A rolling bearing fault diagnosis method based on a joint learning convolutional neural network

Abstract

The application discloses a rolling bearing fault diagnosis method based on a joint learning convolutional neural network, and first maps input signals to a high-dimensional feature space through a joint feature coding network, wherein representative features and fault-related features of the signals are fully coded through convolution in the high-dimensional feature space; then, a signal denoising task and a fault diagnosis task are simultaneously completed based on an attention mechanism coding network, a fault classification network and a decoder network, so that the mechanical system health state can be accurately predicted, and the fault diagnosis performance is improved.

Description

Technical Field

[0001] This invention belongs to the field of mechanical fault diagnosis technology, and more specifically, relates to a rolling bearing fault diagnosis method based on joint learning convolutional neural networks. Background Technology

[0002] Rolling bearings, as core components of rotating machinery systems, are widely used in large-scale mechanical equipment such as wind turbines, high-speed trains, and aerospace equipment. Rolling bearing failure can lead to equipment downtime and industrial production halts, or even serious safety problems and irreparable casualties. Therefore, real-time monitoring of the health status of rolling bearings is of paramount importance to ensure the normal operation of mechanical systems.

[0003] Currently, deep learning technology has gained widespread recognition in the field of fault diagnosis due to its excellent feature learning and automatic decision-making capabilities, especially convolutional neural networks (CNNs), which have been favored by researchers for their advantages in parameter sharing and nonlinear feature learning. Although deep learning algorithms have achieved significant success in fault diagnosis, and CNNs perform well in stable operating environments thanks to their powerful feature learning capabilities, a challenge remains: most mechanical systems typically operate in harsh environments, and the monitored signals are highly susceptible to environmental noise contamination. When signals contain very strong noise, existing common signal diagnosis methods struggle to achieve ideal diagnostic results. Furthermore, noise conditions are constantly changing, and existing methods are only applicable to training and testing under known noise environments, failing to address diagnoses under unknown noise conditions. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a rolling bearing fault diagnosis method based on joint learning convolutional neural networks, which can improve the fault diagnosis performance of rolling bearings in high noise environment and variable load conditions, and simultaneously achieve the tasks of signal denoising and fault diagnosis.

[0005] To achieve the above-mentioned objectives, the present invention provides a method for diagnosing rolling bearing faults based on a joint learning convolutional neural network, characterized by comprising the following steps:

[0006] (1) Acquire acceleration vibration signals;

[0007] Acceleration vibration signals of the rolling bearing under the k-th fault category and the m-th operating condition are collected and denoted as . K represents the total number of fault category numbers, n = 1, 2, ..., N, where N is the total number of sampling points, and m = 1, 2, ..., M, where M represents the total number of rolling bearing operating status numbers;

[0008] (2) Standardization processing of acceleration vibration signals;

[0009]

[0010] in, for Standardized acceleration vibration signal, yes The average value of all sampled data points. yes The standard deviation of all sampled data points;

[0011] The standardized acceleration vibration signals are used to form a training sample set:

[0012]

[0013] (3) Construct a joint learning convolutional neural network model;

[0014] The joint learning convolutional neural network model includes: a joint feature encoding network, a dual attention mechanism encoding network, a fault classification network, and a decoder network;

[0015] The joint feature encoding network consists of four cascaded convolutional layers C1 to C4, and each attention mechanism encoding network consists of four cascaded convolutional + attention modules CA1 to CA4. The output of convolutional layer C1 serves as the input to convolutional layer C2 and convolutional + attention module CA1. The output of convolutional + attention module CA1 is added to the output of convolutional layer C2 and then used as the input to convolutional + attention module CA2, and so on. Finally, the output of convolutional + attention module CA4 of one of the attention mechanism encoding networks is used as the input to the fault classification network to determine the fault category. The output of convolutional + attention module CA4 of another attention mechanism encoding network is used as the input to the decoder network to complete the denoising of the input signal.

[0016] (4) Train the joint learning convolutional neural network model;

[0017] (4.1) Set the maximum number of iterations to EPOCH, and initialize the current number of iterations epoch = 1; give the expected model training error as τ;

[0018] (4.2) Using the sample set X as the training set, let the training sample set contain training samples. from Extract l sample data As training data in a single batch, l << N;

[0019] So, in the epoch-th iteration, the training data The input signal is fed into convolutional layer C1 of the joint feature encoding network. Convolutional layer C1 extracts local features from the input signal. The extracted local features are then fed into convolutional + attention module CA1 and convolutional layer C2. Convolutional + attention module CA1 enhances the weights of the channels of interest in the local features and performs a weighted average with the feature information of other channels. Convolutional layer C2 further extracts features from the local features extracted by convolutional layer C1. The output of convolutional + attention module CA1 is then summed with the output of convolutional layer C2 and used as the input of convolutional + attention module CA2. The output of convolutional layer C2 is used as the input of convolutional layer C3, and so on.

[0020] The output of the convolutional + attention module CA4 is used as the input of the convolutional layer C5 in the fault classification network. The local features extracted by the convolutional layer C5 are transmitted to the GAP calculation module to calculate the probability of each sample data in the input signal corresponding to each fault. Then, the fault category corresponding to the maximum probability is used as the prediction result.

[0021] The local features of the output of the convolution + attention module CA4 in another attention mechanism encoding network are sequentially passed through deconvolution layers D1 to D5. The local features are restored through five deconvolution layers to obtain the denoised signal.

[0022] Calculate the loss function value

[0023]

[0024] in, Represents the τ-th sample data in the training data. The true probability value of the corresponding fault category, Represents the τ-th sample data in the training data. The predicted probability value for the corresponding fault category, Represents the τ-th sample data in the training data. The data values ​​after noise reduction;

[0025] Determine if the current iteration number epoch = EPOCH or loss < τ. If so, stop iterative training and obtain the trained joint learning convolutional neural network model; otherwise, update the joint learning convolutional neural network model with the loss value through the backpropagation algorithm, and then proceed to the next round of training.

[0026] (5) Fault classification and noise reduction of acceleration vibration signals;

[0027] Acceleration vibration signals of rolling bearings under a certain fault category are collected. After standardization processing of the acceleration vibration signals, input data of length l is extracted and input into a joint learning convolutional neural network model to output the corresponding fault category.

[0028] The objective of this invention is achieved as follows:

[0029] This invention presents a rolling bearing fault diagnosis method based on a joint learning convolutional neural network. First, the input signal is mapped to a high-dimensional feature space through a joint feature encoding network. In the high-dimensional feature space, the representative features and fault-related features of the signal are fully encoded through convolution. Then, based on an attention mechanism encoding network, a fault classification network, and a decoder network, the signal denoising task and the fault diagnosis task are completed simultaneously. This method can accurately predict the health status of the mechanical system and improve the performance of fault diagnosis.

[0030] Meanwhile, the rolling bearing fault diagnosis method based on joint learning convolutional neural networks of the present invention also has the following beneficial effects:

[0031] (1) This invention utilizes the powerful learning capability of joint learning convolutional neural networks to effectively restore noise-contaminated signals to relatively pure signals, and can also perform Gaussian denoising and fault diagnosis under unknown noise conditions.

[0032] (2) This invention integrates the vibration signal denoising task and the fault diagnosis task into an end-to-end network framework for the first time. This architecture enables the fault diagnosis task to obtain good noise robustness by sharing network parameters and features, and the denoising task can obtain fault information specific to the fault diagnosis task, thereby obtaining better denoising effect for specific fault signals. Attached Figure Description

[0033] Figure 1 This is a flowchart of the rolling bearing fault diagnosis method based on joint learning convolutional neural network of the present invention;

[0034] Figure 2 This is a schematic diagram of the network framework of a joint learning convolutional neural network; Detailed Implementation

[0035] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be particularly noted that in the following description, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.

[0036] Example

[0037] Figure 1 This is a flowchart of the rolling bearing fault diagnosis method based on joint learning convolutional neural networks of the present invention.

[0038] In this embodiment, as Figure 1As shown, the present invention provides a rolling bearing fault diagnosis method based on a joint learning convolutional neural network, comprising the following steps:

[0039] S1. Obtain acceleration vibration signals;

[0040] Acceleration vibration signals of the rolling bearing under the k-th fault category and the m-th operating condition are collected and denoted as . K represents the total number of fault category numbers, n = 1, 2, ..., N, where N is the total number of sampling points, m = 1, 2, ..., M, where M represents the total number of operating status numbers of the rolling bearing; in this embodiment, the rolling bearing operates under different harsh environments, and the collected acceleration vibration signals include environmental noise.

[0041] S2. Standardization processing of acceleration vibration signals;

[0042]

[0043] in, for Standardized acceleration vibration signal, yes The average value of all sampled data points. yes The standard deviation of all sampled data points;

[0044] The standardized acceleration vibration signals are used to form a training sample set:

[0045]

[0046] S3. Construct a joint learning convolutional neural network model;

[0047] like Figure 2 As shown, the joint learning convolutional neural network model includes: a joint feature encoding network, a dual attention mechanism encoding network, a fault classification network, and a decoder network;

[0048] The joint feature encoding network consists of four cascaded convolutional layers C1 to C4, each employing the leaky ReLU activation function. Specifically, convolutional layer C1 has a kernel size of 16×1, a stride of 4, and a leaky ReLU parameter r of 0.5; convolutional layer C2 has a kernel size of 9×1, a stride of 4, and a leaky ReLU parameter r of 0.5; convolutional layer C3 has a kernel size of 6×1, a stride of 2, and a leaky ReLU parameter r of 0.5; and convolutional layer C4 has a kernel size of 3×1, a stride of 2, and a leaky ReLU parameter r of 0.5.

[0049] like Figure 2 As shown, the dual-attention mechanism encoding network consists of two branches, an upper branch and a lower branch. The structures of the two branches are identical, each composed of four cascaded convolutional + attention modules CA1 to CA4. The upper branch is used for fault classification, and the lower branch is used for signal denoising.

[0050] Among them, the parameter settings of the convolutional layers in the convolution + attention modules CA1 to CA4 are exactly the same, specifically: the convolutional kernel size of the convolutional layer is set to 1×1, the stride is set to 4, and the parameter r of the activation function leaky ReLU is set to 0.5.

[0051] The decoder network consists of five cascaded deconvolutional layers, D1 to D5, each employing the leaky ReLU activation function. Specifically, deconvolutional layer D1 has a kernel size of 3×1, a stride of 2, and a leaky ReLU parameter r of 0.5; deconvolutional layer D2 has a kernel size of 6×1, a stride of 2, and a leaky ReLU parameter r of 0.5; deconvolutional layer D3 has a kernel size of 9×1, a stride of 4, and a leaky ReLU parameter r of 0.5; deconvolutional layer D4 has a kernel size of 12×1, a stride of 4, and a leaky ReLU parameter r of 0.5; and deconvolutional layer D5 has a kernel size of 1×1, a stride of 1, and a leaky ReLU parameter r of 0.5.

[0052] In this embodiment, the input signal is input through convolutional layer C1, and the output signal of convolutional layer C1 is used as the input to convolutional layer C2 and convolutional + attention module CA1. The output of convolutional + attention module CA1 is added to the output of convolutional layer C2 and used as the input to convolutional + attention module CA2, and so on. Finally, the output of convolutional + attention module CA4 of one of the attention mechanism encoding networks is used as the input to the fault classification network to determine the fault category. The output of convolutional + attention module CA4 of another attention mechanism encoding network is used as the input to the decoder network to complete the denoising of the input signal.

[0053] S4. Train the joint learning convolutional neural network model;

[0054] S4.1 Set the maximum number of iterations to EPOCH, and initialize the current iteration number epoch = 1; give the expected model training error as τ;

[0055] S4.2. Using the sample set X as the training set, let the training sample set contain training samples. from Extract l sample data As training data in a single batch, l << N;

[0056] So, in the epoch-th iteration, the training data The input signal is fed into convolutional layer C1 of the joint feature encoding network. Convolutional layer C1 extracts local features from the input signal. The extracted local features are then fed into convolutional + attention module CA1 and convolutional layer C2. Convolutional + attention module CA1 enhances the weights of the channels of interest in the local features and performs a weighted average with the feature information of other channels. Convolutional layer C2 further extracts features from the local features extracted by convolutional layer C1. The output of convolutional + attention module CA1 is then summed with the output of convolutional layer C2 and used as the input of convolutional + attention module CA2. The output of convolutional layer C2 is used as the input of convolutional layer C3, and so on.

[0057] The output of the convolutional + attention module CA4 is used as the input of the convolutional layer C5 in the fault classification network. The local features extracted by the convolutional layer C5 are transmitted to the GAP calculation module to calculate the probability of each sample data in the input signal corresponding to each fault. Then, the fault category corresponding to the maximum probability is used as the prediction result.

[0058] The local features of the output of the convolution + attention module CA4 in another attention mechanism encoding network are sequentially passed through deconvolution layers D1 to D5. The local features are restored through five deconvolution layers to obtain the denoised signal.

[0059] Calculate the loss function value

[0060]

[0061] in, Represents the τ-th sample data in the training data. The true probability value of the corresponding fault category, Represents the τ-th sample data in the training data. The predicted probability value for the corresponding fault category, Represents the τ-th sample data in the training data. The data values ​​after noise reduction;

[0062] Determine if the current iteration number epoch = EPOCH or loss < τ. If so, stop iterative training and obtain the trained joint learning convolutional neural network model; otherwise, update the joint learning convolutional neural network model with the loss value through the backpropagation algorithm, and then proceed to the next round of training.

[0063] S5. Fault classification and noise reduction of acceleration vibration signals;

[0064] Acceleration vibration signals of rolling bearings under a certain fault category are collected. After standardization processing of the acceleration vibration signals, input data of length l is extracted and input into a joint learning convolutional neural network model to output the corresponding fault category.

[0065] To better illustrate the technical effects of the present invention, a sub-experiment included in a specific embodiment is used to experimentally verify the invention. In this experimental verification, a rolling bearing test bench is used to simulate the working process of a rolling bearing. The rolling bearing fault diagnosis test bench used in this embodiment includes a drive motor, a belt drive system, a vertical loading device, a lateral loading device, two fan motors, and a control system. The vertical and lateral load loading devices are designed to simulate the axial and lateral loads borne by the rolling bearing. The two fan motors can generate airflow opposite to the running direction of the rolling bearing. Two accelerometers ensure that vibrations in both the horizontal and vertical directions of the rolling bearing can be detected, and the signal sampling frequency is set to 5120Hz.

[0066] In this embodiment, 10 typical fault states and 1 healthy state were selected, as described in Table 1.

[0067] Table 1. Eleven bearing states used in the embodiments

[0068]

[0069]

[0070] For each fault condition, four operating speeds were designed: 60 km / h, 90 km / h, 120 km / h, and 150 km / h, and four different vertical loads: 56 kN, 146 kN, 236 kN, and 272 kN. Furthermore, to better simulate the complex operating environment of high-speed trains, Gaussian white noise with different signal-to-noise ratios (SNR) was added to the original signal. In this experimental verification, three sets of noise signals with different SNRs (-6 dB, 0 dB, and 6 dB) were set up to simulate the strong, medium, and weak noise conditions of the rolling bearing. When a small noise was added, its impact on the vibration signal was minimal. However, when a large noise was added, the original waveform of the vibration signal was completely destroyed, making it difficult to identify. Since the obtained vibration signal is a very long time series, we used a sliding segmentation method to obtain more training samples. The step size of the sliding segmentation was set to 256, and the length of each sample was set to 2048 to ensure that each sample contains a complete periodic signal. After sliding segmentation, a total of 128,874 training samples and 41,258 test samples were obtained.

[0071] In this experiment, the proposed joint learning convolutional neural network model was implemented using the Keras deep learning framework and Python 3.6, and trained and tested on a server with a GTX 2080 graphics card with 8GB of VRAM. During training, the batch size was set to 256, and the Adam optimizer was used to optimize the network parameters with a learning rate of 0.0001.

[0072] First, this experiment verifies the joint learning architecture in SNR dB The performance at -6dB was analyzed and compared with the results of F-Task-CNN and D-Task-CNN trained separately, as shown in Table 2.

[0073] Table 2. Performance evaluation results of the three methods

[0074] Performance evaluation indicators F-Task-CNN D-Task-CNN JL-CNN ACC 0.715 0.002 N / A 0.838 0.001 SNR N / A 2.348 0.003 2.789 0.005 MSE N / A 0.582 0.001 0.526 0.001

[0075] The experimental results above demonstrate that the joint learning framework improves diagnostic and denoising performance. Compared to F-Task-CNN, the joint learning neural network increases diagnostic accuracy by 12.3%, indicating that with the help of SD-tasks, the network can acquire more valuable information. Furthermore, for SD-tasks measured by SNR and MSE, the joint learning convolutional neural network outperforms D-Task-CNN. In conclusion, the joint learning framework is more suitable for fault diagnosis of rolling bearings.

[0076] Subsequently, this experiment compared the joint learning convolutional neural network with three known fault diagnosis methods and four known denoising methods under noise conditions of -6dB, -3dB, and 0dB, respectively. The results are shown in Table 3. It can be seen that the joint learning convolutional neural network JL-Network performs best in both error diagnosis and denoising. For the error diagnosis task, compared with WDCNN, the accuracy of the joint learning convolutional neural network is improved by 10%, 6.1%, and 2.6% under the three noise conditions, respectively, indicating that the present invention has excellent error diagnosis performance under strong noise conditions. For the denoising task, the joint learning convolutional neural network outperforms the two traditional denoising methods Wavelet and EMD, as well as the other two deep learning denoising methods DAE and CAE.

[0077] Table 3. Experimental results of different methods under three noise conditions

[0078]

[0079] Finally, this experiment analyzed the complexity of the joint learning model and compared it with other models. The training time for 100 epochs and the testing time required to predict a batch size (256 samples) were recorded, and the results are shown in Table 4. It can be seen that because the joint learning convolutional neural network needs to process two tasks simultaneously, it has two different branches and a complex network structure. Compared with other deep learning models, their training times are almost identical. However, due to the complexity of the joint learning neural network, it requires a longer testing time, but only slightly longer than other models, which is sufficient for real-world scenarios.

[0080] Table 4. Time results of different methods

[0081]

[0082] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.

Claims

1. A method for diagnosing rolling bearing faults based on a joint learning convolutional neural network, characterized in that, Includes the following steps: (1) Acquire acceleration vibration signals; Acceleration vibration signals of the rolling bearing under the k-th fault category and the m-th operating condition are collected and denoted as . K represents the total number of fault category numbers, n = 1, 2, ..., N, where N is the total number of sampling points, and m = 1, 2, ..., M, where M represents the total number of rolling bearing operating status numbers; (2) Standardization processing of acceleration vibration signals; in, for Standardized acceleration vibration signal, yes The average value of all sampled data points. yes The standard deviation of all sampled data points; The standardized acceleration vibration signals are used to form a training sample set: (3) Construct a joint learning convolutional neural network model; The joint learning convolutional neural network model includes: a joint feature encoding network, a dual attention mechanism encoding network, a fault classification network, and a decoder network; The joint feature encoding network consists of four cascaded convolutional layers C1 to C4, and each attention mechanism encoding network consists of four cascaded convolutional + attention modules CA1 to CA4. The output of convolutional layer C1 serves as the input to convolutional layer C2 and convolutional + attention module CA1. The output of convolutional + attention module CA1 is added to the output of convolutional layer C2 and then used as the input to convolutional + attention module CA2, and so on. Finally, the output of convolutional + attention module CA4 of one of the attention mechanism encoding networks is used as the input to the fault classification network to determine the fault category. The output of convolutional + attention module CA4 of another attention mechanism encoding network is used as the input to the decoder network to complete the denoising of the input signal. (4) Train the joint learning convolutional neural network model; (4.1) Set the maximum number of iterations to EPOCH, and initialize the current number of iterations epoch = 1; give the expected model training error as τ; (4.2) Using the sample set X as the training set, let the training sample set contain training samples. from Extract l sample data As training data in a single batch, l << N; So, in the epoch-th iteration, the training data The input signal is fed into convolutional layer C1 of the joint feature encoding network. Convolutional layer C1 extracts local features from the input signal. The extracted local features are then fed into convolutional + attention module CA1 and convolutional layer C2. Convolutional + attention module CA1 enhances the weights of the channels of interest in the local features and performs a weighted average with the feature information of other channels. Convolutional layer C2 further extracts features from the local features extracted by convolutional layer C1. The output of convolutional + attention module CA1 is then summed with the output of convolutional layer C2 and used as the input of convolutional + attention module CA2. The output of convolutional layer C2 is used as the input of convolutional layer C3, and so on. The output of the convolutional + attention module CA4 is used as the input of the convolutional layer C5 in the fault classification network. The local features extracted by the convolutional layer C5 are transmitted to the GAP calculation module to calculate the probability of each sample data in the input signal corresponding to each fault. Then, the fault category corresponding to the maximum probability is used as the prediction result. The local features of the output of the convolution + attention module CA4 in another attention mechanism encoding network are sequentially passed through deconvolution layers D1 to D5. The local features are restored through five deconvolution layers to obtain the denoised signal. Calculate the loss function value in, Represents the τ-th sample data in the training data. The true probability value of the corresponding fault category, Represents the τ-th sample data in the training data. The predicted probability value for the corresponding fault category, Represents the τ-th sample data in the training data. The data values ​​after noise reduction; Determine if the current iteration number epoch = EPOCH or loss < τ. If so, stop iterative training and obtain the trained joint learning convolutional neural network model; otherwise, update the joint learning convolutional neural network model with the loss value through the backpropagation algorithm, and then proceed to the next round of training. (5) Fault classification and noise reduction of acceleration vibration signals; Acceleration vibration signals of rolling bearings under a certain fault category are collected. After standardization processing of the acceleration vibration signals, input data of length l is extracted and input into a joint learning convolutional neural network model to output the corresponding fault category.

2. The rolling bearing fault diagnosis method based on joint learning convolutional neural network according to claim 1, characterized in that, The joint feature encoding network comprises four convolutional layers, each employing the leaky ReLU activation function. Specifically, convolutional layer C1 has a kernel size of 16×1, a stride of 4, and a leaky ReLU parameter r of 0.5; convolutional layer C2 has a kernel size of 9×1, a stride of 4, and a leaky ReLU parameter r of 0.5; convolutional layer C3 has a kernel size of 6×1, a stride of 2, and a leaky ReLU parameter r of 0.5; and convolutional layer C4 has a kernel size of 3×1, a stride of 2, and a leaky ReLU parameter r of 0.

5.

3. The rolling bearing fault diagnosis method based on joint learning convolutional neural network according to claim 1, characterized in that, The dual attention mechanism coding network includes two branches, an upper branch and a lower branch. The structures of the upper and lower branches are exactly the same, each consisting of four cascaded convolutional + attention modules CA1 to CA4. The upper branch is used for fault classification, and the lower branch is used for signal denoising. Among them, the parameter settings of the convolutional layers in the convolution + attention modules CA1 to CA4 are exactly the same, specifically: the convolutional kernel size of the convolutional layer is set to 1×1, the stride is set to 4, and the parameter r of the activation function leaky ReLU is set to 0.

5.

4. The rolling bearing fault diagnosis method based on joint learning convolutional neural network according to claim 1, characterized in that, The decoder network comprises five cascaded deconvolutional layers D1 to D5, each employing the leaky ReLU activation function. Specifically, deconvolutional layer D1 has a kernel size of 3×1, a stride of 2, and a leaky ReLU parameter r of 0.5; deconvolutional layer D2 has a kernel size of 6×1, a stride of 2, and a leaky ReLU parameter r of 0.5; deconvolutional layer D3 has a kernel size of 9×1, a stride of 4, and a leaky ReLU parameter r of 0.5; deconvolutional layer D4 has a kernel size of 12×1, a stride of 4, and a leaky ReLU parameter r of 0.5; and deconvolutional layer D5 has a kernel size of 1×1, a stride of 1, and a leaky ReLU parameter r of 0.5.

Images