A rotating machinery state monitoring method based on stack auto-encoding dimension reduction

Abstract

The application provides a rotating machinery state monitoring method based on stack auto-encoding dimension reduction. First, the original vibration signal of the rotating machinery is collected, and multi-domain features of the signal are extracted, including time domain, frequency domain and time-frequency domain features. Then, the extracted multi-domain high-dimensional features are preliminarily screened through a variance selection method, the screened features are subjected to nonlinear dimension reduction by using a stack auto-encoding network, and the dimension-reduced features are divided into a training set and a test set. Next, the training set is input into a fuzzy neural network for training, and after the training is completed, the test set is input, so as to construct a health degree index and a curve capable of reflecting the running state of the rotating machinery. Finally, a fault indication scale is established through a 3 sigma criterion, and once the health degree index exceeds the scale for three times in succession, an early warning is performed, so that the state monitoring of the rotating machinery is realized. The application can discover early faults of the rotating machinery in time, ensures the safe operation of the rotating machinery and reduces the loss caused by the faults.

Description

Technical Field

[0001] The present invention relates to the field of condition monitoring, and in particular to a rotating machinery condition monitoring method based on stack autoencoder dimensionality reduction. Background Art

[0002] Condition monitoring is the inspection and monitoring of the working status of a running piece of machinery or its components to determine whether it is operating normally and whether there are any signs of abnormalities or degradation. It also tracks abnormalities, predicts degradation trends, and determines the extent of degradation and wear. Rotating machinery often operates in harsh and complex environments, making it prone to various failures that can impact equipment lifespan, threaten human safety, and cause inestimable losses. Failure is both a state and a process. Effective monitoring and early warning of rotating machinery failures before they occur, enabling timely downtime for inspection and repair to ensure the safety of equipment and personnel, are of great research significance.

[0003] The core of condition monitoring is to construct health indicators that accurately and effectively reflect the current operating state of rotating machinery. Currently, most methods extract various features to construct multi-domain fusion health indicators. These extracted features are often high-dimensional and contain redundant and interfering components, which affects both the computational efficiency of the model and the fusion effect of the health indicator, necessitating optimization or dimensionality reduction. Many existing models use methods such as principal component analysis (PCA) and linear discriminant analysis (LDA) for feature dimensionality reduction. However, the operating state of rotating machinery is a nonlinear process, and linear dimensionality reduction methods such as PCA and LDA are not applicable.

[0004] The stacked autoencoder network is a deep neural network composed of multiple autoencoders. Using an unsupervised layer-by-layer learning approach, it minimizes reconstruction error, forcing the intermediate hidden layers to learn the full information of the input signal, extracting deeper features and achieving superior nonlinear dimensionality reduction. Compared to other dimensionality reduction methods, the stacked autoencoder network retains more information about the input data and more deeply explores the internal structure of the data. By combining fuzzy neural networks with the 3σ criterion to construct health status indicators and fault indication scales, it can better reflect the operating status of rotating machinery and achieve effective condition monitoring. Summary of the Invention

[0005] Purpose of the invention: The purpose of the present invention is to provide a rotating machinery state monitoring method based on stacked autoencoder dimensionality reduction to address the defects of the prior art, so as to realize the state monitoring of rotating machinery.

[0006] Technical solution: The rotating machinery condition monitoring method based on stacked autoencoder dimensionality reduction described in the present invention specifically includes the following steps:

[0007] (1) Collect the original vibration signal of the rotating unit, extract its time domain, frequency domain and time-frequency domain features, and obtain multi-domain high-dimensional feature vectors;

[0008] (2) The extracted multi-domain high-dimensional features are preliminarily screened by variance selection method, and then a stacked autoencoder network with two hidden layers is used to perform nonlinear dimensionality reduction on the screened features, and the reduced dimensionality features are divided into training set and test set;

[0009] (3) The training set is input into the fuzzy neural network for training. After the training is completed, the test set is input into the trained network to obtain a rotating machinery health index that integrates multi-domain features. The health index is normalized to obtain a rotating machinery health curve with an index between [0, 1] and a downward trend;

[0010] (4) Using the 3σ criterion, calculate the upper and lower thresholds of the rotating machinery health curve and establish a fault indication scale based on the 3σ criterion;

[0011] (5) Once the health index exceeds the fault indication scale three times in a row, it is considered that the rotating machinery has failed and a maintenance warning is issued, thereby realizing the health status monitoring of the rotating machinery.

[0012] Furthermore, the implementation process of step (1) is as follows:

[0013] Extract time domain features from the signal, including peak value, peak-to-peak value, mean value, average amplitude, root mean square amplitude, standard deviation, effective value, skewness index, kurtosis index, peak index, pulse index, margin index, and waveform index. Extract frequency domain features from the signal, including average frequency, centroid frequency, mean square frequency, root mean square frequency, frequency variance, and frequency standard deviation. Perform wavelet packet decomposition on the signal to extract time and frequency domain features, including wavelet energy entropy and wavelet singular spectrum entropy. Finally, construct a multi-domain high-dimensional feature vector x = {x1, x2, ... x n}, where n represents the number of features extracted.

[0014] Furthermore, the implementation process of step (2) is as follows:

[0015] Rotating machinery is often in a normal state during its initial operation, with relatively stable signal amplitudes. Therefore, the amplitudes of the extracted features should also be small in the early stages. Based on this, a variance selection method is used to perform a preliminary screening of the extracted features. The initial variance of each feature is calculated, and features with large variances are filtered out, while those with small variances are retained.

[0016] A stacked autoencoder network is established to encode and decode the initially filtered, multi-domain, high-dimensional feature vector x' using a nonlinear activation function. The objective function is constructed as the root mean square error (RMSE), with a weight decay term added to prevent overfitting. The network is trained via gradient updates to find the optimal weights to minimize reconstruction error. This ensures that the hidden layer output of the stacked autoencoder retains the input data to the greatest extent possible. The number of hidden layer nodes is set to be smaller than the number of input nodes, achieving effective dimensionality reduction.

[0017] The features of some vibration signals collected during the normal operation of the rotating machinery are taken as the training set, and the features of all vibration signals collected are taken as the test set.

[0018] Furthermore, the implementation process of step (3) is as follows:

[0019] The training set is input into a fuzzy neural network. The number of nodes in the input layer is the dimension after feature dimensionality reduction, the number of output nodes is 1, the network membership function is Gaussian, and the loss function is mean square error. This is how the network is trained. After training is complete, the test set is input into the network. The network output is the health index of the rotating machinery. The health index is normalized to obtain a rotating machinery health curve with an index between [0, 1] and a downward trend. Normalization is achieved using the following formula:

[0020]

[0021] In formula (4), H is the health index of rotating machinery output by the fuzzy neural network, and H' is the normalized health index.

[0022] Furthermore, the implementation process of step (4) is as follows:

[0023] Take m rotating machinery health indicators y under the initial normal state of the rotating machinery i , i=1,2,…m, based on the 3σ criterion, calculate m rotating machinery health indexes y i The mean μ and standard deviation σ of , and the fault indication scale is set to [μ-3σ, μ+3σ]; where:

[0024]

[0025]

[0026] is m rotating machinery health index y i , the average value of i=1,2,…m.

[0027] Furthermore, the implementation process of step (5) is as follows:

[0028] The condition monitoring process often includes unstable factors. To avoid false alarms, an early warning is issued when the indicator exceeds the fault indication scale three times in a row, thereby realizing the health status monitoring of rotating machinery.

[0029] Beneficial effects: Compared with the existing technology, the present invention extracts multi-domain features for fusion, including time domain, frequency domain and time-frequency domain. Compared with a single feature indicator, it can more comprehensively reflect the operating status of rotating machinery, especially early faults; the present invention uses a stacked autoencoder network, and on the basis of the extracted multi-domain features, continues to extract its deeper features and performs nonlinear dimensionality reduction, thereby retaining the state information of the original signal to the greatest extent, improving the accuracy of the health status indicator and the computational efficiency of the entire model; the present invention constructs a rotating machinery health status indicator through a fuzzy neural network, and establishes a fault indication scale based on the 3σ criterion, which can monitor and warn the operating status of the rotating machinery more accurately and timely, detect weak faults of the rotating machinery earlier, and has a lower false alarm rate and missed alarm rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 is a flow chart of the method of the present invention;

[0031] Figure 2 This is a physical diagram of a bearing full life cycle test device in an embodiment of the present invention;

[0032] Figure 3 Schematic diagram of the structure of the stacked autoencoding network of the present invention;

[0033] Figure 4 It is a structural diagram of the fuzzy neural network of the present invention;

[0034] Figure 5 is the bearing health status curve of the present invention;

[0035] Figure 6 It is a fault indication scale established based on the 3σ criterion of the present invention;

[0036] Figure 7 This is a partial enlarged view of the health status curve of the present invention crossing the fault indication scale three times in a row. DETAILED DESCRIPTION

[0037] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0038] The present invention proposes a method for monitoring the state of rotating machinery based on stacked autoencoder dimensionality reduction, such as Figure 1 As shown. The full life cycle test data of the bearing is used for condition monitoring. The actual picture of the test equipment is as follows Figure 2 shown.

[0039] The specific steps include:

[0040] Step 1: Collect vibration signals of rotating machinery continuously or at intervals of a fixed time, and the collected vibration signals are time series signals. In order to verify the effectiveness of the method, the collected time domain signals contain the entire process from normal to bearing outer ring fault. The data set used in this example contains 984 samples, and each sample contains 20480 points. Time domain, frequency domain and time-frequency domain features are extracted for each sample to construct an n x 984 dimensional multi-domain high-dimensional feature matrix x = {x1, x2, … x n}, where n represents the number of features extracted for each sample;

[0041] Step 2: At the initial stage of bearing operation, the signal amplitude is relatively stable, and the extracted features should also have a small amplitude at the initial stage. Based on this, the extracted features are preliminarily screened by variance selection method, the variance of each feature at the initial stage is calculated, and some features with large variance are screened out, and features with small variance are retained.

[0042] A stacked auto-encoding network is established, the number of nodes in the first hidden layer is set to 12, and the number of nodes in the second hidden layer is set to 6, and the network structure is as shown in Figure 4 The multi-domain high-dimensional feature vector x' after preliminary screening is encoded and decoded through a nonlinear activation function. The target function is constructed as the root mean square error, and a weight decay term is added to prevent overfitting. The network is trained by gradient update to find the optimal weight to minimize the reconstruction error, so that the output of the hidden layer of the stacked auto-encoding network can retain the information of the input data to the greatest extent, and effective dimension reduction is achieved.

[0043] The output of the second hidden layer of the network is a 6 x 984 dimensional data after dimension reduction. The features of 20 groups of vibration signals collected at the initial stage of normal operation are taken as the training set, and the dimension of the training set is 6 x 20. The features of all the collected vibration signals are taken as the test set, and the dimension of the test set is 6 x 984.

[0044] Step 3: The training set is input into the fuzzy neural network, and the network structure is as shown in Figure 5 The number of nodes in the input layer is 6, i.e. the dimension of the feature after dimension reduction, the number of output nodes is 1, the membership function of the network is selected as Gaussian type, and the loss function is selected as mean square error, so as to train the network. After training, the test set is input into the network, and the output of the network is the health index of the rotating machinery. The health index is normalized to obtain a rotating machinery health curve with a health index between 0 and 1 and a downward trend, as shown in Figure 6 The normalization is realized by the following formula:

[0045]

[0046] In formula (4), H is the health index of rotating machinery output by the fuzzy neural network, and H' is the normalized health index.

[0047] Step 4: Take m rotating machinery health indicators y under the initial normal state of the rotating machinery i , i=1,2,…m, based on the 3σ criterion, calculate m rotating machinery health indexes y i The mean μ and standard deviation σ of , and the fault indication scale is set to [μ-3σ, μ+3σ]; where:

[0048]

[0049]

[0050] is m rotating machinery health index y i , the average value of i=1,2,…m.

[0051] Step 5: The condition monitoring process often includes unstable factors. To avoid false alarms, an early warning is issued when the indicator exceeds the fault indication scale three times in a row, thereby realizing the health condition monitoring of the rotating machinery. Figure 7 As shown in Figure 5, the condition monitoring curve exceeds the above threshold three times in a row at samples 533 to 535. At this time, it can be determined that the equipment has failed, and an early warning can be issued.

[0052] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for monitoring the state of rotating machinery based on stacked autoencoder dimensionality reduction, characterized in that: The monitoring method comprises the following steps: 1) Collect the original vibration signal of the rotating machinery, extract the signal characteristics of the original vibration signal of the rotating machinery, and construct a multi-domain high-dimensional feature vector; 2) Preliminarily screening the multi-domain high-dimensional feature vectors obtained in step 1), performing nonlinear dimensionality reduction on the screened multi-domain high-dimensional feature vectors, and dividing the reduced multi-domain high-dimensional feature vectors into a training set and a test set; 3) inputting the training set in step 2) into the fuzzy neural network for training. After the training is completed, inputting the test set into the trained fuzzy neural network to obtain a rotating machinery health index, and normalizing the rotating machinery health index to obtain a rotating machinery health curve; 4) Calculate the upper and lower thresholds of the rotating machinery health curve obtained in step 3) using the 3σ criterion, and establish a fault indication scale based on the 3σ criterion; 5) When the health index of the rotating machinery exceeds the fault indication scale established in step 4), it is considered that the rotating machinery has failed and a maintenance warning is issued, thereby monitoring the health status of the rotating machinery; The signal features of the original vibration signal of the rotating machinery extracted in step 1) include time domain features, frequency domain features and time-frequency domain features; The time domain features include the following: peak value, peak-to-peak value, mean value, average amplitude, root mean square amplitude, standard deviation, effective value, skewness index, kurtosis index, peak index, pulse index, margin index and waveform index of the original vibration signal of the rotating machinery; The frequency domain features include the following: average frequency, center of gravity frequency, mean square frequency, root mean square frequency, frequency variance and frequency standard deviation of the original vibration signal of the rotating machinery; The time-frequency domain features are obtained by performing wavelet packet decomposition on the original vibration signal of the rotating machinery, including wavelet energy entropy and wavelet singular spectrum entropy; The method for extracting the signal features of the original vibration signal of the rotating machinery is to divide the original vibration signal of the rotating machinery into several segments, extract the time domain features, the frequency domain features and the time-frequency domain features for each segment, and form the same signal features of all segments into a signal feature vector. Then, multiple signal feature vectors construct a multi-domain high-dimensional feature vector x={x1,x2,…x n }, where x1, x2, … x n Represents different signal feature vectors; The specific steps of step 2) are: preliminary screening of the constructed multi-domain high-dimensional feature vector x by variance selection method, and calculation of each signal feature vector x1, x2, ... x n The variance of the rotating machinery under normal operating conditions in the initial stage is calculated, and the calculated variance is sorted from large to small, and the signal feature vectors of 20%-30% in the top column are screened out; a stacked autoencoder network is established, and the multi-domain high-dimensional feature vector x' after preliminary screening is encoded and decoded through a nonlinear activation function. The constructed objective function is the root mean square error, and a weight decay term is added to prevent overfitting. The network is trained through gradient update to find the optimal weight to minimize the reconstruction error, so that the hidden layer output of the stacked autoencoder network can retain the information of the multi-domain high-dimensional feature vector x' after preliminary screening to the greatest extent, and the number of nodes in the hidden layer is set to be less than the number of input nodes, thereby effectively reducing the dimension.

2. The rotating machinery state monitoring method based on stacked autoencoder dimensionality reduction according to claim 1, characterized in that: The specific steps of step 3) are: taking the multi-domain high-dimensional feature vectors of the original vibration signal collected under the normal operating state of the initial rotating machinery after dimensionality reduction as a training set, taking the multi-domain high-dimensional feature vectors of all the collected vibration signals after dimensionality reduction as a test set, inputting the training set into a fuzzy neural network, wherein the number of nodes in the input layer of the fuzzy neural network is the dimension of the original vibration signal after dimensionality reduction, the number of its output nodes is 1, the network membership function of the fuzzy neural network is selected as a Gaussian type, and its loss function is selected as the mean square error, and the network training is performed; After training is completed, the test set is input into the fuzzy neural network, and the network output is the health index of the rotating machinery. The health index of the rotating machinery is normalized to obtain a rotating machinery health curve with an index between [0, 1] and a downward trend.

3. The rotating machinery state monitoring method based on stacked autoencoder dimensionality reduction according to claim 2, characterized in that: The normalization is achieved by the following formula: Among them, H is the rotating machinery health index output by the fuzzy neural network, H min is the minimum rotating machinery health index output by the fuzzy neural network, H max is the maximum rotating machinery health index output by the fuzzy neural network, and H' is the normalized rotating machinery health index.

4. The rotating machinery state monitoring method based on stacked autoencoder dimensionality reduction according to claim 2, characterized in that: The specific steps of step 4) are: take m rotating machinery health indicators y under the initial normal state of the rotating machinery i , i=1,2,…m, based on the 3σ criterion, calculate m rotating machinery health indexes y i The mean μ and standard deviation σ of , and the fault indication scale is set to [μ-3σ, μ+3σ]; where: is m rotating machinery health index y i , the average value of i=1,2,…m.

5. The rotating machinery state monitoring method based on stacked autoencoder dimensionality reduction according to claim 1, characterized in that: In step 5), in order to avoid false alarms, an early warning is issued when the health indicator of the rotating machinery exceeds the fault indication scale three times in a row, thereby realizing the health status monitoring of the rotating machinery.

Images