Bearing fault intelligent diagnosis method based on order tracking and convolutional network
Through the combination of order tracking and one-dimensional convolutional neural network, the modal aliasing problem of bearing fault diagnosis under variable speed conditions is solved, efficient fault feature extraction and accurate diagnosis is achieved, and the engineering practicality and accuracy of bearing fault diagnosis is improved.
Patent Information
- Application Number
- CN202510459168.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-14
- Publication Date
- 2025-07-29
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing technology has modal aliasing phenomenon in bearing fault diagnosis under variable speed conditions, resulting in a decrease in diagnostic accuracy and high dependence on comprehensive data sets, which limits practical applications.
The order tracking technology is used to resample signal and build one-dimensional convolutional neural network. The order tracking algorithm is used to preprocess non-periodic signals, and a one-dimensional multi-layer convolutional neural network is built for self-extracting of fault features, reducing the requirements for the integrity of the training data set.
Accurate fault diagnosis under variable speed conditions is achieved, generalization ability is improved, computational complexity and dependence on data sets are reduced, and engineering practicality is improved.
Smart Images
Figure CN120385506A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to an intelligent diagnosis method for bearing faults, and more specifically to an intelligent diagnosis method for bearing faults based on order tracking and convolutional networks. Background Art
[0002] Rolling bearings, as core components in precision electromechanical systems, often operate in environments subject to significant temperature and humidity fluctuations and endure alternating loads for extended periods. Consequently, bearings are susceptible to defects and damage such as pitting and spalling during service. To ensure the continued safe operation of electromechanical equipment, it is imperative to deploy real-time condition monitoring systems at the edge of the equipment and establish efficient health management methods to accurately detect early-stage bearing failures. This not only enables the timely detection of potential problems but also prevents serious accidents caused by untimely fault detection. Therefore, intelligent bearing fault diagnosis has become a key area of technological development for defense equipment.
[0003] In the study of bearing fault diagnosis based on vibration information, traditional neural networks have demonstrated good results in diagnosing bearing faults under constant speed conditions. Existing neural network models possess powerful feature learning capabilities, which can avoid manual feature extraction, improve computational efficiency, and provide good identification of bearing fault types and their corresponding severity. However, their network structures are often relatively complex, making practical implementation difficult. Furthermore, intelligent classification algorithms based on deep neural networks suffer from poor robustness, requiring that the data in the test and training datasets be identically distributed. Otherwise, the classification accuracy of the intelligent fault diagnosis algorithms drops sharply. When the monitored bearings operate under variable speed conditions, the distribution of fault feature data changes with speed. This speed-induced modal aliasing significantly impacts the performance of the fault diagnosis network, leading to a significant decrease in diagnostic accuracy. Furthermore, in practical industrial applications, the cost of constructing a fully labeled fault dataset covering all speed conditions is extremely high, limiting the feasibility of this approach in real-world industrial scenarios. Therefore, effectively addressing modal aliasing under non-uniform speed conditions and reducing reliance on comprehensive datasets remain key challenges in the field of bearing fault diagnosis. Summary of the Invention
[0004] The present invention provides an intelligent bearing fault diagnosis method based on order tracking and convolutional networks, aiming to achieve accurate diagnosis of bearing faults under variable speed conditions.
[0005] The above objectives are achieved through the following technical solutions:
[0006] An intelligent bearing fault diagnosis method based on order tracking and convolutional network includes the following steps:
[0007] Step 1: Sample the bearing operation vibration signal to obtain the original monitoring signal;
[0008] Step 2: Based on the original signal base sampling frequency and the rotational speed information of the bearing extracted at any moment on the time axis, take the lowest rotational speed in the rotational speed sequence as the base rotational speed;
[0009] Step 3: Take the original data sampling frequency as the base sampling frequency, that is, the signal sampling frequency at the base rotational speed;
[0010] Step 4: Based on the base sampling frequency, base rotational speed, and the rotational speed of the current sample sequence, calculate the current sample resampling frequency, determine the original sampling time nodes of each original vibration point in the original monitoring data sequence, calculate the resampling time nodes of each group of monitoring data according to the sample resampling frequency, and calculate the value of the current sampling time point by interpolation;
[0011] Step 5: Use the current sampling time point planning and interpolation method to resample each group of original samples to obtain the resampled sequence, and then sample according to the input sample data point length requirement to construct the training set and test set;
[0012] Step 6: Neural network construction and diagnosis: Construct a one-dimensional multi-layer convolutional neural network, use a wide convolutional kernel for the first layer, and use small convolutional kernels for subsequent layers to capture features and achieve self-extraction of fault features.
[0013] The beneficial effects of an intelligent bearing fault diagnosis method based on order tracking and convolutional network of the present invention are as follows:
[0014] Based on the order tracking technology, resample the original non-stationary monitoring signal to achieve the purpose of sampling the monitoring signal at a fixed step size, reorganize the discrete monitoring signal, preprocess the non-periodic signal through the order tracking algorithm, solve the problem of the change of spectral characteristics under variable speed conditions, convert the non-steady signal into a steady signal, and overcome the influence of rotational speed fluctuation on the distribution of fault characteristics. On the basis of signal reorganization, use a one-dimensional convolutional neural network to adaptively extract the weak fault feature signals hidden in the resampled vibration signal. It does not include a domain adaptation network, reducing the design and calculation complexity. Avoid the complex operations of manually designing and extracting fault data, improve engineering practicability. Reduce the requirements for the integrity of the training data set and enhance the generalization ability. Finally, realize the "end-to-end" intelligent diagnosis of the monitoring signal under variable rotational speed conditions. Description of the Drawings
[0015] Figure 1 Shows a flowchart of an intelligent bearing fault diagnosis method based on order tracking and convolutional network;
[0016] Figures 2 to 4 Shows the effect diagram of the order tracking algorithm;
[0017] Figure 5 Shows the one-dimensional convolutional fault diagnosis neural network composition diagram;
[0018] Figure 6 Shows the front view of the bearing fault diagnosis system;
[0019] Figure 7 Shows the side view of the bearing fault diagnosis system;
[0020] Figure 8 Shows the network training and testing process diagrams of datasets A, B, and C.
[0021] In the figure: 1, three-phase induction motor; 2, shaft; 3, pedestal bearing; 4, vibration sensor; 5, bearing with measured fault; 6, driver; 7, tachometer. Specific implementation manner
[0022] An intelligent bearing fault diagnosis method based on order tracking and convolutional network, comprising the following steps:
[0023] Step 1: Deploy an acceleration sensor and a photoelectric sensor on the monitored bearing in the fault state, and connect them to the signal acquisition system.
[0024] Step 2: Perform high-frequency vibration sampling on the bearing operation vibration signals in different types of fault (inner race fault, outer race fault, cage fault) states to obtain the original monitoring signal x i (i = 1, 2,..., n). Among them, high frequency refers to f s = 25.6 kHz.
[0025] Step 3: Based on the original signal base sampling frequency f s and the rotational speed information f r,i (i = 1, 2,..., n) extracted at any moment on the time axis of the bearing, to ensure that in the entire rotational speed change range, each order component can be completely sampled and reconstructed, and to avoid the information loss problem caused by insufficient sampling at high rotational speeds, take the lowest rotational speed in the rotational speed sequence f r,i as the base rotational speed, which is based on the lowest rotational speed of all samples, rather than a single sample. That is, f base = min(f r,1 , f r,2 ,..., f r,n )
[0026] Step 4: Take the sampling frequency f s of the original data as the base sampling frequency RF base , and this base sampling frequency is also the signal sampling frequency at the base rotational speed, that is
[0027] RFbase = f s
[0028] Step 5: Calculate the current sample resampling frequency based on the base sampling frequency, base rotational speed, and current sample sequence rotational speed, that is
[0029]
[0030] Step 6: Determine the original monitoring data sequence x i (i = 1, 2,..., n) for each original vibration point's original sampling time node t i,j (i = 1, 2,..., n; j = 1, 2,..., m i ),that is
[0031]
[0032] Step 7: Based on the resampling frequency RF of each group of monitoring data i , calculate the resampling time node t of each group of monitoring data r,i,j , that is
[0033]
[0034] Step 8: Use interpolation to calculate the value at the current sampling time point, that is
[0035]
[0036] Step 9: Based on the above time point planning method and interpolation method, resample each group of original samples x i to obtain the resampled sequence x of each group of samples r,i (i = 1, 2,..., n), and then sample the resampled sequence according to the input sample data point length requirement to form a training set and a test set.
[0037] Step 10: Construct a one-dimensional multi-layer convolutional neural network. The first layer uses a wide convolutional kernel form (64×1 - 128×1) to increase the receptive field of the original data and reduce the influence of random noise; subsequent layers use small convolutional kernels (2×1 - 3×1) to facilitate capturing the detailed features contained in the subsequent data and achieve self-extraction of fault features.
[0038] Specifically, the one-dimensional convolutional neural network for intelligent bearing fault diagnosis in the present invention includes four convolutional layers and two fully connected layers. Among them, the first convolutional layer contains 16 convolutional kernels, each convolutional kernel has a length of 64, and the convolution calculation step size is 8; the convolutional kernels in the second, third, and fourth convolutional layers all have a length of 3, the numbers of convolutional kernels are 32, 64, and 64 respectively, and the step size is 1, which are used to extract the detailed features of the subsequent intermediate layer data. The activation layer uses the ReLU function, and the pooling layer uses the max pooling method. The two fully connected layers contain 1024 and 20 neuron nodes respectively, the input layer contains 2048 neurons, and the output layer contains 4 neurons. During the network training process of the present invention, the stochastic gradient descent method is adopted.
[0039] For network parameter update, the loss function of the fault diagnosis network is:
[0040]
[0041] where p(x) is the distribution of the neural network output result; q(x) is the distribution of the label value.
[0042] Table 1 Specific structural parameters of the fault diagnosis network model
[0043]
[0044] The present invention conducts experimental research based on a bearing fault diagnosis system. The bearing fault diagnosis system includes a three-phase induction motor 1, a shaft 2 fixedly connected to the output shaft of the three-phase induction motor 1. The left side of the shaft 2 is rotatably connected to the first pedestal bearing 3. The bearing of the second pedestal bearing 3 is removed, the right end of the shaft 2 is inserted into the bearing housing of the second pedestal bearing 3, and a vibration sensor 4 is fixedly connected to the upper end of the bearing housing of the second pedestal bearing 3. The outer ring of the measured faulty bearing 5 is fixed in the bearing housing of the second pedestal bearing 3, and the inner ring of the measured faulty bearing 5 is fixedly connected to the shaft 2. The three-phase induction motor 1 is electrically connected to a driver 6, and the driver 6 is used to control the rotation speed of the three-phase induction motor 1. It also includes a tachometer 7 for measuring the rotation speed of the output shaft of the three-phase induction motor 1. An accelerometer is used to measure the rotational acceleration of the bearing.
[0045] The bearings to be tested include three states: healthy, outer ring fault, and inner ring fault. Vibration signals in the three states are collected under four rotational speed conditions of 900 RPM, 1020 RPM, 1140 RPM, and 1260 RPM respectively, and the sampling frequency is 25.6 kHz. The rotational speed of the data bearings in the training set is 900 RPM, and the rotational speeds of the data in the test set are 1020 RPM, 1140 RPM, and 1260 RPM respectively, which are Dataset A, Dataset B, and Dataset C in sequence. The number of samples collected in each state at each rotational speed is 3000.
[0046] In this field, common similar intelligent fault diagnosis methods include 1DCNN network, AlexNet network and VGG network. Among them, the AlexNet network introduces the ReLU activation function to replace the traditional Sigmoid and Tanh functions, effectively solving the problem of gradient disappearance. It uses Max-Pooling to reduce the computational amount while retaining key information. The VGG network (VGG, visual Geometry Group) uses small convolutional kernels to reduce the number of parameters. At the same time, it makes up for the limitation of the receptive field through depth, can capture higher-level features in the image, significantly improves the classification performance, has strong versatility, and performs excellently in a variety of visual tasks.
[0047] Based on the collected bearing fault data, this method is used to diagnose bearing faults. The diagnostic effects of the proposed OT-1DCNN network and the control group algorithms on the data set are shown in Table 2.
[0048] Table 2: Comparison results of different algorithms based on the self-made data set
[0049]
[0050] As can be seen from Table 2, the proposed OT-1DCNN algorithm has a higher fault diagnosis accuracy than the VGG Net network on the three sub-data sets by 2.62%, 0.89% and 13.69% respectively, higher than the 1DCNN network by 3.21%, 7.27% and 11.00% respectively, and higher than the Alex Net network by 1.18%, 2.28% and 13.33% respectively. In terms of the overall classification accuracy, the proposed OT-1DCNN algorithm reaches an accuracy of 92.81%. The comparison experiment results show that the proposed OT-1DCNN algorithm has good cross-domain diagnosis performance for processing fault diagnosis tasks under variable speed conditions.
Claims
1. An intelligent bearing fault diagnosis method based on order tracking and convolutional network, comprising the following steps: Step 1: Sample the bearing operation vibration signal to obtain the original monitoring signal; Step 2: Based on the original signal base sampling frequency and the rotational speed information of the bearing extracted at any moment on the time axis, take the lowest rotational speed in the rotational speed sequence as the base rotational speed; Step 3: Take the original data sampling frequency as the base sampling frequency, that is, the signal sampling frequency at the base rotational speed; Step 4: Based on the base sampling frequency, base rotational speed, and the rotational speed of the current sample sequence, calculate the current sample resampling frequency, determine the original sampling time nodes of each original vibration point in the original monitoring data sequence, calculate the resampling time nodes of each group of monitoring data according to the sample resampling frequency, and calculate the value of the current sampling time point by interpolation; Step 5: Use the current sampling time point planning and interpolation method to resample each group of original samples to obtain a resampled sequence, and then sample according to the input sample data point length requirement to construct a training set and a test set; Step 6: Neural network construction and diagnosis: Construct a one-dimensional multi-layer convolutional neural network. The first layer uses a wide convolutional kernel, and the subsequent layers use small convolutional kernels to capture features to achieve self-extraction of fault features.
2. The intelligent bearing fault diagnosis method based on order tracking and convolutional network according to claim 1. During signal acquisition, when the monitored bearing is in a fault state, the deployed acceleration sensor and photoelectric sensor are connected to the signal acquisition system to perform vibration sampling on the bearing operation vibration signals in different types of fault states, so as to obtain the original monitoring signals x that cover vibration signals and noise signals. i (i = 1, 2,..., n).
3. The intelligent bearing fault diagnosis method based on order tracking and convolutional network according to claim 2, where the original signal base sampling frequency is f s , and the rotational speed information of the bearing extracted at any moment on the time axis is f r,i (i = 1, 2,..., n), and the rotational speed sequence is f r,i , that is, f base = min(f r,1 , f r,2 ,..., f r,n ).
4. The intelligent bearing fault diagnosis method based on order tracking and convolutional network according to claim 3, which combines a sensor to process the vibration signal in real time and outputs a fault classification result. It includes a motor, the inner ring of the measured fault bearing is fixedly connected to the output shaft of the motor, the outer ring of the measured fault bearing is fixedly connected in the bearing seat, and a vibration sensor is installed on the bearing seat. The sample resampling frequency is 5. The intelligent bearing fault diagnosis method based on order tracking and convolutional network according to claim 4, the original sampling time node 6. The resampling time node of the intelligent bearing fault diagnosis method based on order tracking and convolutional network according to claim 5 7. The value at the current sampling time point of the intelligent bearing fault diagnosis method based on order tracking and convolutional network according to claim 6 8. The intelligent bearing fault diagnosis method based on order tracking and convolutional network according to claim 1, the one-dimensional multi-layer convolutional neural network includes four convolutional layers and two fully connected layers. Among them, the first convolutional layer contains 16 convolutional kernels, the length of each convolutional kernel is 64, and the convolutional calculation step size is 8; the lengths of the convolutional kernels in the second, third, and fourth convolutional layers are all 3, the numbers of convolutional kernels are 32, 64, and 64 respectively, and the step size is 1, which is used to extract the detailed features of the subsequent intermediate layer data. The activation layer uses the ReLU function, and the pooling layer uses the maximum pooling method. The two fully connected layers contain 1024 and 20 neuron nodes respectively, the input layer contains 2048 neurons, and the output layer contains 4 neurons.
9. The intelligent bearing fault diagnosis method based on order tracking and convolutional network according to claim 1, during the network training process, the stochastic gradient descent method is adopted.
10. A bearing fault diagnosis system, characterized in that, Use the intelligent bearing fault diagnosis method based on order tracking and convolutional network according to any one of claims 1-9, which combines a sensor to process the vibration signal in real time and outputs a fault classification result.