Paddle shaft rolling bearing fault diagnosis method based on QCNN-Transformer

Abstract

The invention relates to a QCNN-Transform based fault diagnosis method for a propeller shaft rolling bearing. The QCNN-Transform based fault diagnosis method for the propeller shaft rolling bearing comprises the following steps of: firstly, carrying out According to the method, a quadratic convolutional neural network is innovatively introduced, a power function is ingeniously fused in a quadratic neuron calculation process, and an additional nonlinear mapping mechanism is added. During signal processing, a time domain and frequency domain combined noise reduction strategy is adopted, a quadratic convolutional neural network is used for fine processing of signals in the time domain, and Fourier transform is combined in the frequency domain for deep analysis. In order to strengthen the feature extraction capability, secondary convolution and a Transform neural network are organically combined to construct a brand new model architecture. The method also integrates a cross entropy loss algorithm and a cosine annealing algorithm to improve the convergence speed of the model. Compared with a traditional method, the method does not need a complex process and special preprocessing of bearing information, and feature extraction is more convenient. Under data of different noise intensities and load powers, the QCNN-Transformer model constructed by the method can stably and efficiently obtain relatively high classification accuracy, and a new path is developed for the field of fault diagnosis of the propeller shaft rolling bearing.

Description

Technical Field

[0001] The present invention relates to the technical field of rolling bearings, and in particular to a QCNN-Transformer-based propeller shaft rolling bearing fault diagnosis method. Background Art

[0002] With the continuous advancement of industrial automation and intelligence, the operating speed and automation level of rotating machinery have significantly increased. Rolling bearings play a core role in these mechanical systems, and their performance directly impacts the efficiency and stability of the entire system. Ship shafting is a vital component of a ship, transmitting engine power to the propeller, rotating the propeller and propelling the ship. Ship shafting is a type of rotating machinery. Condition monitoring and fault diagnosis of key components in rotating machinery systems, along with the timely development of maintenance strategies, have become an important means and key technology for ensuring the safe and stable operation of equipment.

[0003] Given that rolling bearings operate in closed and complex environments, they are susceptible to aging and wear, which can lead to failures, resulting in safety incidents and significant economic losses. Therefore, it is crucial to develop accurate and efficient rolling bearing fault diagnosis technology, which can not only reduce maintenance costs but also improve the reliability and stability of mechanical equipment. The goal of fault diagnosis is to develop systems that can monitor and identify faults in real time to ensure the continued safe operation of mechanical equipment. To achieve this, diagnostic technology must be able to quickly and continuously analyze vibration signals and demonstrate excellent anti-interference capabilities and generalization under various noise conditions and load changes to minimize the impact of environmental noise and operating condition fluctuations on diagnostic results.

[0004] After searching, the application publication number is CN113052062A, a one-dimensional convolutional neural network bearing fault diagnosis method. Rolling bearings are important components in industrial equipment, and it is very important to monitor and diagnose the health status of rolling bearings. Traditional rolling bearing fault diagnosis methods have low bearing fault identification accuracy, require manual feature extraction, and have poor generalization ability. The one-dimensional convolutional neural network (1D-CNN) method proposed in the present invention can not only accurately diagnose bearing faults, but also overcome the shortcomings of traditional fault diagnosis methods. In the present invention, a network structure in which the number of convolution kernels decreases as the size of the convolution kernel decreases is proposed, which effectively improves the accuracy of rolling bearing fault detection. Experimental results show that the 1D-CNN network structure proposed in the present invention has an average accuracy of 99.2% under a single load and an average accuracy of 98.83% across different loads.

[0005] Patent CN113052062A's bearing fault diagnosis method based on 1D-CNN has problems such as limited feature extraction capability, insufficient noise robustness, limited cross-load generalization capability, and insufficient frequency domain feature analysis. Specifically, 1D-CNN may not be able to fully capture fault features when processing complex nonlinear signals, and its generalization capability is limited under noise and load changes; at the same time, this method mainly relies on time domain signal processing and ignores the importance of frequency domain features. The present invention enhances the feature extraction capability by introducing a quadratic convolutional neural network (QCNN), especially for the processing of nonlinear features, thereby improving the noise robustness and generalization performance of the model. In addition, a joint denoising strategy of time domain and frequency domain is adopted, combined with QCNN and Fourier transform technology, which effectively improves the denoising effect and enables the model to maintain high diagnostic accuracy under different noise intensities. Furthermore, the present invention combines QCNN with Transformer neural network to construct a new model architecture, using Transformer's self-attention and multi-head attention mechanisms to capture long-distance dependencies in the fault evolution process, thereby enhancing the model's generalization capability under cross-load conditions. Finally, the cross entropy loss algorithm and cosine annealing algorithm are introduced to improve the convergence speed and training stability of the model, further enhancing the generalization ability and practical application effect of the model. In summary, the present invention effectively overcomes the technical deficiencies in patent CN113052062A. Summary of the Invention

[0006] The present invention aims to solve the above problems in the prior art. A propeller shaft rolling bearing fault diagnosis method based on QCNN-Transformer is proposed. The technical solution of the present invention is as follows:

[0007] A propeller shaft rolling bearing fault diagnosis method based on QCNN-Transformer, which includes the following steps:

[0008] Step 1: Collect rolling bearing vibration data and standardize the data using the Z-score standardization method. The rolling bearing vibration data is divided into a training set and a validation set, and the data labels are digitally converted using one-hot encoding.

[0009] Step 2: Based on the common failure modes of rolling bearings and their vibration generation principles, a comprehensive noise reduction method combining time-domain and frequency-domain filters is proposed. In the time-domain processing, a quadratic convolutional neural network (QCNN) mechanism is introduced to achieve precise and detailed signal processing. In the frequency-domain processing, Fourier transform technology is used to convert the acquired signal into the frequency domain for analysis and processing.

[0010] Step 3: Build a fault diagnosis model based on the QCNN and Transformer neural network architecture, and add two fully connected layers and a softmax layer at the end of the model;

[0011] Step 4: Use the cross entropy loss algorithm to calculate the loss function Loss, and make the loss function Loss converge through multiple iterations; use the cosine annealing algorithm to optimize the hyperparameters in the training set;

[0012] Step 5: Use the trained model to perform fault diagnosis on the validation set, output the corresponding diagnostic results, and calculate a series of predetermined performance indicators to evaluate the effectiveness and accuracy of the model.

[0013] Furthermore, the mathematical model of Z-score standardization in step 1 is described as:

[0014]

[0015] Where x i represents the i-th observation value in the original data set, μ x is the mean value of the original data set x, σ x is the standard deviation of the original data set x.

[0016] Furthermore, the key component of the time domain filter in step 2 is the secondary convolutional neural network. The mathematical model of the secondary convolution operation is described as:

[0017]

[0018] Where, W q ,W k ,W v is the weight vector, w q ,w k ,w ν ∈R is the weight, b q ,b k ,b ν ∈R is the bias and n represents the length of the input vector.

[0019] Furthermore, the frequency domain filter in step 2 acts as an auxiliary module and is endowed with the ability to directly operate the signal frequency domain; the neural network is used as a filter in the frequency domain, and this process is achieved through Fourier transform, which represents the signal passing through the time domain filter. application Convert the time domain signal into a frequency domain signal, and then use the full neural network to implement frequency domain filtering:

[0020]

[0021] Where, ω f , bf are the weights and biases in the frequency domain, is the filtered signal in the frequency domain, and N is the number of frequency components in the frequency domain.

[0022] Furthermore, the Transformer model in step 3 realizes feature analysis in the fault evolution process by utilizing self-attention and multi-head attention mechanisms; the self-attention mechanism is to input the fault feature vector X=[x1,x2,…,x n ], can be transformed into three vectors: query vector Q, key vector K and value vector V through matrix transformation, namely:

[0023] Q=XW Q , K=XW K , V=XW V

[0024] Where W Q 、W K 、W V are the transformation matrices of Q, K and V respectively.

[0025] Furthermore, the mathematical expression of the multi-head attention mechanism in the Transformer model in step 3 is:

[0026] MultiHead(Q,K,V)=[head1head2head3]W O

[0027] In the formula, head i =self_Attention(QW i Q ,KW i K ,VW i V ) is the calculation result of the self-attention mechanism in space i, W O It is the weight matrix connecting the calculation results of each self-attention module; through multiple W Q 、W K and W V The combination of Q, K and V is calculated in parallel, and the final output is obtained by splicing and linearizing the calculation results of each group.

[0028] Furthermore, the mathematical expression of the cross entropy loss algorithm in step 4 is:

[0029]

[0030] Where n is the number of samples, y ij is the label of the j-th sample corresponding to the i-th category; a ijis the probability that the j-th sample corresponds to the i-th category after the Softmax function; M is the total number of categories.

[0031] Furthermore, the cosine annealing learning rate update formula in step 4 can be expressed as:

[0032]

[0033] Where η(t) represents the learning rate at the tth iteration, η0 is the initial learning rate; η min is the minimum value of the learning rate, t is the current number of iterations, and T is the number of iterations in the entire learning rate adjustment cycle.

[0034] An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the QCNN-Transformer-based propeller shaft rolling bearing fault diagnosis method as described in any one of the above is implemented.

[0035] A non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements a propeller shaft rolling bearing fault diagnosis method based on QCNN-Transformer as described in any one of the above.

[0036] The advantages and beneficial effects of the present invention are as follows:

[0037] The rolling bearing fault diagnosis method proposed in the present invention innovatively introduces a quadratic convolutional neural network. In the calculation process of the quadratic neuron, the power function is cleverly integrated, and this measure can add an additional nonlinear mapping mechanism. In sharp contrast, traditional neurons can only use activation functions to improve their nonlinear performance capabilities after the calculation is completed. The method of the present invention can more effectively deal with the inherent nonlinear characteristics in the bearing vibration signal, significantly enhancing the noise robustness and feature extraction efficiency of the model. In the signal processing link, the present invention adopts a noise reduction strategy that combines the time domain and the frequency domain. At the time domain level, a quadratic convolutional neural network is introduced to achieve refined processing of the signal; in the frequency domain, by combining Fourier transform, the signal is converted to the frequency domain for in-depth analysis and processing, thereby further improving the noise reduction effect. In order to further enhance the feature extraction capability of the model, the present invention organically combines quadratic convolution with the Transformer neural network to construct a new model architecture, which opens up new research ideas and method paths for the field of rolling bearing fault diagnosis.

[0038] The fault diagnosis experiment results show that under diverse data conditions covering different noise intensities and different load powers, the QCNN-Transformer model constructed in the present invention exhibits excellent performance and can stably and efficiently obtain high classification accuracy.

[0039] The present invention innovatively combines a quadratic convolutional neural network with a Transformer neural network. Compared with traditional convolutional neural networks or Transformer neural networks, the neural network model of the present invention has stronger feature extraction capabilities, noise resistance stability and generalization performance; the present invention introduces a cross-entropy loss algorithm and a cosine annealing algorithm in model training, which can effectively improve the model convergence speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a flow chart of a propeller shaft rolling bearing fault diagnosis method based on QCNN-Transformer in a preferred embodiment of the present invention;

[0041] Figure 2 This is the structure diagram of the QCNN-Transformer model;

[0042] Figure 3 This is a comparison chart of the test results of each model when the signal-to-noise ratio of the CWRU dataset is 0dB;

[0043] Figure 4 This is a comparison chart of the test results of each model when the signal-to-noise ratio of the CWRU dataset is -6dB;

[0044] Figure 5 This is a comparison chart of the test results of each model when the signal-to-noise ratio of the propeller shaft rolling bearing dataset is 0dB;

[0045] Figure 6 The following is a comparison chart of the test results of each model when the signal-to-noise ratio of the propeller shaft rolling bearing dataset is -6dB. DETAILED DESCRIPTION

[0046] The following will describe the technical solutions in the embodiments of the present invention in detail with reference to the accompanying drawings. The described embodiments are only a part of the embodiments of the present invention.

[0047] The technical solution of the present invention to solve the above technical problems is:

[0048] like Figure 1 , is a flow chart of a propeller shaft rolling bearing fault diagnosis method based on QCNN-Transformer in a specific embodiment, the method comprising:

[0049] Step 1: Collect rolling bearing vibration data and standardize it using the Z-score method to ensure that the processed data has zero mean and unit standard deviation. Subsequently, the rolling bearing vibration data is divided into training and validation sets, and the data labels are digitized using one-hot encoding. This process lays a solid foundation for subsequent model training and validation.

[0050] The mathematical model of Z-score standardization is described as:

[0051]

[0052] Where x i Represents the i-th observation value in the original data set. μ x is the mean value of the original data set x. x is the standard deviation of the original data set x.

[0053] Step 2: Based on the common failure modes of rolling bearings and their vibration generation principles, a comprehensive noise reduction processing method combining time domain filters and frequency domain filters is proposed. Specifically, in terms of time domain processing, the present invention innovatively introduces the quadratic convolutional neural network (QCNN) mechanism to achieve accurate and refined signal processing; in the frequency domain processing part, Fourier transform technology is used to convert the acquired signal into the frequency domain space for in-depth analysis and processing; the model structure is as follows: Figure 2 shown.

[0054] In this embodiment, the key component of the time domain filter is the quadratic convolutional neural network. The quadratic convolution introduces quadratic neurons, whose unique calculation mechanism significantly improves performance. The mathematical model of the operation is described as follows:

[0055]

[0056] Where, W q ,W k ,W v is the weight vector, w q ,w k ,w ν ∈R is the weight, b q ,b k ,b ν ∈R is the bias and n represents the length of the input vector.

[0057] In this embodiment, the frequency domain filter acts as an auxiliary module and is given the ability to directly operate on the signal frequency domain. The main idea is to use the neural network as a filter in the frequency domain. This process is achieved through Fourier transform, which represents the signal passing through the time domain filter. application Convert the time domain signal into a frequency domain signal, and then use the full neural network to implement frequency domain filtering:

[0058]

[0059] Where, ω f , b f are the weights and biases in the frequency domain, is the filtered signal in the frequency domain.

[0060] Step 3: Build a fault diagnosis model based on the QCNN and Transformer neural network architecture, and add two fully connected layers and a softmax layer at the end of the model to enhance the classification ability and accuracy of the model;

[0061] In this embodiment, the Transformer model effectively implements feature analysis during the fault evolution process by utilizing self-attention and multi-head attention mechanisms. The self-attention mechanism is to analyze the input fault feature vector X = [x1, x2, ..., x n ], can be transformed into three vectors: query vector Q, key vector K and value vector V through matrix transformation, namely:

[0062] Q=XW Q , K=XW K , V=XW V

[0063] Where W Q 、W K 、W V are the transformation matrices of Q, K and V respectively.

[0064] In this embodiment, the mathematical expression of the multi-head attention mechanism in the Transformer model is:

[0065] MultiHead(Q,K,V)=[head1head2head3]W O

[0066] In the formula, head i =self_Attention(QW i Q ,KW i K ,VW i V ) is the calculation result of the self-attention mechanism in space i, W O is the weight matrix connecting the calculation results of each self-attention module. It can be seen that through multiple W Q 、W K and W VThe combination of Q, K and V is calculated in parallel, and the final output is obtained by splicing and linearizing the calculation results of each group.

[0067] Step 4: Use the cross entropy loss algorithm to calculate the loss function Loss, and make the loss function Loss converge through multiple iterations; use the cosine annealing algorithm to optimize the hyperparameters in the training set;

[0068] Step 5: Use the trained model to perform fault diagnosis on the validation set, output the corresponding diagnostic results, and calculate a series of predetermined performance indicators to evaluate the effectiveness and accuracy of the model.

[0069] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0070] In one embodiment of the present invention, a propeller shaft rolling bearing fault diagnosis method based on QCNN-Transformer is verified. A possible embodiment is given below to illustrate its specific implementation plan in a non-limiting manner.

[0071] In this example, the Case Western Reserve University Open Bearing Database was used as a validation dataset to verify the performance of the proposed QCNN-Transformer model in rolling element bearing fault diagnosis. Furthermore, to further validate the model's diagnostic effectiveness under actual operating conditions, a fault diagnosis experiment was designed and conducted on a propeller shaft test bench. This experiment simulated rolling element bearing faults in a real-world operating environment, allowing for a comprehensive evaluation of the model's diagnostic capabilities.

[0072] The bearing data collection system at Case Western Reserve University (CWRU) is driven by a three-phase induction motor. Two accelerometers are used to record the vibration signals of faulty bearings. The system's motor loads are set to 0 HP, 1 HP, 2 HP, and 3 HP, with a speed maintained at 1772 rpm. Data is collected at a sampling frequency of 12 kHz. Using electrospark machining (EDM), single-point defects are created on the inner and outer rings of the bearings, as well as on the rolling elements. The damage sizes for each location range from 7 inches, 14 inches, and 21 inches. Together with the normal state, a total of 10 data samples are generated, as detailed in Table 1.

[0073] The propeller shaft rolling bearing fault diagnosis test platform consists primarily of a torque sensor, an accelerometer, a magnetic powder brake, information acquisition software, an information acquisition card, and a motor. The three most common fault locations in propeller shaft rolling bearings under actual operating conditions are the outer race, inner race, and rolling elements. To obtain signals from the bearings under normal and fault conditions, fault diagnosis tests were conducted using SKF 7307AC rolling bearings. The faulty bearings were artificially machined to vary the fault location and defect size. Data samples were generated for 10 different conditions, based on different damage locations and degrees of damage, as shown in Table 2. In this study, the number of training samples was incrementally increased, starting at 20 and ending at 160, with increments of 20. This small sample size was designed to closely simulate the challenges of data acquisition in real-world scenarios, thereby enhancing the authenticity and reliability of the experiment. Validation samples were set to 100 for each condition. This large validation sample size effectively ensures high accuracy of the experimental results, providing solid data support for the research conclusions.

[0074] Table 1 Parameters of Case Western Reserve University bearing dataset

[0075]

[0076]

[0077] Table 2 Parameters of propeller shaft test bench dataset

[0078]

[0079] The performance of the QCNN-Transformer model depends to some extent on the learning rate parameter. To effectively regulate this, this example uses a cosine annealing algorithm with a periodically decreasing learning rate strategy to balance rapid convergence in the early stages of training with refined optimization in the later stages, aiming to improve the model's generalization and stability.

[0080] The cosine annealing learning rate update formula can be expressed as:

[0081]

[0082] Where η(t) represents the learning rate at the tth iteration. η0 is the initial learning rate. min is the minimum value of the learning rate, which prevents the learning rate from being too small and affecting the training effect. t is the current iteration number. T is the number of iterations in the entire learning rate adjustment cycle.

[0083] This example verifies the effectiveness of the QCNN-Transformer model and compares it with current models that perform well in the field of bearing fault identification, including WDCNN, DCA-BiGRU, QCNN, and Transformer. Among them, WDCNN is a classic convolutional neural network model that is widely used in bearing fault identification tasks and has laid the foundation for many subsequent improved models. DCA-BiGRU is a dual-path convolutional fault diagnosis model based on the attention mechanism and bidirectional gated recurrent unit (BiGRU). It uses BiGRU to achieve the fusion of spatiotemporal features and uses the attention mechanism (DCA) to extract weighted vibration signal features. QCNN is a quadratic convolutional network fault diagnosis model that introduces quadratic neurons on the basis of traditional convolutional neural networks to enhance feature extraction capabilities. In addition, Transformer is a deep learning architecture based on the self-attention mechanism. Its core advantage is that it can efficiently process sequence data and capture long-distance dependencies in the data, thereby providing a new approach for fault feature extraction. By comparing with these advanced models, this study further verified the performance and advantages of QCNN-Transformer in bearing fault identification tasks. The simulation experiments were run in Windows 11 on an Intel(R) Core(TM) i5-8300H CPU @ 2.30 GHz and Intel(R) UHD Graphics 630. The code was written in Python using the PyTorch deep learning framework. Each set of experiments was repeated five times to minimize random errors.

[0084] In the field of fault diagnosis and anomaly detection, the core of data preprocessing is to improve the quality of signal data to ensure that the signal truly reflects the characteristics of the industrial environment. This includes eliminating redundant information, data standardization, and dividing the signal into time series windows or discrete sampling points, and adding noise with a specific signal-to-noise ratio (SNR) to simulate on-site electromagnetic interference and mechanical vibration. The signal-to-noise ratio (SNR) is defined as 10log 10 (p s / p n ), here p s and p n represent the average power of signal and noise respectively.

[0085] In the Case Western Reserve University bearing dataset, when the signal-to-noise ratio is 0dB, the test results of the five models are shown in Table 3. When the signal-to-noise ratio is -6dB, the test results of the five models are shown in Table 4. From the perspective of the gradual increase in the training sample size from 20 to 160, QCNN-Transformer performs well. When SNR = 0dB, the accuracy can reach 100% in most cases, especially when the sample size is large. Under the low signal-to-noise ratio of SNR = -6dB, QCNN-Transformer can also maintain a high accuracy rate, and as the sample size increases, the performance improves steadily, which is better than other comparison methods. Overall, QCNN-Transformer shows strong robustness and good adaptability under different noise conditions and changes in sample size, and can effectively complete related tasks. The comparison image of the test results of the five models is shown in the figure Figure 3 and Figure 4 shown.

[0086] Data collection was completed on the propeller shaft rolling bearing fault diagnosis test platform, and the collected data sets were used for training and verification of the above models respectively. The final test results are shown in Table 5. Under different noise conditions, the QCNN-Transformer model in this paper has significant advantages. When SNR = 0dB, as the number of training samples gradually increases from 20 to 160, the accuracy of the QCNN-Transformer model steadily increases from 80.61% to 98.68%, significantly better than other models. For example, when the sample size is small (20), the accuracy of QCNN-Transformer has far exceeded models such as Transformer (22.01%) and DCABIGRU (11.02%); when the sample size is large (160), its advantage is even more prominent, which is 7.21 percentage points higher than WDCNN's 91.47%. Under the condition of low signal-to-noise ratio SNR = -6dB, the performance of many comparison models has declined significantly, but the QCNN-Transformer model can still maintain a high accuracy. Performance test results of each model based on the propeller shaft bearing fault diagnosis dataset are compared. Figure 5 and Figure 6 shown.

[0087] Table 3 Test results of each model in CWRU dataset when signal-to-noise ratio is 0dB (%)

[0088]

[0089]

[0090] Table 4 Test results of each model in the CWRU dataset at a signal-to-noise ratio of -6dB (%)

[0091]

[0092]

[0093] Table 5 Performance test results of each model based on propeller shaft bearing fault diagnosis dataset (%)

[0094]

[0095]

[0096] The rolling bearing fault diagnosis method proposed in the present invention innovatively introduces a quadratic convolutional neural network. In the calculation process of the quadratic neuron, the power function is cleverly integrated, and this measure can add an additional nonlinear mapping mechanism. In sharp contrast, traditional neurons can only use activation functions to improve their nonlinear performance capabilities after the calculation is completed. The method of the present invention can more effectively deal with the inherent nonlinear characteristics in the bearing vibration signal, significantly enhancing the noise robustness and feature extraction efficiency of the model. In the signal processing link, the present invention adopts a noise reduction strategy that combines the time domain and the frequency domain. At the time domain level, a quadratic convolutional neural network is introduced to achieve refined processing of the signal; in the frequency domain, by combining Fourier transform, the signal is converted to the frequency domain for in-depth analysis and processing, thereby further improving the noise reduction effect. In order to further enhance the feature extraction capability of the model, the present invention organically combines quadratic convolution with the Transformer neural network to construct a new model architecture, which opens up new research ideas and method paths for the field of rolling bearing fault diagnosis. The fault diagnosis experiment results show that under diverse data conditions covering different noise intensities and different load powers, the QCNN-Transformer model constructed in the present invention exhibits excellent performance and can stably and efficiently obtain high classification accuracy.

[0097] The systems, devices, modules or units described in the above embodiments may be implemented by computer chips or entities, or by products with certain functions.

[0098] Computer-readable media includes permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media (transitory media), such as modulated data signals and carrier waves.

[0099] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0100] The above embodiments should be understood as merely illustrating the present invention and not as limiting the scope of protection of the present invention. After reading the contents of the present invention, technicians may make various changes or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A propeller shaft rolling bearing fault diagnosis method based on QCNN-Transformer, characterized in that: The following steps are involved: Step 1: Collect rolling bearing vibration data and standardize the data using the Z-score standardization method. The rolling bearing vibration data is divided into a training set and a validation set, and the data labels are digitally converted using one-hot encoding. Step 2: Based on the common failure modes of rolling bearings and their vibration generation principles, a comprehensive noise reduction method combining time-domain and frequency-domain filters is proposed. In the time-domain processing, a quadratic convolutional neural network (QCNN) mechanism is introduced to achieve precise and detailed signal processing. In the frequency-domain processing, Fourier transform technology is used to convert the acquired signal into the frequency domain for analysis and processing. Step 3: Build a fault diagnosis model based on the QCNN and Transformer neural network architecture, and add two fully connected layers and a softmax layer at the end of the model; Step 4: Use the cross entropy loss algorithm to calculate the loss function Loss, and make the loss function Loss converge through multiple iterations; use the cosine annealing algorithm to optimize the hyperparameters in the training set; Step 5: Use the trained model to perform fault diagnosis on the validation set, output the corresponding diagnostic results, and calculate a series of predetermined performance indicators to evaluate the effectiveness and accuracy of the model.

2. The propeller shaft rolling bearing fault diagnosis method based on QCNN-Transformer according to claim 1 is characterized in that: The mathematical model of Z-score standardization in step 1 is described as: Where x i represents the i-th observation value in the original data set, μ x is the mean value of the original data set x, σ x is the standard deviation of the original data set x.

3. The propeller shaft rolling bearing fault diagnosis method based on QCNN-Transformer according to claim 1 is characterized in that: The key component of the time domain filter in step 2 is the secondary convolutional neural network. The mathematical model of the secondary convolution operation is described as: Where, W q ,W k ,W v is the weight vector, w q ,w k ,w v ∈R is the weight, b q ,b k ,b v ∈R is the bias and n represents the length of the input vector.

4. The propeller shaft rolling bearing fault diagnosis method based on QCNN-Transformer according to claim 1 is characterized in that: The frequency domain filter in step 2 acts as an auxiliary module and is given the ability to directly operate the signal frequency domain; the neural network is used as a filter in the frequency domain. This process is achieved through Fourier transform, and the signal representation of the signal passing through the time domain filter is Apply FFTF(·) to convert the time domain signal into a frequency domain signal, and then use the full neural network to implement frequency domain filtering: Where w f , b f are the weights and biases in the frequency domain, is the filtered signal in the frequency domain, and N is the number of frequency components in the frequency domain.

5. The propeller shaft rolling bearing fault diagnosis method based on QCNN-Transformer according to claim 1 is characterized in that: In step 3, the Transformer model realizes feature analysis in the fault evolution process by using self-attention and multi-head attention mechanisms; the self-attention mechanism is to analyze the input fault feature vector X = [x1, x2, ..., x n ], can be transformed into three vectors: query vector Q, key vector K and value vector V through matrix transformation, namely: Q=XW Q ,K=XW K ,V=XW V Where W Q 、W K 、W V are the transformation matrices of Q, K and V respectively.

6. The QCNN-Transformer-based propeller shaft rolling bearing fault diagnosis method according to claim 5, characterized in that: The mathematical expression of the multi-head attention mechanism in the Transformer model in step 3 is: MultiHead(Q,K,V)=[head1 head2 head3]W O Where, is the calculation result of the self-attention mechanism in space i, W O It is the weight matrix connecting the calculation results of each self-attention module; through multiple W Q 、W K and W V The combination of Q, K and V is calculated in parallel, and the final output is obtained by splicing and linearizing the calculation results of each group.

7. The QCNN-Transformer-based propeller shaft rolling bearing fault diagnosis method according to claim 1, characterized in that: The mathematical expression of the cross entropy loss algorithm in step 4 is: Where n is the number of samples, y ij is the label of the jth sample corresponding to the i-th category; a ij is the probability that the j-th sample corresponds to the i-th category after the Softmax function; M is the total number of categories.

8. The QCNN-Transformer-based propeller shaft rolling bearing fault diagnosis method according to claim 1, characterized in that: The cosine annealing learning rate update formula in step 4 can be expressed as: Where η(t) represents the learning rate at the tth iteration, and η0 is the initial learning rate; η min is the minimum value of the learning rate, t is the current number of iterations, and T is the number of iterations in the entire learning rate adjustment cycle.

9. An electronic device, characterized in that: It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the propeller shaft rolling bearing fault diagnosis method based on QCNN-Transformer as described in any one of claims 1 to 8.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the QCNN-Transformer-based propeller shaft rolling bearing fault diagnosis method according to any one of claims 1 to 8 is implemented.

Images