A local fault diagnosis method for rotating machinery based on mechanism and convolutional measurement network

Abstract

The present invention discloses a method for diagnosing local faults of rotating machinery based on a mechanism and a convolutional measurement network, comprising the following steps: step S1, constructing a data set for model training using the mechanism of local faults of rotating machinery; step S2, constructing a deep convolutional denoising autoencoder network in which the number of network layers is determined by the required signal compression rate and the hidden layer corresponds to the frequency of the original signal; step S3, collecting mechanical vibration signals and speed signals at the equipment end, and calculating the corresponding characteristic frequencies when local faults of gears or bearings occur at different positions of the equipment; step S4, intercepting a well-trained encoding subnetwork to replace the observation matrix in traditional compressed sensing to perform compression measurement on the vibration signal of the rotating machinery to obtain a compressed domain signal; step S5, remotely transmitting the compressed domain signal; step S6, directly extracting features of the compressed domain signal at the receiving end, and determining the fault problem of the equipment based on the extracted fault feature information.

Description

Technical Field

[0001] The present invention belongs to the field of rotating machinery fault diagnosis, and more specifically, relates to a rotating machinery local fault diagnosis method based on mechanism and convolution measurement network. Background Art

[0002] Condition monitoring of rotating machinery is crucial for its proper operation. However, collected vibration signals often contain significant noise, making feature extraction difficult. Furthermore, the sheer volume of data also places a heavy burden on long-distance transmission. Therefore, compressing the vibration response signal while preserving fault signatures is crucial. Recently, compressed sensing (CS), developed from sparse decomposition, can achieve signal compression and reconstruct high-dimensional original signals from low-dimensional observations. It has found application in fault diagnosis. Compressed sensing of mechanical vibration signals based on optimized classification segments the signal based on signal energy and reconstructs bearing fault signatures using a learned dictionary and basis pursuit algorithm (Compressed Sensing of Mechanical Vibration Signals Based on Optimized Classification, Wang Qiang, Zhang Peilin, Wang Huaiguang, Wu Dinghai, Zhang Yunqiang, Department of Vehicle and Electrical Engineering, Ordnance Engineering College, Shijiazhuang). A feature proxy and convex optimization reconstruction algorithm for compressed rolling bearing fault signals utilizes a feature proxy and a convex optimization algorithm to effectively reconstruct bearing fault signals. The sparse and convolutional characteristics of localized rolling bearing fault signals are analyzed to learn the fault impact pattern. For the compressed bearing observation signal, a proxy containing the impact moment features is constructed, and a target optimization function is established for the proxy. The Fast Iterative Shrinkage Threshold Algorithm (FISTA) is used to directly extract sparse coefficients from the proxy. The fault signal is then reconstructed by convolving the learned model with the sparse coefficients. (Feature Proxy and Convex Optimization Reconstruction Algorithm for Rolling Bearing Compressed Fault Signals, Lin Huibin, Deng Lifa, School of Mechanical and Automotive Engineering, South China University of Technology, Guangzhou, Guangdong). The aforementioned literature uses compressed sensing methods for signal reconstruction, which involves iterations of optimal solutions, making real-time reconstruction difficult. Furthermore, due to noise, traditional methods using a measurement matrix for linear observation of the signal cannot achieve significant signal compression while ensuring sufficient reconstruction accuracy. With the rapid development of artificial intelligence, deep learning can be introduced into compressed sensing to address these issues. Summary of the Invention

[0003] To address the low compression efficiency and slow signal reconstruction processes of existing remote fault diagnosis methods based on compressed sensing, this paper proposes a method for localized fault diagnosis of rotating machinery based on a mechanism and convolutional measurement network. This method can compress data in real time to facilitate long-distance transmission and directly extract features from the compressed data to rapidly complete remote diagnosis of mechanical faults. This method can be used to address the challenges of rapidly extracting fault features and the difficulty of transmitting large amounts of data over long distances.

[0004] The present invention is achieved through at least one of the following technical solutions.

[0005] A method for local fault diagnosis of rotating machinery based on mechanism and convolutional measurement network includes the following steps:

[0006] S1. Build a dataset for model training using the localized failure mechanism of rotating machinery.

[0007] S2. Construct a deep convolutional denoising autoencoder network. The number of network layers of the deep convolutional denoising autoencoder network is determined by the required signal compression rate. The hidden layer corresponds to the frequency of the original signal. The deep convolutional denoising autoencoder network is trained.

[0008] S3. Collect the mechanical vibration signal and speed signal of the equipment end, and calculate the corresponding characteristic frequency when the gear or bearing local fault occurs at different positions of the equipment;

[0009] S4, intercepting the encoding sub-network of the fully trained deep convolutional denoising autoencoder network to perform compression measurement on the vibration signal of the rotating machinery to obtain a compressed domain signal;

[0010] S5, remotely transmitting the compressed domain signal;

[0011] S6. Hilbert demodulate the compressed domain signal directly at the receiving end, and use the extracted fault feature information to determine the fault problem of the equipment.

[0012] Furthermore, the step S1 specifically includes:

[0013] S11. Establish a fault impact component based on a mathematical model of a local fault signal of a rotating machinery to obtain a noise-free sample;

[0014] S12, adding Gaussian white noise to the noise-free sample to obtain a noisy sample;

[0015] S13. Use the noisy samples as input and the fault impact components as sequence annotations to complete the construction of the dataset.

[0016] Furthermore, the deep convolutional denoising autoencoder network includes an encoding subnetwork and a decoding subnetwork, wherein the encoding subnetwork includes a convolutional layer and a pooling layer, and the decoding subnetwork includes a convolutional layer and an upsampling layer.

[0017] Furthermore, the deep convolutional denoising autoencoder network specifically includes:

[0018] The dimension of each convolutional layer output is equal to the dimension of the input, and the parameters of the convolution kernel are obtained according to the following formula:

[0019]

[0020] Where: h represents the size of the convolution kernel, p represents the size of the padding, s represents the size of the stride, d represents the dimension of the convolution layer input, and d' represents the dimension of the convolution layer output.

[0021] Furthermore, the maximum pooling layer of the deep convolutional denoising self-encoding network plays a role in signal compression and is set after the convolution layer of the encoding sub-network.

[0022] Furthermore, the number of convolution kernels in the encoding sub-network decreases layer by layer in multiples of 2 until the number of convolution kernels in the last layer is 1.

[0023] Furthermore, the step S3 specifically includes:

[0024] S31, collecting the vibration acceleration signal of the device and the rotation speed of the input shaft;

[0025] S32. Analyze the possible fault types of gears and bearings based on the equipment structure, and calculate the fault characteristic frequency based on the equipment operating conditions.

[0026] Furthermore, the step S4 specifically includes:

[0027] S41, using the data set of step S1 to train the deep convolutional denoising autoencoder network, and intercepting the fully trained encoding sub-network;

[0028] S42. Use the encoding sub-network to perform compression measurement on the rotating machinery vibration signal to obtain a compressed domain signal.

[0029] Furthermore, in step S5, in order to alleviate the pressure of long-distance transmission to the greatest extent, only the compressed domain signal and the speed condition information are transmitted wirelessly over long distances.

[0030] Furthermore, the step S6 specifically includes:

[0031] S61. Perform Hilbert demodulation on the compressed domain signal to obtain a fault characteristic frequency of the device;

[0032] S62. Compare the characteristic frequencies of possible faults of the equipment to determine the fault location of the equipment, thereby achieving the purpose of equipment health status assessment.

[0033] The present invention has the following advantages and effects compared with the prior art:

[0034] (1) This method uses the fault mechanism to construct a data set, which solves the problem of difficulty in obtaining a large number of samples, especially noise-free samples, for model training in actual engineering;

[0035] (2) The constructed deep convolutional denoising autoencoder network uses convolutional layers for feature extraction and pooling layers as a means of signal compression. The parameters of the pooling layer are adjusted according to the expected compression rate, and good feature extraction effects can still be achieved at extremely low compression rates.

[0036] (3) The deep convolutional measurement network trained with the proposed dataset has good generalization performance. After one training, it can be directly used for remote diagnosis of local faults of rotating machinery of different types and working conditions.

[0037] (4) The proposed deep convolutional measurement network is used to replace the random matrix of the traditional compressed sensing method for signal observation, which can effectively retain the fault characteristics while significantly reducing the compression rate, and fault diagnosis can be achieved in the compressed domain. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. The drawings constitute part of this application, but are only non-limiting examples of the inventive concept and are not intended to make any limitation.

[0039] Figure 1 It is a flowchart of the implementation of a method for diagnosing local faults of rotating machinery based on a mechanism and a convolutional measurement network according to the present invention;

[0040] Figure 2 is a flow chart of the data set construction method in the method of the present invention;

[0041] Figure 3 is the time domain waveform of the original vibration signal collected in the embodiment of the present invention;

[0042] Figure 4 is a compressed signal extracted by a deep convolutional measurement network in an embodiment of the present invention;

[0043] Figure 5 is a frequency domain diagram of fault features extracted using a traditional compressed sensing method in an embodiment of the present invention;

[0044] Figure 6 This is a frequency domain diagram of fault features extracted by the deep convolutional measurement network in an embodiment of the present invention. DETAILED DESCRIPTION

[0045] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0046] A method for diagnosing local faults of rotating machinery based on a mechanism and a convolutional measurement network comprises the following steps:

[0047] S1. Build a dataset for model training using the localized failure mechanism of rotating machinery.

[0048] S2. Construct a deep convolutional denoising autoencoder network, which consists of a symmetrical encoder subnetwork and a decoder subnetwork. The encoder subnetwork consists of convolutional layers and pooling layers, and the decoder subnetwork consists of convolutional layers and upsampling layers. The number of layers in the deep convolutional denoising autoencoder network is determined by the required signal compression rate, and the hidden layers correspond to the frequency of the original signal. The deep convolutional denoising autoencoder network is trained.

[0049] S3. Collect the mechanical vibration signal and speed signal of the equipment end, and calculate the corresponding characteristic frequency when the gear or bearing local fault occurs at different positions of the equipment;

[0050] S4. The encoding subnetwork (i.e., the deep convolutional measurement network) of the fully trained deep convolutional denoising autoencoder network is intercepted to replace the measurement matrix in traditional compressed sensing to perform compressive measurement on the rotating machinery vibration signal to obtain a compressed domain signal; wherein the encoding subnetwork is the encoding part of the autoencoder network;

[0051] S5, remotely transmitting the compressed domain signal;

[0052] S6. Directly extract features of the compressed domain signal at the receiving end, and use the extracted fault feature information to determine the fault problem of the device.

[0053] The step S1 specifically includes:

[0054] S11. Based on the mathematical model of local fault signals of rotating machinery in existing literature (Structural Damping Values ​​as a Function of Dynamic Resonance Stress and Defination Levels, JD Stevenson, Vice President and General Manager, Structural Mechanics Associates, Cleveland, Ohio, USA), establish the fault impulse component and obtain a noise-free sample.

[0055] S12, adding a certain amount of Gaussian white noise to the noise-free sample to obtain a noisy sample;

[0056] S13. Use the noisy samples as input and the fault impact components as sequence annotations to complete the construction of the dataset.

[0057] The step S2 specifically includes:

[0058] S21. The dimension of the output of each convolutional layer is equal to the dimension of the input, and the parameters of the convolution kernel are reasonably designed according to the following formula.

[0059]

[0060] Where: h represents the size of the convolution kernel, p represents the size of the padding, s represents the size of the stride, d represents the dimension of the convolution layer input, and d' represents the dimension of the convolution layer output.

[0061] S22 and the maximum pooling layer play the role of signal compression and are set after the convolution layer of the encoding sub-network.

[0062] S23. The number of convolution kernels in the encoding sub-network decreases layer by layer in multiples of 2 until the number of convolution kernels in the last layer is 1.

[0063] As a preferred embodiment, the decoding subnetwork is constructed using convolutional layers and upsampling layers, adopting a symmetrical structure with the encoding subnetwork. The specific number of layers in the encoding subnetwork is determined by the actual required compression ratio Δ. Each additional pooling layer also adds a corresponding convolutional layer for feature extraction. Table 1 shows the hyperparameters of the proposed deep convolutional denoising autoencoder network at a compression ratio of 6.25%.

[0064] Table 1 Hyperparameters of the proposed deep convolutional denoising autoencoder network when the compression rate is 6.25%

[0065]

[0066] As a preferred embodiment, step S3 specifically includes:

[0067] S31, collecting the vibration acceleration signal of the device and the rotation speed of the input shaft;

[0068] S32. Analyze the possible fault types of gears and bearings (such as bearing inner race fault, bearing outer race fault, etc.) based on the equipment structure, and calculate their fault characteristic frequencies according to the formula based on the equipment operating conditions (Sliding window denoising K-Singular Value Decomposition and its application on rolling bearing impact fault diagnosis Honggang Yang, Huibin Lin*, Kang Ding, School of Mechanical and Automotive Engineering, South China University of Technology, Guangzhou 510640, China).

[0069] Furthermore, in step S4:

[0070] S41. Use the data set of step S1 to train the deep convolutional denoising autoencoder network, and intercept the fully trained encoding sub-network (i.e., the deep convolutional measurement network).

[0071] S42. Use a deep convolutional measurement network to perform compression measurement on the vibration signal of the rotating machinery to obtain a compressed domain signal.

[0072] Furthermore, the step S5 specifically includes: in the step S5, in order to alleviate the long-distance transmission pressure to the greatest extent, only the compressed domain signal and the speed condition information are transmitted over long distances wirelessly.

[0073] Furthermore, the step S6 specifically includes:

[0074] S61. Perform Hilbert demodulation on the compressed domain signal to obtain a fault characteristic frequency of the device;

[0075] S62. Compare the obtained fault characteristic frequency with the characteristic frequencies of possible faults of the equipment to determine the fault location of the equipment, thereby achieving the purpose of evaluating the health status of the equipment.

[0076] Based on deep learning and compressed sensing theory, this invention utilizes the corresponding relationship between autoencoder networks and compressed sensing to propose a deep convolutional measurement network that can compress and reduce noise on signals. It also uses the fault mechanism to construct a data set, solving the problem of difficulty in collecting a large number of fault signals (especially noise-free signals) for model training in actual engineering. The deep convolutional measurement network obtained by this method can replace the observation matrix in traditional compressed sensing to perform compressed measurement of rotating machinery fault signals, thereby reducing the pressure of data transmission and storage. It can also realize fault feature extraction in the compressed domain, thereby achieving more accurate and rapid remote diagnosis of mechanical faults. To verify the feasibility and correctness of the proposed method, this embodiment uses rolling bearings with localized faults in rotating machinery as the research object, and compares the diagnostic effects of the traditional compressed sensing algorithm and the rotating machinery localized fault diagnosis method based on the mechanism and convolutional measurement network of the present invention. The measurement matrix of the comparison method is a commonly used Gaussian random matrix, the sparse dictionary and its parameter selection method refer to the shift-invariant K-SVD (ShiftInvariant K-Singular Value Decomposition) dictionary, and the reconstruction algorithm uses the CoSaMP algorithm commonly used in compressed sensing. The rolling bearing model used in the experiment is NUP311EN, with a pitch diameter of 85mm, a rolling element diameter of 18mm, 13 rolling elements, and a contact angle of 0. The inner ring fault is machined by wire cutting to create a 1mm deep and 0.2mm wide groove. The rotation frequency of the shaft where the faulty bearing is located is f n =8.33Hz, the characteristic frequency of the inner race fault is f i =65.69Hz.

[0077] Figure 4 This is the compressed signal extracted from the vibration signal using a deep convolutional measurement network. As can be seen from the figure, the original 98,304 points have been compressed to a 6,144-point signal, reducing the data volume by approximately 94%. Furthermore, the proposed method effectively preserves the impact location information in the compressed domain signal. Using these compressed signals for long-distance transmission effectively alleviates transmission pressure and allows for direct assessment of equipment health based on these compressed signals, quickly identifying fault locations.

[0078] Figure 5 and Figure 6 The demodulation spectra of the characteristic signals extracted by the comparative method and the method of the present invention are shown in the figure. As can be seen from the figure, the frequency shift f in the demodulation spectrum obtained by the proposed method is n , inner race fault characteristic frequency f i The first four harmonics and the corresponding modulation sidebands can be clearly distinguished. Compared with the demodulation spectrum of the characteristic signal extracted by the contrast method, the characteristic frequency is more prominent, and the amplitude shows a regular decreasing trend, which can intuitively diagnose the failure of the bearing inner ring.

[0079] In summary, the method for diagnosing local faults of rotating machinery based on the mechanism and convolutional measurement network described in the present invention has the following advantages when used for fault diagnosis: (1) The dataset construction method based on the fault mechanism solves the problem of difficulty in obtaining a large number of samples, especially noise-free samples, for training in actual engineering; (2) The deep convolutional measurement network trained with the proposed dataset has good generalization and can be directly used for remote diagnosis of local faults of rotating machinery of different types and working conditions after one training; (3) The proposed deep convolutional measurement network is used to replace the random matrix of the traditional compressed sensing method for signal observation, which can effectively retain the fault characteristics while significantly reducing the compression rate, and fault diagnosis can be achieved in the compressed domain.

[0080] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be considered as equivalent replacement methods and are included in the scope of protection of the present invention.

Claims

1. A local fault diagnosis method for rotating machinery based on mechanism and convolutional measurement network, characterized in that: The following steps are involved: S1. Constructing a data set for model training using a local fault mechanism of rotating machinery, including the following steps: S11. Establishing a fault impact component based on a mathematical model of a local fault signal of rotating machinery to obtain a noise-free sample; S12, adding Gaussian white noise to the noise-free sample to obtain a noisy sample; S13, using the noisy samples as input and the fault impact components as sequence annotations to complete the construction of the data set; S2. Construct a deep convolutional denoising autoencoder network. The number of network layers of the deep convolutional denoising autoencoder network is determined by the required signal compression rate. The hidden layer corresponds to the frequency of the original signal. The deep convolutional denoising autoencoder network is trained. The deep convolutional denoising autoencoder network includes an encoder subnetwork and a decoder subnetwork. The encoder subnetwork includes convolutional layers and pooling layers, while the decoder subnetwork includes convolutional layers and upsampling layers. The maximum pooling layer of the deep convolutional denoising autoencoder network plays a role in signal compression. It is set after the convolutional layer of the encoder subnetwork. The number of convolution kernels in the encoder subnetwork decreases layer by layer in multiples of 2 until the last convolution layer has 1. The deep convolutional denoising autoencoder network specifically includes: The dimension of each convolutional layer output is equal to the dimension of the input, and the parameters of the convolution kernel are obtained according to the following formula: Where: represents the size of the convolution kernel, Indicates the size of the padding. Indicates the size of the stride. represents the dimension of the convolutional layer input, Represents the dimension of the convolutional layer output; S3. Collect the mechanical vibration signal and speed signal from the equipment end and calculate the corresponding characteristic frequencies when local gear or bearing failures occur at different locations of the equipment. Specifically, S31, collecting the vibration acceleration signal of the device and the rotation speed of the input shaft; S32. Analyze the possible fault types of gears and bearings based on the equipment structure, and calculate their fault characteristic frequencies based on the equipment operating conditions; S4. Intercepting the encoding sub-network of the fully trained deep convolutional denoising autoencoder network to perform compression measurement on the rotating machinery vibration signal to obtain a compressed domain signal, specifically including: S41, using the data set of step S1 to train the deep convolutional denoising autoencoder network, and intercepting the fully trained encoding sub-network; S42. Using the encoding sub-network to perform compression measurement on the rotating machinery vibration signal to obtain a compression domain signal; S5, remotely transmitting the compressed domain signal; S6. Hilbert demodulate the compressed domain signal directly at the receiving end, and use the extracted fault feature information to determine the fault problem of the equipment.

2. The method for local fault diagnosis of rotating machinery based on mechanism and convolutional measurement network according to claim 1, characterized in that: In step S5, in order to reduce the pressure of long-distance transmission to the greatest extent, only the compressed domain signal and the speed condition information are transmitted wirelessly over long distances.

3. The method for local fault diagnosis of rotating machinery based on mechanism and convolutional measurement network according to any one of claims 1 to 2, characterized in that: The step S6 specifically includes: S61. Perform Hilbert demodulation on the compressed domain signal to obtain a fault characteristic frequency of the device; S62. Compare the characteristic frequencies of possible faults of the equipment to determine the fault location of the equipment, thereby achieving the purpose of equipment health status assessment.

Images