Bearing degradation starting point detection method based on unsupervised streaming update threshold

Abstract

The invention relates to the technical field of bearing detection, and discloses a bearing degradation starting point detection method based on an unsupervised streaming update threshold, which comprises the following steps: acquiring a real-time vibration signal of a bearing; inputting the collected real-time vibration signal into a pre-trained Gaussian process regression model for prediction to obtain a predicted vibration signal; calculating a mahalanobis distance between the predicted vibration signal and the real-time vibration signal, wherein the mahalanobis distance is used for representing the performance state of the bearing; and inputting the calculated mahalanobis distance into a pre-trained DSPOT model, updating a threshold value, and comparing a relative value of the input mahalanobis distance with the updated threshold value and the threshold point exceeding the threshold value to obtain a degradation starting point of the bearing. The detection method has no requirement for input signal feature distribution, only needs to determine the initial threshold value according to the signals in the normal operation stage, and then dynamically updates the threshold value according to the actual operation condition, so that the degradation trend starting point is accurately determined, and the method does not depend on specific working conditions and is wide in application.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bearing detection, and relates to a bearing degradation starting point detection method based on unsupervised flow updating threshold. BACKGROUND

[0002] As an important core component of rotating machinery, rolling bearings are widely used in key fields such as aerospace equipment, numerical control machine tools, and wind power generation equipment. The running state of the rolling bearings is directly related to the performance and reliability of the entire equipment. Due to long-term bearing of complex mechanical load and environmental influence, if fatigue damage or failure occurs, not only will the equipment function be interrupted, but also serious industrial production accidents may be caused, and even life and property safety may be threatened. Therefore, the determination of the degradation starting point of the rolling bearing during operation has important scientific research value and practical application significance, and is one of the key technologies to ensure the reliability and economic benefit of the equipment.

[0003] Generally, the operation of the bearing is considered to be divided into two stages, namely the normal operation stage and the linear degradation stage, and the junction point of the two stages is considered to be the degradation starting point of the bearing, that is, the conversion time point of the normal stage and the degradation stage. The determination of the degradation starting point of the bearing not only has important significance for the maintenance of the mechanical device, but also the prediction accuracy of the degradation starting point will affect the construction of the label of the whole cycle life of the bearing, and then affect the prediction accuracy of the remaining life.

[0004] Most of the existing performance degradation starting point detection methods use criterion or other methods to fuse criterion, which puts strict requirements on the signal feature distribution of the bearing. The signal feature distribution must follow the normal distribution. However, in actual operation conditions or signal sampling, such strict conditions are often difficult to meet. The vibration signal is affected by background noise in actual operation, so the vibration signal cannot strictly follow the normal distribution, resulting in errors in the prediction of the degradation starting point.

[0005] In addition, most of the existing performance degradation starting point detection methods have the problem of determining a fixed threshold value using various algorithms. When the continuous vibration signals of the bearing are greater than the threshold value, the time point is determined as the degradation starting point. However, in the actual running environment of the bearing, the vibration signal amplitude of the normal running stage of the bearing may change due to changes in the working conditions, and the fixed threshold value cannot be adjusted in time with the changes in the working conditions, resulting in incorrect identification of the degradation starting point. SUMMARY

[0006] In view of the problems that the performance degradation starting point prediction method of the rolling bearing in the prior art has high requirements for the feature distribution of the vibration signal and cannot flexibly adapt to complex tasks such as variable working conditions due to the use of a fixed threshold, the application provides a bearing degradation starting point detection method based on an unsupervised stream updating threshold, which can avoid misjudgment of some time points caused by interference signals and better meet the requirements of the actual operation of the bearing.

[0007] To achieve the above object, the application provides the following technical scheme.

[0008] In the first aspect, the application provides a bearing degradation starting point detection method based on an unsupervised stream updating threshold, which comprises the following steps.

[0009] Collecting real-time vibration signals of the bearing;

[0010] Inputting the collected real-time vibration signals into a pre-trained Gaussian process regression model for prediction to obtain predicted vibration signals;

[0011] Calculating Mahalanobis distances between the predicted vibration signals and the real-time vibration signals, wherein the Mahalanobis distances are used to represent the performance state of the bearing;

[0012] Inputting the calculated Mahalanobis distances into a pre-trained DSPOT model for threshold updating, comparing the relative values of the input Mahalanobis distances with the updated threshold and the threshold point of the super-threshold, and obtaining the degradation starting point of the bearing.

[0013] In combination with the first aspect, further, the step of inputting the collected real-time vibration signals into the pre-trained Gaussian process regression model for prediction to obtain the predicted vibration signals comprises the following steps.

[0014] Inputting the collected real-time vibration signal data into a pre-constructed autoencoder to extract vibration signal features and obtain extracted real-time vibration signal features;

[0015] Inputting the extracted real-time vibration signal features into a pre-trained Gaussian process regression model for prediction to obtain the predicted vibration signals.

[0016] In combination with the first aspect, further, the step of inputting the collected real-time vibration signal data into the pre-constructed autoencoder to extract vibration signal features and obtain the extracted real-time vibration signal features comprises the following steps.

[0017] The autoencoder comprises one hidden layer, two encoding layers and two decoding layers corresponding to the encoding layers; the collected real-time vibration signals are input into the encoding layers, pass through two fully connected layers and an activation function to enter the hidden layer for feature extraction, and the data output by the hidden layer is the extracted real-time vibration signal features.

[0018] With reference to the first aspect, further, the training method of the autoencoder is:

[0019] Collecting the vibration signals of the entire life cycle of the bearing, the vibration signals of the entire life cycle being used for training of the autoencoder;

[0020] Inputting the collected vibration signals of the entire life cycle into the encoding layer, and passing through two layers of full connection layers and activation functions of the encoding layer to enter the hidden layer; after the hidden layer, passing through two layers of full connection layers and activation functions symmetrical to the encoding layer to obtain the data after decoding; calculating the error between the data after decoding and the data input into the decoding layer, and reducing the error to the minimum through iterative training to obtain the optimal autoencoder, that is, the trained autoencoder.

[0021] With reference to the first aspect, further, the inputting of the extracted real-time vibration signal features into the pre-trained Gaussian process regression model for prediction to obtain the predicted vibration signal comprises:

[0022] Collecting the vibration signals of the entire life cycle of the bearing;

[0023] Inputting the collected vibration signals of the entire life cycle into the autoencoder for feature extraction to obtain the extracted vibration signal features of the entire life cycle;

[0024] Dividing the extracted vibration signal features of the entire life cycle into a training set;

[0025] Inputting the divided training set into the Gaussian process regression model for training to obtain the trained Gaussian process regression model;

[0026] Inputting the extracted real-time vibration signal features into the trained Gaussian process regression model for prediction to obtain the predicted vibration signal.

[0027] With reference to the first aspect, further, the training method of the Gaussian process regression model is:

[0028] Dividing the entire running stage of the bearing into four stages, using the vibration signal data of the first stage as a training set , The time series of the first stage, that is, the time series of the training set, two dimensions respectively representing the degradation stage and the time point of the bearing; The features of the vibration signal of the first quarter stage are used as the label of the training set, ; The vibration signal features of the entire life cycle of the bearing;

[0029] The Gaussian process assumes that the function values corresponding to any finite samples obey a joint Gaussian distribution, which is expressed as:

[0030] ;

[0031] ;

[0032] wherein, is a Gaussian process function; is noise, denotes a standard normal distribution, denotes the variance of the standard normal distribution; denotes a Gaussian process constructed by the mean function and the RBF kernel function ; is the mean function, which is set to 0; denotes the RBF kernel function, which is used to measure the correlation between samples, and is expressed as:

[0033] ;

[0034] wherein, is an output amplitude factor, is a characteristic length scale, denotes the Euclidean distance between two different data points in , and denotes the time series of the test set;

[0035] The objective of training the Gaussian process regression model is to learn the hyperparameters of the Gaussian process , wherein denotes the observed noise variance, and the hyperparameters are adjusted by maximizing the marginal log-likelihood function, the objective function is:

[0036] ;

[0037] wherein, denotes the unit matrix, is the number of time points.

[0038] By using the gradient descent method to continuously optimize the hyperparameters, the relationship between the samples is modeled, and the trained Gaussian process regression model is obtained.

[0039] In combination with the first aspect, further, the Mahalanobis distance between the predicted vibration signal and the real-time vibration signal is calculated, comprising:

[0040] ;

[0041] wherein, is the joint covariance matrix, ; is the vibration amplitude of the extracted real-time vibration signal feature at the time point. is the amplitude of the predicted vibration signal mean at the th time point predicted by the Gaussian process regression model; is the extracted real-time vibration signal feature; is the predicted vibration signal mean predicted by the Gaussian process regression model; is the time point; is the Mahalanobis distance.

[0042] With reference to the first aspect, further, the calculated Mahalanobis distance is input into the pre-trained DSPOT model to update the threshold value, and the input Mahalanobis distance is compared with the updated threshold value and the threshold point exceeding the threshold value to obtain the degradation starting point of the bearing, comprising:

[0043] If the relative value of the newly input Mahalanobis distance is an abnormal value, no update is performed; if the relative value of the newly input Mahalanobis distance is a normal value, only the window mean value is updated , the DSPOT model is not updated; if the relative value of the newly input Mahalanobis distance is an extreme value, the DSPOT model is updated, the window mean value is updated , and the threshold value is updated ;

[0044] The updated calculation formula is as follows:

[0045] The window mean value at the th time point is expressed as:

[0046] ;

[0047] wherein, is the relative value of the Mahalanobis distance of the th observation data closest to the th time point, i.e., the relative value of the Mahalanobis distance of the th observation data closest to the th time point, ; is the relative value of the Mahalanobis distance at the th time point.

[0048] The relative value of the Mahalanobis distance at the th time point is:

[0049] ;

[0050] The GPD distribution is refitted using the new extreme point, and , are calculated, so as to complete the update of the threshold value:

[0051] ;

[0052] wherein, is the number of all extreme points up to the current time point, is the number of all time points at present, denotes the updated threshold value, denotes the estimation result of the shape parameter in the refitted GPD distribution, denotes the estimation result of the scale parameter in the refitted GPD distribution, is the quantile probability, the super-threshold threshold point .

[0053] In combination with the first aspect, further, the training method of the DSPOT model, namely the process of initializing the DSPOT model, is as follows:

[0054] Using the DSPOT model, it is first necessary to initialize the DSPOT model, initialize the super-threshold threshold point , initialize the window mean , and initialize the threshold value . The initialization data is the first vibration signal feature, that is, is the number of initialization data, and the value of the present application is preferably 200. First, the expression of the initial window mean is:

[0055] ;

[0056] wherein, ; denotes the first vibration signal feature, ; denotes the extracted Mahalanobis distance of the kth time point; it needs to be explained that because each time point has the vibration signal feature corresponding to the time point and the Mahalanobis distance corresponding to the time point, the kth vibration signal feature and the kth time point are one-to-one corresponding.

[0057] The relative value of the Mahalanobis distance of the kth time point is calculated:

[0058] ;

[0059] wherein, denotes the Mahalanobis distance of the kth time point;

[0060] ​​​​​​Initialization of super-threshold threshold , For the high quantile point in the initialization data, set it as the 98% empirical quantile.

[0061] If the relative value of the Mahalanobis distance at the kth time point is greater than , , then the difference between the relative value of the Mahalanobis distance at the kth time point and is recorded as , which conforms to the Generalized Pareto Distribution (GPD), and the parameters are , , and the third parameter , is the location parameter, indicating the lower limit or threshold of the distribution, which is usually set to empty. The maximum likelihood function is used to estimate , , and the estimated result is , :

[0062] ;

[0063] wherein represents the maximum likelihood function; , represents the difference between the relative value of the Mahalanobis distance at the kth time point greater than , and , , is the number of time points greater than , i.e. the number of extreme points, is the shape parameter of the GPD distribution, which affects the tail thickness, is the scale parameter of the GPD distribution, which affects the scalability of the distribution.

[0064] The initial threshold is calculated using the estimated result :

[0065] ;

[0066] wherein is the quantile probability, preferably .

[0067] The initialization of the DSPOT model is completed through the above operations.

[0068] In combination with the first aspect, further, the method for determining the degradation starting point of the bearing is: ​​

[0069] The Mahalanobis distance of the last time point is transmitted into the DSPOT model, and each newly transmitted time point is compared with the , current time point, and if it is less than , it is a normal value, if it is greater than and less than , it is an extreme value, and if it is greater than , it is an abnormal value, and the starting point of continuous abnormalities is the starting point of bearing degradation.

[0070] In a second aspect, the present application provides a bearing degradation starting point detection system based on unsupervised streaming updating threshold, comprising:

[0071] a signal acquisition module configured to acquire real-time vibration signals of the bearing;

[0072] a predicted vibration signal module configured to input the acquired real-time vibration signals into a pre-trained Gaussian process regression model for prediction to obtain a predicted vibration signal;

[0073] a Mahalanobis distance module configured to calculate the Mahalanobis distance between the predicted vibration signal and the real-time vibration signal, the Mahalanobis distance being used to represent the performance state of the bearing;

[0074] a DSPOT model module configured to input the calculated Mahalanobis distance into a pre-trained DSPOT model, update the threshold value, and compare the relative value of the input Mahalanobis distance with the updated threshold value and the super-threshold threshold point to obtain the degradation starting point of the bearing.

[0075] Compared with the prior art, the present application provides a bearing degradation starting point detection method based on unsupervised streaming updating threshold, which has the following beneficial effects:

[0076] (1) The detection method of the present application does not have strict requirements for the feature distribution of the input signal, only needs to determine the initial threshold value according to the normal operation stage signal, and then dynamically updates the threshold value according to the actual operation condition, so as to accurately determine the degradation trend starting point, can avoid the misjudgment of some time points caused by interference signals, is more in line with the requirements of the actual operation condition of the bearing, does not depend on a specific working condition, and is widely applicable.

[0077] (2) In the detection process, the present application discards the method of using a fixed threshold value to detect abnormal points to detect degradation, but instead uses a streaming dynamic updating threshold value method to judge whether a new input signal deviates from the overall trend to determine whether it is abnormal, which can well avoid the misjudgment of some time points caused by interference signals, and is more in line with the requirements of the actual operation condition of the bearing.

[0078] (3) The abnormality detection method of the multiple model fusion constructs the performance degradation index of the bearing, and can more accurately and effectively reflect the degradation condition of the bearing. BRIEF DESCRIPTION OF DRAWINGS

[0079] Figure 1 The flowchart of the degradation starting point detection method of the application is shown in the figure.

[0080] Figure 2 The principle diagram of the bearing life strengthening test bed used in example 1 is shown in the figure.

[0081] Figure 3 The original vibration signal of bearing 1 collected in example 1 is shown in the figure.

[0082] Figure 4 The original vibration signal of bearing 2 collected in example 1 is shown in the figure.

[0083] Figure 5 The network structure diagram of the self-encoder used in example 1 is shown in the figure.

[0084] Figure 6 The vibration signal feature map of bearing 1 extracted by using the self-encoder in example 1 is shown in the figure.

[0085] Figure 7 The vibration signal feature map of bearing 2 extracted by using the self-encoder in example 1 is shown in the figure.

[0086] Figure 8 The vibration signal trajectory diagram of bearing 1 predicted by using the Gaussian process regression model in example 1 is shown in the figure.

[0087] Figure 9 The vibration signal trajectory diagram of bearing 2 predicted by using the Gaussian process regression model in example 1 is shown in the figure. Figure 8 The enlarged diagram of the red circle in the figure.

[0088] Figure 10 The vibration signal trajectory diagram of bearing 2 predicted by using the Gaussian process regression model in example 1 is shown in the figure.

[0089] Figure 11 The vibration signal trajectory diagram of bearing 2 predicted by using the Gaussian process regression model in example 1 is shown in the figure. Figure 10 The enlarged diagram of the red circle in the figure.

[0090] Figure 12 The Mahalanobis distance representing the performance condition of bearing 1 calculated in example 1 is shown in the figure, and the degradation starting point determined by using the DSPOT method flow dynamic threshold updating method is shown in the figure.

[0091] Figure 13 The enlarged diagram of the red circle in the figure. Figure 12 The enlarged diagram of the red circle in the figure.

[0092] Figure 14A schematic diagram of the Mahalanobis distance representing the performance condition of the bearing 2 calculated in Example 1, and the degradation starting point determined by the DSPOT method with the flow dynamic updating threshold value;

[0093] Figure 15 For Figure 15 The enlarged schematic view at the center of the circle. DETAILED DESCRIPTION

[0094] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0095] Example 1

[0096] As Figure 1 shown, the present embodiment proposes a bearing degradation starting point detection method based on unsupervised flow updating threshold value, including the following steps:

[0097] Step 1: Use ABLT-1A type bearing life strengthening test bench to collect vibration signals of rolling bearings in the whole life cycle.

[0098] ABLT-1A type bearing life strengthening test bench is used to simulate two different working conditions, and the vibration signals of rolling bearings in the whole life cycle under the two working conditions are collected correspondingly to establish a whole life data set, and the test data of the two groups are marked as Bearing1 and Bearing2 respectively.

[0099] As Figure 2As shown in the above bearing life enhancement test bench, it includes a vibration acceleration sensor, a transmission system, a loading system and a data acquisition system. The transmission system includes a motor, a shaft coupling and a transmission shaft, and each bearing is arranged on the transmission shaft. The data acquisition system includes a vibration acceleration sensor and a thermocouple arranged on the bearing, a current sensor arranged on the motor and an upper computer; when the bearing full life experiment is carried out, the bearing is driven to rotate by the transmission system, and at the same time, a certain load is applied to the bearing in the radial direction by the loading system to simulate different working conditions of the bearing in actual operation. The test bench can simultaneously carry out full life cycle test on four bearings, i.e. bearing 1, bearing 2, bearing 2 and bearing 4; among them, bearing 2 and bearing 3 share one vibration acceleration sensor. Signals are collected by four thermocouples, three vibration acceleration sensors and a current sensor, and data is saved by the upper computer. The bearing used in this embodiment is a 6008 single-row deep groove ball bearing, and two bearings of the same type are used to carry out two groups of full life tests under two working conditions. Test condition 1: rotating speed 3000r / min, bearing load 18000N, signal sampling frequency 25.6KHz, each time collecting 1s, recording 25600 samples every 1min. Test condition 2: rotating speed 1500r / min, bearing load 5000N, signal sampling frequency 25.6KHz, each time collecting 1s, recording 25600 samples every 1min. Two groups of full life tests are carried out respectively, and the complete failure state is reached after running for 1824min and 2160min respectively. The test data of the two groups are marked as Bearing1 and Bearing2 respectively.

[0100] Figure 3 and Figure 4 are the vibration signals of bearing 1 and bearing 2 respectively. As can be seen from Figure 3 and Figure 4 , the vibration signals of the two bearings are stable and have small amplitude in the normal running stage, and after a long time of running and wear, the bearings are degraded, the vibration signals are no longer stable, the amplitude is therefore large, and the complete failure is reached. Through experimental analysis, only the first 1000 data per second can represent the vibration at this time point, so the shape of the input data is , wherein is the number of time points; 1000 is the number of vibration signals collected at the time point, i.e. the feature dimension; represents a real number set; represents the vibration signal of the bearing full life cycle.

[0101] Step 2: using the self-encoder to automatically extract the vibration signal features of the bearing full life cycle to obtain the low-dimensional vibration signal features after extraction.

[0102] As shown in Figure 5As shown, the multi-layer autoencoder is used to extract deep features of the bearing vibration signals in the whole life cycle. Figure 5 The reconstructed signals will be error calculated with the corresponding original signals to measure the accuracy of the feature extraction of the autoencoder and further guide the training of the autoencoder. The number of layers of the autoencoder has a great influence on the result of feature extraction. Too few layers may result in failure to extract representative features, and too many layers may result in slow training speed or overfitting, etc. In the present embodiment, the autoencoder is set to have 2 encoding layers (parameters 1000→100), 1 hidden layer (parameters 10), and 2 decoding layers corresponding to the encoding layers. The vibration signals in the whole life cycle of the bearing are mapped to a low-dimensional feature space using the autoencoder: wherein d is the dimension of the extracted low-dimensional features, and d = 10; represents the vibration signal features, which are the extracted low-dimensional features. After the data features are extracted by the autoencoder, the shape of the extracted data is The first 5 dimensions of the extracted 10-dimensional features are visualized as shown in Figure 6 and Figure 7 . By observing the extracted bearing vibration signal features, it can be found that the change trend of the vibration signal features is consistent with the vibration signals in the whole life cycle of the bearing , i.e., the amplitude of the vibration signal features increases continuously with the degradation of the bearing.

[0103] Step 3: Use the Gaussian process regression model to predict the vibration signal of the rolling bearing under normal operation.

[0104] In the present embodiment, the mean of the predicted vibration signal obtained by the Gaussian process regression model is used to calculate the Mahalanobis distance, rather than directly using the sampling data to estimate the mean as in the prior art, because the mean directly estimated from the sampling data contains noise interference, and the mean of the predicted vibration signal ​​​​​​​​​​​​​​​The interference is largely avoided, so the calculated Mahalanobis distance is more stable and reliable.

[0105] The training method of the Gaussian process regression model is as follows:

[0106] The entire running stage of the bearing is divided into four stages, and the vibration signal data of the first stage is used as the training set , and the data of the last three stages is used as the test set, wherein is the time series of the first stage, that is, the time series of the training set, and the two dimensions respectively represent the degradation stage and the time point of the bearing; indicates the characteristics of the first quarter of the vibration signal, which is used as the training set label, .

[0107] The Gaussian process assumes that the function values corresponding to any finite samples follow a joint Gaussian distribution, which can be expressed as:

[0108] ;

[0109] ;

[0110] wherein is a Gaussian process function; is noise, indicates a standard normal distribution, indicates the variance of the standard normal distribution; indicates a Gaussian process constructed by the mean function and the RBF kernel function ; is the mean function, which is set to 0; indicates the RBF kernel function, which is used to measure the correlation between samples, and is usually expressed as:

[0111] ;

[0112] wherein is the output amplitude factor, is the characteristic length scale, indicates the Euclidean distance between two different data points in , indicates the time series matrix of the stage to be predicted, that is, the time series of the test set.

[0113] The goal of training the Gaussian process regression model is to learn the hyperparameters of the Gaussian process, wherein indicates the observation noise variance, and the hyperparameters are adjusted by maximizing the marginal log-likelihood function, and the objective function is:

[0114] ;

[0115] wherein, denotes the unit matrix.

[0116] By using the gradient descent method to continuously optimize the hyperparameters, the modeling of the relationship between the samples is completed, and a trained Gaussian process regression model is obtained.

[0117] After completing the modeling of the Gaussian process regression model, the time points of the rolling bearing full life cycle vibration signal are input into the built Gaussian process regression model for prediction, and the mean and variance of the predicted vibration signal trajectory are obtained.

[0118] ;

[0119] ;

[0120] wherein, denotes the distance matrix between the bearing full life cycle time series and the training set time series; denotes the autocorrelation matrix of the bearing full life cycle time series, .

[0121] The vibration signal trajectory of the rolling bearing predicted by the Gaussian process regression model is shown in Figures 8 to 11 . The red curve is the mean of the predicted vibration signal trajectory, the pink shaded part is the 95% confidence interval, and the blue points represent normal points and the red points represent abnormal points. The vibration signal trajectory predicted by the Gaussian process regression model is the vibration signal trajectory of the rolling bearing under normal operation, so the predicted trajectory is relatively stable, and it can be found that the vibration signal deviates from the predicted trajectory more and more in the degradation stage of the rolling bearing. The Mahalanobis distance is calculated to construct the bearing performance index.

[0122] Step 4: Calculate the Mahalanobis distance between the real-time vibration signal of the collected bearing and the predicted vibration signal predicted by the Gaussian process regression model to represent the performance of the bearing.

[0123] The Mahalanobis distance comprehensively considers the correlation of each different dimension time and the covariance matrix, so it can better measure the distance between the samples at different time points and the vibration signal in the normal operation stage, and better reflect the relative position between the abnormal value and the normal data, that is, better reflect the performance degradation of the bearing. The Mahalanobis distance (MD) between the real-time vibration signal features extracted in step 2 and the predicted vibration signal predicted by the Gaussian process regression is calculated to represent the degradation of the bearing. The shape of the obtained Mahalanobis distance is .

[0124] ;

[0125] wherein, is the joint covariance matrix, ; is the vibration amplitude of the extracted real-time vibration signal feature at the th time point; is the vibration amplitude of the predicted vibration signal mean value predicted by the Gaussian process regression model at the th time point; is the extracted real-time vibration signal feature; is the predicted vibration signal mean value predicted by the Gaussian process regression model; is the time point; is the Mahalanobis distance.

[0126] Step 5: Adopting the DSPOT model to flow dynamically update the threshold value, and determining the degradation starting point of the rolling bearing.

[0127] Step 51, initializing the DSPOT model.

[0128] The DSPOT model based on the extreme value theory overcomes the defects of the traditional method which needs a large number of labels or fixed threshold value, and can dynamically adjust the threshold value according to the running trend of the bearing, so as to realize the sensitive capture of the abnormal rising trend in the data stream.

[0129] Using the DSPOT model, firstly, the DSPOT model needs to be initialized, the super-threshold threshold point is initialized, the window mean value is initialized, and the threshold value is initialized. The initialization data is the first vibration signal features, that is, is the number of initialization data, and the value of of the present application is preferably 200. Firstly, the expression of the initial window mean value is:

[0130] ;

[0131] wherein, ; represents the first vibration signal features, ; represents the extracted Mahalanobis distance at the th time point; it needs to be explained that: because each time point has the vibration signal feature corresponding to the time point and the Mahalanobis distance corresponding to the time point, the The characteristics of the vibration signal are similar to those of the first vibration signal. Each point in time corresponds to a specific point in time.

[0132] Calculate the first ( The relative values ​​of Mahalanobis distances at ) time points :

[0133]

[0134] in, This represents the Mahalanobis distance at the k-th time point;

[0135] Initialize the threshold point , To initialize the high quantiles in the data, set them to the 98th percentile empirical quantile.

[0136] If the relative value of the Mahalanobis distance at the k-th time point Greater than Then the relative value of the Mahalanobis distance at the k-th time point is... With The difference is denoted as It conforms to the Generalized Pareto Distribution (GPD), with the following parameters: , The GPD distribution also requires a third parameter. , This is a location parameter representing the lower bound or threshold of the distribution; it is usually left blank. Maximum likelihood function is used to estimate... , The estimation results were obtained. , :

[0137] ;

[0138] in, Represents the maximum likelihood function; , indicating greater than The The relative values ​​of Mahalanobis distances at various time points and The difference, greater than The number of time points (i.e., the number of extreme points). The shape parameters of the GPD distribution affect the tail thickness. This is the scale parameter of the GPD distribution, which affects the scalability of the distribution.

[0139] Using the estimation results , Calculate the initial threshold :

[0140] ;

[0141] in, For quantile probabilities, the preferred method is... .

[0142] The above operations complete the initialization of the DSPOT model.

[0143] Step 52: Perform anomaly detection.

[0144] Define the new input as greater than The The relative values ​​of Mahalanobis distances at various time points This is an outlier, greater than Less than The The relative values ​​of Mahalanobis distances at various time points It is an extreme value, less than The The relative values ​​of Mahalanobis distances at various time points This is a normal value. Because a new threshold is obtained after each new data point is input. The relative value of the Mahalanobis distance at the point is compared with the threshold value at the point. If the relative value of the Mahalanobis distance is greater than the threshold value at the point, it is an abnormal point. The starting point of continuous abnormality is the starting point of the degradation of the rolling bearing.

[0145] Step 51 completes the initialization of the DSPOT model. Next, the relative values ​​of the Mahalanobis distances at new time points are continuously input into the DSPOT model. If the newly input relative value of the Mahalanobis distance is an outlier, no update is performed; if the newly input relative value of the Mahalanobis distance is a normal value, only the window mean is updated. If the relative value of the newly input Mahalanobis distance is an extreme value (the corresponding time point is the extreme point), then update the DSPOT model and update the window mean. and update threshold The updated calculation formula is as follows:

[0146] No. Window mean at each time point The expression is:

[0147] ;

[0148] in, It is the distance from the first The most recent time point The observation data, i.e., the distance from the first observation data. The most recent time point The relative values ​​of the Mahalanobis distances, ; It is the first The relative values ​​of Mahalanobis distances at each time point.

[0149] No. The relative values ​​of Mahalanobis distances at various time points for:

[0150] ;

[0151] Refit the GPD distribution using the new extreme points and calculate , This completes the threshold update:

[0152] ;

[0153] in, It represents the number of all extreme points up to the current time point. It represents the number of all points in time at present. This indicates the updated threshold. This represents the estimated shape parameters in the refitted GPD distribution. This represents the estimated scale parameter in the refitted GPD distribution.

[0154] The above anomaly detection steps achieve dynamic threshold updates in a streaming manner, which can effectively adapt to the detection of the degradation initiation point of rolling bearings under different operating conditions. Furthermore, because no [further details are needed], [the system] does not use [other methods]. The criteria determine the threshold, which effectively avoids the requirement that vibration signals must meet a normal distribution.

[0155] like Figures 12 to 15 As shown, the green line represents the Mahalanobis distance between the calculated actual vibration trajectory and the predicted vibration trajectory. The amplitude of the Mahalanobis distance also increases continuously with the degradation of the rolling bearing. The red dashed line represents the continuously updated threshold. It can be observed that the first 200 data points are used to initialize the parameters. After the initial threshold, when encountering extreme values, the DSPOT model will consider new extreme value distributions and continuously adjust the threshold to detect the starting point of the rolling bearing degradation.

[0156] In order to verify the superiority of the present application, two groups of bearing vibration signals in working condition 1 and three groups of bearing vibration signals in working condition 3 provided by PHM2012 data mining challenge are selected to carry out the test. The two bearings in the experimental working condition 1 are Bearing1_1 and Bearing1_2, the running condition is that the rotating speed is 1800r / min, the bearing load is 4000N, the signal sampling frequency is 25.6KHz, 0.1s is collected each time, and 2560 sampling points are recorded every 10s. The three bearings in the experimental working condition 3 are Bearing3_1, Bearing3_2 and Bearing3_3, the running condition is that the rotating speed is 1500r / min, the bearing load is 5000N, the sampling frequency is 25.6KHz, 0.1s is collected each time, and 2560 sampling points are recorded every 10s. Table 1 is the comparison table of the results of the degradation starting point predicted by the present application and the degradation starting point predicted by the criterion, and it can be found that the prediction effect of the degradation starting point of the present application is better than that of the criterion method. In the case of rapid degradation, as shown by Bearing1_2, Bearing3_1, Bearing3_2 and Bearing1, although the traditional method often encounters the interference of abnormal points, the two methods predict the degradation starting point similarly. On the contrary, in the process of slow bearing degradation, the DSPOT model of the present application shows a significant advantage over the traditional method, as shown by Bearing1_1, Bearing3_3 and Bearing2. The degradation starting point detection method of the present application effectively avoids the misjudgment of the degradation starting point by some interference. In the gradual degradation mode, when the amplitude of the Mahalanobis distance gradually changes, the traditional method cannot effectively alleviate the interference of the larger amplitude, resulting in the determination of the degradation starting point too early. However, the present application dynamically adjusts the threshold in real time, which helps to more accurately determine the degradation starting point.

[0157] Table 1

[0158]

[0159] Example 2

[0160] Based on the same inventive concept as example 1, this embodiment introduces a bearing degradation starting point detection system based on unsupervised streaming threshold updating, which comprises:

[0161] The signal acquisition module is configured to acquire real-time vibration signals of the bearing;

[0162] The predicted vibration signal module is configured to input the acquired real-time vibration signal into the pre-trained Gaussian process regression model for prediction to obtain a predicted vibration signal;

[0163] ​A Mahalanobis distance module configured to calculate a Mahalanobis distance between the predicted vibration signal and the real-time vibration signal, the Mahalanobis distance being used to represent the performance state of the bearing;

[0164] A DSPOT model module configured to input the calculated Mahalanobis distance into a pre-trained DSPOT model, update a threshold value, and compare the input Mahalanobis distance with the updated threshold value and a super-threshold threshold point to obtain a degradation starting point of the bearing.

[0165] It should be noted that the relational terms herein such as first and second and the like are used solely to distinguish one from another entity or action without necessarily requiring or implying any actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. The terms "includes", "including", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0166] Although the embodiments of the present application have been shown and described, it would be appreciated by those skilled in the art that changes, modifications, substitutions, and variations can be made to the embodiments without departing from the principles and spirit of the application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for bearing degradation onset point detection based on unsupervised streaming update threshold, characterized in that, The method comprises the following steps: collecting real-time vibration signals of the bearing; inputting the collected real-time vibration signals into a pre-trained Gaussian process regression model for prediction to obtain predicted vibration signals; calculating Mahalanobis distance between the predicted vibration signals and the real-time vibration signals, wherein the Mahalanobis distance is used to represent the performance state of the bearing; inputting the calculated Mahalanobis distance into a pre-trained DSPOT model for threshold updating, and comparing the relative value of the inputted Mahalanobis distance with the updated threshold and the threshold point of the threshold to obtain the degradation starting point of the bearing.

2. The method for bearing degradation onset point detection based on unsupervised streaming update threshold according to claim 1, wherein, The method of inputting the collected real-time vibration signals into the pre-trained Gaussian process regression model for prediction to obtain the predicted vibration signals comprises the following steps: inputting the collected real-time vibration signal data into a pre-constructed autoencoder to extract vibration signal features to obtain extracted real-time vibration signal features; inputting the extracted real-time vibration signal features into the pre-trained Gaussian process regression model for prediction to obtain the predicted vibration signals.

3. The method for bearing degradation onset point detection based on unsupervised streaming update threshold according to claim 2, wherein, The method of inputting the collected real-time vibration signal data into the pre-constructed autoencoder to extract vibration signal features to obtain the extracted real-time vibration signal features comprises the following steps: The autoencoder comprises one hidden layer, two encoding layers and two decoding layers corresponding to the encoding layers; the collected real-time vibration signals are inputted into the encoding layers, and then are inputted into the hidden layer through two fully connected layers and an activation function for feature extraction; the data outputted by the hidden layer is the extracted real-time vibration signal features.

4. The method for bearing degradation onset point detection based on unsupervised streaming update threshold according to claim 3, wherein, The training method of the autoencoder comprises the following steps: collecting vibration signals of the whole life cycle of the bearing, wherein the vibration signals of the whole life cycle of the bearing are used for training the autoencoder; inputting the collected vibration signals of the whole life cycle into the encoding layers, and then inputting the vibration signals into the hidden layer through two fully connected layers and an activation function of the encoding layers; the data outputted by the hidden layer is the data after decoding; calculating the error between the data after decoding and the data inputted into the decoding layers, and reducing the error to the minimum through iterative training to obtain an optimal autoencoder, i.e., a trained autoencoder.

5. The method for bearing degradation onset point detection based on unsupervised streaming update threshold as claimed in claim 2, wherein, The method of inputting the extracted real-time vibration signal features into the pre-trained Gaussian process regression model for prediction to obtain the predicted vibration signals comprises the following steps: collecting vibration signals of the whole life cycle of the bearing; inputting the collected vibration signals of the whole life cycle into the autoencoder for feature extraction to obtain extracted vibration signal features of the whole life cycle; dividing the extracted vibration signal features of the whole life cycle into a training set; inputting the divided training set into the Gaussian process regression model for training to obtain a trained Gaussian process regression model; inputting the extracted real-time vibration signal features into the trained Gaussian process regression model for prediction to obtain the predicted vibration signals.

6. The method for bearing degradation onset point detection based on unsupervised streaming update threshold according to claim 5, wherein, The training method of the Gaussian process regression model comprises the following steps: The entire running stage of the bearing is divided into four stages, and the vibration signal data of the first stage is used as the training set , The time series of the first stage, that is, the time series of the training set, has two dimensions respectively representing the degradation stage and the time point of the bearing; The features of the first quarter of the vibration signal are represented as the training set label, ; The vibration signal features of the entire life cycle of the bearing are represented; The Gaussian process assumes that the function values corresponding to any finite samples follow a joint Gaussian distribution, which is expressed as: ； ； where, is a Gaussian process function; is noise, denotes a standard normal distribution, denotes a variance of a standard normal distribution; denotes a Gaussian process constructed by a mean function and an RBF kernel function ; is a mean function, set to 0; denotes an RBF kernel function, used to measure the correlation between samples, denoted as: ； wherein, is an output amplitude factor, is a characteristic length scale, denotes the Euclidean distance between two different data points in denotes a time series of the test set; The goal of training a Gaussian process regression model is to learn the hyperparameters of the Gaussian process where denotes the observation noise variance, the hyperparameters are adjusted by maximizing the marginal log-likelihood function, the objective function is: ； wherein denotes the identity matrix, is the number of time points; The relationship between the samples is modeled by continuously optimizing the hyperparameters using the gradient descent method to obtain the trained Gaussian process regression model.

7. The unsupervised stream-based updating threshold-based bearing degradation onset point detection method of claim 2, wherein, The method of calculating the Mahalanobis distance between the predicted vibration signals and the real-time vibration signals comprises the following steps: ； wherein, is the joint covariance matrix, ; is the amplitude of the vibration signal feature at the time point ; is the amplitude of the predicted vibration signal mean at the time point ; is the extracted real-time vibration signal feature; is the predicted vibration signal mean; is the time point; is the Mahalanobis distance.

8. The method for bearing degradation onset point detection based on unsupervised streaming update threshold according to claim 2, wherein, The calculated Mahalanobis distance is input into a pre-trained DSPOT model, threshold updating is performed, and the input Mahalanobis distance is compared with the updated threshold and the super-threshold threshold point to obtain a bearing degradation starting point, comprising: If the relative value of the newly incoming Mahalanobis distance is an outlier, no update is performed; if the relative value of the newly incoming Mahalanobis distance is a normal value, only the window mean is updated , the DSPOT model is not updated; if the relative value of the newly incoming Mahalanobis distance is an extreme value, the DSPOT model is updated, the window mean is updated , and the threshold is updated ; The updated calculation formula is as follows: The first window mean of the time points The expression is: ； wherein is the relative value of the Mahalanobis distance at the th time point; ; The relative value of the Mahalanobis distance at the first time point is : ； The GPD distribution is refitted with the new extreme point, and the threshold is updated as follows: , , ； wherein is the number of all extreme points up to the current time point, is the number of all time points up to the current time point, denotes the updated threshold value, denotes the estimate of the shape parameter in the refitted GPD distribution, denotes the estimate of the scale parameter in the refitted GPD distribution, is the quantile probability, the threshold value above which the threshold point .

9. The unsupervised stream-based updating threshold-based bearing degradation onset point detection method of claim 8, wherein, The determination method of the bearing degradation starting point is: After The Mahalanobis distance at each time point is input into the DSPOT model, and the distance between each newly input time point and the current time point is calculated. , Compared to, it is less than This is a normal value, greater than and less than It is an extreme value, greater than If it is an outlier, the starting point of a series of outliers is the starting point of bearing degradation.

10. A system for bearing degradation onset point detection based on unsupervised streaming update threshold, characterized in that, Comprising: A signal acquisition module configured to acquire a real-time vibration signal of the bearing; A predicted vibration signal module configured to input the acquired real-time vibration signal into a pre-trained Gaussian process regression model for prediction to obtain a predicted vibration signal; A Mahalanobis distance module configured to calculate a Mahalanobis distance between the predicted vibration signal and the real-time vibration signal, the Mahalanobis distance being used to represent a performance state of the bearing; A DSPOT model module configured to input the calculated Mahalanobis distance into a pre-trained DSPOT model, perform threshold updating, and compare a relative value of the input Mahalanobis distance with the updated threshold and the super-threshold threshold point to obtain a bearing degradation starting point.

Images