A mechanical equipment health status evaluation method based on SAE-SA

Abstract

The present invention discloses a method for evaluating the health status of mechanical equipment based on SAE-SA. The method collects vibration signals from key components of mechanical equipment; uses a kurtosis index to screen the signal channel; performs feature extraction after signal noise reduction processing, and then fuses the extracted eigenvalues using a stacked autoencoder; linearly combines the fused features, uses a genetic algorithm to adaptively select the parameters for the linear combination, and uses the combination of coefficients with the best attenuation trend and monotonicity as the monitoring point health index curve; finally, weighted averages the health indexes of all monitoring points to obtain a comprehensive health index for the mechanical equipment. The present invention can accurately assess the health status of key components in mechanical equipment and provide guidance for predictive maintenance of mechanical equipment.

Description

Technical Field

[0001] The present invention relates to a method for evaluating the health status of mechanical equipment, which is characterized by using a SAE-SA algorithm to evaluate the health status of the mechanical equipment. Background Art

[0002] In modern industrial production, although the proportion of CNC equipment is increasing, and effective equipment health management can be achieved through the use of CNC systems and multi-sensor data collected during the processing process, a considerable amount of mechanical equipment used in industrial production still lacks advanced CNC equipment. Such equipment often has a long service life and operates in high-temperature, high-pressure, and high-frequency vibration environments for a long time, making it more prone to equipment downtime caused by failures. Therefore, developing a universal method to monitor and evaluate the health status and remaining life of various types of mechanical equipment in real time will facilitate the establishment of reasonable and effective maintenance plans, reduce unnecessary equipment downtime, and save significant maintenance costs.

[0003] After searching the literature and patents of existing technologies, it was found that the common methods for assessing the health status of mechanical equipment are as follows:

[0004] Method 1: Chinese patent application number CN105718876B, titled "A Method for Evaluating the Health Status of a Ball Screw," combines Laplace feature dimensionality reduction with Mahalanobis distance analysis to establish a nonlinear mapping relationship between sensor signal sample points and health values in feature space under different health states, thereby quantitatively evaluating the degree of ball screw performance degradation. This evaluation method requires the use of various features collected from a brand new device as baseline data, but this condition does not exist in actual production.

[0005] Method 2: Chinese patent application number: CN109255201A, titled "A Ball Screw Pair Health Assessment Method Based on SOM-MQE." This method first collects vibration signals from the bearing supports at both ends of the screw assembly and the preferred channel on the nut. These signals are preprocessed and feature extracted. The extracted feature values are then mapped to health values using SOM-MQE. An adaptive threshold detection algorithm is used to determine the alarm threshold of the health value curve. When the health value exceeds the alarm threshold, an automatic alarm is triggered. This assessment method focuses on failure warnings and only categorizes the equipment's health status into two conditions: faulty and healthy, without further refinement.

[0006] Method 3: Chinese patent application number: CN109141881B, titled "A Rotating Machinery Health Assessment Method Using a Deep Autoencoder Network." This method directly extracts features from the raw equipment signals using a deep autoencoder and constructs health indicators using an unsupervised SOM algorithm. This assessment method uses a deep autoencoder to directly extract features from the raw signals, which can easily miss important hidden features.

[0007] Therefore, the present invention is committed to developing a scientific and intuitive method for evaluating the condition of mechanical equipment, which can reflect the health status of mechanical equipment at all stages of its life cycle as much as possible while avoiding the complexity of practical application. Summary of the Invention

[0008] In view of the above shortcomings, the present invention provides a mechanical equipment health status evaluation method based on SAE-SA, which can evaluate the health status of the mechanical equipment according to the signals collected from the key components of the mechanical equipment.

[0009] The specific steps are as follows:

[0010] Step 1: Collect vibration signals from key components of the mechanical equipment, especially rotating components. The vibration signals are collected using a three-axis acceleration sensor installed at the key components of the mechanical equipment.

[0011] Step 2: Select the channel signal that best represents the degradation trend of the corresponding component from the three channel signals output by each three-channel vibration sensor as the original signal for health assessment;

[0012] Step 3: Preprocess the original sensor signal to remove noise and interference;

[0013] Step 4: Add a time window to the signal, extract the time domain features, frequency domain features, and time-frequency domain features of the signal within the time window, combine all features of the sample to generate a feature data set, and standardize the sample feature data;

[0014] Step 5: Input the sample feature dataset into the pre-built stacked autoencoder model (SAE) for feature fusion.

[0015] Step 6: Obtain the optimal linear combination coefficient through the genetic algorithm (SA), perform linear combination on the fusion feature data set, and obtain the health value mapping of the detection object.

[0016] Step 7: Take the weighted average of the health indexes of the key components of the mechanical equipment to obtain the comprehensive health index of the mechanical equipment.

[0017] Furthermore, the signal screening method used in step 2 is to calculate the kurtosis value kur of the vibration signal of each channel respectively, and select the vibration signal of the channel with the largest kurtosis value. The larger the kurtosis value, the more shocks and spikes the signal contains, and the vibration signal of this channel can better reflect the health status of the monitoring point;

[0018] The kurtosis value is calculated as follows:

[0019]

[0020] Among them, x i is the amplitude of a certain point in the signal, μ is the mean value of the signal in this segment, N is the signal length, and rms is the root mean square value of the signal in this segment. The calculation method is as follows:

[0021]

[0022] Furthermore, the preprocessing method of step 3 is to use empirical mode decomposition (EMD) to decompose the original signal into several eigencomponents, calculate the correlation coefficient between the original signal and each eigencomponent, select the five components with the highest correlation and combine them to eliminate a large amount of noise components;

[0023] Furthermore, the time window length used to segment the signal in step 4 is 512, and the entire signal is divided into m signal segments of length 512. 14 time domain features, 12 frequency domain features, and 8 time-frequency domain features are extracted for each time window signal. The time domain features include root mean square value, kurtosis, skewness, maximum value, minimum value, peak-to-peak value, maximum absolute value, absolute average value, root square amplitude, margin, waveform index, pulse index, peak index and energy index, a total of 14; the frequency domain features are the features extracted after Fourier transform of each time window signal to obtain frequency domain data, including center of gravity frequency, root mean square frequency, standard deviation frequency, frequency variance, spectrum root mean square value, spectrum mean, spectrum skewness, spectrum kurtosis, energy spectrum density, power spectrum density, amplitude spectrum, phase spectrum, a total of 12; each time window signal is subjected to three-layer wavelet packet decomposition, and db4 wavelet is selected to obtain the wavelet reconstruction coefficients of the last node, a total of 8. The 34 signal features of each time window signal segment are combined to form the original feature data Then the original feature data of each time window are combined into the original feature data set Q = [X1, X2, ... X i …,X m ] T and standardized.

[0024] Furthermore, the feature fusion process in step 5 is based on stacked autoencoders (SAE), which is composed of multiple autoencoders (AE) stacked together.

[0025] A single AE layer can be viewed as two neural networks.

[0026] The first neural network transforms the input data into an input containing multi-dimensional features through the Sigmoid function and the model parameter set θ = (w, b). Convert to hidden representation (also called hidden feature vector) h=[h1,h2,…,h p ] T (p≤34).

[0027] h=sig m(wx+b) (3)

[0028]

[0029] Where w is the weight matrix and b is the bias.

[0030] The second neural network maps the hidden vector h back to the reconstructed vector through the Sigmoid function and the model parameter set θ′={w′,b′} As shown below:

[0031]

[0032] Where θ′={w′,b′} is the model parameter set. The reconstruction vectors of all m samples constitute the reconstruction vector set

[0033] Finally, the mean square error is used to define the reconstruction error between the reconstructed vector and the original feature vector. i …,X m ] T The i-th eigenvector of The reconstruction error is defined as:

[0034]

[0035] In the formula is the i-th input sample X in the sample set i The n-th dimension feature, is the i-th sample of the reconstructed vector set The n-th dimension feature of . m is the number of samples in the sample set, and 34 is the dimension of a single sample.

[0036] The stacked autoencoder (SAE) consists of multiple stacked AEs, where the input of the latter AE comes from the hidden layer of the previous AE. Each AE aims to minimize the input-output reconstruction error.

[0037] Design a SAE stacked by 6 AEs, with the number of hidden layers being 28, 24, 20, 16, 8, and 4 respectively. That is, by training the stacked autoencoder model, the data with 34-dimensional features is converted into a fused feature vector with 4-dimensional fused features. All m samples are fused to form a fusion feature vector set h = [h 1 ,h 2 ,...,h i ,...,h m ] T .

[0038] Furthermore, the goal of step 6 is to linearly combine the above four fusion features to obtain the health index curve:

[0039] HI=a1h1+a2h2+a3h3+a4h4 (7)

[0040] The health index curve is usually required to have 1) monotonicity 2) trend 3) decay correlation.

[0041] A genetic algorithm is used to adaptively select the parameters (a1, a2, a3, and a4) for the linear combination of the four fusion features. Each individual in the population is used as a linear combination parameter vector. The monotonicity and decay trend of the combined health curve are used as the fitness function of the genetic algorithm.

[0042] Monotonic fitness function:

[0043]

[0044] score mono =I(·) (9)

[0045] I(·) is an indicator function, which is 1 when the condition in the brackets is met and 0 otherwise. The smaller the value, the better the monotonicity.

[0046] Decaying trend fitness function:

[0047]

[0048] T is a custom decay function that decays over time:

[0049] T i =mi (11)

[0050] score corr The closer it is to 1 or 1, the better the time series correlation is. :

[0051] According to the changing characteristics and value range of the two fitness functions, a combined fitness function score is designed:

[0052] score=-score mono *score corr (12)

[0053] It can be seen that the higher the score, the better the monotonicity of the combined HI and the more obvious the decay trend.

[0054] Each generation of the evolutionary process consists of the following steps:

[0055] 1) Set iteration parameters: population size S, maximum evolutionary generations N, crossover probability P c , mutation probability P m , value range Bound;

[0056] 2) Calculate the fitness value of each individual in the population, that is, the absolute value of the trend evaluation value;

[0057] 3) Selection operation, using the roulette wheel method to select the next generation of individuals;

[0058] 4) Crossover operation, randomly selecting two individuals to exchange genes;

[0059] 5) Mutation operation: randomly select an individual for gene mutation;

[0060] After the evolution is completed, the optimal solution and optimal value are output, that is, the coefficient combination with the best attenuation trend and monotonicity and the health index curve.

[0061] Furthermore, in step 7, the health index of each key component monitoring point is weighted and averaged to obtain the comprehensive health index of the mechanical equipment:

[0062]

[0063] The present invention uses the SAE-SA algorithm to assess the health status of mechanical equipment. First, multi-channel vibration information of multiple key component monitoring points of the mechanical equipment is collected, and the channels are screened using the kurtosis index; after signal noise reduction processing, feature extraction is performed, and the extracted feature values are then fused using a stacked autoencoder (SAE); the fused features are linearly combined, and the parameters of the linear combination are adaptively selected using a genetic algorithm. The coefficient combination with the best attenuation trend and monotonicity is used as the monitoring point health index curve; finally, the health index of all monitoring points is weighted and averaged to obtain the comprehensive health index of the mechanical equipment. Compared with the existing technology, the present invention has the following beneficial technical effects:

[0064] 1. This invention collects vibration signals from key mechanical equipment components by installing acceleration sensors at monitoring points on key components, independently of the equipment's numerical control system, to obtain health indices. This in turn generates a comprehensive health index, more comprehensively reflecting changes in the health of the equipment and offering broad applicability.

[0065] 2. The signal features extracted by the present invention are richer, and the stacked autoencoder method is used to fuse multiple features into four types, eliminating redundant information while retaining the typical characteristics of the signal as much as possible.

[0066] 3. Based on feature extraction and fusion, the present invention linearly combines four fusion features that retain the vast majority of signal features. A new fitness function is designed to ensure that the optimal health index curve calculated by the genetic algorithm has a good attenuation trend and monotonicity. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 This is a flow chart of a method for evaluating the health status of mechanical equipment based on SAE-SA according to an embodiment of the present invention;

[0068] Figure 2 1 is a structural diagram of a stacked autoencoder model of a mechanical equipment health status evaluation method based on SAE-SA according to an embodiment of the present invention;

[0069] Figure 3 This is a detailed flow chart of a genetic algorithm for optimizing parameters of a fusion feature linear combination of a mechanical equipment health status evaluation method based on SAE-SA according to an embodiment of the present invention;

[0070] Figure 4 It is a schematic diagram of a mechanical equipment testing device system according to an embodiment of the present invention;

[0071] Figure 5 is a graph of the comprehensive health index of mechanical equipment calculated by an embodiment of the present invention; DETAILED DESCRIPTION

[0072] like Figure 1 As shown, the SAE-SA-based mechanical equipment health status assessment method described in this embodiment includes the following steps:

[0073] Step 1: Collect vibration signals at key components of mechanical equipment, especially rotating components of mechanical equipment. The vibration signals are collected by three-axis acceleration sensors installed at monitoring points of key components of mechanical equipment, such as Figure 4 As shown;

[0074] Step 2: Calculate the kurtosis value kur of the vibration signal of each channel respectively, select the signal of the channel with the largest kur value as the signal that best represents the degradation trend of the corresponding component, and use it as the original signal for constructing the health index;

[0075] Step 3: Preprocess the original signal by using empirical mode decomposition (EMD) to decompose the original signal into several eigencomponents, calculate the correlation coefficient between the original signal and each eigencomponent, select the five components with the highest correlation and combine them to eliminate noise and interference;

[0076] Step 4: Since the signal is high-frequency and has a large amount of data, a time window is added to the signal to facilitate analysis. The time window length is set to 512, and the entire signal is divided into m signal segments of length 512.

[0077] Extract the time domain features, frequency domain features, and time-frequency domain features of the signal within the time window. The time domain features include root mean square value, kurtosis, skewness, maximum value, minimum value, peak-to-peak value, maximum absolute value, absolute mean value, root square amplitude, margin, waveform index, pulse index, peak index, and energy index, a total of 14. The frequency domain features are the features extracted after Fourier transforming the signals in each time window to obtain frequency domain data, including center of gravity frequency, root mean square frequency, standard deviation frequency, frequency variance, spectrum root mean square value, spectrum mean, spectrum skewness, spectrum kurtosis, energy spectrum density, power spectrum density, amplitude spectrum, and phase spectrum, a total of 12. Perform three-layer wavelet packet decomposition on the signal in each time window, select db4 wavelet, and obtain the wavelet reconstruction coefficients of the last node, a total of 8. The 34 feature combinations of all m time windows generate a sample feature data set with a shape of [m,34], which is then standardized.

[0078] Step 5: Input the sample feature dataset into the pre-built stacked autoencoder model (SAE) for feature fusion. Figure 2 As shown in the figure, it is formed by stacking 6 AE layers, with the number of hidden layers being 28, 24, 20, 16, 8, and 4 respectively. The resulting fused feature dataset is of shape [m, 4]. The four fused features of each sample retain the original features to the greatest extent possible.

[0079] Step 6: Figure 3 As shown, the optimal linear combination coefficient is obtained through the genetic algorithm (SA), the fusion feature data set is linearly combined, and the health value mapping HI of the detection object is obtained.

[0080] The goal of designing the fitness function is to make the combined health index curve monotonic and have a decaying trend.

[0081] Monotonic fitness function:

[0082]

[0083] score mono =I(·) (15)

[0084] I(·) is an indicator function, which is 1 when the condition in the brackets is met and 0 otherwise. The smaller the value, the better the monotonicity.

[0085] Decaying trend fitness function:

[0086]

[0087] T is a custom decay function that decays over time:

[0088] T i =mi (17)

[0089] score corr The closer it is to 1 or 1, the better the time series correlation is. :

[0090] According to the changing characteristics and value range of the two fitness functions, a combined fitness function score is designed:

[0091] score=-score mono *score corr (18)

[0092] It can be seen that the higher the score, the better the monotonicity of the combined HI and the more obvious the decay trend.

[0093] Set the iteration parameters: population size S = 1000, maximum evolutionary generations N = 150, crossover probability P c =0.8, mutation probability P m =0.6, value range Bound=[-1, 1];

[0094] In each evolutionary iteration of genetics, the following steps are repeated: calculating the fitness value of each individual in the population, that is, the absolute value of the trend evaluation value; selection operation, using the roulette method to select the next generation of individuals; crossover operation, randomly selecting two individuals for gene exchange; mutation operation, randomly selecting an individual for gene mutation;

[0095] After the evolution is completed, the optimal solution and optimal value are output, that is, the coefficient combination with the best attenuation trend and monotonicity and the health index curve. The health index value is quantified in the range of [0, 1] to quantify the evaluation results of mechanical equipment performance degradation, which is more intuitive and effective.

[0096] Step 7: Take the weighted average of the health index of each key component monitoring point of the mechanical equipment to obtain the comprehensive health index of the mechanical equipment, such as Figure 5 shown.

[0097]

Claims

1. A mechanical equipment health status evaluation method based on SAE-SA, characterized in that: Use SAE-SA to evaluate the health status of mechanical equipment; The specific steps are as follows: Step 1: Collect vibration signals of key components of the mechanical equipment, especially rotating components of the mechanical equipment; the vibration signals are collected by a three-axis acceleration sensor installed at the key components of the mechanical equipment; Step 2: Select the channel signal that best represents the degradation trend of the corresponding component from the three channel signals output by each three-channel vibration sensor as the original signal for health assessment; Step 3: Preprocess the original sensor signal to remove noise and interference; Step 4: Add a time window to the signal, extract the time domain features, frequency domain features, and time-frequency domain features of the signal within the time window, combine all features of the sample to generate a feature data set, and standardize the sample feature data; Step 5: Input the sample feature dataset into the pre-built stacked autoencoder model SAE for feature fusion; Step 6: Obtain the optimal linear combination coefficient through the genetic algorithm SA, perform linear combination on the fusion feature data set, and obtain the health value mapping of the detection object; Step 7: Take the weighted average of the health indexes of the key components of the mechanical equipment to obtain the comprehensive health index of the mechanical equipment.

2. The SAE-SA-based mechanical equipment health status evaluation method according to claim 1, characterized in that: The signal screening method adopted in step 2 is to calculate the kurtosis value of the vibration signal of each channel respectively, and select the vibration signal of the channel with the largest kurtosis value.

3. The SAE-SA-based mechanical equipment health status evaluation method according to claim 1, characterized in that: The preprocessing method of step 3 is to use empirical mode decomposition to decompose the original signal into several eigencomponents, calculate the correlation coefficient between the original signal and each eigencomponent respectively, select the five components with the highest correlation and combine them to perform signal noise reduction.

4. The SAE-SA-based mechanical equipment health status evaluation method according to claim 1, characterized in that: The time window length of the signal used in step 4 is 512, and the entire signal is divided into m signal segments of length 512; the time domain features, frequency domain features, and time-frequency domain features of each time window signal are extracted; Time domain features include root mean square value, kurtosis, skewness, maximum value, minimum value, peak-to-peak value, maximum absolute value, absolute average value, root mean square amplitude, margin, waveform index, pulse index, peak index and energy index, a total of 14; Frequency domain features are features extracted after Fourier transform of each time window signal to obtain frequency domain data, including centroid frequency, root mean square frequency, standard deviation frequency, frequency variance, spectrum root mean square value, spectrum mean, spectrum skewness, spectrum kurtosis, energy spectrum density, power spectrum density, amplitude spectrum, and phase spectrum, a total of 12; time-frequency domain features are decomposed using three-layer wavelet packets, using db4 wavelet to obtain the wavelet reconstruction coefficients of the last node, a total of 8.

5. The SAE-SA-based mechanical equipment health status evaluation method according to claim 1, characterized in that: The feature fusion method described in step 5 is a stacked autoencoder, which is composed of 6 stacked autoencoders, and the number of hidden layers is 28, 24, 20, 16, 8, and 4 respectively. The input of the latter autoencoder layer comes from the hidden layer of the previous autoencoder layer; the data with 34-dimensional features is Become a fused feature vector with 4-dimensional fused features All m samples are fused to form a fusion feature vector set h = [h 1 ,h 2 ,...,h i ,...,h m ] T .

6. The SAE-SA based mechanical equipment health status evaluation method according to claim 1, characterized in that: The linear combination method of the health index curve described in step 6 is a genetic algorithm; A genetic algorithm is used to adaptively select the parameters (a1, a2, a3, a4) of the linear combination of the above four fusion features. Each individual in the population is used as a linear combination parameter vector. The monotonicity and attenuation trend of the combined health curve are used as the fitness function of the genetic algorithm. Monotonic fitness function: score mono =I(·) (2) I(·) is an indicator function, which is 1 when the condition in the brackets is met; otherwise, it is 0. The smaller the value, the better the monotonicity. Decaying trend fitness function: T is a custom decay function that decays over time: T i =m-i (4) score corr The closer it is to 1 or 1, the better the time series correlation; According to the changing characteristics and value range of the two fitness functions, a combined fitness function score is designed: score=-score mono *score corr (5) It can be seen that the higher the score, the better the monotonicity of the combined HI and the more obvious the decay trend; Each generation of the evolutionary process consists of the following steps: 1) Set iteration parameters: population size S, maximum evolutionary generations N, crossover probability P c , mutation probability P m , value range Bound; 2) Calculate the fitness value of each individual in the population, that is, the absolute value of the trend evaluation value; 3) Selection operation, using the roulette wheel method to select the next generation of individuals; 4) Crossover operation, randomly selecting two individuals to exchange genes; 5) Mutation operation: randomly select an individual for gene mutation; After the evolution is completed, the optimal solution and optimal value are output, that is, the coefficient combination with the best attenuation trend and monotonicity and the health index curve.

7. The SAE-SA based mechanical equipment health status evaluation method according to claim 1, characterized in that: The method for combining the health index of each key component monitoring point to obtain the comprehensive health index of the mechanical equipment is weighted averaging.

Images