelet basis functions and wavelet decomposition layer numbers to obtain a number of wavelet coefficients, and processing the number of wavelet coefficients based on a preset threshold value; The processed wavelet coefficients are reconstructed to obtain a one-dimensional time domain signal.

[0027] It can be understood that the frequency characteristics corresponding to the embodiments of the present invention are frequency ranges; It can be understood that the embodiments of the present invention may include determining the wavelet basis function and the number of wavelet decomposition layers based on the frequency characteristics: if the fault frequency is mainly concentrated in the medium and low frequency bands, and the signal contains more transient impact components, then a wavelet basis function with better time-frequency localization characteristics can be selected, such as the db4 wavelet basis; and determining the appropriate number of decomposition layers based on the frequency range and the complexity of the signal, so that the wavelet decomposition can effectively separate the frequency band where the fault characteristic frequency components are located, which is convenient for subsequent wavelet coefficient processing and signal reconstruction, so as to better obtain a one-dimensional time domain signal that can reflect the fault state of the reducer. This is a prior art and will not be repeated here.

[0028] It can be understood that the preset threshold in the embodiment of the present invention can obtain the initial threshold by using the threshold calculation method based on Stein unbiased risk estimation, analyzing the signal strength and noise distribution in any real-time signal, and adjusting the initial threshold based on the signal strength and noise distribution analysis results to obtain the preset threshold; The signal strength can be quantified by calculating the root mean square value (RMS) of the real-time signal. The calculation formula of the root mean square value (RMS) of the real-time signal is a prior art and will not be described in detail here. The noise distribution can be estimated by using statistical methods. For stationary noise, it can be assumed to obey Gaussian distribution and the noise parameters can be preliminarily estimated by calculating the mean and variance of the signal.

[0029] It can be understood that adjusting the initial threshold based on the signal strength and noise distribution analysis results may include: According to the signal strength analysis results, if the overall signal strength is strong and the local changes are not large, it means that the effective components in the signal are relatively prominent. At this time, the initial threshold can be appropriately increased to more strictly remove noise and reduce the impact on the effective signal. For example, multiply the initial threshold by an adjustment coefficient greater than 1 (such as 1.2); If the signal strength is weak or there is a local weak signal area, in order to avoid misjudging useful weak signals as noise removal, the initial threshold should be appropriately lowered. The initial threshold can be multiplied by an adjustment coefficient less than 1 (such as 0.8).

[0030] In the case of uneven noise distribution, if the noise is strong in a certain frequency band, and this frequency band overlaps with the frequency band where the fault characteristic frequency is located, then the threshold should be appropriately increased in this frequency band; if the noise in a certain frequency band is weak and has little effect on fault diagnosis, the threshold can be appropriately lowered to retain more signal details. For example, the threshold adjustment method of frequency band weighting is adopted, and the threshold of the frequency band with strong noise is multiplied by (such as 1.2), and the threshold of the frequency band with weak noise is multiplied by (such as 0.8).

[0031] It can be understood that the embodiment of the present invention processes a plurality of wavelet coefficients based on a preset threshold by setting the coefficients smaller than the threshold to zero and shrinking the wavelet coefficients larger than the threshold according to a preset formula to remove the noise component in the signal, wherein the preset formula is: , where is the processed wavelet coefficient, is the wavelet coefficient, and λ is the preset threshold.

[0032] See also Figure 3 As shown, the steps for analyzing the accuracy and integrity of several one-dimensional time domain signals include: Step S21, comparing the plurality of one-dimensional time domain signals with a preset time domain signal range, determining the abnormal proportion of the abnormal one-dimensional time domain signals based on the comparison result, and determining the accuracy of the plurality of one-dimensional time domain signals based on the comparison result of the abnormal proportion and the preset proportion; Step S22, judging the waveform integrity of the signal waveform diagram determined based on the plurality of one-dimensional time domain signals, and determining the integrity of the plurality of one-dimensional time domain signals based on the waveform integrity.

[0033] It can be understood that the preset time domain signal range described in the embodiment of the present invention can be determined by statistical analysis of a large number of normal operation samples to determine the distribution range of parameters such as amplitude, and adopting a probability distribution method, such as assuming that the signal amplitude obeys a normal distribution, and determining a confidence interval (such as a 95% confidence interval) as the boundary of the preset time domain signal range according to the mean μ and variance σ² obtained by sample statistics. For example, for the amplitude range, the lower limit may be μ-kσ, and the upper limit may be μ+kσ (where k is a coefficient determined according to the confidence level, and for a 95% confidence interval, k≈1.96); The preset proportion is 5%.

[0034] It can be understood that, in the embodiment of the present invention, when the abnormality proportion is less than or equal to the preset proportion, some one-dimensional time domain signals are accurate, and when the abnormality proportion is greater than the preset proportion, some one-dimensional time domain signals are inaccurate.

[0035] It can be understood that the embodiments of the present invention can determine the waveform integrity of the signal waveform diagram by calculating the amplitude difference between adjacent sampling points and setting a difference threshold. If the number of adjacent sampling point differences that exceeds the threshold exceeds a preset ratio, it is considered that the waveform may have discontinuities or jumps, and the waveform integrity is damaged, wherein the difference threshold is determined by the average amplitude variation range of the signal, such as 10% of the average amplitude, and the preset ratio is 5%.

[0036] It can be understood that the embodiment of the present invention converts the one-dimensional time domain signal into the actual spectral kurtosis diagram by performing fast spectral kurtosis calculation on the adjusted one-dimensional time domain signal, and processing the time domain signal in segments, and the length of each segment of the signal is determined according to the sampling frequency and the signal characteristics, such as taking 1024 sampling points as one segment; The power spectral density of each signal is calculated, and then the spectral kurtosis value of each frequency point is calculated based on the spectral kurtosis algorithm, and finally these spectral kurtosis values ​​are plotted into an actual spectral kurtosis diagram. The technical means for converting the actual spectral kurtosis diagram are existing technologies and will not be repeated here.

[0037] See also Figure 4 As shown, the steps of constructing a single fault feature set based on a number of one-dimensional time domain signals, a number of actual spectral kurtosis diagrams, a number of real-time temperature values, and a number of real-time pressure values ​​include: Step S31, calculating time domain characteristic parameters of a plurality of said one-dimensional time domain signals; Step S32, extracting key features in any of the actual spectral kurtosis graphs; Step S33, calculating the actual temperature change rate corresponding to the real-time temperature values, and collecting statistics of the actual temperature intervals corresponding to the real-time temperature values; Step S34, calculating the actual pressure average and the actual pressure fluctuation amplitude corresponding to the real-time pressure values; Step S35, normalize the various features obtained by the above calculation and combine them into a single fault feature vector.

[0038] It can be understood that the time domain characteristic parameters described in the embodiment of the present invention include the mean, variance, root mean square value, peak value, kurtosis, and skewness of several one-dimensional time domain signals. The mean reflects the average level of the signal, the variance reflects the degree of discreteness of the signal, the root mean square value comprehensively considers the energy of the signal, the peak value can highlight the maximum amplitude part of the signal, the kurtosis is sensitive to the impact characteristics of the signal, and the skewness can measure the asymmetry of the signal distribution. The corresponding calculation formula is the existing technology and will not be repeated here.

[0039] It can be understood that the embodiment of the present invention compares the mean, variance, root mean square value, peak value, kurtosis, and skewness corresponding to the one-dimensional time domain signal with the corresponding preset value library, and judges a single fault based on the comparison result, wherein the preset value library is established according to the mean, variance, root mean square value, peak value, kurtosis, and skewness corresponding to several single faults. When the mean, variance, root mean square value, peak value, kurtosis, and skewness differ from the corresponding values ​​in the library within ±0.3, the corresponding fault is regarded as a single fault, otherwise, it is a compound fault.

[0040] It can be understood that the preset frequency points in the embodiment of the present invention are calculated based on the fault characteristic frequencies corresponding to various parts. For example, for a bearing outer ring fault, its fault characteristic frequency f BPFO There may be an obvious peak on the spectral kurtosis diagram. The characteristic frequency formula of the bearing outer ring fault obtained by theoretical calculation is f BPFO =n / 2×(1-d / D×cosα)×f r, where n is the number of rolling elements, d is the rolling element diameter, D is the bearing pitch diameter, α is the contact angle, and f r is the rotation frequency of the shaft.

[0041] It can be understood that, in the embodiment of the present invention, if a frequency point is found in the spectral kurtosis diagram to be consistent with the characteristic frequency of the bearing inner ring fault and the spectral kurtosis value of the frequency point is relatively high, and there is no obvious abnormal spectral kurtosis peak in other frequency bands, then it can be judged as a single fault of the bearing inner ring.

[0042] Specifically, the step of extracting key features in any of the actual spectral kurtosis graphs includes: Setting a spectral kurtosis threshold, and taking an area in the actual spectral kurtosis graph that is larger than the spectral kurtosis threshold as the spectral kurtosis optimal frequency band; The actual spectral kurtosis mean, actual spectral kurtosis variance and actual frequency band width within the spectral kurtosis optimal frequency band are calculated as key features.

[0043] Specifically, the steps of classifying several single fault feature sets based on the clustering algorithm include: Calculate the actual data density of any single fault feature sample and the remaining single fault feature samples based on the clustering algorithm; A single fault feature sample whose actual data density is greater than a set density threshold is used as a cluster consisting of a cluster core point and its neighborhood; For any boundary point, if it is in the neighborhood of a cluster core point, it will be divided into the corresponding cluster, otherwise, it will be marked as a noise point.

[0044] Specifically, the embodiment of the present invention can automatically identify the number and shape of clusters in the data in this way without pre-specifying the number of clusters, effectively process non-spherical distributed data, improve the accuracy and adaptability of clustering, better distinguish different single fault feature sets, and obtain several single fault categories with obvious distinction. .

[0045] Specifically, the steps of building a single fault diagnosis model include: Build a neural network model based on a deep learning framework; Train the neural network model based on the training set to obtain an initial model; The model performance of the initial model is tested based on the test set to determine a single fault diagnosis model based on the test results.

[0046] It is understandable that the embodiment of the present invention uses a deep learning framework to build a neural network model, such as a multi-layer perceptron (MLP) structure. A single fault feature vector is used as the input layer data, and multiple hidden layers are set. The number of neurons in the hidden layer is adjusted according to the data characteristics and model complexity, for example, a decreasing structure is used, which is 128, 64, and 32 neurons respectively. The activation function uses the ReLU function to improve the nonlinear expression ability of the model. The output layer uses the softmax function for classification, and outputs the probability distribution corresponding to a single fault category. During the training process, the cross entropy loss function is used to measure the difference between the model prediction result and the true label, and the stochastic gradient descent (SGD) optimization algorithm is used to adjust the model parameters in combination with the momentum term. The momentum term ranges from 0.9 to 0.99, and the initial value of the learning rate is set to 0.01. The learning rate decay strategy is adopted. After a certain number of training rounds (such as 10 rounds), the learning rate decays to 0.9 times the original. At the same time, the L2 regularization term is introduced, and the penalty coefficient is set to 0.001 to prevent the model from overfitting. Through multiple iterations of training, until the loss value of the model on the validation set converges, a trained single fault diagnosis model is obtained.

[0047] Specifically, the steps of performing preliminary diagnosis on a composite fault sample set according to a single fault diagnosis model include: The single fault feature vector corresponding to each sample in the composite fault sample set is sequentially input into the trained single fault diagnosis model; The model calculates each input feature vector and outputs the actual probability value corresponding to each single fault category; The actual probability value is compared with the preset probability value to determine an initial diagnosis result based on the comparison result.

[0048] It is understandable that the embodiment of the present invention sets the probability threshold to 0.5. When the output probability of a single fault category is greater than the threshold, it is considered that the sample may have the single fault component, and its probability value is recorded. For example, for a composite fault sample, the model outputs the probability of "bearing fault" as 0.6 and the probability of "gear fault" as 0.7, then it is preliminarily determined that the sample has these two single fault components, and these preliminary diagnosis results are sorted into a list form, including the name of the fault component, the corresponding probability value and other information, to provide a basis for the subsequent determination of the final diagnosis result of the composite fault.

[0049] Specifically, the steps of determining the final diagnosis result of the composite fault include: Extract the association pattern between single fault feature samples and compound fault types based on several historical fault data, and determine the compound fault inference rule based on the association pattern; Match any single fault feature sample with the composite fault inference rule one by one to obtain a matching result; The final diagnosis result is determined based on the matching result and the initial diagnosis result.

[0050] Specifically, the step of adjusting the preset sampling frequency based on the evaluation result includes: Identify the correct classification results in the compound fault classification recognition results, calculate the proportion of correct classification results, and obtain the classification accuracy; The classification accuracy is compared with the preset accuracy, and the preset sampling frequency is adjusted based on the comparison result.

[0051] It is understandable that the preset accuracy rate of the embodiment of the present invention is 90%.

[0052] It can be understood that, in the embodiment of the present invention, when the classification accuracy is less than the preset accuracy, the preset sampling frequency is increased; When the classification accuracy is greater than or equal to the preset accuracy, the accuracy change trend is monitored; when the accuracy change trend decreases, the preset sampling frequency is increased.

[0053] So far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will fall within the protection scope of the present invention.

[0054] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A zero-sample reducer composite fault diagnosis method based on fast spectral kurtosis, characterized in that: include: Step S1, respectively setting a plurality of accelerometers, a plurality of temperature sensors and a plurality of pressure sensors at the output end and the input end of a plurality of faulty reduction gearboxes, collecting data based on a preset sampling frequency, and obtaining a plurality of real-time signals, a plurality of real-time temperature values ​​and a plurality of real-time pressure values; Step S2, processing the plurality of real-time signals to obtain a plurality of one-dimensional time domain signals, analyzing the accuracy and integrity of the plurality of one-dimensional time domain signals, adjusting the one-dimensional time domain signals according to the analysis results, and converting them into actual spectral kurtosis diagrams; Step S3, constructing a single fault feature set based on a number of one-dimensional time domain signals, a number of actual spectral kurtosis diagrams, a number of real-time temperature values, and a number of real-time pressure values, and classifying the number of single fault feature sets based on a clustering algorithm to obtain a number of single fault categories; Step S4, based on any single fault category, dividing the corresponding fault samples into a training set and a test set to construct a single fault diagnosis model; Step S5, combining and simulating several single fault categories to construct a composite fault sample set, setting several labels corresponding to several composite fault samples one by one, and performing preliminary diagnosis on the composite fault sample set according to the single fault diagnosis model to obtain the initial diagnosis result of the corresponding single fault feature sample in any composite fault sample; Step S6, determining the final diagnosis result of the composite fault according to the initial diagnosis result of each single fault feature sample in the composite fault sample and the preset composite fault inference rule, evaluating the composite fault classification and identification result, and adjusting the preset sampling frequency based on the evaluation result.

2. The zero-sample reducer composite fault diagnosis method based on fast spectral kurtosis according to claim 1 is characterized in that: The step of processing a plurality of said real-time signals to obtain a one-dimensional time domain signal comprises: Determine the frequency characteristics corresponding to any of the real-time signals, and determine the wavelet basis function and the number of wavelet decomposition layers based on the frequency characteristics; Performing wavelet decomposition processing on any of the real-time signals based on wavelet basis functions and wavelet decomposition layer numbers to obtain a number of wavelet coefficients, and processing the number of wavelet coefficients based on a preset threshold value; The processed wavelet coefficients are reconstructed to obtain a one-dimensional time domain signal.

3. The zero-sample reducer composite fault diagnosis method based on fast spectral kurtosis according to claim 2 is characterized in that: The steps to analyze several 1D time domain signals for accuracy and completeness include: Comparing the one-dimensional time domain signals with a preset time domain signal range, determining the abnormal proportion of the abnormal one-dimensional time domain signals based on the comparison result, and determining the accuracy of the one-dimensional time domain signals based on the comparison result of the abnormal proportion and the preset proportion; The waveform integrity of a signal waveform diagram determined based on a plurality of one-dimensional time domain signals is determined, and the integrity of the plurality of one-dimensional time domain signals is determined based on the waveform integrity.

4. The zero-sample reducer composite fault diagnosis method based on fast spectral kurtosis according to claim 3 is characterized in that: The steps of constructing a single fault feature set based on a number of one-dimensional time domain signals, a number of actual spectral kurtosis graphs, a number of real-time temperature values, and a number of real-time pressure values ​​include: Calculating time domain characteristic parameters of a plurality of said one-dimensional time domain signals; Extracting key features in any of the actual spectral kurtosis graphs; Calculating actual temperature change rates corresponding to a number of the real-time temperature values, and performing statistics on actual temperature intervals corresponding to a number of the real-time temperature values; Calculating an actual pressure average and an actual pressure fluctuation amplitude corresponding to a plurality of the real-time pressure values; The various features calculated above are normalized and combined into a single fault feature vector.

5. The zero-sample reducer composite fault diagnosis method based on fast spectral kurtosis according to claim 4 is characterized in that: The step of extracting key features in any of the actual spectral kurtosis graphs comprises: Setting a spectral kurtosis threshold, and taking an area in the actual spectral kurtosis graph that is larger than the spectral kurtosis threshold as the spectral kurtosis optimal frequency band; The actual spectral kurtosis mean, actual spectral kurtosis variance and actual frequency band width within the spectral kurtosis optimal frequency band are calculated as key features.

6. The zero-sample reducer composite fault diagnosis method based on fast spectral kurtosis according to claim 5 is characterized in that: The steps of classifying several single fault feature sets based on the clustering algorithm include: Calculate the actual data density of any single fault feature sample and the remaining single fault feature samples based on the clustering algorithm; A single fault feature sample whose actual data density is greater than a set density threshold is used as a cluster consisting of a cluster core point and its neighborhood; For any boundary point, if it is in the neighborhood of a cluster core point, it will be divided into the corresponding cluster, otherwise, it will be marked as a noise point.

7. The zero-sample reducer composite fault diagnosis method based on fast spectral kurtosis according to claim 6 is characterized in that: The steps to build a single fault diagnosis model include: Build a neural network model based on a deep learning framework; Train the neural network model based on the training set to obtain an initial model; The model performance of the initial model is tested based on the test set to determine a single fault diagnosis model based on the test results.

8. The zero-sample reducer composite fault diagnosis method based on fast spectral kurtosis according to claim 7 is characterized in that: The steps of making a preliminary diagnosis of a composite fault sample set according to a single fault diagnosis model include: The single fault feature vector corresponding to each sample in the composite fault sample set is sequentially input into the trained single fault diagnosis model; The model calculates each input feature vector and outputs the actual probability value corresponding to each single fault category; The actual probability value is compared with the preset probability value to determine an initial diagnosis result based on the comparison result.

9. The zero-sample reducer composite fault diagnosis method based on fast spectral kurtosis according to claim 8 is characterized in that: The steps to determine the final diagnosis result of the compound fault include: Extract the association pattern between single fault feature samples and compound fault types based on several historical fault data, and determine the compound fault inference rule based on the association pattern; Match any single fault feature sample with the composite fault inference rule one by one to obtain a matching result; The final diagnosis result is determined based on the matching result and the initial diagnosis result.

10. The zero-sample reducer composite fault diagnosis method based on fast spectral kurtosis according to claim 9, characterized in that: The step of adjusting the preset sampling frequency based on the evaluation result comprises: Identify the correct classification results in the compound fault classification recognition results, calculate the proportion of correct classification results, and obtain the classification accuracy; The classification accuracy is compared with the preset accuracy, and the preset sampling frequency is adjusted based on the comparison result.

Images