A method for predicting the remaining useful life of bearings based on label distribution

Abstract

The present invention discloses a bearing life prediction based on label distribution, including dividing a data set into a training set and a test set, preprocessing the bearing vibration signal, normalizing the vibration signal, and storing it as a two-dimensional grayscale image; extracting features from the two-dimensional grayscale image using histogram of oriented gradients and gray-level co-occurrence matrix, and performing feature fusion; processing the label of the bearing's used time using a probability function for label distribution, and finally modeling the bearing's used time through a kernel extreme learning machine to predict the used time of an unknown image, and subtracting the bearing's service life from its used time to predict its remaining service life. The feature fusion of the histogram of oriented gradients and the gray-level co-occurrence matrix obtains a more comprehensive representation; label distribution learning can describe the life more accurately, and combining it with the kernel extreme learning machine effectively shortens the running time while ensuring better accuracy, achieving a better prediction effect compared with traditional methods.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence for predicting the service life of mechanical bearings, and particularly relates to a method for predicting the remaining service life of bearings based on a label distribution algorithm. Background Art

[0002] Due to the intelligent complexity of modern mechanical equipment and the diversity of its working environment, the harsh and changeable working conditions often easily lead to the rapid aging of some equipment components. This greatly increases the possibility of equipment failure. With the integration and intelligence of mechanical equipment, once a certain equipment fails, it is extremely easy to cause casualties or lead to the sudden shutdown of the entire system, resulting in huge economic losses. Therefore, it is necessary to conduct fault prediction and health management (Prognostics and Health Management, PHM) on key machinery and equipment, perform health assessment based on the real-time data of the equipment, and take certain predictive maintenance measures for the system through artificial intelligence and other methods, so as to extend the life cycle of the system or reduce other risks.

[0003] At present, PHM has been widely emphasized in many Western developed countries and has been applied to various industrial scenarios, such as aerospace systems. PHM technology encompasses professional knowledge in multiple fields including machinery, computer science, and electricity, and has obvious practical significance and great application potential. Within the framework of PHM, the prediction of the remaining useful life (RUL) is an important component. During the process of predicting the remaining useful life, by using the monitoring information of the system, the remaining useful life of the system is predicted before a major failure occurs, thereby providing information for preventive maintenance. For example, when the remaining useful life of the equipment reaches the warning value, the task assignment is re-performed to reduce the working intensity of the equipment, so as to increase the running time of the entire system, or the equipment is inspected and repaired in advance to reduce the probability of major failures. At present, there is no unified classification method for the prediction methods of the remaining useful life, so the classification methods in many literatures are different. According to different classification methods, there are knowledge-based models, expected life models, artificial neural networks, physical model prediction methods; there are model-based, data-driven, and hybrid prediction methods; there are also traditional artificial intelligence and deep learning prediction methods. In recent years, with the rise of artificial intelligence, label distribution learning, as a new learning method, has gradually been applied to fields such as age estimation and crowd counting. In label learning, multi-label learning can solve the problem of label polysemy, and the learning algorithm based on multi-labels can solve data classification and imbalance problems. However, multi-label learning cannot represent the degree of description of each sample and its related labels, and cannot give the correlation degree between examples and labels. The concept of label distribution is used to describe the degree of description of labels, and each degree of description corresponds to a real value, and the set of degrees of description is the probability distribution value of all labels in the example. Currently, in label distribution learning, scholars have proposed a variety of label distribution learning algorithms. According to different algorithm design strategies, they can be divided into the following three types: (1) The algorithm based on problem transformation, PT-Bayes (Problem Transformation Bayes), uses the posterior probability of each label as the degree of description of the label and predicts the label distribution output. (2) Algorithms based on algorithm adaptation, such as AA-KNN (Algorithm Adaptation K Nearest Neighbors), AA-BP (Algorithm Adaptation Backpropagation), etc., transform the existing algorithms and optimize the model to directly output the label distribution. (3) Specialized algorithms proposed for label distribution learning, such as SA-IIS (Specialized Algorithms Improved Iterative Scaling), SA-BFGS (Specialized Algorithms Broyden Fletcher Goldfarb Shanno).The above five algorithms are all relatively effective label distribution algorithms in existing strategies. Among them, the specialization performance is relatively good. However, the specialization algorithm uses the maximum entropy model to train through iterative solution and other methods, and then uses the model to predict the distribution law of labels. It takes a long time to process a large dataset. Although algorithms such as PT-Bayes or AA-KNN can shorten the processing time, the classification effect is not very satisfactory. Geng et al. added a conditional probability neural network to label distribution learning, but it is only applicable to completely ordered labels and cannot solve general label learning tasks. At the same time, in these algorithms, the selection of parameters has a huge impact on the experimental results, and the adjustment of parameters may cause the algorithm to fall into local optimality or overfitting. To solve the shortcoming that the feedforward neural network is prone to local optimality, Huang et al. proposed the Extreme Learning Machine (ELM) algorithm. This algorithm is a fast single-hidden layer neural network. When training, the number of hidden layer nodes is set in advance, and the entire training process does not require iteration, only a least squares problem with the minimum norm needs to be solved. Compared with other neural networks, the parameter setting of the extreme learning machine is simple, and it has a speed and generalization ability that cannot be compared with other algorithms. In the extreme learning machine, the matrix H is generated by random assignment. Due to random uncertainty, the obtained H is different each time, and the output weight β is also different, resulting in unstable fluctuations in the results. To enhance the stability of the extreme learning machine, the kernel function is introduced into the extreme learning machine, and the kernel matrix Ω is used to replace the random matrix H in the extreme learning machine. After the kernel parameters are set, the stable network output weight is obtained by solving, avoiding the result fluctuations caused by random parameters, and enhancing the stability of the model on the basis of the extreme learning machine. Existing remaining useful life prediction methods include: first, methods based on physical models, second, methods based on statistical models, and third, methods using artificial intelligence. The first method is to establish relevant mathematical models to represent the true state of the bearing. If it is a non-linear complex system, its structure is generally not very simple. Usually, it is difficult to establish an accurate mathematical model or physical model, and the generalization ability of this method is relatively weak. The second method is a method based on a statistical model, also known as a method based on an empirical model. By establishing a statistical model based on empirical knowledge and presenting the RUL prediction result in the form of a probability density function according to the statistical results, this method is widely used but requires a large number of experiments and consumes a large amount of human resources. The third method uses artificial intelligence methods, which can handle complex mechanical system prediction problems that are difficult to describe by physical models or statistical models. Traditional artificial intelligence methods have problems such as incomplete feature extraction, difficult data collection, and generalization. Deep learning has problems such as large computational complexity caused by complex structures, continuous parameter adjustment, and feature selection. Summary of the Invention

[0004] Aiming at the problems of low efficiency in obtaining image data, low accuracy in feature extraction, complex iterative operations, and large time consumption in traditional life prediction methods. The present invention uses label distribution learning based on kernel extreme learning machine to predict the bearing life. The histogram of oriented gradients and gray-level co-occurrence matrix are used to extract the edge features and local texture features of the bearing image respectively, and the features extracted by the two are fused to combine the edge and local features, so that the features of the image are more comprehensively characterized, greatly improving the prediction accuracy. Label distribution learning can describe the service life more accurately; combining label distribution learning with kernel extreme learning machine effectively shortens the running time while ensuring good accuracy, and achieves better prediction results compared with traditional methods.

[0005] A method for predicting the remaining service life of a bearing based on label distribution provided by the present invention:

[0006] It includes the following steps:

[0007] Step a, divide the data set of bearing information into a training set and a test set, preprocess the bearing vibration signal, normalize the vibration signal, and store it as a two-dimensional grayscale image;

[0008] Step b, extract features from the grayscale image obtained in step a by using the histogram of oriented gradients method and the gray-level co-occurrence matrix;

[0009] Step c, perform label distribution processing on the bearing service life label by using a probability function;

[0010] Step d, merge the processed label distribution training set and test set in step c with the features to form a new feature data set and process it with a kernel extreme learning machine, calculate the kernel matrix and weights, multiply the obtained outputs, obtain the predicted output label distribution, and obtain the maximum probability predicted service life;

[0011] Step e, calculate the accuracy rates obtained by the histogram of oriented gradients and gray-level co-occurrence matrix features through the kernel extreme learning machine respectively, then assign higher weights to those with higher classification accuracy rates and lower weights to those with lower classification accuracy rates to obtain the fused features, and process the fused features with the kernel extreme learning machine;

[0012] Step f, add the expected value to the service life prediction, and jointly determine the service life with the maximum probability predicted service life in step d;

[0013] Step g, subtract the predicted bearing service life obtained in step f from its service life to obtain its remaining service life.

[0014] Further, in step a, the min-max normalization method is used to normalize the bearing signal, and the result is mapped between [0, 255]. The specific formula is:

[0015]

[0016] where X is the original data, Xnorm is the result after normalization, Xmax is the maximum value in the original data, and Xmin is the minimum value in the original data. In the conversion process, a direct conversion method is used to randomly intercept a segment of N 2 signal from the one-dimensional signal, fill the signal values row by row into an N×N matrix, and then normalize the matrix to obtain the corresponding grayscale image.

[0017] Further, in step b, the gradient magnitude and gradient direction of the image are calculated. The gradient of the pixel point (x, y) is:

[0018] G x (x, y) = F(x + 1, y) - F(x - 1, y)

[0019] G y (x, y) = F(x, y + 1) - F(x, y - 1)

[0020] where G x (x, y)) represents the horizontal direction gradient of the pixel point, and G y (x, y) represents the vertical direction gradient of the pixel point; for the image pixel point (x, y), its gradient value G(x, y) is:

[0021]

[0022] The gradient direction θ(x, y) is:

[0023]

[0024] The input image is divided into several small cells using 8*8pixle / Cell. The gradient direction θ(x, y) of the cell is evenly divided into 9 intervals within the range of [0, π]. Any pixel gradient direction within the cell will fall within one of the 9 intervals. Taking the gradient magnitude of the pixel point as the weight, the weighted histogram of each gradient direction is calculated to obtain the feature vector of the cell direction gradient histogram. Every 4 (2*2) cells form a block. After concatenating all the direction gradient histogram features within the block and performing normalization, the feature vector of the block direction gradient histogram is obtained. Finally, all the feature vectors of the block direction gradient histograms are concatenated to obtain the concatenated feature, which is the structural feature of the bearing image;

[0025] Gray-level co-occurrence matrix feature extraction includes: First, perform normalization processing on the gray-level matrix:

[0026]

[0027] In the formula, P(i,j) represents how many adjacent paired points with gray-levels i and j in the image, θ represents four direction values of 0°, 45°, 90°, and 135°, d represents the separation distance from a pixel point with gray-level i, leaving a certain fixed position. In the gray-level co-occurrence matrix, there are a total of four different directions and two offset points, obtaining 14 possible attributes; Four texture parameters are used as texture features, including: energy, contrast, entropy, and correlation to obtain a higher recognition rate of bearing fault images, and the average value and variance are taken as the bearing features extracted therefrom.

[0028] Further, in step c, it is assumed that the probability distribution of the bearing usage duration satisfies a normal probability density function f(x) centered on the true usage duration, with the mathematical expectation being μ and the variance being δ 2 , for a true usage duration of x, the description degree of each usage duration label y to x In this way, the single-label of the usage duration is converted into a unified label distribution; for samples with a relatively short usage duration of less than one year, the normal probability density function is used to perform label distribution processing on the usage duration. If the sum of the probabilities of each usage duration is less than 1, normalization is used to make the sum of the probability values equal to 1.

[0029] Further, in step d, the output model of the labeled distribution extreme learning machine is β is the weight vector, h(x i ) is the vector that maps x i from the output space to the L-dimensional feature space; when solving, it is necessary to minimize the training error and make the norm of the output weight the smallest. After simplification, it is rewritten as min: s.t: ζ i = t i - f(x i ), i = 1, 2, ……, N. Among them, ||β|| 2 represents the structural risk, ζ i is the error between the output corresponding to the training sample and the true value, represents the empirical risk, C is the regularization parameter, and according to the Karush-Kuhn-Tucker conditions, the output weight β is obtained: The labeled output function is expressed as Since in the extreme learning machine, the matrix H is generated by random assignment, and different Hs are obtained each time, resulting in different output weight vectors β. By introducing the kernel function Ω into the extreme learning machine to replace the random matrix H, the stability of the extreme learning machine model is increased. The kernel matrix is defined using the Mercer condition: Ω ELM = h(x i )·h(x j ) = K(x i , x j ), K(x i , x j ) = exp(-γ||x i - x j || 2 ), and the output function is the kernel function, c is the regularization parameter, Ω is the kernel matrix. Substitute the training set and test set data into the calculation to obtain Ω test and Ω train ; Multiply the kernel matrix Ω test obtained in step d by the weight vector β to get the predicted output label distribution of the test set

[0030]

[0031] From the label set, the position with the maximum probability is the maximum probability usage duration, Time p .

[0032] Furthermore, in step e, the histogram of oriented gradients and the gray-level co-occurrence matrix are respectively input to obtain their respective classification accuracies:

[0033]

[0034] Then, the weight of the one with a higher classification accuracy accounts for 52% to 60%, and the weight of the one with a lower classification accuracy accounts for 48% to 40%. The finally fused feature is:

[0035] λ = [(m×χ) + (n×ω)]

[0036] where m and n are the features extracted from the histogram of oriented gradients and the gray-level co-occurrence matrix respectively, and χ and ω are the weights of the histogram of oriented gradients and the gray-level co-occurrence matrix respectively; After fusing the features extracted from the histogram of oriented gradients and the gray-level co-occurrence matrix, input them into the kernel extreme learning machine for training to obtain the predicted output label distribution.

[0037] Further, in step f, due to the influence of classifier performance and feature extraction, there will be errors in the final predicted distribution result. Therefore, the expected usage duration is added to the usage duration prediction estimate. In the age marker distribution, the normal probability density function is used to describe the probability of different usage durations. Here, μ is the position of the maximum probability in the usage duration distribution and also the expectation of the normal distribution. The predicted distribution usage duration can be expressed as Time pred：

[0038] Time pred =α*Time p +(1-α)*Time E

[0039] Time p is the usage duration with the maximum probability, which is the usage duration corresponding to the maximum probability in the usage duration marker distribution prediction. Time E is the expected value usage duration, which is the sum of the products of all usage durations and their corresponding predicted probabilities. α is the weight value that determines the ratio of the usage duration with the maximum probability to the expected value usage duration, and its value range is from 0 to 1.

[0040] Further, in step g, the remaining service life of the bearing is obtained by subtracting the predicted usage duration obtained in f from the service life years of the bearing.

[0041] Beneficial effects: The present invention proposes a method for predicting bearing life based on marker distribution, which has the following advantages:

[0042] (1) During the data conversion process, a direct conversion method is adopted, that is, a fixed-length signal is randomly intercepted from the one-dimensional vibration signal and filled into the corresponding matrix row by row, reducing data loss caused by traditional signal processing methods and improving the overall operation efficiency of the model.

[0043] (2) The edge features and local texture features of the bearing image are respectively extracted by using the histogram of oriented gradients and the gray-level co-occurrence matrix, and the features extracted by the two are fused, so that the edge and local features are combined, and the features of the image are more comprehensively characterized, greatly improving the prediction accuracy.

[0044] (3) The marker distribution learning is applied to the research of bearing life, and the probability density function is used to represent the bearing usage duration marker. The true usage duration marker is represented by the maximum probability, and the probability values of other durations decrease as the distance from the true duration marker increases. Using the duration marker distribution can more accurately describe the duration marker.

[0045] (4) The extreme learning machine algorithm is introduced, and the kernel extreme learning machine algorithm in the regression mode improves the operation speed. The setting of the kernel parameter can solve for a more stable network output weight value, avoiding the result fluctuation caused by random parameters, and enhancing the stability of its model based on the extreme learning machine.

[0046] (5) It is represented by the maximum probability usage duration and the expected value duration together, balancing the prediction results in the duration estimation, reducing the errors caused by reasons such as the classifier, and making the prediction results more accurate. Description of the Drawings

[0047] Figure 1 is a general algorithm flowchart for bearing life prediction based on label distribution;

[0048] Figure 2 is the basic schematic diagram of converting vibration signals into grayscale images;

[0049] Figure 3 is a schematic diagram of calculating the gray-level co-occurrence matrix;

[0050] Figure 4 is the flowchart of feature extraction of histogram of oriented gradients;

[0051] Figure 5 is the label distribution diagram;

[0052] Figure 6 is the flowchart of the label distribution learning algorithm of the kernel extreme learning machine; Detailed Implementation Modes

[0053] The present invention will be further described below with reference to the accompanying drawings:

[0054] The present invention mainly aims at the problems of low efficiency in obtaining image data, low accuracy in feature extraction, complex iterative operations and large time consumption in traditional life prediction methods. The present invention conducts bearing life prediction by using label distribution learning based on the kernel extreme learning machine. The features of the histogram of oriented gradients and the gray-level co-occurrence matrix are fused to extract more comprehensive characterizations; label distribution learning can accurately describe the importance degree of each label, and combining it with the kernel extreme learning machine can ensure better accuracy while also shortening the running time better.

[0055] For the bearing life prediction based on label distribution of the present invention, the prediction data uses the test bench data set of PRONOSTIA, as Figure 1 gives the main process of mechanical bearing life prediction. It mainly includes the following contents:

[0056] (1) Since the bearing is in a normal operating environment and its lifespan can reach thousands or tens of thousands of hours, and the actual degradation process is quite slow, it will take a long time to collect data on the entire degradation process of the bearing. Therefore, this experiment uses the publicly available PHM2012 dataset to study the remaining lifespan of the bearing. The experimental data in the dataset comes from the test bench of PRONOSTIA. In the first working condition, it runs at 1800 revolutions per minute with a load of 4000 N, and 7 bearings are tested; in the second working condition, the rotational speed is 1650 revolutions per minute and the load is 4200 N, and 7 bearings are also tested; in the third working condition, it is 1500 revolutions per minute with a load of 5000 N, and 3 bearings are tested. The data of the first two bearings in each working condition is complete and is set as training data, that is, all the data generated from the start of the bearing operation to the end of degradation. The data of the remaining bearings is set as the training dataset, which lacks the data of the last period of time.

[0057] (2) The basic schematic diagram of converting the vibration signal into a grayscale image is as Figure 2 shown. The original signals of the training set and the test set are preprocessed. According to the min-max normalization method, normalization is performed, and the result is mapped between [0, 255]. The specific formula is:

[0058]

[0059] where X is the original data, X nrom is the result after normalization, X max is the maximum value in the original data, X min is the minimum value in the original data. Here, in order to eliminate decimals, the round function is used to round the values. For the normalized data, the direct conversion method is adopted. The one-dimensional signal is intercepted according to a fixed length and filled into the two-dimensional matrix row by row to be converted into the corresponding two-dimensional grayscale image.

[0060] (3) The obtained grayscale images are subjected to feature extraction using the histogram of oriented gradients method and the gray-level co-occurrence matrix. Figure 3 is the schematic diagram of the calculation of the gray-level co-occurrence matrix. Figure 4 is the flow chart of the feature extraction of the histogram of oriented gradients.

[0061] (4) The probability function is used to mark the distribution of the bearing usage time. It is assumed that the probability distribution of the bearing usage time satisfies the normal probability density function f(x) centered on the actual usage time, with the mathematical expectation μ and the variance δ 2 , for an actual usage time of x, the description degree of each usage time mark y to x Thus, the single mark of the usage time is converted into a unified mark distribution, as Figure 5 shown.

[0062] (5) Combine the processed labeled distribution dataset with the features to form a new feature dataset and process it using a kernel extreme learning machine to obtain the maximum probability predicted usage duration. Figure 6 It is the algorithm flowchart of the kernel extreme learning machine.

[0063] (6) Calculate the accuracy of the prediction results obtained by the two feature extraction methods, reasonably allocate weights to obtain the fused features, and put them into the kernel extreme learning machine again for prediction.

[0064] (7) Add the expected value to the usage duration prediction, and jointly determine the usage duration with the maximum probability predicted usage duration.

[0065] (8) Subtract the predicted bearing usage duration from the number of years of the bearing service life to obtain the predicted bearing life.

[0066] A method for predicting the remaining service life of a bearing based on label distribution provided by the present invention:

[0067] It includes the following steps:

[0068] Step a: Divide the dataset of bearing information into a training set and a test set, preprocess the bearing vibration signal, normalize the vibration signal, and store it as a two-dimensional grayscale image.

[0069] Step b: Extract features from the grayscale image obtained in step a using the Histogram of Oriented Gradient (HOG) method and the Gray-Level Co-occurrence Matrix (GLCM).

[0070] Step c: Perform label distribution processing on the bearing usage duration label using a probability function.

[0071] Step d: Combine the processed labeled distribution training set and test set in step c with the features to form a new feature dataset and process it using a Kernel Based Extreme Learning Machine (KELM), calculate the kernel matrix and weights, multiply the obtained outputs to obtain the predicted output label distribution, and obtain the maximum probability predicted usage duration.

[0072] Step e: Calculate the accuracy rates obtained by the kernel extreme learning machine for the features of the Histogram of Oriented Gradient and the Gray-Level Co-occurrence Matrix respectively, then assign higher weights to those with higher classification accuracy rates and lower weights to those with lower classification accuracy rates to obtain the fused features, and process the fused features using the kernel extreme learning machine.

[0073] Step f, add the expected value to the predicted usage duration, and jointly determine the usage duration with the usage duration predicted with the highest probability in Step d;

[0074] Step g, subtract the predicted bearing usage duration obtained in Step f from its service life to obtain its remaining service life.

[0075] Furthermore, in Step a, the min-max normalization method is used to normalize the bearing signal,

[0076] and map its result to the range [0, 255]. The specific formula is:

[0077]

[0078] where X is the original data, Xnorm is the result after normalization, Xmax is the maximum value in the original data, Xmin is the minimum value in the original data. Here, to eliminate decimals, the round function is used to round the numerical values. During the conversion process, the direct conversion method is adopted. A segment of signal of length N 2 is randomly intercepted from the one-dimensional signal, and the signal values are filled into the N×N matrix row by row, that is, every N signal values can be filled into one row of the matrix, and then the matrix is normalized to obtain the corresponding grayscale image.

[0079] Furthermore, in Step b, calculate the gradient magnitude and gradient direction of the image. The gradient of the pixel point (x, y) is:

[0080] G x (x,y) = F(x + 1, y) - F(x - 1, y)

[0081] G y (x,y) = F(x, y + 1) - F(x, y - 1)

[0082] where G x (x,y)) represents the horizontal direction gradient of the pixel point, and G y (x,y) represents the vertical direction gradient of the pixel point; for the image pixel point (x, y), its gradient value G(x, y) is:

[0083]

[0084] The gradient direction θ(x, y) is:

[0085]

[0086] The input image is divided into several small cells with 8*8 pixels / Cell. The gradient direction θ(x,y) of the cell is evenly divided into 9 intervals (bins) within the range of [0, π]. Any pixel gradient direction within the cell will fall within one of the 9 intervals. Taking the gradient magnitude of the pixel as the weight, calculate the weighted histogram of each gradient direction to obtain the feature vector of the cell's histogram of oriented gradients. Combine every 4 (2*2) cells into a block, concatenate all the histogram of oriented gradient features within the block and then perform normalization to obtain the feature vector of the block's histogram of oriented gradients. Finally, concatenate the histogram of oriented gradient features of all blocks to obtain the concatenated features, which are the structural features of the bearing image;

[0087] The extraction of the gray-level co-occurrence matrix features includes: First, perform normalization on the gray-level matrix:

[0088]

[0089] In the formula, P(i,j) represents how many adjacent pairs of points with gray-levels i and j in the image, θ represents four direction values of 0°, 45°, 90°, and 135°, d represents the separation distance from a pixel point with gray-level i, leaving a certain fixed position. In the gray-level co-occurrence matrix, there are a total of four different directions and two offset points, obtaining 14 possible attributes; To obtain a higher recognition rate of bearing fault images, four texture parameters are used as texture features, including: energy (Asm), contrast (Con), entropy, and correlation (Cor) to obtain a higher recognition rate of bearing fault images. Take the average value and variance as the bearing features extracted.

[0090] Further, in step c, assume that the probability distribution of the bearing usage duration satisfies a normal probability density function f(x) centered on the true usage duration, with the mathematical expectation μ and variance δ 2 , for a true usage duration of x, the description degree of each usage duration label y to x In this way, the single-label of the usage duration is converted into a unified label distribution; For samples with a usage duration less than one year, which are relatively short, use the normal probability density function to perform label distribution processing on the usage duration. If the sum of the probabilities of each usage duration is less than 1, perform normalization to make the sum of the probability values equal to 1, which can ensure the accuracy of the maximum value and the expected value in subsequent predictions.

[0091] Further, in step d, the output model of the label distribution extreme learning machine is β is the weight vector, h(x i ) is x iA vector mapping from the output space to the L-dimensional feature space; when solving, it is necessary to minimize the training error and minimize the norm of the output weights. After simplification, it is rewritten as min: s.t: ζ i = t i - f(x i ), i = 1, 2, ……, N. Among them, ||β|| 2 represents the structural risk, ζ i is the error between the output corresponding to the training sample and the true value, represents the empirical risk, C is the regularization parameter. According to the Karush-Kuhn-Tucker conditions, the output weight β is obtained: The labeled output function is expressed as Since in the extreme learning machine, the matrix H is generated by random assignment, and each time the obtained H is different, the output weight β is also different. To increase its stability, the kernel function Ω is introduced into the extreme learning machine to replace the random matrix H, which increases the stability of the extreme learning machine model. The kernel matrix is defined using the Mercer condition: Ω ELM = h(x i )·h(x j ) = K(x i , x j ), K(x i , x j ) = exp(-γ||x i - x j || 2 ). The output function is obtained K(x i , x j ) is the kernel function, c is the regularization parameter, Ω is the kernel matrix. The training set and test set data are substituted into the calculation to obtain Ω test and Ω train ; The kernel matrix Ω test obtained in step d is multiplied by the weight β to obtain the predicted output label distribution of the test set

[0092]

[0093] From the label set, the position with the highest probability is the maximum probability usage duration, Time p .

[0094] Furthermore, in step e, the histogram of oriented gradients and the gray-level co-occurrence matrix are respectively input to obtain their respective classification accuracies:

[0095]

[0096] Then, the weight with a high classification accuracy accounts for 52% to 60%, and the weight with a low classification accuracy accounts for 48% to 40%. The finally fused features are as follows:

[0097] λ = [(m × χ) + (n × ω)]

[0098] where m and n are the features extracted from the histogram of oriented gradients and the gray-level co-occurrence matrix respectively, and χ and ω are the weights of the histogram of oriented gradients and the gray-level co-occurrence matrix respectively; the features extracted from the histogram of oriented gradients and the gray-level co-occurrence matrix are fused and then input into the kernel extreme learning machine for training to obtain the predicted output label distribution.

[0099] Further, in step f, due to the influence of the classifier performance and feature extraction, there will be errors in the final predicted distribution result. Therefore, the expected usage duration is added to the usage duration prediction estimate. In the age label distribution, the normal probability density function is used to describe the probability of different usage durations. Here, μ is the position of the maximum probability in the usage duration distribution and also the expectation of the normal distribution. The predicted distribution usage duration can be expressed as Time pred :

[0100] Time pred = α * Time p + (1 - α) * Time E

[0101] Time p is the maximum probability usage duration, which is the usage duration corresponding to the maximum probability of the usage duration label distribution. Time E is the expected value usage duration, which is the sum of the products of all usage durations and their corresponding predicted probabilities. α is the weight value that determines the ratio of the maximum probability usage duration to the expected value usage duration, and its value range is from 0 to 1.

[0102] Further, in step g, the remaining service life of the bearing is obtained by subtracting the predicted usage duration obtained in f from the service life years of the bearing.

[0103] The above is only the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A bearing life prediction method based on label distribution, characterized in that, it includes the following steps: Step a, divide the data set of bearing information into a training set and a test set, preprocess the bearing vibration signal, normalize the vibration signal, and store it as a two-dimensional grayscale image; Step b, extract features from the grayscale image obtained in Step a using the histogram of oriented gradients method and the gray-level co-occurrence matrix; Step c, perform label distribution processing on the bearing usage duration label using a probability function; Step d, merge the processed label distribution training set and test set in Step c with the features to form a new feature data set and process it using a kernel extreme learning machine, calculate the kernel matrix and weights, multiply the obtained outputs, obtain the predicted output label distribution, and obtain the maximum probability predicted usage duration; Step e, calculate the accuracy rates obtained by the histogram of oriented gradients and the gray-level co-occurrence matrix features through the kernel extreme learning machine respectively, then assign higher weights to those with higher classification accuracy rates and lower weights to those with lower classification accuracy rates to obtain the fused features, and process the fused features using the kernel extreme learning machine; Step f, add the expected value to the usage duration prediction and jointly determine the usage duration with the maximum probability predicted usage duration in Step d; Step g, subtract the predicted bearing usage duration obtained in Step f from its service life to obtain its remaining service life.

2. The bearing life prediction method based on label distribution according to claim 1, characterized in that, in the said Step a, the min-max normalization method is used to normalize the bearing signal, and its result is mapped to between [0, 255], and the specific formula is: Among them, X is the original data, Xnorm is the result after normalization, Xmax is the maximum value in the original data, Xmin is the minimum value in the original data. During the conversion process, a direct conversion method is adopted. A signal segment of N 2 is randomly intercepted from the one-dimensional signal, and the signal values are filled into the N×N matrix row by row, and then the matrix is normalized to obtain the corresponding grayscale image.

3. The bearing life prediction method based on label distribution according to claim 1, characterized in that, in the said Step b, calculate the gradient magnitude and gradient direction of the image, and the gradient of the pixel point (x, y) is: G x (x,y) = F(x + 1,y) - F(x - 1,y) G y (x,y) = F(x,y + 1) - F(x,y - 1) Among them, G x (x, y)) represents the horizontal gradient of the pixel point, and G y (x, y) represents the vertical gradient of the pixel point; for the image pixel point (x, y), its gradient value G(x, y) is: The gradient direction θ(x, y) is: Divide the input image into several small cells with 8*8pixle / Cell, divide the cell gradient direction θ(x, y) evenly into 9 intervals within the range of [0, π], and the gradient direction of any pixel within the cell will surely fall within the 9 intervals. Take the gradient magnitude of this pixel point as the weight, calculate the weighted histogram of each gradient direction, obtain the feature vector of the cell histogram of oriented gradients, form a block with every 4 (2*2) cells, normalize after concatenating all the histogram of oriented gradients features within the block to obtain the feature vector of the block histogram of oriented gradients, and finally concatenate all the block histogram of oriented gradients features to obtain the concatenated feature, which is the structural feature of the bearing image; The extraction of gray-level co-occurrence matrix features includes: first, normalize the gray-level matrix: Where P(i,j) represents how many adjacent pairs of points in the image have gray-levels of i and j, θ represents the four direction values of 0°, 45°, 90°, and 135°, d represents the separation distance from a pixel point with gray-level i, leaving a certain fixed position. In the gray-level co-occurrence matrix, there are a total of four different directions and two offset points, obtaining 14 possible attributes; four texture parameters are used as texture features, including: energy, contrast, entropy, and correlation to obtain a higher recognition rate of bearing fault images, and the average value and variance are taken as the bearing features extracted therefrom.

4. The bearing life prediction method based on label distribution according to claim 1, characterized in that In step c, it is assumed that the probability distribution of the bearing service life satisfies the normal probability density function f(x) centered on the true service life, with the mathematical expectation μ and the variance δ 2 , for a true service life of x, the degree of description of each service life label y for x In this way, the single label of the service life is converted into a unified label distribution; for samples with a shorter service life of less than one year, the normal probability density function is used to process the label distribution of the service life. If the sum of the probabilities of each service life is less than 1, normalization is used to make the sum of the probability values equal to 1.

5. The bearing life prediction method based on label distribution according to claim 1, characterized in that In the said step d, the output model of the labeled distribution extreme learning machine is β is the weight vector, and h(x i ) is the vector that maps x i from the output space to the L-dimensional feature space; when solving, it is necessary to minimize the training error and minimize the norm of the output weights, which is simplified and rewritten as min: s.t: ζ i = t i - f(x i ), i = 1, 2,......., N where, ||β|| 2 represents the structural risk, ζ i is the error between the output corresponding to the training sample and the true value, represents the empirical risk, C is the regularization parameter, and according to the Karush-Kuhn-Tucker conditions, the output weight β is obtained: The labeled output function is expressed as Since in the extreme learning machine, the matrix H is generated by random assignment, and each obtained H is different, and the output weight β is also different. The kernel function Ω is introduced into the extreme learning machine to replace the random matrix H, which increases the stability of the extreme learning machine model. The kernel matrix is defined using the Mercer condition: Ω ELM = h(x i )·h(x j ) = K(x i , x j ), K(x i , x j ) = exp(-γ||x i - x j || 2 ), and the output function K(x i , x j ) is the kernel function, c is the regularization parameter, Ω is the kernel matrix, and the training set and test set data are substituted for calculation to obtain Ω test and Ω train ; Multiply the kernel matrix Ω test obtained in step d by the weight β to obtain the predicted output label distribution of the test set From the tag set, the position with the highest probability is the maximum probability usage duration, Time p .

6. The bearing life prediction method based on label distribution according to claim 1, characterized in that in step e, the histogram of oriented gradients and the gray-level co-occurrence matrix are respectively input to obtain their respective classification accuracies: Then, the weight of the one with a high classification accuracy accounts for 52% to 60%, and the weight of the one with a low classification accuracy accounts for 48% to 40%. The finally fused feature is: λ = [(m×χ)+(n×ω)] where m and n are the features extracted from the histogram of oriented gradients and the gray-level co-occurrence matrix respectively, and χ and ω are the weights of the histogram of oriented gradients and the gray-level co-occurrence matrix respectively; The features extracted from the histogram of oriented gradients and the gray-level co-occurrence matrix are fused and then input to a kernel extreme learning machine for training to obtain the predicted output label distribution.

7. The bearing life prediction method based on label distribution according to claim 1, characterized in that In step f, due to the influence of classifier performance and feature extraction, there will be errors in the final predicted distribution result. Therefore, the expected usage duration is added to the usage duration prediction estimate. In the age label distribution, the normal probability density function is used to describe the probability of different usage durations. Here, μ is the position of the maximum probability in the usage duration distribution and also the expectation of the normal distribution. The predicted distribution usage duration can be expressed as Time pred : Time pred = α * Time p + (1 - α) * Time E Time p is the maximum probability usage duration, which is the usage duration corresponding to the maximum probability predicted by the usage duration marker distribution, Time E is the expected value usage duration, which is the sum of the products of all usage durations and their corresponding predicted probabilities. α is a weight value that determines the ratio of the maximum probability usage duration to the expected value usage duration, and its value range is from 0 to 1.

8. The bearing life prediction method based on label distribution according to claim 1, characterized in that in step g, the remaining service life of the bearing is obtained by subtracting the predicted service life obtained in f from the number of years of the bearing's service life.

Images