Rolling bearing fault diagnosis method based on deep learning

Abstract

A rolling bearing fault diagnosis method based on deep learning is implemented according to the following steps: S1, establishing standard values according to time domain signal features, frequency domain signal features and envelope spectrum signal features; s2, inputting an original vibration signal; s3, respectively demodulating a time domain signal feature, a frequency domain signal feature and an envelope spectrum signal feature according to the original vibration signal; and S4, comparing the time domain signal feature, the frequency domain signal feature and the envelope spectrum signal feature demodulated from the original vibration signal with respective standard values, and judging whether the rolling bearing has a fault or not. The method has the beneficial effects that the standard value is established by integrating three types of signal features of the time domain, the frequency domain and the envelope spectrum, fault judgment is carried out, intervention links of artificial feature extraction and selection are reduced depending on the self-adaptive learning capability of deep learning, the influence of human experience difference on a diagnosis result is avoided, and the diagnosis accuracy is improved. The accuracy and efficiency of fault diagnosis are improved, and intelligent diagnosis of rolling bearing faults is realized.

Description

Technical Field

[0001] This invention relates to a bearing fault analysis method, and more particularly to a rolling bearing fault diagnosis method based on deep learning. Background Technology

[0002] Rolling bearings, as a key fundamental component, play an indispensable role in numerous industrial sectors. Their applications are extremely wide-ranging, including but not limited to engines and transmissions in automobile manufacturing, aircraft engines in the aerospace field, motors and pumps on industrial production lines, and main shaft systems in wind turbine generators. During operation, rolling bearings, through their structure of inner ring, outer ring, rolling elements, and cage, effectively reduce friction between rotating parts, withstand radial and axial loads, and ensure smooth and efficient equipment operation. Typically, bearings are installed between a shaft and a housing, operating under high-speed rotation or heavy-load conditions, and may be exposed to complex environments such as high temperature, humidity, or contamination to support the normal operation of mechanical systems. Their performance directly affects the reliability and lifespan of the entire equipment.

[0003] However, rolling bearings are subjected to various complex loads and environmental influences during operation, leading to faults of different types and degrees. These faults often generate characteristic signals or pulse signals in the vibration signal. However, due to the nonlinear and non-stationary characteristics of the signals themselves, as well as the influence of factors such as noise, interference, and mode aliasing, fault signals are difficult to extract and identify from the original signals. Traditional fault diagnosis methods usually require manual feature extraction and selection, and also require operators to have extensive prior knowledge and experience. This not only limits the promotion and application of fault diagnosis methods, but also makes them susceptible to human factors, leading to inaccurate diagnostic criteria. Therefore, research on fault diagnosis methods based on deep learning is essential. It can reduce human intervention through adaptive learning, improve the accuracy and efficiency of fault diagnosis, and thus achieve intelligent diagnosis. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a deep learning-based method for diagnosing rolling bearing faults that solves the problems of difficult fault signal extraction and excessive manual intervention.

[0005] To achieve the above objectives, the technical solution of this invention is as follows: A rolling bearing fault diagnosis method based on deep learning, implemented according to the following steps: S1, establish standard values ​​based on time-domain signal characteristics, frequency-domain signal characteristics, and envelope spectrum signal characteristics respectively; S2, input the original vibration signal; S3, based on the original vibration signal, demodulate the time-domain signal characteristics, frequency-domain signal characteristics, and envelope spectrum signal characteristics respectively; S4 compares the time-domain signal characteristics, frequency-domain signal characteristics, and envelope spectrum signal characteristics demodulated from the original vibration signal with their respective standard values ​​and determines whether the rolling bearing has a fault.

[0006] The beneficial effects of this invention are as follows: This method establishes standard values ​​by integrating three types of signal features—time domain, frequency domain, and envelope spectrum—and then performs fault judgment. This effectively overcomes the shortcomings of existing technologies where fault signals are difficult to extract and identify due to nonlinearity, non-stationarity, and interference factors. Simultaneously, relying on the adaptive learning capability of deep learning, it reduces the intervention of manual feature extraction and selection, avoids the influence of human experience differences on diagnostic results, and improves the accuracy and efficiency of fault diagnosis, achieving intelligent diagnosis of rolling bearing faults. As a preferred approach, vibration signals of the bearing under different operating conditions can be collected first, and their corresponding time domain, frequency domain, and envelope spectrum features can be extracted. The logic for establishing standard values ​​of each feature can be optimized through the iterative learning process of deep learning, making the standard values ​​more closely match the bearing characteristics under actual operating conditions. As another preferred approach, in the fault judgment stage, preliminary anomaly screening is first completed based on the deviation between the time domain signal features and the standard values. Then, frequency domain features are used to match the characteristic frequencies of each component of the bearing, and envelope spectrum features are combined to locate the excitation source of modulation-type faults, thereby improving the targeting and accuracy of fault diagnosis in a hierarchical manner.

[0007] Furthermore, the establishment of time-domain signal characteristic standard values ​​in step S1 is carried out according to the following steps: S111, collect vibration time-domain signals; S112, through time-domain superposition learning, selects the maximum value in the time-domain signal, and simultaneously calculates the RMS value in the time-domain signal of each sample; S113 is converted into the corresponding vibration acceleration level Z, and the largest value is selected as the standard value of the time domain signal characteristics.

[0008] This technical solution establishes time-domain feature standard values ​​by acquiring vibration time-domain signals and combining time-domain superposition learning, RMS value calculation, and vibration acceleration level conversion. This comprehensively captures extreme fluctuations and statistical characteristics in the time-domain signals, avoiding the limitations of a single feature dimension on the reliability of the standard values. The established time-domain feature standard values ​​better reflect the vibration characteristic boundaries of the bearing under normal operating conditions. As a preferred approach, time-domain vibration signals of the bearing under different operating conditions such as rated speed and load fluctuations can be acquired, and multiple rounds of time-domain superposition learning can be performed. The RMS value of each sample is recorded synchronously and converted into vibration acceleration levels, with the maximum value selected as the time-domain feature standard value. As another preferred approach, during the time-domain superposition learning process, the acquired time-domain signals are first preprocessed in segments, and then the maximum signal value within each segment is extracted. The final time-domain feature standard value is determined by combining the results of each segment, enhancing the representativeness of the standard value for the overall characteristics of the time-domain signal.

[0009] Furthermore, the standard values ​​for frequency domain signal characteristics established in step S1 are established according to the following steps: S121, the time-domain signal is transformed by FFT to obtain the frequency-domain signal; S122, Segment the frequency bands according to the spectrum diagram and learn the characteristic frequency features in each frequency band; S123, select the value with the largest characteristic frequency feature in each frequency band as the standard value of frequency domain signal feature.

[0010] This scheme converts the time-domain signal into a frequency-domain signal using FFT transformation. It then establishes frequency-domain standard values ​​by combining frequency band segmentation and feature frequency learning. This effectively separates signal features within different frequency ranges, avoiding interference from mode aliasing in fault feature extraction, and making the frequency-domain standard values ​​more accurately correspond to the frequency characteristics of normal bearing operation. As a preferred approach, the frequency-domain signal after FFT transformation is segmented according to the inherent characteristic frequency ranges of the bearing's inner ring, outer ring, and rolling elements. Clustering of characteristic frequencies is performed on the signals within each frequency band, and the largest eigenvalue in each cluster is selected as the frequency-domain standard value for the corresponding frequency band. As another preferred approach, after frequency band segmentation, the characteristic frequencies corresponding to different speeds are learned in a targeted manner, based on the bearing's operating speed variation range, further refining the operational condition adaptability of the frequency-domain standard values.

[0011] Furthermore, the standard values ​​for the envelope spectrum signal features in step S1 are established according to the following steps: S131, Perform Hilbert transform on the time-domain signal to extract the envelope spectrum signal; S132, the envelope spectrum signal is obtained after FFT transformation; S132, learn the characteristic frequency features of each frequency band, and select the value with the largest characteristic frequency feature in each frequency band as the standard value of the envelope spectrum signal feature for envelope spectrum analysis.

[0012] This scheme obtains the envelope spectrum signal by combining Hilbert transform and FFT transform, and learns the characteristic frequencies of each frequency band to establish standard values. This effectively demodulates fault characteristic signals modulated in other frequency bands, compensating for the insufficient capture of modulation-type fault signals by time-domain and frequency-domain features, and making the signal characteristics covered by the standard values ​​more comprehensive. As a preferred approach, the time-domain signal is first preprocessed with bandpass filtering before performing Hilbert transform to extract the envelope signal. Then, an FFT transform is used to obtain the envelope spectrum. The envelope spectrum is then segmented according to the modulation frequency range, and the characteristic frequencies of each band are learned. As another preferred approach, during the feature learning process of the envelope spectrum, the focus is on the frequency bands related to bearing fault modulation. The maximum characteristic frequency value within this band is specifically selected as the standard value of the envelope spectrum, improving the accuracy of identifying modulation-type fault characteristics.

[0013] Furthermore, step S3 is implemented according to the following steps: S31, through neural network processing, the predicted values ​​of time-domain signal features, frequency-domain signal features and envelope spectrum signal features are obtained respectively; S32, the predicted values ​​of time-domain signal features, frequency-domain signal features, and envelope spectrum signal features are input into the loss function along with the standard values ​​of time-domain signal features, frequency-domain signal features, and envelope spectrum signal features to obtain the loss value; S33 updates the weights of each signal feature using the loss value to make the predicted value closer to the standard value.

[0014] This technical solution obtains predicted values ​​of three types of signal features through neural networks, and optimizes the prediction results by combining loss function calculation and weight update. This continuously improves the matching degree between predicted values ​​and standard values, enhancing the stability and accuracy of fault diagnosis. As a preferred approach, a convolutional neural network is used to perform feature mapping processing on the three types of signal features to obtain corresponding predicted values. The predicted values ​​and corresponding standard values ​​are then input into a mean squared error loss function to calculate the loss value. Based on the loss value, the weights corresponding to each feature in the network are updated using a backpropagation algorithm. As another preferred approach, during the weight update process, differentiated update step sizes are assigned to the three types of features (time domain, frequency domain, and envelope spectrum). Priority is given to optimizing the weights corresponding to features with higher correlation to the fault, accelerating the convergence of predicted values ​​to standard values. Attached Figure Description

[0015] Figure 1 This is a flowchart of an embodiment of the present invention; Figure 2 This is a flowchart illustrating deep learning in an embodiment of the present invention; Figure 3 This is an example diagram of time-domain learning in an embodiment of the present invention; Figure 4 This is an example diagram of frequency domain learning in an embodiment of the present invention; Figure 5 This is an example diagram of envelope spectrum learning in an embodiment of the present invention. Detailed Implementation

[0016] An embodiment of the present invention provides a rolling bearing fault diagnosis method based on deep learning, implemented according to the following steps: S1, establish standard values ​​based on time-domain signal characteristics, frequency-domain signal characteristics, and envelope spectrum signal characteristics respectively; The standard values ​​of time-domain signal characteristics are established according to the following steps: S111, collect vibration time-domain signals; S112, through time-domain superposition learning, selects the maximum value in the time-domain signal, and simultaneously calculates the RMS value in the time-domain signal of each sample; S113 is converted into the corresponding vibration acceleration level Z, and the largest value is selected as the standard value of the time domain signal characteristics.

[0017] The specific calculation of the standard value of the time-domain signal characteristics is as follows: Selecting a sufficient number of samples, we first acquire vibration time-domain signals. Through time-domain superposition learning, we select the maximum value in the time-domain signal and simultaneously calculate the RMS value of each sample's time-domain signal, converting it into the corresponding vibration acceleration level Z (dB). The sample vibration signal is... x i ( i =1,2,…, n , n (Number of sampling points).

[0018] Therefore, we know that: RMS = sqrt(( x 1 2 + x 2 2 +…+ x n 2 ) / n ).

[0019] And the vibration acceleration level Z = 20lg(RMS / a 0), of which a 0 is the reference value for acceleration. a 0 = 9.81 × 10 -3 m / s 2 .

[0020] The largest value among the learned results is selected as the standard value Z of the time-domain signal feature.max .

[0021] The standard values ​​of frequency domain signal characteristics are established according to the following steps: S121, the time-domain signal is transformed by FFT to obtain the frequency-domain signal; S122, Segment the frequency bands according to the spectrum diagram and learn the characteristic frequency features in each frequency band; S123, select the value with the largest characteristic frequency feature in each frequency band as the standard value of frequency domain signal feature.

[0022] The specific calculation of the standard value of frequency domain signal characteristics is as follows: After completing the feature learning of the time-domain signal, the time-domain signal is transformed by FFT to obtain the frequency-domain signal. Based on the spectrum, the frequency bands are divided into 5-50Hz, 50-300Hz, 300-1800Hz, 1800-5000Hz, 5000-7500Hz, and 7500-10000Hz. Alternatively, frequency bands can be segmented for learning according to specific requirements, such as learning the peak characteristics of each band and the corresponding characteristic frequencies of the inner ring, outer ring, steel balls, and cage. The characteristic frequencies corresponding to the inner ring, outer ring, steel balls, and cage are calculated based on the parameters of the corresponding bearing model. According to the learning results, the largest value in each frequency band is selected as the standard value of the frequency-domain signal characteristics.

[0023] The standard values ​​of envelope spectrum signal characteristics are established according to the following steps: S131, Perform Hilbert transform on the time-domain signal to extract the envelope spectrum signal; S132, the envelope spectrum signal is obtained after FFT transformation; S132, learn the characteristic frequency features of each frequency band, and select the value with the largest characteristic frequency feature in each frequency band as the standard value of the envelope spectrum signal feature for envelope spectrum analysis.

[0024] The specific calculation of the standard values ​​of the envelope spectrum signal characteristics is as follows: The envelope of the time-domain signal is extracted by performing a Hilbert transform, and then the envelope spectrum is obtained by performing an FFT transform. Envelope spectrum analysis aims to obtain more comprehensive signal characteristics by demodulating features modulated in other frequency bands. Similarly, the peak characteristics of each frequency band, as well as the corresponding characteristic frequencies of the inner ring, outer ring, steel ball, and cage, can be learned. Based on the learning results, the largest value in each frequency band is selected as the characteristic value for envelope spectrum analysis.

[0025] S2, input the original vibration signal; S3, based on the original vibration signal, demodulate the time-domain signal characteristics, frequency-domain signal characteristics, and envelope spectrum signal characteristics respectively; S31, through neural network processing, the predicted values ​​of time-domain signal features, frequency-domain signal features and envelope spectrum signal features are obtained respectively; S32, the predicted values ​​of time-domain signal features, frequency-domain signal features, and envelope spectrum signal features are input into the loss function along with the standard values ​​of time-domain signal features, frequency-domain signal features, and envelope spectrum signal features to obtain the loss value; S33 updates the weights of each signal feature using the loss value to make the predicted value closer to the standard value.

[0026] The calculation process for step S3 is as follows: The time-domain signal, frequency-domain signal, and envelope spectrum signal after processing the original vibration signal are used as input values ​​for deep learning. After processing by the neural network, the predicted value is obtained. P = f ( WA ' + b ) in, W As weight, A For input values, b For bias, f Here is the activation function.

[0027] in W In the initial learning process, equal weights are used as the initial values. The loss function refers to the deviation between the predicted value and the standard value. A smaller deviation indicates more accurate given weights with less impact on the result; during the learning process, the weights will gradually decrease. Conversely, if the deviation has a greater impact on the result, the weights will gradually increase. Through continuous learning with multiple samples, the neural network can be trained to obtain more accurate outputs and improve the accuracy of judgments. The specific calculation process is as follows: Where y i p represents the actual output value. i This is the predicted value of the function, where N represents the total number of output values, and i represents the corresponding number: Where w i ∂ represents the weight of the corresponding parameter, ∂ refers to the gradient, and η represents the learning rate, which decays exponentially and ranges from (0,1).

[0028] Step S3 is performed repeatedly in the early stages of deep learning when the predicted value deviates too much from the standard value, so that the predicted value is close to or the same as the standard value.

[0029] S4 compares the time-domain signal characteristics, frequency-domain signal characteristics, and envelope spectrum signal characteristics demodulated from the original vibration signal with their respective standard values ​​and determines whether the rolling bearing has a fault.

[0030] After determining that the rolling bearing is not faulty, it is classified according to GB / T 32333-2015 "Methods and Technical Conditions for Measuring Vibration (Acceleration) of Rolling Bearings".

[0031] The above embodiments are merely one preferred embodiment of the present invention. Ordinary variations and substitutions made by those skilled in the art within the scope of the technical solution of the present invention are all included within the protection scope of the present invention.

Claims

1. A deep learning-based method for diagnosing rolling bearing faults, implemented according to the following steps: S1, establish standard values ​​based on time-domain signal characteristics, frequency-domain signal characteristics, and envelope spectrum signal characteristics respectively; S2, input the original vibration signal; S3, based on the original vibration signal, demodulate the time-domain signal characteristics, frequency-domain signal characteristics, and envelope spectrum signal characteristics respectively; S4 compares the time-domain signal characteristics, frequency-domain signal characteristics, and envelope spectrum signal characteristics demodulated from the original vibration signal with their respective standard values ​​and determines whether the rolling bearing has a fault.

2. The deep learning-based rolling bearing fault diagnosis method according to claim 1, characterized in that: The establishment of time-domain signal characteristic standard values ​​in step S1 is carried out according to the following steps: S111, collect vibration time-domain signals; S112, through time-domain superposition learning, selects the maximum value in the time-domain signal, and simultaneously calculates the RMS value in the time-domain signal of each sample; S113 is converted into the corresponding vibration acceleration level Z, and the largest value is selected as the standard value of the time domain signal characteristics.

3. The deep learning-based rolling bearing fault diagnosis method according to claim 2, characterized in that: The standard values ​​for frequency domain signal characteristics established in step S1 are established according to the following steps: S121, the time-domain signal is transformed by FFT to obtain the frequency-domain signal; S122, Segment the frequency bands according to the spectrum diagram and learn the characteristic frequency features in each frequency band; S123, select the value with the largest characteristic frequency feature in each frequency band as the standard value of frequency domain signal feature.

4. The deep learning-based rolling bearing fault diagnosis method according to claim 3, characterized in that: The standard values ​​for the envelope spectrum signal features in step S1 are established according to the following steps: S131, Perform Hilbert transform on the time-domain signal to extract the envelope spectrum signal; S132, the envelope spectrum signal is obtained after FFT transformation; S132, learn the characteristic frequency features of each frequency band, and select the value with the largest characteristic frequency feature in each frequency band as the standard value of the envelope spectrum signal feature for envelope spectrum analysis.

5. The deep learning-based rolling bearing fault diagnosis method according to any one of claims 1-4, characterized in that: Step S3 is implemented according to the following steps: S31, through neural network processing, the predicted values ​​of time-domain signal features, frequency-domain signal features and envelope spectrum signal features are obtained respectively; S32, the predicted values ​​of time-domain signal features, frequency-domain signal features, and envelope spectrum signal features are input into the loss function along with the standard values ​​of time-domain signal features, frequency-domain signal features, and envelope spectrum signal features to obtain the loss value; S33 updates the weights of each signal feature using the loss value to make the predicted value closer to the standard value.

Images