A method for zero-fault sample diagnosis of rolling bearings

Abstract

This invention relates to a zero-fault sample diagnosis method for rolling bearings, belonging to the field of intelligent fault diagnosis of rolling bearings. The method includes: (1) collecting rolling bearing vibration data; (2) calculating the fault frequencies of various key components of the rolling bearing; (3) using Hilbert demodulation and Fourier transform to extract the spectrum of the rolling bearing vibration data, and then introducing a grayscale image conversion method, combining the fault frequencies, and constructing multiple grayscale image training sets based on the normal state vibration data of the rolling bearing; (4) establishing multiple feature encoders based on deep residual shrinkage networks, designing a novel contrast loss function, and optimizing the feature encoders using the grayscale image training sets; (5) constructing a rolling bearing fault detection model based on the feature encoders to achieve intelligent diagnosis of rolling bearings under zero-fault samples. This invention eliminates the dependence on fault samples and achieves high-precision intelligent diagnosis of rolling bearings under zero-fault samples.

Description

Technical Field

[0001] This invention belongs to the field of intelligent fault diagnosis of rolling bearings, and specifically relates to a method for diagnosing zero-fault samples of rolling bearings. Background Technology

[0002] With the deep integration of new-generation information technology and advanced manufacturing technology, big data-driven intelligent fault diagnosis methods have sparked an innovative wave of industrial intelligent operation and maintenance in the field of mechanical equipment health status monitoring. This has significant theoretical research value and application value for promoting the intelligent transformation and upgrading of my country's manufacturing equipment and high-quality industrial development. Big data-driven intelligent methods have significant advantages in massive data processing, state feature learning, and high-precision pattern recognition. By establishing a mapping relationship between equipment operating status and real-time monitoring data, they solve practical engineering problems such as mechanical equipment condition detection, diagnosis, and prediction, and have received widespread attention from industry and academia.

[0003] As a crucial basic component of mechanical equipment, the health status of rolling bearings significantly impacts the safe and reliable operation of such equipment. However, in practice, rolling bearings operate under normal conditions for extended periods, resulting in imbalanced monitoring data across various states. This makes it impossible to construct complete training data, posing numerous challenges to data-driven methods. Under imbalanced data conditions, intelligent models cannot be effectively trained. While methods such as data migration and data generation have seen significant advancements, these methods still require overcoming the incompatibility between source and target domain data, and the training conditions are stringent, making it difficult to guarantee the model's application effectiveness. Furthermore, existing intelligent fault diagnosis methods often employ offline data for training. The inherent differences between training data and real-time monitoring data make offline-built fault diagnosis models susceptible to numerous unknown operating conditions, leading to decreased diagnostic performance. To fundamentally address these issues, there is an urgent need to propose an intelligent diagnostic method for rolling bearings with zero-fault samples. This method would overcome the dependence of data-driven methods on complete rolling bearing fault samples, improving the versatility of data-driven methods in practical engineering applications. Summary of the Invention

[0004] The purpose of this invention is to address the problem of unbalanced fault samples in existing data-driven fault diagnosis methods by providing a zero-fault sample diagnosis method for rolling bearings, thereby achieving intelligent diagnosis under zero-fault sample conditions for rolling bearings.

[0005] To achieve the above objectives, the technical solution of the present invention is: a method for zero-fault sample diagnosis of rolling bearings, comprising:

[0006] Vibration data of rolling bearings under normal conditions are collected, and the spectrum of the vibration data of rolling bearings under normal conditions is obtained by combining Hilbert demodulation and Fourier transform.

[0007] The failure frequencies of key components of rolling bearings are calculated, and a spectrum-grayscale image conversion method is proposed to obtain a grayscale image training set based on the vibration data of rolling bearings under normal conditions.

[0008] Multiple feature encoders based on deep residual shrinkage networks were established. A contrastive loss function was designed based on the failure frequency of each key component of the rolling bearing. The feature encoders were then optimized using a grayscale image training set.

[0009] A rolling bearing fault detection model based on feature encoders is constructed. By combining multiple rolling bearing fault detection models based on feature encoders under different fault frequencies, intelligent diagnosis of rolling bearings under zero-fault samples is achieved.

[0010] In one embodiment of the present invention, an accelerometer is used to collect vibration data of a rolling bearing under normal conditions.

[0011] By combining the parameters of the rolling bearing, including the inner ring, outer ring, rolling elements, and cage, and using the knowledge of the failure frequency of each key component of the rolling bearing, the failure frequency of each key component of the rolling bearing is calculated.

[0012] In one embodiment of the present invention, the specific calculation formula for calculating the failure frequency of each key component of the rolling bearing is as follows:

[0013]

[0014] in, , , , These are the failure frequencies of the inner ring, outer ring, rolling elements, and cage, respectively. Where d represents the operating frequency of the rolling bearing, d represents the diameter of the balls, and D represents the diameter of the pitch circle of the rolling bearing. This is the ball contact angle.

[0015] In one embodiment of the present invention, for the calculated fault frequencies of each key component of the rolling bearing, fault frequency band data is extracted from the spectrum of vibration data under normal rolling bearing conditions. Furthermore, based on different fault overtones, the spectrum corresponding to different fault overtones is extracted, and the spectrum extraction interval is represented as follows: Where F is the fault frequency, Z Take a positive integer. From empirical formula Sure.

[0016] In one embodiment of the present invention, the spectrum-to-grayscale image conversion method, that is, converting the spectrum truncation interval into a grayscale image, specifically involves: truncating the spectrum of the vibration data of the rolling bearing under normal conditions, with a truncation length of... If a spectral peak is found within the spectral cutoff interval, all spectral values ​​within the interval are replaced with the peak and converted to pixels. Conversely, if no peak is found, the spectral values ​​within the corresponding cutoff interval are converted to pixels. The conversion formula is as follows:

[0017]

[0018] in, Representing coordinates The pixel at point M represents the length of the grayscale image.

[0019] In one embodiment of the present invention, the specific implementation of establishing multiple feature encoders based on deep residual shrinkage networks, designing a contrastive loss function based on the failure frequencies of key components of the rolling bearing, and optimizing the feature encoders using a grayscale image training set is as follows:

[0020] Construct a feature encoder based on a deep residual shrinkage network, denoted as E; use E to extract features from m groups of grayscale images obtained at fault harmonics corresponding to the fault frequency. The feature extraction process is represented as follows:

[0021]

[0022] in, Indicates the frequency of failure Z A grayscale image under multiple fault frequency multiplication, where Enc() represents the feature extraction network;

[0023] Let the vibration data of two connected rolling bearings under normal conditions be... and Extracted using E respectively and Features of the corresponding grayscale image and Calculate the cosine similarity between pairwise high-dimensional features. Construct a comparison matrix (Matrix) based on the results:

[0024]

[0025]

[0026] By comparing the diagonal elements of the matrix Matrix to find the maximum value and the off-diagonal elements to find the minimum value, two loss functions are constructed:

[0027]

[0028] By observing and comparing the matrix Matrix, we can see that the diagonal elements have the maximum value, while the off-diagonal elements have the minimum value, represented as follows: Hot encoding yields a cross-entropy loss function:

[0029]

[0030] Among them, each of the m groups of grayscale images obtained under the fault frequency corresponding to the fault frequency includes N grayscale images, where n represents the nth grayscale image; This is obtained from the comparison matrix Matrix, specifically by finding the position of the maximum value in each row. This represents the i-th row in the comparison matrix Matrix, and y is the hot code of each row in the comparison matrix Matrix; for example, the hot code of the first row is to set the first element to 1 and the rest to 0.

[0031] Finally, a contrastive loss function is obtained, expressed as follows:

[0032] .

[0033] The deep residual shrinkage network adopts ResNet-18 and consists of a deep residual network, a soft thresholding function, and an attention mechanism.

[0034] In one embodiment of the present invention, a rolling bearing fault detection model based on feature encoders trained at different fault frequencies is constructed. The vibration data of the online monitored rolling bearing is characterized by features, and a cosine similarity is calculated between the cosine similarity and the vibration data of the reference rolling bearing under normal conditions. If the cosine similarity exceeds a threshold, the rolling bearing is considered to have a fault corresponding to the corresponding fault frequency; otherwise, it is determined that the rolling bearing has not had a fault corresponding to the corresponding fault frequency. By combining multiple rolling bearing fault detection models based on feature encoders at different fault frequencies, fault diagnosis results for the online monitored rolling bearing at all fault frequencies are obtained.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] 1. Envelope analysis of rolling bearing vibration signals is performed using Hilbert demodulation technology, which improves the ability to resolve non-stationary vibration signals. Combined with Fourier transform, the spectral characteristics of rolling bearing monitoring vibration data can be fully displayed.

[0037] 2. By introducing knowledge of rolling bearing failure mechanisms, namely the failure frequency of key components, it is possible to analyze normal vibration data samples from different perspectives and obtain diverse grayscale image training samples.

[0038] 3. A deep residual shrinkage network model was adopted, which utilizes the powerful feature extraction capability of the deep residual shrinkage network model to enable high-dimensional characterization of the vibration data of rolling bearings in online monitoring, so as to capture subtle changes in the operating state of rolling bearings.

[0039] 4. A novel contrastive loss function was designed, which enables effective training of a deep residual shrinkage network using only the vibration data of rolling bearings under normal conditions. This method of training the network frees the data-driven intelligent model from dependence on complete fault samples of rolling bearings, effectively solving the problem of extremely unbalanced rolling bearing state data in practice.

[0040] 5. By combining multiple failure frequencies and designing corresponding failure detection processes, end-to-end intelligent fault diagnosis of rolling bearings under the guidance of zero-failure samples can be achieved. This meets the application requirements for building intelligent fault diagnosis methods for in-service rolling bearings under actual engineering conditions, without the need for reliability testing, and features low-cost construction. Attached Figure Description

[0041] Figure 1 This invention provides a training process for a rolling bearing zero-fault sample depth residual shrinkage network.

[0042] Figure 2 yes Figure 1 The intelligent diagnostic method for zero-fault samples of rolling bearings involves the loss function descent process during the training of a deep residual shrinkage network under different fault frequencies.

[0043] Figure 3 and Figure 4 It is a confusion matrix of fault diagnosis results involved in the intelligent diagnostic method for zero-fault samples of rolling bearings. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0045] Please see Figure 1 , Figure 2 , Figure 3 and Figure 4 The present invention provides a method for zero-fault sample diagnosis of rolling bearings, which mainly includes the following steps:

[0046] Step one: Considering the high reliability design of rolling bearings and their long-term operation under normal conditions, an accelerometer is used to acquire vibration data of the rolling bearing under normal conditions at a certain sampling frequency.

[0047] Specifically, by installing sensors close to the rolling bearing, it is possible to monitor critical components of the rolling bearing that are prone to failure, such as the inner ring, outer ring, rolling elements, and cage. Based on the operating frequency of the rolling bearing balls and the vibration signal sampling requirements, the sampling frequency of the accelerometer is set to collect vibration data of the rolling bearing under normal conditions.

[0048] Step two: Measure the structural dimensional parameters of the rolling bearing, such as ball diameter, pitch circle diameter, and contact angle. Introduce the knowledge of rolling bearing failure mechanisms, calculate the failure frequency of key components such as the inner ring, outer ring, rolling elements, and cage, and arrange them according to their magnitude.

[0049] Specifically, taking the HRB 6001 2RZ rolling bearing as an example, according to the official website, the ball diameter is 12mm, the pitch circle diameter is 28mm, the contact angle is calculated to be 15 degrees, and the operating speed of the rolling bearing is 2000 r / min. Calculations show that the failure frequencies of the inner ring, outer ring, rolling elements, and cage of this type of rolling bearing are 164.96Hz, 68.37Hz, 32.22Hz, and 9.76Hz, respectively.

[0050] Step 3: Construct a deep residual shrinking network model. This model originates from deep residual networks, with the key difference being the modification of the residual module. A squeeze-and-excitation network is introduced, incorporating two crucial operations: squeeze and excitation. Simultaneously, a soft threshold is used to learn the feature weights after the convolutional operation. The soft threshold calculation formula is as follows:

[0051]

[0052] in, It is a constant. This is the input to the soft threshold, which is the output after the convolution operation.

[0053] Step four involves performing Hilbert demodulation and Fast Fourier Transform on the normal-state vibration data samples to obtain the spectral data. Next, two fault frequencies are selected, for example... , The spectral range is calculated, and spectral data of 1x, 2x, and 3x the fault frequency range are extracted from the spectral data respectively; finally, the data of the spectral range is converted into grayscale images.

[0054] Specifically, and Substitute into the formula

[0055]

[0056] The interval is 7. A spectral data segment with a length of 576 is extracted, based on... The inner and outer ring fault frequencies are replaced by F, and Z is set to 1, 2, and 3 respectively, resulting in three sets of spectral data. First, the data in the non-truncated interval is set to zero, and peak values ​​are searched within the spectral interval. If a peak value is five times the average value, it is considered that a spectral peak exists within that interval. Next, all data within the interval is replaced by the peak value; otherwise, the data is normalized to the 0-1 range. Finally, the spectral data is converted into pixels to obtain the corresponding grayscale images. Therefore, based on the inner and outer ring fault frequencies, three sets of grayscale images can be obtained, constructing two training sets, each consisting of three sets of grayscale images.

[0057] Step 5: Design a contrastive loss function based on normal vibration data to train the feature encoder constructed based on a deep residual shrinkage network.

[0058] Step 5.1, specifically, let the feature encoder based on the deep residual shrinkage network be E, and the grayscale image set obtained based on the fault frequency F be... 1, 2, and 3 represent the 1st, 2nd, and 3rd harmonics of the fault frequency F, and n represents the total number of samples of vibration data under normal rolling bearing conditions. Next, E is used to... High-dimensional feature characterization was performed on (Z=1, 2, 3) to obtain vibration data samples. Corresponding high-dimensional features , , Continuous vibration data samples Corresponding high-dimensional features , , .calculate The cosine similarity between each pair of pairs is used to construct a comparison matrix.

[0059] Step 5.2, specifically, introduces two assumptions: 1. Two consecutive vibration data samples under normal conditions have high similarity; 2. Spectral data at different fault overtones are uncorrelated. Based on these two assumptions, a novel loss function is finally obtained:

[0060]

[0061] Using the obtained loss function, the training process of the two feature encoders is as follows: Figure 2 As shown.

[0062] Step Six: Select inner ring faults and outer ring faults to verify the proposed intelligent diagnostic method for zero-fault samples of rolling bearings. First, construct a process for detecting inner ring faults and outer ring faults; then, combine the two detection processes to finally achieve intelligent diagnosis of rolling bearings in normal condition, as well as inner ring faults and outer ring faults.

[0063] Step 6.1, through , Two training datasets were constructed, each containing 3000 grayscale images, all derived from the vibration data of rolling bearings under normal conditions. These two training datasets were used to optimize two feature encoders based on deep residual contraction networks, which were then applied to the detection of corresponding fault states. Specifically, the detection threshold was set to a range of [0.02-0.2]. Based on the fault frequency corresponding to the inner race fault, three sets of grayscale images corresponding to the vibration sample to be detected were obtained. The trained feature encoders were used for high-dimensional representation, and the cosine similarity between the high-dimensional features and the high-dimensional features of the baseline normal-state vibration data samples was calculated. The similarity calculation result was compared with the detection threshold; if the threshold was exceeded, the rolling bearing was considered to have an inner race fault.

[0064] Step 6.2: If the test result is normal, there are two possibilities: either the bearing is in a normal state, or it is in another fault state. Therefore, the outer ring fault detection process is used to make another judgment. If the detection threshold is not exceeded, the rolling bearing is in a normal state; otherwise, it indicates an outer ring fault.

[0065] Step 6.3: Construct a test set containing 3000 vibration data samples. Each test set contains 1000 samples each representing normal condition, inner ring fault, and outer ring fault. First, perform inner ring fault detection. If the detection result is abnormal, the sample is determined to originate from an inner ring fault state. If the detection result is normal, proceed to the outer ring fault detection process. If the detection result is abnormal, the sample originates from an outer ring fault state; otherwise, it is considered a normal state.

[0066] Step 6.4, as described above, can also be performed first for outer ring fault detection and then for inner ring fault detection, so as to achieve intelligent diagnosis of zero-fault samples of rolling bearings.

[0067] Based on the above steps, with the detection threshold set to 0.1, the confusion matrix of the intelligent diagnostic results for zero-fault samples of rolling bearings, which first detects inner ring faults, is as follows: Figure 3 As shown in the figure. First, the confusion matrix of the intelligent diagnostic results of zero-fault samples of rolling bearings for outer ring fault detection is as follows: Figure 4 As shown.

[0068] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for diagnosing zero-fault samples in rolling bearings, characterized in that, include: Vibration data of rolling bearings under normal conditions are collected, and the spectrum of the vibration data of rolling bearings under normal conditions is obtained by combining Hilbert demodulation and Fourier transform. The failure frequencies of key components of rolling bearings are calculated, and a spectrum-grayscale image conversion method is proposed to obtain a grayscale image training set based on the vibration data of rolling bearings under normal conditions. Multiple feature encoders based on deep residual shrinkage networks were established. A contrastive loss function was designed based on the failure frequency of each key component of the rolling bearing. The feature encoders were then optimized using a grayscale image training set. A rolling bearing fault detection model based on feature encoders is constructed. By combining multiple rolling bearing fault detection models based on feature encoders under different fault frequencies, intelligent diagnosis of rolling bearings under zero-fault samples is achieved. The spectrum-to-grayscale image conversion method converts a truncated spectrum interval into a grayscale image. Specifically, it involves truncating the spectrum of the rolling bearing's vibration data under normal conditions, with a truncated length of L = M. 2 If a spectral peak is found within the spectral cutoff interval, all spectral values ​​within the interval are replaced with the peak and converted to pixels. Conversely, if no peak is found, the spectral values ​​within the corresponding cutoff interval are converted to pixels. The conversion formula is as follows: Where P(j,k) represents the pixel at coordinates (j,k), and M represents the length of the grayscale image; The specific implementation of establishing multiple feature encoders based on deep residual shrinkage networks, designing a contrastive loss function based on the failure frequencies of key components of the rolling bearing, and optimizing the feature encoders using a grayscale image training set is as follows: Construct a feature encoder based on a deep residual shrinkage network, denoted as E; use E to extract features from m groups of grayscale images obtained at fault harmonics corresponding to the fault frequency. The feature extraction process is represented as follows: f Z =Enc(img z ) Among them, img z This represents the grayscale image at Z times the fault frequency, and Enc() represents the feature extraction network. Let the vibration data of two connected rolling bearings under normal conditions be s. i-1 With s i Extract s using E respectively i-1 With s i The corresponding grayscale image features f1, f2, ..., f z ,…,f m With f′1, f′2, ..., f′ z ,…,f′ m Calculate the cosine similarity cc(·) between each pair of high-dimensional features, and construct a comparison matrix Matrix based on the results: By comparing the diagonal elements of the matrix Matrix to find the maximum value and the off-diagonal elements to find the minimum value, two loss functions are constructed: By observing and comparing the matrix Matrix, we find that the diagonal elements have the maximum value and the off-diagonal elements have the minimum value, represented as an m×m Hot code, resulting in a cross-entropy loss function: Among them, each of the m groups of grayscale images obtained at the fault frequency corresponding to the fault frequency includes N grayscale images, where n represents the nth grayscale image; p is obtained from the contrast matrix Matrix, specifically by obtaining the position of the maximum value of each row element, r i y represents the i-th row in the comparison matrix Matrix, and y is the hot code of each row in the comparison matrix Matrix; Finally, a contrastive loss function is obtained, expressed as follows:

2. The method for zero-fault sample diagnosis of rolling bearings according to claim 1, characterized in that, Vibration data of rolling bearings under normal conditions are collected using an accelerometer.

3. The method for zero-fault sample diagnosis of rolling bearings according to claim 1, characterized in that, By combining the parameters of the rolling bearing, including the inner ring, outer ring, rolling elements, and cage, and using the knowledge of the failure frequency of each key component of the rolling bearing, the failure frequency of each key component of the rolling bearing is calculated.

4. A method for diagnosing zero-fault samples of rolling bearings according to claim 1 or 3, characterized in that, The specific calculation formula for calculating the failure frequency of each key component of the rolling bearing is as follows: Among them, F i F o F b F c These are the failure frequencies of the inner ring, outer ring, rolling elements, and cage, respectively. r Where d is the operating frequency of the rolling bearing, d represents the diameter of the ball, D represents the diameter of the pitch circle of the rolling bearing, and α is the ball contact angle.

5. The method for zero-fault sample diagnosis of rolling bearings according to claim 4, characterized in that, For the calculated fault frequencies of each key component of the rolling bearing, fault frequency band data is extracted from the spectrum of vibration data under normal rolling bearing conditions. Furthermore, based on different fault octaves, the spectrum corresponding to different fault octaves is extracted. The spectrum extraction interval is represented as [ZF-Zε, ZF+Zε], where F is the fault frequency, Z is a positive integer, and ε is obtained from an empirical formula. Sure.

6. The method for zero-fault sample diagnosis of rolling bearings according to claim 1, characterized in that, The deep residual shrinkage network adopts ResNet-18 and consists of a deep residual network, a soft thresholding function, and an attention mechanism.

7. The method for zero-fault sample diagnosis of rolling bearings according to claim 1, characterized in that, Based on different feature encoders trained at different fault frequencies, a rolling bearing fault detection model based on feature encoders is constructed. The vibration data of the online monitored rolling bearing is characterized by features, and the cosine similarity is calculated with the vibration data of the benchmark rolling bearing under normal conditions. If the cosine similarity calculation result exceeds the threshold, the rolling bearing is considered to have a fault corresponding to the corresponding fault frequency; otherwise, the rolling bearing is determined not to have a fault corresponding to the corresponding fault frequency. By combining multiple rolling bearing fault detection models based on feature encoders at different fault frequencies, the fault diagnosis results of the online monitored rolling bearing at all fault frequencies are obtained.

Images