Bearing fault detection method based on self-encoder reconstruction error

Abstract

The invention discloses a bearing fault detection method based on an auto-encoder reconstruction error. The method belongs to the technical field of mechanical system fault detection, and comprises the following operation steps: selecting a sensor type, determining an installation position and setting related parameters; acquiring one-dimensional synchronous vibration data of k groups of bearings under normal working conditions, and respectively preprocessing the one-dimensional synchronous vibration data; training and testing an auto-encoder model by using the preprocessed data; obtaining one-dimensional synchronous vibration data of the real-time working condition of the bearing, and performing preprocessing; reconstructing the preprocessed real-time working condition data by using the trained auto-encoder model, and calculating a reconstruction error E; judging whether a fault occurs or not; according to the bearing fault detection method based on the self-encoder reconstruction error, the bearing fault can be effectively and reliably detected; the invention belongs to an artificial intelligence algorithm, and needs less manual expert knowledge; when the self-encoder model is trained, only the data in the healthy state of the bearing is used, and the problem of insufficient data volume during model training is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of mechanical system fault detection and relates to a bearing fault detection method based on autoencoder reconstruction error. Background Art

[0002] With the rapid development of industrial manufacturing, the importance of bearings, as core components in aerospace, high-end machine tools, and high-end medical equipment, has become increasingly prominent. Bearings play a critical role in mechanical equipment and can directly impact the equipment's operating efficiency and stability. During operation, bearings are susceptible to problems such as foreign matter, load impact, and poor lubrication, which can increase wear, reduce service life, and even pose safety hazards. Furthermore, the long-term operation of rolling bearings under harsh operating conditions such as high speeds and heavy loads inevitably leads to various failures. Bearing failures not only cause equipment downtime but can also result in significant economic losses. Therefore, research on bearing fault detection is not only of theoretical value but also of significant economic significance for practical industrial applications.

[0003] Bearing fault detection is essentially a pattern recognition problem, primarily involving signal acquisition, feature extraction and selection, and classification and recognition. Vibration signals are currently the most commonly used signal acquisition method. When a rolling bearing fails, the working surface of the part collides with the faulty area, causing the bearing's vibration signal to change. Therefore, analyzing bearing faults using the bearing's vibration signal can achieve effective detection results. Feature extraction, by extracting fault features in the time, frequency, and time-frequency domains, can analyze the bearing's operating status and enable fault detection and diagnosis. For example, extracting bearing time-frequency features using empirical mode decomposition can suppress the impact of disturbance noise on bearing fault detection and diagnosis.

[0004] In recent years, principal component analysis (PCA) has been widely used in fault detection. PCA transforms multiple variables in raw data into a small number of uncorrelated variables, eliminating redundant information and thus improving the accuracy of fault detection and classification. However, limitations of PCA still hinder its effectiveness. First, PCA relies on linear transformations, making it difficult to capture the complex nonlinear characteristics of bearing vibration signals. Second, the number of principal components selected depends on empirical experience. Retaining too few principal components can result in loss of fault-sensitive information, while retaining too many can introduce noise and fail to effectively eliminate redundant information.

[0005] To address these issues, researchers proposed using an autoencoder to analyze the signal. An autoencoder is an unsupervised learning model based on a neural network. It automatically learns data features through nonlinear transformations, effectively extracting complex features from bearing vibration signals. The autoencoder's nonlinear feature extraction capability enables it to capture complex features such as shocks and harmonics, while also avoiding information loss or noise introduced by improperly selecting the number of principal components through adaptive feature selection.

[0006] This invention discloses a bearing fault detection method based on autoencoder reconstruction error. This method does not rely on human expert knowledge or historical fault data, but only uses historical fault-free data for model training. In actual production, historical fault-free data is easy to obtain, while historical fault data is more difficult to obtain. Personnel generally do not allow the system to operate under fault conditions. This method does not use historical fault data, effectively solving the problem of insufficient data for model training. Summary of the Invention

[0007] Purpose of the invention: In order to overcome the shortcomings of the existing methods mentioned in the background technology, the present invention requests to disclose a bearing fault detection method based on autoencoder reconstruction error, firstly, the one-dimensional synchronous vibration data is expanded into multi-dimensional synchronous data, then the potential features of the data are learned and the original data is reconstructed through the encoder and decoder, the reconstruction error is calculated, and it is used as the fault feature, and finally the fault detection is performed according to the obtained fault feature and the pre-set threshold.

[0008] Technical solution: The bearing fault detection method based on autoencoder reconstruction error of the present invention comprises the following steps:

[0009] Step (1): Select the sensor type, determine the installation location and set relevant parameters;

[0010] Step (2): obtain k groups of one-dimensional synchronous vibration data of bearings under normal working conditions and preprocess them respectively;

[0011] Step (3): Use the preprocessed data to train and test the autoencoder model;

[0012] Step (4): Obtain one-dimensional synchronous vibration data of the bearing in real time and perform preprocessing;

[0013] Step (5): Use the trained autoencoder model to reconstruct the preprocessed real-time working condition data and calculate the reconstruction error E;

[0014] Step (6): Determine whether a fault occurs.

[0015] Furthermore, in step (1), the sensor types are vibration sensors and speed sensors; the vibration sensor is installed near the bearing to be detected to measure the vibration data of the bearing; the speed sensor is installed near the rotating shaft to measure the rotating shaft speed data; the relevant parameters include: the number of nodes in the input layer and the output layer are both M, the number of hidden layers L, the number of nodes in each hidden layer, the number of training iterations θ, the loss function setting value δ, the number of synchronous sampling points per rotation of the rotating shaft M, the number of rotations of the rotating shaft m, and the fault threshold ρ; the number of nodes in each hidden layer can be equal or unequal.

[0016] Furthermore, in step (2), based on the preset number of synchronous sampling points M per shaft revolution and the shaft revolution number m, the one-dimensional synchronous vibration data of the bearing is synchronously collected with the spindle speed signal as a reference. The one-dimensional synchronous vibration data of the k groups of bearings collected under normal working conditions can be expressed as:

[0017] Y=[Y1,Y2,Y3,...,Y i ,...,Y k ],

[0018] Among them, Y i =[y i (1),y i (2),y i (3),...,y i (mM)];

[0019] The normalization method is used to standardize the k groups of one-dimensional synchronous vibration data under normal working conditions. The standardized one-dimensional synchronous vibration data Y of the i-th group of normal working conditions is i ′ can be expressed as:

[0020]

[0021] Where, For data Y i The average value of Represents data Y i The standard deviation of

[0022] The standardized k groups of normal working conditions one-dimensional synchronous vibration data are expanded into M-dimensional data respectively; for the i-th group of standardized normal working conditions one-dimensional synchronous vibration data Y i ′, the expanded M-dimensional data Z i As shown below:

[0023]

[0024] For the expanded M-dimensional data Z i Perform zero-centering as follows:

[0025]

[0026] in, For M-dimensional data Z i The average value of the h-th row of data;

[0027] The k groups of M-dimensional data after zero-centering are divided into training set Z1 and test set Z2, which are used to train and test the autoencoder model respectively.

[0028] Furthermore, the implementation process of step (3) is as follows:

[0029] First, the preprocessed training set Z1 is input into the autoencoder model to be trained for iterative training. The training stops when the preset loss function value is reached, and the trained autoencoder model is obtained. The preprocessed test set Z2 is input into the trained autoencoder model to test its reconstruction ability. The autoencoder model consists of an encoder and a decoder, and both the encoder and decoder parts use the logsig activation function.

[0030] Among them, the encoder part is expressed as follows:

[0031] z=f enc (Z1) = f logsig (W1Z1+b1),

[0032] Among them, f enc (·) represents the encoder model, W1 is the encoder weight matrix, b1 is the encoder bias vector, f logsig It is the logsig activation function.

[0033] The decoder part is shown as follows:

[0034] Z1 p =f dec (z) = f logsig (W2z+b2),

[0035] Among them, f dec (·) represents the decoder model, W2 is the weight matrix of the decoder, and b2 is the bias vector of the decoder.

[0036] The loss function δ is expressed as follows:

[0037]

[0038] Among them, N represents the number of training set samples, Z1 t Represented as the t-th sample of the input training set Z1, It represents the t-th sample of the reconstructed multi-dimensional synchronous data Z1;

[0039] Use the test set Z2 to test the trained autoencoder model. Once the trained autoencoder model passes the test, it can be applied to bearing fault detection.

[0040] Furthermore, the implementation process of step (4) is as follows:

[0041] According to the preset number of synchronous sampling points M per shaft revolution and the shaft revolution m, the synchronous vibration data of the bearing is synchronously collected with the spindle speed signal as the reference signal. The collected real-time one-dimensional synchronous vibration data can be expressed as:

[0042] G=[g(1),g(2),g(3),...,g(mM)];

[0043] The obtained real-time working condition one-dimensional synchronous vibration data G is standardized. The data after standardization is G1, which is expressed as follows:

[0044]

[0045] in, is the average value of data G, σ G represents the standard deviation of data G;

[0046] The standardized data G1 is expanded into multi-dimensional synchronized data G2 as shown below:

[0047]

[0048] The expanded M-dimensional data G2 is zero-centered as follows:

[0049]

[0050] in, is the average value of the h-th row of data in the M-dimensional data G2.

[0051] Furthermore, the implementation process of step (5) is as follows:

[0052] The expanded multi-dimensional synchronous data G3 is input into the trained autoencoder model, the multi-dimensional synchronous data G3 is reduced in dimension and reconstructed, and the reconstruction error E is calculated, which is expressed as follows:

[0053]

[0054] Among them, G3 t is the tth sample of the multidimensional synchronous data G3, f dec (f enc (G3 t )) represents the tth sample of the reconstructed multi-dimensional synchronous data G3.

[0055] Furthermore, the implementation process of step (6) is as follows:

[0056] The calculated reconstruction error E is regarded as a fault feature and combined with the set threshold ρ to determine whether a bearing fault occurs. If the reconstruction error E is greater than or equal to the set threshold ρ, the bearing is considered to be faulty; otherwise, it is considered to be in a healthy state. At this time, return to step (4) to continue fault detection.

[0057] Beneficial effects

[0058] Compared with the prior art, the method of the present invention mainly includes the following beneficial effects:

[0059] (1) The bearing fault detection method based on autoencoder reconstruction error disclosed in the patent of this invention can effectively and reliably detect bearing faults.

[0060] (2) The bearing fault detection method based on autoencoder reconstruction error disclosed in the patent of this invention is an artificial intelligence algorithm and requires little human expert knowledge.

[0061] (3) The bearing fault detection method based on autoencoder reconstruction error disclosed in the patent of this invention only uses data when the bearing is in a healthy state when training the autoencoder model, which solves the problem of insufficient data during model training.

[0062] (4) The bearing fault detection method based on the autoencoder reconstruction error disclosed in the patent of this invention is not affected by the change of the spindle speed and can be applied to the bearing fault detection work at different speeds. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 This is a flow chart of the bearing fault detection method based on autoencoder reconstruction error disclosed in the present invention;

[0064] Figure 2 This is a synchronous vibration data diagram of a healthy rolling bearing according to an embodiment of the present invention;

[0065] Figure 3 This is a synchronous vibration data diagram when a rolling bearing fails in an embodiment of the present invention;

[0066] Figure 4 It is the detection result of the method of the present invention. DETAILED DESCRIPTION

[0067] The present invention will be described in further detail below with reference to the accompanying drawings.

[0068] The present invention provides a bearing fault detection method based on autoencoder reconstruction error, the process of which is as follows: Figure 1 As shown, the specific steps include:

[0069] Step (1): Select the sensor type, determine the installation location and set relevant parameters;

[0070] In the step (1), the sensor types are a vibration sensor and a speed sensor; the vibration sensor is installed near the bearing to be detected to measure the vibration data of the bearing; the speed sensor is installed near the rotating shaft to measure the rotating shaft speed data; the detection related parameters include: the number of nodes in the input layer and the output layer are both M, the number of hidden layers L, the number of nodes in each hidden layer, the number of training iterations θ, the loss function setting value δ, the number of synchronous sampling points per rotation of the rotating shaft M, the number of rotations of the rotating shaft m, and the fault threshold ρ; the number of nodes in each hidden layer can be equal or unequal.

[0071] In this embodiment, the number of nodes in the input layer and the output layer is M=100, the number of hidden layers is L=1, the number of nodes in the hidden layer is h=16, the number of training iterations is θ=500, the loss function setting value is δ=0.34, the number of synchronous sampling points per shaft revolution M is set to 100, the number of shaft revolutions m is set to 30, and the fault threshold is ρ=0.3814.

[0072] Step (2): obtain k groups of one-dimensional synchronous vibration data of bearings under normal working conditions and preprocess them respectively;

[0073] According to the preset number of synchronous sampling points M per shaft revolution and the shaft revolution m, the one-dimensional synchronous vibration data of the bearing is synchronously collected with the spindle speed signal as a reference. The one-dimensional synchronous vibration data of the k groups of bearings collected under normal working conditions can be expressed as:

[0074] Y=[Y1,Y2,Y3,...,Y i ,...,Y k ],

[0075] Among them, Y i =[y i (1),y i (2),y i (3),...,y i (mM)];

[0076] This embodiment uses the rolling bearing simulation signal constructed in the existing literature for further explanation. 100 sets of one-dimensional synchronous vibration data of the bearings under normal working conditions are collected. The number of synchronous sampling points M per shaft revolution is set to 100 and the number of shaft revolutions m is set to 30. The one-dimensional synchronous vibration data of the 100 sets of bearings under normal working conditions can be expressed as:

[0077] Y=[Y1,Y2,Y3,...,Y i ,...,Y 100 ],

[0078] Among them, Yi =[y i (1),y i (2),y i (3),...,y i (3000)];

[0079] The standardized processing method is used to standardize the one-dimensional synchronous vibration data of 100 groups of bearings under normal working conditions. Among them, the standardized one-dimensional synchronous vibration data Y of the normal working condition of the i-th group is i ′ can be expressed as:

[0080]

[0081] Where, For data Y i The average value of Represents data Y i The standard deviation of

[0082] The standardized one-dimensional synchronous vibration data of 100 groups of bearings under normal working conditions are expanded into M-dimensional data. For the i-th group of standardized one-dimensional synchronous vibration data under normal working conditions Y i ′, the expanded M-dimensional data Z i As shown below:

[0083]

[0084] For the expanded M-dimensional data Z i Perform zero-centering as follows:

[0085]

[0086] in, For M-dimensional data Z i The average value of the h-th row of data;

[0087] The expanded 100 sets of M-dimensional data are divided into training set Z1 and test set Z2 with a division ratio of 7:3, which are used to train and test the autoencoder model respectively.

[0088] Step (3): Use the preprocessed data to train and test the autoencoder model;

[0089] First, the preprocessed training set Z1 is input into the autoencoder model to be trained for iterative training. The training stops when the preset loss function value is reached, and the trained autoencoder model is obtained. The preprocessed test set Z2 is input into the trained autoencoder model to test its reconstruction ability. The autoencoder model consists of an encoder and a decoder, and both the encoder and decoder parts use the logsig activation function.

[0090] Among them, the encoder part is expressed as follows:

[0091] z=f enc (Z1) = f logsig (W1Z1+b1),

[0092] Among them, f enc (·) represents the encoder model, W1 is the encoder weight matrix, b1 is the encoder bias vector, f logsig It is the logsig activation function.

[0093] The decoder part is shown as follows:

[0094] Z1 p =f dec (z) = f logsig (W2z+b2),

[0095] Among them, f dec (·) represents the decoder model, W2 is the weight matrix of the decoder, and b2 is the bias vector of the decoder.

[0096] The loss function δ is expressed as follows:

[0097]

[0098] Among them, N represents the number of training set samples, Z1 t Represented as the t-th sample of the input training set Z1, It represents the t-th sample of the reconstructed multi-dimensional synchronous data Z1.

[0099] In this example, the preprocessed training set Z1 is input into the autoencoder model to be trained for iterative training. The number of iterations θ=500 is stopped when the preset loss function setting value δ=0.34 is reached, and the trained autoencoder model is obtained. The preprocessed test set Z2 is input into the trained autoencoder model. The difference in reconstruction error between the training set and the test set is 0.0057, indicating that the autoencoder model has good reconstruction ability.

[0100] Step (4): Obtain one-dimensional synchronous vibration data of the bearing in real time and perform preprocessing;

[0101] According to the preset number of synchronous sampling points M per shaft revolution and the shaft revolution m, the synchronous vibration data of the bearing is synchronously collected with the spindle speed signal as a reference. The collected real-time one-dimensional synchronous vibration data can be expressed as:

[0102] G=[g(1),g(2),g(3),...,g(3000)];

[0103] The obtained real-time working condition one-dimensional synchronous vibration data G is standardized. The data after standardization is G1, which is expressed as follows:

[0104]

[0105] Where, is the average value of data G, σ G represents the standard deviation of data G;

[0106] The standardized data G1 is expanded into multi-dimensional synchronized data G2 as shown below:

[0107]

[0108] The expanded M-dimensional data G2 is zero-centered as follows:

[0109]

[0110] Step (5): Use the trained autoencoder model to reconstruct the preprocessed real-time working condition data and calculate the reconstruction error E;

[0111] The expanded multi-dimensional synchronous data G3 is input into the trained autoencoder model, the multi-dimensional synchronous data G3 is reduced in dimension and reconstructed, and the reconstruction error E is calculated, which is expressed as follows:

[0112]

[0113] Among them, G3 t is the tth sample of the multidimensional synchronous data G3, f dec (f enc (G3 t )) represents the tth sample of the reconstructed multi-dimensional synchronous data G3.

[0114] Step (6): Determine whether a fault occurs.

[0115] The calculated reconstruction error E is regarded as a fault feature and combined with the set threshold ρ to determine whether a bearing fault occurs; if the reconstruction error E is greater than or equal to the set threshold ρ, the bearing is considered to be faulty; otherwise, it is considered to be in a healthy state; at this time, return to step (4) and continue detection.

[0116] Figure 2 and 3 The synchronous vibration data of the rolling bearing when it is healthy and when it fails are shown. It is difficult to identify the rolling bearing failure from the figure. Figure 4 The figure shows the fault detection results of the method of the present invention. The hollow circles in the figure represent the reconstruction error E when the bearing is healthy, and the squares represent the reconstruction error E when the bearing is faulty. Figure 4As can be seen from the figure, there is a significant difference in the reconstruction error E between the healthy and faulty states of the bearing. Therefore, in this embodiment, the threshold is set to 0.3814 according to the 3σ rule. When the reconstruction error E is greater than or equal to 0.3814, the bearing is judged to be faulty. Otherwise, the bearing is judged to be healthy, and the process returns to step (4) to continue testing.

Claims

1. The bearing fault detection method based on autoencoder reconstruction error of the present invention specifically comprises the following steps: Step (1): Select the sensor type, determine the installation location and set relevant parameters; Step (2): obtain k groups of one-dimensional synchronous vibration data of bearings under normal working conditions and preprocess them respectively; Step (3): Use the preprocessed data to train and test the autoencoder model; Step (4): Obtain one-dimensional synchronous vibration data of the bearing in real time and perform preprocessing; Step (5): Use the trained autoencoder model to reconstruct the preprocessed real-time working condition data and calculate the reconstruction error E; Step (6): Determine whether a fault occurs.

2. The bearing fault detection method based on autoencoder reconstruction error according to claim 1 is characterized in that: In step (1), the sensor types are vibration sensors and speed sensors; the vibration sensor is installed near the bearing to be detected to measure the vibration data of the bearing; the speed sensor is installed near the rotating shaft to measure the rotating shaft speed data; the relevant parameters include: the number of nodes in the input layer and the output layer are both M, the number of hidden layers L, the number of nodes in each hidden layer, the number of training iterations θ, the loss function setting value δ, the number of synchronous sampling points per rotation of the rotating shaft M, the number of rotations of the rotating shaft m, and the fault threshold ρ; the number of nodes in each hidden layer can be equal or unequal.

3. The bearing fault detection method based on autoencoder reconstruction error according to claim 1 is characterized in that: In step (2), based on the preset number of synchronous sampling points M per shaft revolution and the shaft revolution m, the one-dimensional synchronous vibration data of the bearing is synchronously collected with the spindle speed signal as a reference. The one-dimensional synchronous vibration data of the k groups of bearings collected under normal working conditions can be expressed as: <h2 style=";text-align:left;direction:ltr">Y=[Y1,Y2,Y3,...,Y<h2 style=";text-align:left;direction:ltr"> i <h2 style=";text-align:left;direction:ltr"> ,...,Y<h2 style=";text-align:left;direction:ltr"> k <h2 style=";text-align:left;direction:ltr"> ], Among them, Y i = [y i (1), y i (2), y i (3),..., y i (mM)]; The normalization method is used to standardize the k groups of one-dimensional synchronous vibration data under normal working conditions. The standardized one-dimensional synchronous vibration data Y of the i-th group of normal working conditions is i ′ can be expressed as: Where, For data Y i The average value of Represents data Y i The standard deviation of The k groups of normal working conditions one-dimensional synchronous vibration data are expanded into M-dimensional data respectively; for the i-th group of normal working conditions one-dimensional synchronous vibration data Y i ′, the expanded M-dimensional data Z i As shown below: For the expanded M-dimensional data Z i Perform zero-centering as follows: in, For M-dimensional data Z i The average value of the h-th row of data; The k groups of M-dimensional data after zero-centering are divided into training set Z1 and test set Z2, which are used to train and test the autoencoder model respectively.

4. The bearing fault detection method based on autoencoder reconstruction error according to claim 1, characterized in that: In step (3), first, the preprocessed training set Z1 is input into the autoencoder model to be trained for iterative training, and the training stops when the preset loss function setting value is reached, thereby obtaining a trained autoencoder model; The preprocessed test set Z2 is input into the trained autoencoder model to test its reconstruction ability; the autoencoder model consists of an encoder and a decoder, and both the encoder and decoder parts use the logsig activation function; Among them, the encoder part is expressed as follows: z=f enc (Z1)=f logsig (W1Z1+b1), Among them, f enc (·) represents the encoder model, W1 is the encoder weight matrix, b1 is the encoder bias vector, f logsig It is the logsig activation function; The decoder part is shown as follows: <h2 style=";text-align:left;direction:ltr">Z1<h2 style=";text-align:left;direction:ltr"> p <h2 style=";text-align:left;direction:ltr"> =f<h2 style=";text-align:left;direction:ltr"> dec <h2 style=";text-align:left;direction:ltr"> (z)=f<h2 style=";text-align:left;direction:ltr"> logsig <h2 style=";text-align:left;direction:ltr"> (W2z+b2), Among them, f dec (·) represents the decoder model, W2 is the weight matrix of the decoder, and b2 is the bias vector of the decoder; The loss function δ is expressed as follows: Among them, N represents the number of training set samples, Z1 t Represented as the t-th sample of the input training set Z1, It represents the t-th sample of the reconstructed multi-dimensional synchronous data Z1; Use the test set Z2 to test the trained autoencoder model. Once the trained autoencoder model passes the test, it can be applied to bearing fault detection.

5. The bearing fault detection method based on autoencoder reconstruction error according to claim 1, characterized in that: In step (4), according to the preset number of synchronous sampling points M per shaft revolution and the shaft revolution number m, the spindle speed signal is used as the reference signal to synchronously collect the synchronous vibration data of the bearing. The collected real-time working condition one-dimensional synchronous vibration data can be expressed as: G=[g(1),g(2),g(3),...,g(mM)]; The obtained real-time working condition one-dimensional synchronous vibration data G is standardized. The data after standardization is G1, which is expressed as follows: in, is the average value of data G, σ G represents the standard deviation of data G; The standardized data G1 is expanded into multi-dimensional synchronized data G2 as shown below: The expanded M-dimensional data G2 is zero-centered as follows: in, is the average value of the h-th row of data in the M-dimensional data G2.

6. The bearing fault detection method based on autoencoder reconstruction error according to claim 1, characterized in that: In step (5), the expanded multidimensional synchronous data G3 is input into the trained autoencoder model, the multidimensional synchronous data G3 is reduced in dimension and reconstructed, and the reconstruction error E is calculated, which is expressed as follows: Among them, G3 t is the tth sample of the multidimensional synchronous data G3, f dec (f enc (G3 t )) represents the tth sample of the reconstructed multi-dimensional synchronous data G3.

7. The bearing fault detection method based on autoencoder reconstruction error according to claim 1, characterized in that: In step (6), the calculated reconstruction error E is regarded as a fault feature, and combined with the set threshold ρ to determine whether a bearing fault occurs; if the reconstruction error E is greater than or equal to the set threshold ρ, the bearing is considered to be faulty; otherwise, it is considered to be in a healthy state; at this time, return to step (4) and continue fault detection.

Images