CN108760266A - The virtual degeneration index building method of mechanical key component based on learning distance metric - Google Patents

Abstract

A method for constructing virtual degradation indicators of key mechanical components based on distance metric learning. Firstly, the time domain, frequency domain and time-frequency domain characteristics of vibration signals of key mechanical components are extracted, and at the same time, the degradation status of key mechanical components is divided according to the frequency spectrum and power spectrum. Comprehensively evaluate the correlation, monotonicity and predictability of each index, select the index whose performance is better than the root mean square value to form the eigenvector of the key mechanical parts and perform distance metric learning, and then use the eigenvector in the normal stage to train the optimized self-organization Map the neural network, input the newly acquired vibration signal data, and calculate the distance between its feature vector and the weight vector of the corresponding activation node, so as to establish the virtual degradation index of enhanced minimum quantization error; Degradation information of key components of excavation machinery equipment is beneficial to improve the accuracy of remaining life prediction.

Description

Construction Method of Virtual Degradation Index of Mechanical Key Components Based on Distance Metric Learning

technical field

The invention belongs to the technical field of remaining life prediction and health management of mechanical equipment, and in particular relates to a method for constructing virtual degradation indicators of key mechanical components based on distance metric learning.

Background technique

Mechanical equipment often works in a complex and changeable environment, and its key components fail frequently. With the development of modern technology, the coupling relationship between key components of mechanical equipment is getting closer and closer. Once a component fails, it will cause the entire The failure or even paralysis of the mechanical system will cause serious economic losses and even casualties. Therefore, it is imminent to predict the remaining life of key mechanical components so that they can be maintained before failure and ensure the safe service of mechanical equipment.

The remaining life prediction of key components of mechanical equipment mainly includes monitoring signal acquisition, physical degradation index extraction, index evaluation and optimization, virtual degradation index construction and remaining life evaluation. In addition to being affected by the selected prediction model, the accuracy of remaining life prediction results is also closely related to the degradation indicators used. Excellent degradation indicators need to have good correlation, monotonicity and predictability, but affected by the quality of the original signal and signal processing methods, the physical degradation indicators directly extracted from the monitoring signal are often only valid for a certain stage of the degradation process. It is more sensitive and it is difficult to maintain a good trend throughout the degradation process. At the same time, the working environment of mechanical equipment is complex and changeable, and the physical degradation index is greatly affected by the working conditions, which is not conducive to the expression of degradation information of key mechanical components. The above shortcomings will lead to a reduction in the accuracy of early health monitoring and remaining life of key mechanical components. Therefore, it is very important to construct a degradation index with excellent comprehensive performance for the accuracy of remaining life prediction of key mechanical components.

Contents of the invention

In order to overcome the above shortcomings of the prior art, the present invention provides a method for constructing a virtual degradation index of key mechanical components based on distance metric learning. An index with superior comprehensive performance is selected from a set of alternative physical degradation indexes through index evaluation. The learning and self-organizing neural network algorithm maps the preferred physical degradation index to a single virtual degradation index, characterizes the degree of deviation of the key mechanical parts from the normal state, and describes the degradation process of the key mechanical parts.

In order to achieve the above object, the technical scheme that the present invention takes is:

A method for constructing virtual degradation indicators of key mechanical components based on distance metric learning, comprising the following steps:

1) Perform Fourier transform sequentially on the vibration signals collected during the life cycle of key mechanical components to obtain the frequency spectrum and power spectrum of the vibration signal. The degradation process of components is divided into three stages: normal operation, fault development and severe degradation;

2) Extract the physical degradation indicators of the vibration signal from the time domain, frequency domain, and time-frequency domain respectively to form a set of alternative physical degradation indicators; the specific steps are as follows:

2.1) Extract the time-domain degradation indicators of the vibration signal, which are mean value, standard deviation, variance, skewness, kurtosis, maximum value, minimum value, peak-to-peak value, average amplitude, root mean square value, waveform index, peak index , pulse index, margin index, skewness index, and kurtosis index, which are recorded as F1-F16;

2.2) Extract the frequency-domain degradation index of the vibration signal, which are frequency-domain average energy, frequency-domain energy variance, mean frequency, root-mean-square frequency, and spectrum dispersion degree, and record them as F17-F21;

2.3) Extract the time-frequency domain degradation index of the vibration signal, obtain the first 8 eigenmode components of the vibration signal through empirical mode decomposition, and calculate the energy of the 8 eigenmode components and the energy entropy of the empirical mode decomposition in turn, and Record it with F22-F30;

3) Establish a comprehensive evaluation method for degradation indicators, including correlation criteria, monotonicity criteria, and predictability criteria, and quantitatively evaluate the performance of the 30 physical degradation indicators extracted in step 2) according to formulas (1) to (4) ;

3.1) Correlation criterion based on Spearman correlation coefficient

Among them, Y=(y ₁ ,y ₂ ,…,y _L ) and t=(t ₁ ,t ₂ ,…,t _L ) are the degradation index sequence and time sequence respectively, and L is the length of the degradation index sequence, and are the average values of the degradation index sequence and the time series, respectively. This criterion describes the correlation between the degradation index and the time series. The closer the value is to 1, the stronger the correlation between the two;

3.2) Monotonicity criterion

Among them, Y=(y ₁ ,y ₂ ,…,y _L ) is the degradation index sequence, L is the length of the degradation index sequence, ε(x) is the unit step function, and this criterion describes the monotonous increase or monotonous decrease of the degradation index The closer the value is to 1, the better the monotonicity of the degradation index and the closer it is to the actual degradation of key mechanical components;

3.3) Predictive criteria

in, are the mean values of the degradation index at the initial time and failure time respectively, y _f is the degradation index value, σ(y _f ) is the standard deviation of the degradation index at the failure time, this criterion describes the variation range of the degradation index in the whole life cycle and The dispersion at the failure time, the closer the value is to 1, the larger the variation range of the degradation index and the smaller the standard deviation at the failure time, the more consistent the variation range and failure threshold of the index between different individuals, the more suitable for remaining life prediction;

3.4) Comprehensive evaluation criteria

Comprehensive evaluation criteria The above evaluation criteria are linearly weighted and combined, as shown in the following formula (4):

SM＝ω ₁ Corr(Y)+ω ₂ Mon(Y)+ω ₃ Pro(Y) (4)

Among them, SM is the comprehensive evaluation criterion, Y is the index sequence, ω ₁ +ω ₂ +ω ₃ =1, and ω ₁ ,ω ₂ ,ω ₃ ∈[0,1] are used to represent the weight of the three evaluation criteria;

4) According to the size of the comprehensive evaluation criterion SM in step 3) of the 30 physical degradation indicators, select the degradation indicators with excellent performance to form the eigenvectors of the vibration signals of key mechanical components;

5) The vibration signal eigenvectors of key mechanical components in the three stages have different labels, and use this information for distance metric learning to obtain a distance metric matrix suitable for measuring the state-space similarity in the degradation process of key mechanical components;

6) Utilize the learned distance metric matrix to measure the similarity of the self-organizing map neural network competition learning stage, so that the network is optimized;

7) Use the vibration signal eigenvector in the normal stage to train the self-organizing map neural network, determine the neural network structure and parameters, and use the Mexican sombrero function for weight adjustment;

8) In the test phase, the feature vector of the newly acquired vibration signal is used as the input of the self-organizing map neural network, and the distance between the real-time feature vector and the corresponding activation node weight vector is calculated according to formula (5), and the virtual degradation index enhancement of the key mechanical parts is obtained minimum quantization error;

Among them, EMQE is the enhanced minimum quantization error index, X is the feature vector of the newly acquired vibration signal, m is the weight vector corresponding to the activation node in the self-organizing map neural network, and A is the distance metric matrix obtained through distance metric learning.

The beneficial effects of the present invention are:

The present invention constructs virtual degradation indicators of key mechanical components based on distance metric learning and self-organizing mapping neural network, integrates state information of various physical degradation indicators, and scientifically divides the degradation stages of key mechanical components and optimizes physical degradation indicators with excellent performance , get the eigenvectors of different degradation stages, and obtain the distance metric matrix suitable for measuring the similarity of the state space of key mechanical parts through distance metric learning, and calculate the relationship between real-time eigenvectors and network activation node weight vectors through the improved self-organizing map neural network The distance is obtained to minimize the quantization error by enhancing the virtual degradation index of mechanically critical components. This index can well characterize the degradation process of key mechanical components, and applying it to remaining life prediction and health management can effectively improve the accuracy of prediction results.

Description of drawings

Fig. 1 is the flowchart of the method of the present invention.

Fig. 2 shows the two physical degradation indexes with the best overall performance of the rolling bearing B1 _of the embodiment.

Fig. 3 shows the two physical degradation indexes with the worst comprehensive performance of the rolling bearing B1 _of the embodiment.

Fig. 4 is a comparison of the enhanced minimum quantization error index and the minimum quantization error index of the rolling bearing B1 _of the embodiment.

Detailed ways

The present invention will be further elaborated below in conjunction with the accompanying drawings and embodiments.

Referring to Figure 1, a method for constructing virtual degradation indicators of key mechanical components based on distance metric learning includes the following steps:

1) Perform Fourier transform sequentially on the vibration signals collected during the life cycle of key mechanical components to obtain the frequency spectrum and power spectrum of the vibration signal. The degradation process of components is divided into three stages: normal operation, fault development and severe degradation;

2) Extract the physical degradation indicators of the vibration signal from the time domain, frequency domain, and time-frequency domain respectively to form a set of candidate physical degradation indicators. Specific steps are as follows:

2.1) Extract the time-domain degradation indicators of the vibration signal, which are mean value, standard deviation, variance, skewness, kurtosis, maximum value, minimum value, peak-to-peak value, average amplitude, root mean square value, waveform index, peak index , pulse index, margin index, skewness index, and kurtosis index, which are recorded as F1-F16;

2.2) Extract the frequency-domain degradation index of the vibration signal, which are frequency-domain average energy, frequency-domain energy variance, mean frequency, root-mean-square frequency, and spectrum dispersion degree, and record them as F17-F21;

2.3) Extract the time-frequency domain degradation index of the vibration signal, obtain the first 8 eigenmode components of the vibration signal through empirical mode decomposition, and calculate the energy of the 8 eigenmode components and the energy entropy of the empirical mode decomposition in turn, and Record it with F22-F30;

3) Establish a comprehensive evaluation method for degradation indicators, including correlation criteria, monotonicity criteria, and predictability criteria, and quantitatively evaluate the performance of the 30 physical degradation indicators extracted in step 2) according to formulas (1) to (4) ;

3.1) Correlation criterion based on Spearman correlation coefficient

Among them, Y=(y ₁ ,y ₂ ,…,y _L ) and t=(t ₁ ,t ₂ ,…,t _L ) are the degradation index sequence and time sequence respectively, and L is the length of the degradation index sequence, and are the average values of the degradation index sequence and the time series, respectively. This criterion describes the correlation between the degradation index and the time series. The closer the value is to 1, the stronger the correlation between the two;

3.2) Monotonicity criterion

Among them, Y=(y ₁ ,y ₂ ,…,y _L ) is the degradation index sequence, L is the length of the degradation index sequence, ε(x) is the unit step function, and this criterion describes the monotonous increase or monotonous decrease of the degradation index The closer the value is to 1, the better the monotonicity of the degradation index and the closer it is to the actual degradation of key mechanical components;

3.3) Predictive criteria

in, are the mean values of the degradation index at the initial time and failure time respectively, y _f is the degradation index value, σ(y _f ) is the standard deviation of the degradation index at the failure time, this criterion describes the variation range of the degradation index in the whole life cycle and The dispersion at the failure time, the closer the value is to 1, the larger the variation range of the degradation index and the smaller the standard deviation at the failure time, the more consistent the variation range and failure threshold of the index between different individuals, the more suitable for remaining life prediction;

3.4) Comprehensive evaluation criteria

Comprehensive evaluation criteria The above evaluation criteria are linearly weighted and combined, as shown in the following formula (4):

SM＝ω ₁ Corr(Y)+ω ₂ Mon(Y)+ω ₃ Pro(Y) (4)

Among them, SM is the comprehensive evaluation criterion, Y is the index sequence, ω ₁ +ω ₂ +ω ₃ =1, and ω ₁ ,ω ₂ ,ω ₃ ∈[0,1] are used to represent the weight of the three evaluation criteria;

4) According to the size of the comprehensive evaluation criteria SM of 30 physical degradation indicators in 3), select the degradation indicators with excellent performance to form the eigenvectors of the vibration signals of key mechanical components;

5) The vibration signal eigenvectors of key mechanical components in the three stages have different labels, and use this information for distance metric learning to obtain a distance metric matrix suitable for measuring the state-space similarity in the degradation process of key mechanical components;

6) Utilize the learned distance metric matrix to measure the similarity of the self-organizing map neural network competition learning stage, so that the network is optimized;

7) Use the vibration signal eigenvector in the normal stage to train the self-organizing map neural network, determine the neural network structure and parameters, and use the Mexican sombrero function for weight adjustment;

8) In the test phase, the feature vector of the newly acquired vibration signal is used as the input of the self-organizing map neural network, and the distance between the real-time feature vector and the corresponding activation node weight vector is calculated according to formula (5), and the virtual degradation index enhancement of the key mechanical parts is obtained minimum quantization error;

Among them, EMQE is the enhanced minimum quantization error index, X is the feature vector of the newly acquired vibration signal, m is the weight vector corresponding to the activation node in the self-organizing map neural network, and A is the distance metric matrix obtained through distance metric learning.

Rolling bearings are a common key mechanical component. In order to further prove the effectiveness of the virtual degradation index construction method for key mechanical components based on distance metric learning, the full life cycle data of rolling bearings obtained through accelerated life experiments on the PRONOSTIA test bench are used for verification. The test bench mainly includes three parts: rotation, loading and testing. The rotating part includes an asynchronous motor, a gearbox and a rotating shaft. The loading part accelerates the degradation of the bearing through the radial force applied to the test bearing, and the testing part collects vibration acceleration signals. , where the sampling frequency f _s of the vibration acceleration signal is 25600Hz, the number of sampling points N is 2560, the duration of each sampling is 0.1s, and the time interval between two adjacent samplings is 10s. The vibration signals of 7 groups of rolling bearings were collected under the working condition of rotating speed 1800rpm and load 4000N, which are recorded as B ₁ ~ B ₇ respectively.

Firstly, the time domain, frequency domain and time-frequency domain degradation indicators of rolling bearings are extracted, and the degradation stages of bearings are divided into three types: normal, fault development and severe degradation according to the frequency spectrum and power spectrum; secondly, each indicator is comprehensively evaluated using formula (4) The correlation, monotonicity and predictability of are set in the present invention The performance of each extracted physical index was quantified, and the evaluation results are shown in Table 1.

Table 1

According to the results of the comprehensive evaluation, Fig. 2 shows the _first two physical degradation indexes of the bearing B1 comprehensive evaluation criterion SM optimal, which are the third eigenmode component energy index and the fourth eigenmode component energy Indicators, Figure 3 shows the worst two physical degradation indicators _of the bearing B1 comprehensive evaluation criteria, which are the mean frequency index and the mean value index. According to the principle that the comprehensive performance is not lower than the root mean square value index, nine physical degradation indexes are determined to form the eigenvector X _t = (x _t,1 ,x _t,2 ,…,x _t,9 ) of the rolling bearing state space, Among them, x _t,1 , x _t,2 ,…,x _t,9 respectively represent the energy index of the third eigenmode component, the energy index of the fourth eigenmode component, the root mean square frequency index, Spectrum dispersion index, empirical mode decomposition energy entropy index, peak-to-peak index, minimum index, maximum index and root mean square index. Then, the feature vectors of three different stages are used for distance metric learning to obtain a distance metric matrix A with a dimension of 9×9, which is used to measure the similarity between the input vector and the weight vector in the self-organizing map neural network. Finally, use the eigenvectors of the normal stage to train the self-organizing map neural network, the trained neural network represents the state space of the rolling bearing in the normal stage, use the newly acquired eigenvector of the vibration signal of the rolling bearing as the input of the network, and calculate the input vector The distance between the weight vector and the corresponding activation node, constructs a virtual degradation index to enhance the minimum quantization error. Fig. 4 is a comparison diagram of the minimum quantization error index and the enhanced minimum quantization error index constructed by using the Euclidean distance for similarity measurement _on the bearing B1. Table 2 shows the evaluation results of the two indicators. It can be seen that the correlation, monotonicity and comprehensive evaluation criteria of the enhanced minimum quantization error index are significantly better than the minimum quantization error index, but slightly inferior to the minimum quantization error index in terms of predictive criteria. Quantified error indicators, which may be the reason for the low number of bearings. To sum up, the method proposed by the present invention can better reflect the degradation process of key mechanical components and help improve the accuracy of remaining life prediction.

Table 2

The method for constructing virtual degradation indexes of key mechanical components based on distance metric learning proposed by the present invention can be applied to the construction of virtual degradation indexes of various key components of mechanical equipment. In practical applications, implementers can specifically extract corresponding physical degradation indicators to form corresponding candidate physical degradation indicator sets according to the degradation characteristics of key components of various types of mechanical equipment. The method proposed by the invention helps to improve the accuracy of remaining life prediction of key components of mechanical equipment. It should be pointed out that without departing from the concept of the present invention, adjustments and deformations made to the method should also be regarded as the protection scope of the present invention.

Claims (1)

1. A method for constructing virtual degradation indicators of key mechanical components based on distance metric learning, which is characterized in that it is suitable for the construction of virtual degradation indicators of various key components of mechanical equipment. For specific degradation characteristics, the corresponding physical degradation indicators are extracted in a targeted manner to form an alternative physical degradation indicator set, including the following steps: 1) Perform Fourier transform sequentially on the vibration signals collected during the life cycle of key mechanical components to obtain the frequency spectrum and power spectrum of the vibration signal. The degradation process of components is divided into three stages: normal operation, fault development and severe degradation; 2) Extract the physical degradation indicators of the vibration signal from the time domain, frequency domain, and time-frequency domain respectively to form a set of alternative physical degradation indicators; the specific steps are as follows: 2.1) Extract the time-domain degradation indicators of the vibration signal, which are mean value, standard deviation, variance, skewness, kurtosis, maximum value, minimum value, peak-to-peak value, average amplitude, root mean square value, waveform index, peak index , pulse index, margin index, skewness index, and kurtosis index, which are recorded as F1-F16; 2.2) Extract the frequency-domain degradation index of the vibration signal, which are frequency-domain average energy, frequency-domain energy variance, mean frequency, root-mean-square frequency, and spectrum dispersion degree, and record them as F17-F21; 2.3) Extract the time-frequency domain degradation index of the vibration signal, obtain the first 8 eigenmode components of the vibration signal through empirical mode decomposition, and calculate the energy of the 8 eigenmode components and the energy entropy of the empirical mode decomposition in turn, and Record it with F22-F30; 3) Establish a comprehensive evaluation method for degradation indicators, including correlation criteria, monotonicity criteria, and predictability criteria, and quantitatively evaluate the performance of the 30 physical degradation indicators extracted in step 2) according to formulas (1) to (4) ; 3.1) Correlation criterion based on Spearman correlation coefficient Among them, Y=(y ₁ ,y ₂ ,…,y _L ) and t=(t ₁ ,t ₂ ,…,t _L ) are the degradation index sequence and time sequence respectively, and L is the length of the degradation index sequence, and are the average values of the degradation index sequence and the time series, respectively. This criterion describes the correlation between the degradation index and the time series. The closer the value is to 1, the stronger the correlation between the two; 3.2) Monotonicity criterion Among them, Y=(y ₁ ,y ₂ ,…,y _L ) is the degradation index sequence, L is the length of the degradation index sequence, ε(x) is the unit step function, and this criterion describes the monotonous increase or monotonous decrease of the degradation index The closer the value is to 1, the better the monotonicity of the degradation index and the closer it is to the actual degradation of key mechanical components; 3.3) Predictive criteria in, are the mean values of the degradation index at the initial time and failure time respectively, y _f is the degradation index value, σ(y _f ) is the standard deviation of the degradation index at the failure time, this criterion describes the variation range of the degradation index in the whole life cycle and The dispersion at the failure time, the closer the value is to 1, the larger the variation range of the degradation index and the smaller the standard deviation at the failure time, the more consistent the variation range and failure threshold of the index between different individuals, the more suitable for remaining life prediction; 3.4) Comprehensive evaluation criteria Comprehensive evaluation criteria The above evaluation criteria are linearly weighted and combined, as shown in the following formula (4): SM＝ω ₁ Corr(Y)+ω ₂ Mon(Y)+ω ₃ Pro(Y) (4) Among them, SM is the comprehensive evaluation criterion, Y is the index sequence, ω ₁ +ω ₂ +ω ₃ =1, and ω ₁ ,ω ₂ ,ω ₃ ∈[0,1] are used to represent the weight of the three evaluation criteria; 4) According to the size of the comprehensive evaluation criterion SM in step 3) of the 30 physical degradation indicators, select the degradation indicators with excellent performance to form the eigenvectors of the vibration signals of key mechanical components; 5) The vibration signal eigenvectors of key mechanical components in the three stages have different labels, and use this information for distance metric learning to obtain a distance metric matrix suitable for measuring the state-space similarity in the degradation process of key mechanical components; 6) Utilize the learned distance metric matrix to measure the similarity of the self-organizing map neural network competition learning stage, so that the network is optimized; 7) Use the vibration signal eigenvector in the normal stage to train the self-organizing map neural network, determine the neural network structure and parameters, and use the Mexican sombrero function for weight adjustment; 8) In the test phase, the feature vector of the newly acquired vibration signal is used as the input of the self-organizing map neural network, and the distance between the real-time feature vector and the corresponding activation node weight vector is calculated according to formula (5), and the virtual degradation index enhancement of the key mechanical parts is obtained minimum quantization error; Among them, EMQE is the enhanced minimum quantization error index, X is the feature vector of the newly acquired vibration signal, m is the weight vector corresponding to the activation node in the self-organizing map neural network, and A is the distance metric matrix obtained through distance metric learning.