A bearing fault diagnosis adaptive input method and system based on deep learning

Abstract

This invention provides a deep learning-based adaptive input method and system for bearing fault diagnosis, belonging to the field of fault diagnosis technology. The method includes: acquiring parameters; and calculating the adaptive input length N based on the acquired parameters. a According to the adaptive input length N a This invention employs a method to extract the first L frequency samples from the envelope spectrum and uses these L extracted frequency samples as input to a convolutional neural network, thus achieving adaptive input for bearing fault diagnosis. By calculating the adaptive input length of the deep learning model and extracting frequency samples from the envelope spectrum, and then using these extracted frequency samples as input to the convolutional neural network, the invention ensures that the deep learning model maintains stable fault diagnosis accuracy under different application conditions. This solves the problem of large fluctuations in fault diagnosis accuracy caused by the fixed input length in existing deep learning-based bearing fault diagnosis methods across different application scenarios.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology, and in particular relates to an adaptive input method and system for bearing fault diagnosis based on deep learning. Background Technology

[0002] Deep learning-based bearing fault diagnosis algorithms primarily use time-domain, frequency-domain, or time-frequency signals as model input. However, noise can overwhelm time-domain signals, degrading model diagnostic performance. Therefore, more and more researchers are using frequency-domain or time-frequency signals as input to deep learning models. These frequency-domain or time-frequency signal-based bearing fault diagnosis deep learning models have a fixed input length. However, bearing structural parameters and sampling frequencies vary across different application scenarios. A fixed input length resulting in a fixed frequency resolution may prevent the extraction of fault feature information, thus affecting the model's diagnostic accuracy and stability in different application scenarios. Summary of the Invention

[0003] To address the aforementioned shortcomings in existing technologies, this invention provides an adaptive input method and system for bearing fault diagnosis based on deep learning, which solves the problem that existing bearing fault diagnosis methods based on deep learning models use fixed input lengths, resulting in significant fluctuations in diagnostic accuracy under different application scenarios.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0005] This solution provides a deep learning-based adaptive input method for bearing fault diagnosis, comprising the following steps:

[0006] S1. Parameter Acquisition: The accelerometer is mounted on the bearing to be sampled, and the sampling frequency f is acquired. s Bearing speed f r The input length L of the model, the structural parameters of the bearing, and the harmonics k of the bearing fault characteristic frequencies;

[0007] S2. Calculate the adaptive input length: Based on the acquired parameters, calculate the adaptive input length N. a ;

[0008] S3. Calculate adaptive input data: Based on the adaptive input length N a The first L frequency samples are extracted from the envelope spectrum, and these L frequency samples are used as the input to the convolutional neural network to complete the adaptive input for bearing fault diagnosis.

[0009] Further, step S1 includes the following steps:

[0010] S101. Data Acquisition: Install the accelerometer on the bearing to be sampled, and collect data according to the sampling frequency f. s To collect the raw vibration signal during bearing operation and record the bearing speed f r ;

[0011] S102. Determine the input length of the model: Based on the designed input size of the convolutional neural network, determine the input length L of the model;

[0012] S103. Determine bearing parameters: Obtain the structural parameters of the bearing, wherein the structural parameters include: the number of rollers z, the rolling diameter d, the pitch diameter D, and the contact angle α.

[0013] S104. Determine the octave of the bearing fault characteristic frequency: Determine the octave k of the bearing fault characteristic frequency according to actual needs.

[0014] Furthermore, the adaptive length N a The expression is as follows:

[0015]

[0016] Furthermore, step S3 includes the following steps:

[0017] S301, Adaptive Sampling: The original vibration signal is sampled according to an adaptive input length N. a Perform resampling, truncate N a One original vibration signal;

[0018] S302. Calculate the envelope spectrum: Calculate the truncated N. a The envelope spectrum of the original vibration information;

[0019] S303. Extracting Data Points: Extract the first L frequency sample points from the envelope spectrum calculation results, and use the extracted L frequency sample points of the envelope spectrum as the input of the convolutional neural network to complete the adaptive input for bearing fault diagnosis.

[0020] This invention also provides a deep learning-based adaptive input system for bearing fault diagnosis, comprising:

[0021] The parameter acquisition module is used to mount the accelerometer onto the bearing to be sampled and to acquire the sampling frequency f. s Bearing speed f r The input length L of the model, the structural parameters of the bearing, and the harmonics k of the bearing fault characteristic frequencies;

[0022] The adaptive input length calculation module is used to calculate the adaptive input length N based on the acquired parameters. a ;

[0023] The adaptive input data calculation module is used to calculate the adaptive input length N. a The first L frequency samples are extracted from the envelope spectrum, and these L frequency samples are used as the input to the convolutional neural network to complete the adaptive input for bearing fault diagnosis.

[0024] The beneficial effects of this invention are:

[0025] (1) In order to solve the problem that the existing bearing fault diagnosis method based on deep learning model uses a fixed input length, which causes the fault diagnosis accuracy of the deep learning model to fluctuate greatly under different application scenarios, this invention calculates the adaptive input length of the model and extracts frequency sample points through the envelope spectrum, and uses the frequency sample points as the input of the convolutional neural network, so that the deep learning model can maintain the stability of fault diagnosis accuracy under different application conditions.

[0026] (2) This invention utilizes the envelope spectrum to achieve global demodulation of bearing vibration signals in the frequency domain, demodulates the fault characteristic frequency from the high-frequency part, and restores its original position in the spectrum, thereby improving the accuracy of bearing fault diagnosis.

[0027] (3) By calculating the adaptive input length and mapping it to various variables under different application environments, this invention improves the stability of the model's fault diagnosis accuracy under different application conditions. Attached Figure Description

[0028] Figure 1 This is a flowchart of the method of the present invention.

[0029] Figure 2 This is a schematic diagram of the process framework of the present invention.

[0030] Figure 3 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0031] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0032] Example 1

[0033] like Figures 1-2 As shown, this invention provides an adaptive input method for bearing fault diagnosis based on deep learning, the implementation of which is as follows:

[0034] S1. Parameter Acquisition: The accelerometer is mounted on the bearing to be sampled, and the sampling frequency f is acquired. s Bearing speed f r The model's input length L, bearing structural parameters, and the harmonics k of the bearing fault characteristic frequencies are implemented as follows:

[0035] S101. Data Acquisition: Install the accelerometer on the bearing to be sampled, and collect data according to the sampling frequency f. s To collect the raw vibration signal during bearing operation and record the bearing speed f r ;

[0036] S102. Determine the input length of the model: Based on the designed input size of the convolutional neural network, determine the input length L of the model;

[0037] S103. Determine bearing parameters: Obtain the structural parameters of the bearing, wherein the structural parameters include: the number of rollers z, the rolling diameter d, the pitch diameter D, and the contact angle α.

[0038] S104. Determine the octave of the bearing fault characteristic frequency: Determine the octave k of the bearing fault characteristic frequency according to actual needs;

[0039] S2. Calculate the adaptive input length: Based on the acquired parameters, calculate the adaptive input length N. a ;

[0040] S3. Calculate adaptive input data: Based on the adaptive input length N a The method involves extracting the first L frequency samples from the envelope spectrum and using these L frequency samples as the input to a convolutional neural network to achieve adaptive input for bearing fault diagnosis. The implementation method is as follows:

[0041] S301, Adaptive Sampling: The original vibration signal is sampled according to an adaptive input length N. a Perform resampling, truncate N a One original vibration signal;

[0042] S302. Calculate the envelope spectrum: Calculate the truncated N. a The envelope spectrum of the original vibration information;

[0043] S303. Extracting Data Points: Extract the first L frequency sample points from the envelope spectrum calculation results, and use the extracted L frequency sample points of the envelope spectrum as the input of the convolutional neural network to complete the adaptive input for bearing fault diagnosis.

[0044] In this embodiment, the bearing structural parameters are assumed to be as follows: z represents the number of rolling elements, d represents the diameter of the rolling elements, D represents the bearing pitch diameter, α is the contact angle, and the bearing speed under a certain application condition is assumed to be f. r The sampling frequency is f s The structural parameters of bearings, as well as their rotational speed and sampling frequency, vary under different application conditions. Since the input length of the deep learning-based bearing fault diagnosis network is fixed, in order to maintain stable diagnostic accuracy under different application conditions, the designed network must establish a mapping relationship between its fixed input length and the variables in different application environments. This mapping relationship is the key to this invention.

[0045] In this embodiment, localized damage to the bearing periodically generates a series of transient pulses. The bearing characteristic frequencies generated by these repetitive transient pulses are important features for characterizing the presence of bearing faults and distinguishing fault types. These bearing fault characteristic frequencies can be calculated based on the bearing's structural parameters. The fault characteristic frequencies of the bearing's inner ring, outer ring, rolling elements, and cage are respectively denoted by f1. in ,f out ,f ball ,f cage The calculation formulas are as follows:

[0046]

[0047]

[0048]

[0049]

[0050] For ease of explanation, this invention uses a 1D convolutional neural network as an example, assuming that a well-designed 1D convolutional neural network accepts a fixed input length of 1×L. A commonly used deep learning model based on frequency features performs a Fast Fourier Transform (FFT) on the original vibration information of length 2L to obtain frequency features of length L, and then uses the extracted frequency features as the model input. Since the length L is a fixed value, if the sampling frequency in a certain application scenario is extremely high, the resolution of the frequency domain features will be too coarse, potentially missing fault feature frequencies. This would prevent the model from extracting effective fault features, thus reducing the diagnostic accuracy of the model in this application scenario.

[0051] Bearing fault signals exhibit significant modulation characteristics, which shifts the fault characteristic frequencies to higher frequencies, increasing the difficulty of bearing fault diagnosis. Envelope spectrum analysis employs Hilbert transform to achieve global demodulation of the bearing vibration signal in the frequency domain, that is, demodulating the fault characteristic frequencies from the high-frequency portion to restore their original positions in the spectrum. Therefore, this invention does not use FFT to obtain frequency domain features, but instead obtains frequency domain features through envelope spectrum analysis demodulation.

[0052] Since the model's input length is L, and the input signal type is the frequency domain characteristic after envelope spectrum analysis, the first L frequency sample points need to be obtained through the envelope spectrum. Because the envelope spectrum is a demodulated signal, to ensure all faults are diagnosed, the frequency band (BW) of these L envelope spectrum frequency sample points only needs to include the largest bearing fault characteristic frequency, which is the largest characteristic frequency of the bearing inner ring fault. Therefore:

[0053] BW = kf in (5)

[0054] In the formula, k represents the octet of the bearing fault characteristic frequency to be selected. Since the input length of the designed 1D convolutional neural network is L, meaning it includes L frequency sample points, and assuming the frequency resolution is Δf, the bandwidth BW of these L frequency sample points can be calculated using the following formula:

[0055] BW=LΔf (6)

[0056] The formula for calculating frequency resolution is as follows:

[0057]

[0058] In the formula, N a This represents the adaptive input length to be calculated. Combining formulas (1), (5), (6), and (7), the adaptive input length N can be obtained. a The calculation formula is as follows:

[0059]

[0060] The formula shows that the adaptive input length includes parameters such as bearing speed, sampling frequency, and bearing structure, thus possessing the characteristic of adapting to different application scenarios.

[0061] In this embodiment, the application process of the adaptive input length algorithm is as follows: 1) First, obtain the adaptive input length N using formula (8). a ;2) Then extract N from the original time-domain signal a 3) Then calculate this N original vibration information; a4) Extract the first L frequency samples of the envelope spectrum as the input of the 1D convolutional neural network. Since the extracted L frequency samples are calculated through the envelope spectrum and belong to the demodulated signal, it can be seen from the above derivation that this demodulated signal contains all fault feature frequencies, so it can ensure that all fault feature frequencies are extracted. It can be seen from formula (7) that the frequency resolution will change according to the adaptive input length, thereby ensuring that all fault feature frequencies are extracted, thus ensuring the diagnostic accuracy of the 1D convolutional neural network proposed for bearings in various application environments.

[0062] Example 2

[0063] like Figure 3 As shown, this invention provides a deep learning-based adaptive input system for bearing fault diagnosis, comprising:

[0064] The parameter acquisition module is used to mount the accelerometer onto the bearing to be sampled and to acquire the sampling frequency f. s Bearing speed f r The input length L of the model, the structural parameters of the bearing, and the harmonics k of the bearing fault characteristic frequencies;

[0065] The adaptive input length calculation module is used to calculate the adaptive input length N based on the acquired parameters. a ;

[0066] The adaptive input data calculation module is used to calculate the adaptive input length N. a The first L frequency samples are extracted from the envelope spectrum, and these L frequency samples are used as the input to the convolutional neural network to complete the adaptive input for bearing fault diagnosis.

[0067] like Figure 3 The bearing fault diagnosis adaptive input system based on deep learning provided in the embodiment shown can execute the technical solution shown in the bearing fault diagnosis adaptive input method based on deep learning in the above embodiment. Its implementation principle and beneficial effects are similar, and will not be repeated here.

[0068] In this embodiment of the invention, the functional units can be divided according to the deep learning-based adaptive input method for bearing fault diagnosis. For example, each function can be divided into its own functional unit, or two or more functions can be integrated into one processing unit. The integrated unit can be implemented in hardware or as a software functional unit. It should be noted that the unit division in this invention is illustrative and represents only a logical division; other division methods may be used in actual implementation.

[0069] In this embodiment of the invention, the deep learning-based adaptive input system for bearing fault diagnosis, in order to realize the principle and beneficial effects of the deep learning-based adaptive input method for bearing fault diagnosis, includes hardware structures and / or software modules corresponding to the execution of various functions. Those skilled in the art should readily recognize that, in conjunction with the illustrative units and algorithm steps described in the embodiments disclosed herein, the invention can be implemented in hardware and / or a combination of hardware and computer software. Whether a function is executed by hardware or computer software depends on the specific application and design constraints of the technical solution. Different methods can be used to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0070] Through the above design, although the model input remains fixed, the fixed input length includes adaptive information for the application scenario, taking into account factors such as bearing structural parameters, bearing speed, and sampling frequency. This ensures that the same network can still extract bearing fault feature information in different application scenarios without changing the network structure. This guarantees the stability of the diagnostic accuracy of the same bearing fault diagnosis model in different application scenarios, and solves the problem that existing deep learning-based bearing fault diagnosis methods using fixed input lengths cause large fluctuations in fault diagnosis accuracy in different application scenarios.

Claims

1. A deep learning-based adaptive input method for bearing fault diagnosis, characterized in that, Includes the following steps: S1. Parameter Acquisition: The accelerometer is mounted on the bearing to be sampled, and the sampling frequency is acquired. Bearing speed Input length of the model The structural parameters of the bearing and the harmonics of the bearing fault characteristic frequencies. ; Step S1 includes the following steps: S101, Data Acquisition: Mount the accelerometer onto the bearing to be acquired, and collect data according to the sampling frequency. To collect raw vibration signals during bearing operation and record bearing speed. ; S102. Determine the input length of the model: Based on the designed input dimensions of the convolutional neural network, determine the input length of the model. ; S103. Determine bearing parameters: Obtain the structural parameters of the bearing, wherein the structural parameters include: the number of rollers in the bearing. Rolling diameter of bearing Bearing pitch diameter and contact angle ; S104. Determine the octaves of the bearing fault characteristic frequency: Based on actual requirements, determine the octaves of the bearing fault characteristic frequency. ; S2. Calculate the adaptive input length: Based on the acquired parameters, calculate the adaptive input length. ; S3. Calculate adaptive input data: based on the adaptive input length Extracting the front portion through envelope spectrum 100 frequency sample points, and extract the envelope spectrum. A number of frequency sample points are used as input to the convolutional neural network to complete the adaptive input for bearing fault diagnosis.

2. The deep learning-based adaptive input method for bearing fault diagnosis according to claim 1, characterized in that, The adaptive length The expression is as follows: 。 3. The deep learning-based adaptive input method for bearing fault diagnosis according to claim 2, characterized in that, Step S3 includes the following steps: S301, Adaptive Sampling: The original vibration signal is sampled according to an adaptive input length. Perform resampling, truncation One original vibration signal; S302. Calculate the envelope spectrum: Calculate the truncated envelope spectrum. The envelope spectrum of the original vibration information; S303. Extracting Data Points: Extracting the first part of the envelope spectrum calculation results. 100 frequency sample points, and extract the envelope spectrum. A number of frequency sample points are used as input to the convolutional neural network to complete the adaptive input for bearing fault diagnosis.

4. A bearing fault diagnosis adaptive input system based on deep learning, characterized in that, include: The parameter acquisition module is used to mount the accelerometer onto the bearing to be sampled and to acquire the sampling frequency. Bearing speed Input length of the model The structural parameters of the bearing and the harmonics of the bearing fault characteristic frequencies. ; Includes the following steps: Data acquisition: The accelerometer is mounted on the bearing to be sampled, and data is collected according to the sampling frequency. To collect raw vibration signals during bearing operation and record bearing speed. ; Determine the input length of the model: Based on the designed input dimensions of the convolutional neural network, determine the input length of the model. ; Determine bearing parameters: Obtain the structural parameters of the bearing, wherein the structural parameters include: the number of rollers in the bearing. Rolling diameter of bearing Bearing pitch diameter and contact angle ; Determine the octaves of the bearing fault characteristic frequencies: Based on actual requirements, determine the octaves of the bearing fault characteristic frequencies. ; The adaptive input length calculation module is used to calculate the adaptive input length based on the acquired parameters. ; The adaptive input data calculation module is used to calculate the adaptive input length. Extracting the front portion through envelope spectrum 100 frequency sample points, and extract the envelope spectrum. A number of frequency sample points are used as input to the convolutional neural network to complete the adaptive input for bearing fault diagnosis.

Images