Bearing fault diagnosis classification method and system based on characteristic mode decomposition and Transformer deep learning model
Through the combination of feature modal decomposition and Transformer deep learning model, the problem of modal aliasing and feature extraction in bearing fault diagnosis is solved, efficient fault diagnosis is achieved, and classification accuracy is improved.
Patent Information
- Application Number
- CN202510461941.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-14
- Publication Date
- 2025-07-11
AI Technical Summary
The existing bearing fault diagnosis methods have problems with modal aliasing and insufficient theoretical foundation in feature extraction, especially the performance of the EMD method is limited, and the VMD method fails to fully capture the pulse and impact characteristics of the signal, making it difficult to effectively separate different components.
The method based on feature modal decomposition and Transformer deep learning model is adopted to extract modal signals through feature modal decomposition, and the Transformer deep learning model is used for global feature extraction and classification, and combined with feedforward neural network to optimize feature vectors to realize bearing fault diagnosis.
The accuracy and efficiency of bearing fault diagnosis have been improved, the feature extraction capability has been significantly improved, and the classification accuracy has exceeded 95%, which has solved the limitations of existing methods in modal decomposition and feature extraction.
Smart Images
Figure CN120296604A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of bearing fault diagnosis and classification. Background Art
[0002] Bearing faults are one of the most common faults in industrial equipment, and effective detection of them is crucial for extending the service life of the equipment. In recent years, the research on bearing faults based on vibration signals has developed rapidly. Completely and effectively extracting representative features from the original signal is the key to solving such problems, and existing methods have not fully solved this problem.
[0003] Fault diagnosis technology is crucial for equipment maintenance, extending the service life of equipment, improving safety and efficiency. Bearings are crucial in rotating machinery and have a significant impact on system performance and reliability. By analyzing bearing signals, potential problems can be identified, the bearing structure can be optimized, and its fault resistance can be improved. Fault diagnosis methods are mainly divided into two categories: mechanism model-based methods and data-driven methods. Although high-precision models can produce good results, their complex establishment process limits practical applications. However, with the development of deep learning and industrial big data, data-driven methods have been widely used.
[0004] Fault diagnosis usually includes three steps: feature extraction, feature selection, and fault classification. Feature extraction is crucial, and its purpose is to extract fault-related features from complex signals. Traditional methods rely on time-frequency domain transformation, which limits the information dimension of the extracted features and complicates subsequent tasks. Modal decomposition has become an effective method for signal multi-component analysis. Common methods include empirical mode decomposition (EMD), ensemble empirical mode decomposition (EEMD), variational mode decomposition (VMD), and local mean decomposition (LMD). The EMD method proposed in the prior art adaptively decomposes the signal into intrinsic mode functions and can effectively process non-linear and non-stationary signals, but there are mode mixing and end effects; mode mixing is alleviated by averaging after adding white noise sequences, but the computational complexity is relatively high. The VMD method detects the central frequency and bandwidth through variational solution and avoids mode mixing. However, it is sensitive to preset parameters. The LMD method decomposes the signal into product functions and separates the envelope signal and pure frequency modulation signal through iterative processing. Although modal decomposition methods have promoted the development of signal processing research, they still face challenges such as computational complexity.
[0005] At present, the research on bearing faults mainly focuses on the following three aspects. The first is signal processing. The complex operating environment and noise hinder the effective extraction of fault features, and existing methods still have limitations in feature extraction. The second is artificial intelligence-related methods. Machine learning has promoted the development of fault classification by using models such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs) with a large amount of bearing data. The third is multi-method fusion, which improves prediction and classification performance by integrating multiple models and has better generalization ability than a single model.
[0006] Traditional EMD methods face challenges such as mode mixing and insufficient theoretical basis. The use of white noise in EEMD methods reduces their performance. In contrast, although the VMD method is theoretically robust, it fails to fully capture the pulse and shock characteristics of signals. Due to the impulsiveness and periodicity of signals, it is difficult for existing methods to separate different components. To improve signal processing ability and adaptability, the present invention proposes a bearing fault diagnosis and classification method and system based on feature mode decomposition and Transformer deep learning model. Summary of the Invention
[0007] The present invention addresses the challenges that existing EMD methods face, such as mode mixing and insufficient theoretical basis. The use of white noise in EEMD reduces its performance. In contrast, although the VMD method is theoretically robust, it fails to fully capture the pulse and shock characteristics of signals. Due to the impulsiveness and periodicity of signals, it is difficult for existing methods to separate different components, etc.
[0008] To solve the above technical problems, the present invention is implemented through the following technical solutions:
[0009] The present invention proposes a bearing fault diagnosis and classification method based on feature mode decomposition and Transformer deep learning model. The bearing fault diagnosis and classification is realized based on feature mode decomposition and Transformer deep learning model, that is, FMD-Transformer. The method includes the following steps:
[0010] Step 1: Preprocess the input signal to generate data samples, and extract modal signals through feature mode decomposition;
[0011] Step 2: Concatenate the time-frequency domain features extracted from each feature mode to form a one-dimensional feature vector, and use the Transformer deep learning model to further extract features from the one-dimensional feature vector of the modal signal. The Transformer deep learning model is based on an encoder, and the Transformer deep learning model realizes global feature extraction through an attention mechanism. The Transformer deep learning model includes stacked Transformer encoders and a fully connected layer;
[0012] Step 3: Optimize the one-dimensional feature vector described in Step 2 using a feedforward neural network. The output generated by the iterative generation of the Transformer deep learning model is used as the feature vector of the input sample, and the bearing fault diagnosis classification result is obtained through the processing of the fully connected layer.
[0013] Furthermore, a preferred implementation is provided. In Step 1, the method for preprocessing the mechanical energy of the input signal to generate data samples and extracting modal signals through empirical mode decomposition is as follows:
[0014]
[0015] where u k (n) is the decomposition mode number k, f k is the filter number k, T s is the sampling measurement period, and M is the shift order;
[0016] That is, the filtered signal is:
[0017]
[0018] where k represents the filter number and * represents the convolution operation;
[0019] Update the filter parameters through the original signal X and u k After the iteration is completed, construct a CC k×k matrix for every two modes. The two modes include extracting the signal mode, removing redundant modes, and discarding the mode with the smallest CK value in the two modes. Repeat the above process to complete the empirical mode decomposition;
[0020]
[0021] Furthermore, a preferred implementation is provided. The method for the Transformer deep learning model to achieve global feature extraction through the attention mechanism in Step 2 is as follows:
[0022]
[0023] where Q, K, and V represent the query vector matrix, key vector matrix, and value vector matrix respectively, and d k is the dimension of the key vector matrix.
[0024] Furthermore, a preferred implementation is provided. In Step 2, the processing steps of using residual connection and layer normalization in the Transformer deep learning model.
[0025] Further, a preferred implementation is provided. The method of connecting the time-frequency domain features extracted from each feature mode using the Transformer deep learning model in step 2 is as follows:
[0026] f(x) = Concat(TTF(FMD(Split(x))))) (6)
[0027] y = Transformer(Feature(x)) (7)
[0028] Among them, TTF represents the feature extraction in the traditional time-frequency domain, and Concat represents the connection of feature vectors in different modes.
[0029] Further, a preferred implementation is provided. The time-frequency domain features extracted from each feature mode using the Transformer deep learning model in step 2 include modal energy, root mean square (RMS) value, peak factor, pulse width, spectral peak frequency, spectral center frequency, and spectral bandwidth.
[0030] Further, a preferred implementation is provided. Step 2 further includes a step of using a multi-head attention mechanism for the Transformer encoder to assign different weights to the time-frequency domain features extracted from each feature mode.
[0031] Solution 2: A bearing fault diagnosis and classification system based on feature mode decomposition and the Transformer deep learning model. The bearing fault diagnosis and classification system is implemented based on feature mode decomposition and the Transformer deep learning model, that is, FMD-Transformer. The system includes:
[0032] A data preprocessing and modal feature extraction module, which is used to preprocess the input signal to generate data samples and extract modal signals through feature mode decomposition;
[0033] A modal feature selection module, which is used to splice the time-frequency domain features of each feature mode to form a one-dimensional feature vector, and further extract features from the one-dimensional feature vector of the modal signal using the Transformer deep learning model. The Transformer deep learning model is implemented based on an encoder. The Transformer deep learning model realizes global feature extraction through an attention mechanism. The Transformer deep learning model includes a stacked Transformer encoder and a fully connected layer;
[0034] The fault classification module is used to optimize the one-dimensional feature vector described in the modal feature selection module by using a feedforward neural network. The output generated by the iterative generation of the Transformer deep learning model is used as the feature vector of the input sample, and the bearing fault diagnosis and classification result is obtained through the processing of the fully connected layer.
[0035] Solution 3: A computer device includes a memory and a processor. A computer program is stored in the memory. When the processor runs the computer program stored in the memory, the processor executes the method described in any one of Solution 1.
[0036] Solution 4: A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the steps of the method described in any one of Solution 1 are implemented.
[0037] The advantages of the present invention are as follows:
[0038] For the bearing fault diagnosis and classification method and system based on feature modal decomposition and Transformer deep learning model described in the present invention, an FMD-Transformer model for bearing fault diagnosis is introduced, and feature modal decomposition (FMD) is used as an alternative method to empirical modal decomposition (EMD), ensemble empirical modal decomposition (EEMD), and VMD. The original complex signal is decomposed into multiple simple basic modal signals, thereby realizing effective feature extraction. The ability of the Transformer deep learning model to process long-distance information effectively processes one-dimensional bearing vibration signals.
[0039] The experimental results of the present invention show that the classification accuracy of the method described in the present invention on the Ottawa bearing vibration data set exceeds 95%. The present invention will be combined with the neural network model, providing new ideas and directions for bearing fault research. The present invention emphasizes the importance of efficient modal decomposition methods in feature extraction, indicating that the model accuracy can be improved by studying various target signals and adopting different modal decomposition techniques. In addition, the results show that within a certain range, the model accuracy will increase with the increase in the number of modes.
[0040] The present invention is also applicable to the technical field of bearing fault analysis. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] Figure 1 It is a schematic diagram of the FMD-Transformer model in the bearing fault diagnosis and classification method based on feature modal decomposition and Transformer deep learning model described in Embodiment 1.
[0042] Figure 2Schematic flowchart of the bearing fault diagnosis and classification method based on feature mode decomposition and Transformer deep learning model described in Embodiment 1.
[0043] Figure 3 Schematic diagram showing the influence of the number of modes on the accuracy in Embodiment 11. Specific embodiments
[0044] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments.
[0045] Embodiment 1. This embodiment provides a bearing fault diagnosis and classification method based on feature mode decomposition and Transformer deep learning model. The bearing fault diagnosis and classification is implemented based on feature mode decomposition and Transformer deep learning model, that is, FMD-Transformer. The method includes the following steps:
[0046] Step 1: Preprocess the input signal to generate data samples, and extract modal signals through feature mode decomposition;
[0047] Step 2: Concatenate the time-frequency domain features extracted from each feature mode to form a one-dimensional feature vector. Use the Transformer deep learning model to further extract features from the one-dimensional feature vector of the modal signal. The Transformer deep learning model is based on an encoder. The Transformer deep learning model realizes global feature extraction through an attention mechanism. The Transformer deep learning model includes stacked Transformer encoders and a fully connected layer;
[0048] Step 3: Optimize the one-dimensional feature vector described in Step 2 using a feedforward neural network. The output generated iteratively by the Transformer deep learning model is used as the feature vector of the input sample, and the bearing fault diagnosis and classification result is obtained through processing by the fully connected layer.
[0049] Embodiment 2. This embodiment further limits the bearing fault diagnosis and classification method based on feature mode decomposition and Transformer deep learning model described in Embodiment 1. The method for preprocessing the mechanical energy of the input signal to generate data samples and extracting modal signals in Step 1 is as follows:
[0050]
[0051] Among them, u k(n) is the decomposition mode number k, f k is the filter number k, T s is the sampling measurement period, M is the shift order;
[0052] That is, the filtered signal is:
[0053]
[0054] where k represents the filter number, * represents the convolution operation;
[0055] Through the original signals X and u k Update the filter parameters. After the iteration is completed, construct a CC k×k matrix for every two modes. The two modes include the extracted signal mode, removing the redundant mode, and discarding the mode with the smallest CK value in the two modes. Repeat the above process to complete the eigenmode decomposition;
[0056]
[0057] Embodiment 3. This embodiment further limits the method for bearing fault diagnosis and classification based on eigenmode decomposition and Transformer deep learning model described in Embodiment 2. In step 2
[0058] The method for the Transformer deep learning model to achieve global feature extraction through the attention mechanism is:
[0059]
[0060] where Q, K, and V represent the query vector matrix, key vector matrix, and value vector matrix respectively, and dk is the dimension of the key vector matrix.
[0061] Embodiment 4. This embodiment further limits the processing steps of using residual connection and layer normalization in the Transformer deep learning model in step 2 of the method for bearing fault diagnosis and classification based on eigenmode decomposition and Transformer deep learning model described in Embodiment 1.
[0062] Embodiment 5. This embodiment further limits the method for connecting the time-frequency domain features extracted from each eigenmode by using the Transformer deep learning model in step 2 of the method for bearing fault diagnosis and classification based on eigenmode decomposition and Transformer deep learning model described in Embodiment 1. The method is:
[0063] f(x) = Concat( TTF( FMD( Split (x))))) (6)
[0064] y = Transformer(Feature(x)) (7)
[0065] Among them, TTF represents the feature extraction in the traditional time-frequency domain, and Concat represents the concatenation of feature vectors of different modalities.
[0066] Embodiment 6: This embodiment further limits a bearing fault diagnosis and classification method based on feature modal decomposition and Transformer deep learning model described in Embodiment 5. In step 2, the time-frequency domain features extracted from each feature mode by using the Transformer deep learning model include modal energy, root mean square (RMS) value, peak factor, pulse width, spectral peak frequency, spectral center frequency, and spectral bandwidth.
[0067] Embodiment 7: This embodiment further limits a bearing fault diagnosis and classification method based on feature modal decomposition and Transformer deep learning model described in Embodiment 2. Step 2 further includes a step of using the multi-head attention mechanism of the Transformer encoder to assign different weights to the time-frequency domain features extracted from each feature mode.
[0068] Embodiment 8: This embodiment proposes a bearing fault diagnosis and classification system based on feature modal decomposition and Transformer deep learning model. The bearing fault diagnosis and classification system is implemented based on feature modal decomposition and Transformer deep learning model, that is, FMD-Transformer. The system includes:
[0069] A data preprocessing and modal feature extraction module, which is used to preprocess the input signal to generate data samples and extract modal signals through feature modal decomposition.
[0070] A modal feature selection module, which is used to perform further feature extraction on the one-dimensional feature vector formed by splicing the time-frequency domain features of each feature mode by using the Transformer deep learning model. The Transformer deep learning model is implemented based on an encoder. The Transformer deep learning model realizes global feature extraction through an attention mechanism. The Transformer deep learning model includes stacked Transformer encoders and a fully connected layer.
[0071] A fault classification module, which is used to optimize the one-dimensional feature vector described in the modal feature selection module by using a feedforward neural network. The output iteratively generated by the Transformer deep learning model is used as the feature vector of the input sample, and the bearing fault diagnosis and classification result is obtained through the processing of the fully connected layer.
[0072] Embodiment Nine: A computer device includes a memory and a processor. A computer program is stored in the memory. When the processor runs the computer program stored in the memory, the processor executes the method described in any one of Embodiments One to Seven.
[0073] Embodiment Ten: A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the steps of the method described in any one of Embodiments One to Seven are implemented.
[0074] Embodiment Eleven: The embodiments proposed in this embodiment are used to explain the above Embodiments One to Ten, and specifically include the following content:
[0075] To improve the signal processing ability and adaptability, in the bearing fault diagnosis and classification method based on feature mode decomposition and Transformer deep learning model of this embodiment, feature mode decomposition FMD is introduced in this embodiment. Feature mode decomposition FMD decomposes the original signal into basic mode components with time-frequency characteristics for analysis. It has three key characteristics:
[0076] (i) FMD can effectively capture the impulses and periodicity of the signal, thereby accurately decomposing the fault modes;
[0077] (ii) It uses finite impulse response (FIR) filters to extract relevant signal patterns while removing redundant patterns, improving the thoroughness of decomposition;
[0078] (iii) FMD can decompose mechanical fault information without prior knowledge. The signal processing process can be represented by mathematical formula (1):
[0079]
[0080] where, u k (n) is the decomposition mode number k, f k is the filter number k, T s is the sampling measurement period, and M is the shift order.
[0081] Compared with the existing initialization methods, the Hanning window can locate the fault period after two iterations, significantly improving the efficiency. The FIR filter is initialized by the Hanning window. The filtered signal can be obtained through the following formula (2):
[0082] u k = X * f k
[0083]
[0084] where k represents the filter number, and * represents the convolution operation. Through the original signal X and u k Update the filter parameters. After the iteration is completed, a CC k×k matrix is constructed for every two modes, and the mode with a smaller CK value in the two modes is discarded. Repeat the above process. After obtaining the preset number of modes, the FMD mode decomposition is completed.
[0085]
[0086] VMD solves the mode aliasing problem in digital mode decomposition. However, it gives priority to bandwidth minimization rather than fault characteristics, resulting in poor performance in bearing fault signal decomposition. In contrast, while considering the impulse characteristics and periodicity of the original signal, FMD effectively extracts mechanical fault characteristics. Compared with VMD, FMD has more advantages in mechanical signal decomposition and analysis. In this embodiment, FMD is introduced into bearing fault classification, and due to its advanced fault feature extraction ability, the accuracy of the diagnosis result is improved.
[0087] This model establishes dependencies between the input and output, as well as between the encoder and decoder, without the need for traditional RNN and CNN architectures.
[0088] The Transformer deep learning model realizes the ability of global feature extraction through stacked attention mechanisms. The attention mechanism is a process of mapping a query and a set of key-value pairs to an output, called scaled dot-product attention, and its calculation process is represented by formula (4):
[0089]
[0090] The implementation of dot-product attention can be completed through a highly optimized matrix multiplication algorithm, which improves the speed and reduces the memory usage in practical applications. Different from a single attention function, multi-head attention enables the model to simultaneously focus on information at different positions in different subspaces, thereby improving the overall model performance. The multi-head attention (MHA) mechanism can be represented by formula (5):
[0091] MultiHead(Q,K,V)=Concat(head1,...,head h )W O
[0092] wherehead i =Attention(QW i Q ,KW i K ,VW i V ) (5)
[0093] An encoder architecture is adopted and combined with an attention mechanism to solve the sequence dependence problem inherent in traditional RNN networks, and it performs excellently in long-distance feature extraction
[21] . Due to its excellent feature extraction ability, Transformer has been widely applied and studied in the fields of natural language processing and computer vision, and remarkable results have been achieved.
[0094] This embodiment proposes a model for bearing fault diagnosis, called FMD-Transformer, which integrates two methods, FMD and Transformer. The model architecture is as Figure 1 shown.
[0095] In the data preprocessing and modal feature extraction stage, the original signal is sampled at a specified frequency to generate data samples. The basic modal signals are extracted from them through FMD, and the time-frequency domain features extracted from each mode are concatenated to form a one-dimensional feature vector. The neural network architecture consists of a series of stacked Transformer encoders and a fully connected layer. The encoder uses a multi-head attention mechanism to explore the relationships between segments of the input vector, assigns different weights to each position to highlight key components. To solve the problem of gradient disappearance, accelerate the model convergence and minimize the risk of overfitting, this embodiment adopts residual connections and layer normalization in the model. The feed-forward neural network is used to further optimize the sequence and enhance the model's expression and learning ability. The output generated by the Transformer encoder through multiple iterations is used as the feature vector of the input sample, and then the classification result is obtained through the processing of the fully connected layer. The detailed parameters of the proposed model network structure are shown in Table 1.
[0096] Table 1. Parameters of each model
[0097]
[0098] The input signal needs to be preprocessed to generate input data samples. The basic modal signals are extracted by using FMD for further analysis. The FMD processing flow is as Figure 2 shown, and the key parameter configurations are shown in Table 1. Each basic mode is analyzed in the time-frequency domain, and consistent features are selected to form the corresponding feature vector. The learning and reasoning processes of each sample depend on the feature vector rather than the sample itself. The subsequent steps follow the standard deep learning model based on Transformer, which will not be elaborated here. The forward propagation process can be expressed as:
[0099] f(x) = Concat(TTF(FMD(Split(x)))) (6)
[0100] y = Transformer(Feature(x)) (7)
[0101] Among them, TTF represents the feature extraction in the traditional time-frequency domain, and Concat represents the concatenation of feature vectors of different modalities.
[0102] The feature extraction of modal signals in the time-frequency domain mainly includes modal energy, root mean square (RMS) value, peak factor, pulse width, spectral peak frequency, spectral center frequency, and spectral bandwidth. These features are calculated according to the formulas in Table 2.
[0103] Table 2. Feature Extraction Methods in the Time-Frequency Domain
[0104]
[0105] i is the sampling point, and x i is the signal amplitude at each sampling point, Fre is the spectral sequence of the modal signal, Mag is the amplitude sequence corresponding to the spectral sequence, and θ is the threshold preset according to the dataset.
[0106] The experimental settings and dataset are as follows: This study focuses on bearing faults and uses a vibration signal dataset from Ottawa, Canada. The dataset contains signals under four different speed conditions: A (acceleration), B (deceleration), C (acceleration first and then deceleration), and D (deceleration first and then acceleration). At each speed, there are three fault states: healthy, inner race defect, and outer race defect. The signal acquisition for each speed and fault state is repeated three times, resulting in a total of 36 signals. Each signal is sampled at a frequency of 200,000 Hz for 10 seconds. Each signal in the dataset is segmented into multiple segments with an interval of 0.05 seconds, generating 200 samples for each signal. A total of 7200 samples are generated for 36 bearing signals. Among them, 7 time-frequency features are extracted for each sample as the feature vector. The 7200 samples are divided into a training set and a test set in a ratio of 7:3.
[0107] The following experiments are conducted in 64-bit PyCharm with a computer configuration of I5-12400F 2.5 GHz (6 cores) and 16 GB of memory. During the experiment, the learning rate is 0.001, and the batch size is 512.
[0108] FMD-Transformer Experiment: As a new signal pattern decomposition technology, FMD has received extensive attention. In this paper, by combining FMD with backpropagation network (BP), CNN, and long short-term memory network (LSTM), the above dataset is used for bearing fault classification to evaluate its effectiveness. Similar to FMD-Transformer, in this embodiment, the data preprocessing and signal feature extraction processes described will use the obtained signal feature vectors as the inputs of the BP, CNN, and LSTM networks for subsequent classification tasks. For comparison, the control group generates feature vectors from the mean, variance, extreme values, and standard deviation in the time domain. TTF is used to represent the feature extraction in the traditional time-frequency domain and is compared with FMD.
[0109] Table 3 shows the classification accuracy and loss obtained by combining TTF, FMD with various classification models. Obviously, when using TTF for feature extraction, the classification performance of LSTM is significantly better than other models and even exceeds that of Transformer. We speculate that, on the one hand, the proposed LSTM model is specifically designed to solve the problem of gradient vanishing or explosion faced by traditional RNN networks when dealing with long sequence data, making the LSTM network more advantageous than other neural network architectures in dealing with sequence signals such as bearing vibration signals. On the other hand, the advantage of Transformer lies in its ability to extract global information. For the feature vectors derived from bearing vibration signals, additional preprocessing techniques may be required to fully utilize the capabilities of Transformer.
[0110] When using FMD for feature extraction, the classification accuracy of all models is generally higher than that of TTF. When applying FMD, the classification loss is significantly reduced, which is consistent with the classification results in Table 3, indicating that FMD is more effective in extracting key fault features from such data, thus providing more valuable inputs for the model.
[0111] Table 3. Comparison of Classification Results and Total Loss
[0112]
[0113]
[0114] Table 4. Classification Accuracy under Different Fault Modes
[0115]
[0116] To further evaluate the effectiveness of FMD-Transformer in the classification task of this dataset, this study conducted additional experiments. Table 4 shows the classification accuracies of the TTF-Transformer and FMD-Transformer models on various fault types. Both models achieved an accuracy of 99% in identifying inner race faults. However, in the case of healthy conditions and outer race faults, the accuracy of the FMD-Transformer model was significantly better than that of the TTF-Transformer model. This indicates that the combination of FMD feature extraction technology and the Transformer architecture can significantly improve the model performance, especially in the identification of specific fault types.
[0117] FMD is a multi-parameter modulation decomposition technique. This embodiment particularly focuses on the correlation between its parameters and the decomposition effect. Based on the experimental results of Miao et al.
[19] , FMD shows robustness when selecting the parameter k in the range of [2, 20], where it is recommended to select k in the range of [5, 10]. In the model proposed in this embodiment, the number of modes directly determines the size of the vibration signal feature vector. Therefore, to study the relationship between the number of modes and the bearing fault classification accuracy, this embodiment conducted a series of experiments by changing the number of modes in the range of [2, 15].
[0118] Figure 3 Shows the variation of the results of different network models with the number of modes. As the number of modes n increases from 2 to 10, the accuracy of each model improves. This indicates that within a certain range, as n increases, FMD decomposes the signal more thoroughly and the feature extraction is more effective. However, when n > 10, the network performance tends to be stable with only minor fluctuations, which is consistent with the initial FMD research results. Therefore, we can conclude that increasing the number of modes can improve the model performance before reaching a certain threshold, but it will also increase the computational complexity. So, the selection of the number of modes needs to be balanced between the decomposition performance and the computational complexity.
[0119] This embodiment focuses on the research of bearing faults. Although variational mode decomposition (VMD) solves the problem of mode mixing, it lacks precision in fault feature decomposition. This paper introduces the FMD-Transformer model for bearing fault diagnosis, taking feature mode decomposition (FMD) as an alternative to empirical mode decomposition (EMD), ensemble empirical mode decomposition (EEMD), and VMD. FMD has excellent analysis and decomposition capabilities for mechanical fault signals, effectively overcoming the limitations of VMD, which is crucial for feature extraction in fault diagnosis. Experimental results show that FMD significantly improves feature extraction, and FMD-Transformer achieves an accuracy of over 95% in classification tasks. This study emphasizes the importance of efficient mode decomposition methods in feature extraction, indicating that model accuracy can be improved by studying various target signals and adopting different mode decomposition techniques. In addition, the results show that within a certain range, the model accuracy increases with the increase in the number of modes.
[0120] Regarding Embodiment 9 of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device. More specific examples (non-exhaustive list) of computer-readable media include the following: an electrical connection portion (electronic device) having one or N wirings, a portable computer disk cartridge (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium can even be paper or other suitable media on which the program can be printed, as the program can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpretation, or other appropriate processing as necessary, and then stored in a computer memory. It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiment, the N steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application-specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.
[0121] Those skilled in the art can understand that the above is only the preferred embodiment of the present invention. The features described in each embodiment and / or claim of the present disclosure can be combined or combined in various ways, even if such combinations or combinations are not explicitly recorded in the present disclosure. It is not used to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or equivalently replace some of the technical features. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.
[0122] Although the preferred embodiments of the present invention have been described, those skilled in the art can make additional changes and modifications once they know the basic creative concept. Therefore, the appended claims are intended to be interpreted to include the preferred embodiments and all changes and modifications falling within the scope of the present invention. Obviously, those skilled in the art can make various changes and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these modifications and variations.
Claims
1. A bearing fault diagnosis and classification method based on feature mode decomposition and Transformer deep learning model, characterized in that Bearing fault diagnosis and classification are realized based on feature mode decomposition and the Transformer deep learning model, that is, FMD-Transformer. The method includes the following steps: Step 1: Preprocess the input signal to generate data samples, and extract modal signals through feature mode decomposition; Step 2: Concatenate the time-frequency domain features extracted from each feature mode to form a one-dimensional feature vector. Use the Transformer deep learning model to further extract features from the one-dimensional feature vector of the modal signal. The Transformer deep learning model is based on an encoder. The Transformer deep learning model realizes global feature extraction through an attention mechanism. The Transformer deep learning model includes stacked Transformer encoders and a fully connected layer; Step 3: Use a feedforward neural network to optimize the one-dimensional feature vector in Step 2. The output generated iteratively by the Transformer deep learning model is used as the feature vector of the input sample, and the bearing fault diagnosis and classification result is obtained through processing by the fully connected layer.
2. The bearing fault diagnosis and classification method based on feature mode decomposition and Transformer deep learning model according to claim 1, wherein The method for preprocessing the mechanical energy of the input signal to generate data samples and extracting modal signals in Step 1 is as follows: where u k (n) is the decomposition mode number k, f k is the filter number k, T s is the sampling measurement period, and M is the shift order; That is, the filtered signal is: where k represents the filter number, and * represents the convolution operation; Through the original signals X and u k Update the filter parameters. After the iteration is completed, construct a CC k×k matrix for every two modes. The two modes include the extracted signal mode, the redundant mode removal, and the mode with the smallest CK value in the two modes is discarded. Repeat the above process to complete the eigenmode decomposition; 3. The bearing fault diagnosis and classification method based on characteristic mode decomposition and Transformer deep learning model according to claim 1, characterized in that, The method for the Transformer deep learning model to realize global feature extraction through the attention mechanism in Step 2 is as follows: Among them, Q, K, and V represent the query vector matrix, the key vector matrix, and the value vector matrix respectively, and d k is the dimension of the key vector matrix.
4. The bearing fault diagnosis and classification method based on feature mode decomposition and Transformer deep learning model according to claim 1, characterized in that The processing steps of using residual connection and layer normalization in the Transformer deep learning model in Step 2.
5. The bearing fault diagnosis and classification method based on feature mode decomposition and Transformer deep learning model according to claim 1, wherein The method for using the Transformer deep learning model to concatenate the time-frequency domain features extracted from each feature mode in Step 2 is as follows: f(x) = Concat(TTF(FMD(Split(x)))) (6) y = Transformer(Feature(x)) (7) where TTF represents the feature extraction in the traditional time-frequency domain, and Concat represents the concatenation of different modal feature vectors.
6. The bearing fault diagnosis and classification method based on characteristic mode decomposition and Transformer deep learning model according to claim 1, characterized in that The time-frequency domain features extracted from each feature mode using the Transformer deep learning model in Step 2 include modal energy, root mean square (RMS) value, peak factor, pulse width, spectral peak frequency, spectral center frequency, and spectral bandwidth.
7. The bearing fault diagnosis and classification method based on feature mode decomposition and Transformer deep learning model according to claim 1, characterized in that, Step 2 also includes the step of using a multi-head attention mechanism for the Transformer encoder to assign different weights to the time-frequency domain features extracted from each feature mode.
8. A bearing fault diagnosis and classification system based on characteristic mode decomposition and Transformer deep learning model, characterized in that The bearing fault diagnosis and classification system is realized based on feature mode decomposition and the Transformer deep learning model, that is, FMD-Transformer. The system includes: A data preprocessing and modal feature extraction module, which is used to preprocess the input signal to generate data samples and extract modal signals through feature mode decomposition; The modal feature selection module is used to perform further feature extraction on the one-dimensional feature vector formed by splicing the time-frequency domain features of each feature modality by using a Transformer deep learning model. The Transformer deep learning model is based on an encoder, and the Transformer deep learning model realizes global feature extraction through an attention mechanism. The Transformer deep learning model includes stacked Transformer encoders and a fully connected layer; The fault classification module is used to optimize the one-dimensional feature vector of the modal feature selection module by using a feedforward neural network. The output iteratively generated by the Transformer deep learning model is used as the feature vector of the input sample, and the bearing fault diagnosis classification result is obtained through the processing of the fully connected layer.
9. A computer device, comprising a memory and a processor, characterized in that, A computer program is stored in the memory. When the processor runs the computer program stored in the memory, the processor executes the method according to any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.