Complex signal fault diagnosis method based on multi-mode feature extraction and domain adaptation

Abstract

The invention discloses a complex signal fault diagnosis method based on multi-mode feature extraction and domain adaptation. The complex signal fault diagnosis method comprises the following steps: step S10, loading data; step S20, data preprocessing; step S30, carrying out multi-mode feature extraction; step S40, carrying out data normalization processing; step S50, making a data loader; step S60, performing algorithm evaluation and structural design; and step S70, algorithm training and performance evaluation. According to the method, the model training and reasoning efficiency are greatly optimized, and the speed and the real-time performance of an algorithm output result are remarkably improved. According to the method, the operation state of the wind power gear box bearing can be evaluated and monitored more accurately, a scientific basis is provided for state prediction and maintenance of equipment, the service life of the equipment can be prolonged, the operation and maintenance cost can be reduced, new power can be injected for intelligent development of the wind power industry, and the economic benefit is increased. And efficient operation and sustainable development of the green energy industry are promoted.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis methods, and in particular to a complex signal fault diagnosis method based on multi-mode feature extraction and domain adaptation. Background Art

[0002] With the increasing complexity of industrial equipment and the diversification of operating environments, the need for condition monitoring and fault diagnosis of key components has become increasingly important. In key mechanical equipment such as wind turbine gearboxes, bearings are core components, and their operating status directly affects the safety and stability of the equipment. However, traditional fault diagnosis methods have significant limitations when dealing with complex signals, cross-domain data, and unlabeled scenarios.

[0003] At present, many fault diagnosis methods rely on supervised learning algorithms for data in a single domain. These methods usually assume that the source domain and target domain data have the same distribution characteristics and require a large amount of manually annotated data. However, in actual industrial applications, different data sets often show significant distribution heterogeneity due to differences in sampling frequency, equipment type, and operating conditions. This difference makes it difficult for traditional methods to be directly applied to cross-domain scenarios, especially between high sampling frequency data and low sampling frequency data. Due to the significant differences in frequency resolution and feature patterns, there are difficulties in feature mapping and evaluation, which significantly limits the performance of the algorithm. In addition, target domain data often lacks label information in practical applications, which further exacerbates the difficulty of effective fault diagnosis in unlabeled scenarios. Therefore, how to efficiently use source domain data to achieve adaptive analysis of target domain data, especially to achieve distribution alignment and feature mapping at different sampling frequencies, has become a research hotspot and core challenge in the current field of fault diagnosis.

[0004] In recent years, variational mode decomposition (VMD) technology has been widely used in the field of signal processing. VMD technology can decompose complex signals into multiple mode signals, providing an effective means for feature extraction. However, in cross-domain fault diagnosis scenarios, it is difficult to solve complex distribution difference problems by relying solely on single feature processing technologies such as VMD technology, especially between high-frequency and low-frequency data. The significant difference in feature distribution further limits its adaptability. At the same time, deep learning technology, especially convolutional neural network (CNN), has demonstrated powerful capabilities in feature learning of high-dimensional complex data, but when dealing with cross-domain tasks, it is difficult to fully align the data distribution differences between the source domain and the target domain by relying solely on CNN for feature learning. Especially in the process of aligning data with different sampling frequencies, the limitations of CNN are more prominent, resulting in insufficient support for fault diagnosis in cross-frequency scenarios.

[0005] In response to the above problems, domain adaptation technology has gradually become an important means to solve cross-domain fault diagnosis. Among them, variance difference representation (VDR), as an efficient distribution alignment method, can effectively reduce the distribution difference between the source domain and the target domain through feature alignment, thereby significantly improving the cross-domain adaptability of the model. However, existing research still faces certain limitations when combining VDR with deep learning technology, especially in the fault diagnosis scenario of multi-mode signals, and its application exploration is still insufficient. In addition, current research on the distribution alignment problem between data with different sampling frequencies is relatively lacking, which further limits the potential of domain adaptation technology in cross-frequency fault diagnosis.

[0006] Therefore, the applicant has found a solution to the above-mentioned problem through beneficial exploration and research, and the technical solution to be introduced below is produced in this context. Summary of the invention

[0007] The technical problem to be solved by the present invention is to provide a complex signal fault diagnosis method based on multi-mode feature extraction and domain adaptation in view of the deficiencies of the prior art.

[0008] The technical problem to be solved by the present invention can be achieved by adopting the following technical solutions:

[0009] A complex signal fault diagnosis method based on multi-mode feature extraction and domain adaptation comprises the following steps:

[0010] Step S10, data loading: performing data loading processing on the laboratory simulation bearing vibration signal data set and the real sampled bearing vibration signal data set;

[0011] Step S20, data preprocessing: filtering the vibration signal data loaded in step S10;

[0012] Step S30, multi-mode feature extraction: Decomposing the vibration signal data filtered in step S20 into multiple feature signals by using a signal processing method based on variational mode decomposition, and performing statistical feature extraction processing on these feature signals, and then integrating the extracted features to replace the original vibration signal data, thereby generating a multi-mode feature signal with higher discrimination and diversity;

[0013] Step S40, data normalization processing: normalizing the multi-mode feature signal generated in step S30;

[0014] Step S50, data loader production: on the one hand, the laboratory simulation bearing vibration signal data set collected at high frequency is used as a training set, and the real sampling bearing vibration signal data set collected at low frequency is used as a test set to verify the performance of the model in mapping from high frequency to low frequency data in a cross-domain scenario; on the other hand, the real sampling bearing vibration signal data set collected at low frequency is used as a training set, and the laboratory simulation bearing vibration signal data set collected at high frequency is used as a test set to verify the performance of the model in mapping from low frequency to high frequency data;

[0015] Step S60, algorithm evaluation and structural design: convolutional neural network is selected as the structural design of the model, and variance difference representation is used as the measurement method of distribution alignment. During the model training process, the deep learning ability of CNN and the domain alignment technology of VDR are combined to enable the model to make full use of cross-domain data features and achieve high-precision fault diagnosis performance;

[0016] Step S70, algorithm training and performance evaluation: Use the data set processed in step S50 to train the model designed in step S60.

[0017] In a preferred embodiment of the present invention, in step S10, the data loading includes the following steps:

[0018] Step S11, downloading the specified laboratory simulation bearing vibration signal data set and the real sampling bearing vibration signal data set, and ensuring that the data sets are complete and in a standardized format;

[0019] Step S12, classify and load the downloaded laboratory simulation bearing vibration signal data set and the real sampling bearing vibration signal data set according to the vibration signal data and the corresponding label data, so as to provide a clear data structure for subsequent processing.

[0020] In a preferred embodiment of the present invention, the bearing vibration signal data loaded in step S10 has a fixed length, each piece of data contains the same number of sampling points, covers a variety of fault states, each state contains several pieces of data, the data distribution is balanced and the sample size is sufficient.

[0021] In a preferred embodiment of the present invention, in step S20, the data preprocessing includes the following steps:

[0022] Step S21, detecting the bearing vibration signal data loaded in step S10, and clearing possible null value data to ensure the validity of each data sample;

[0023] Step S22, eliminating outliers and noise in the bearing vibration signal data to avoid interference of extreme values ​​on model training;

[0024] Step S23, checking the data length of the bearing vibration signal data, confirming that all samples have the same number of sampling points, and removing data samples with abnormal lengths.

[0025] In a preferred embodiment of the present invention, in step S30, the multi-modal feature extraction includes the following steps:

[0026] Step S31, performing frequency domain evaluation on the signals in the healthy state and the faulty state to ensure that the signal strength does not exceed the specified threshold set by the method, so as to ensure the availability of the signal features and the effectiveness of the analysis;

[0027] Step S32, using VMD method to process the original signal, decompose the original signal into several modes based on the threshold, and extract statistical features from each decomposed mode signal, including multiple statistical feature indicators, to form multi-dimensional feature data. The objective function of VMD is as follows;

[0028]

[0029] Among them, u k (t) represents the kth modal signal obtained by decomposition; w k represents the center frequency of the mode uk(t); α represents the balance parameter, which is used to control the mode smoothness; represents the derivative with respect to time t; δ(t) represents the unit pulse signal; j is an imaginary unit, representing the Hilbert transform of the signal;

[0030] Step S33, integrates the signal features of all modes to generate a new data vector to replace the original signal, providing a richer and more discriminative feature representation for subsequent modeling and training.

[0031] In a preferred embodiment of the present invention, in step S40, the data normalization process is specifically as follows:

[0032] The newly generated multidimensional data vector is normalized by using the minimum-maximum normalization method with translation term to linearly scale the eigenvalues ​​to the interval [0,1]. The normalization formula is as follows:

[0033]

[0034] Where x represents the original eigenvalue; min(x) represents the minimum value among the original eigenvalues; max(x) represents the maximum value among the original eigenvalues; and x' represents the normalized eigenvalue.

[0035] In a preferred embodiment of the present invention, in step S50, the data loader is made, comprising the following steps:

[0036] Step S51: In order to verify the performance of data mapping from high frequency to low frequency, the laboratory simulation bearing vibration signal data set collected at high frequency is used as a training set, and its rich frequency range and detailed information are used to train the model; at the same time, the real sampled bearing vibration signal data set used at low frequency is used as a test set, and cross-domain prediction is performed on it through the model;

[0037] Step S52, in order to verify the performance of mapping from low-frequency to high-frequency data, the real sampling fault data set collected at low frequency is used as the training set, and its relatively simplified features are used to train the model; at the same time, the laboratory fault simulation data set collected at high frequency is used as the test set to test the performance of the model under high-frequency data conditions.

[0038] In a preferred embodiment of the present invention, in step S60, the algorithm evaluation and structure design includes the following steps:

[0039] Step S61, selecting a convolutional neural network (CNN) as the structural design of the model;

[0040] Step S62, using variance difference representation as a measurement method for distribution alignment, the objective function of VDR is as follows:

[0041]

[0042] in, Represents the variance information of the feature map, which is mapped to the Reproducing Kernel Hilbert Space (RKHS) through the kernel function; represents the tensor product of two Hilbert spaces, which is used to represent variance information; κ(x,·) represents the kernel function, which is used to map the input to a high-dimensional feature space; E x～p(x) represents the expectation of the source domain distribution p(x); E y～q(y) Represents the expectation of the target domain distribution q(y);

[0043] Step S63, during the model training process, the deep feature extraction capability of CNN and the domain alignment optimization technology of VDR are combined to build a fault diagnosis model that can efficiently utilize cross-domain data features.

[0044] In a preferred embodiment of the present invention, in step S70, the algorithm training and performance evaluation includes the following steps:

[0045] Step S71, using the data set processed in step S50 as input data to train the algorithm network designed in step S60;

[0046] Step S72: During the training process, four key loss functions and evaluation indicators are designed to comprehensively monitor the training progress and performance of the network, namely, VDR loss, classification loss, training set accuracy, and test set accuracy; at the end of each round of iteration, these four indicators are dynamically calculated to intuitively judge the optimization effect of the model;

[0047] Step S73, after each round of iteration is completed, the current performance of the model is evaluated based on the calculated loss function value and accuracy, and the parameter status of the network is saved in real time. If the performance of a certain round of iteration reaches the optimal value, the status is recorded as the final model.

[0048] Due to the adoption of the above technical scheme, the beneficial effect of the present invention is that the present invention fully combines the feature extraction capabilities of variational mode decomposition (VMD) and convolutional neural network (CNN), as well as the distribution alignment capability of variance difference representation (VDR), to effectively deal with the complex signal fault diagnosis problem in cross-domain unlabeled scenarios. At the same time, by solving the difficult-to-map characteristic distribution differences between high sampling frequency and low sampling frequency data, the diagnostic performance of the algorithm in the unlabeled target domain is significantly improved. The present invention also achieves substantial optimization in model training and reasoning efficiency, significantly improving the speed and real-time performance of the algorithm output results. The present invention can more accurately evaluate and monitor the operating status of wind turbine gearbox bearings, providing a scientific basis for equipment status prediction and maintenance. The present invention provides a more adaptive and robust technical support for the status monitoring and fault diagnosis of key industrial equipment such as wind turbine gearboxes, making up for the shortcomings of traditional methods in cross-domain distribution alignment and feature mapping, which not only helps to extend the service life of equipment and reduce operation and maintenance costs, but also injects new impetus into the intelligent development of the wind power industry, and promotes the efficient operation and sustainable development of the green energy industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0050] Figure 1 It is a schematic diagram of the process of the present invention.

[0051] Figure 2(a) is an example of a laboratory simulation bearing data signal.

[0052] Figure 2(b) is an example of a real sampled bearing data signal.

[0053] Figure 3(a) shows an example of VMD decomposition of some real sampled fault data Figure 1.

[0054] Figure 3(b) is an example of VMD decomposition of some real sampled fault data in Figure 2.

[0055] Figure 3(c) is an example of VMD decomposition of some real sampled fault data.

[0056] Figure 3(d) is an example of VMD decomposition of some real sampled fault data.

[0057] Figure 4(a) is an example of the decomposition and integration signal of real sampled bearing data.

[0058] Figure 4(b) is an example of the decomposition and integration signal of laboratory simulation bearing data.

[0059] Figure 5(a) is an example of the integrated normalized signal of real sampled bearing data.

[0060] Figure 5(b) is an example of the integrated normalized signal of laboratory simulation bearing data.

[0061] Figure 6(a) is a t-SNE comparison example of real sampled bearing data.

[0062] Figure 6(b) is an example of t-SNE comparison of laboratory simulation bearing data.

[0063] Figure 7(a) shows the performance of different methods on the training set and test set.

[0064] Figure 7(b) is an example diagram showing the comparison of training time of different methods on the training set. DETAILED DESCRIPTION

[0065] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the present invention is further explained below with reference to specific diagrams.

[0066] See also Figure 1 , the figure shows a complex signal fault diagnosis method based on multi-modal feature extraction and domain adaptation, which includes the following steps:

[0067] Step S10, data loading: data loading processing is performed on the laboratory simulation bearing vibration signal data set and the real sampling bearing vibration signal data set, completing the data loading operation and providing a comprehensive and diverse data basis for subsequent feature extraction and algorithm training.

[0068] Step S20, data preprocessing: filtering the vibration signal data loaded in step S10 to ensure the integrity and consistency of the data, which covers a variety of fault states, each of which contains multiple samples. In order to improve the data quality, the present invention comprehensively cleans up the data, including removing possible null values, removing outliers and data samples of abnormal length, ensuring that the data structure is standardized and the content is reliable. Through these preprocessing steps, a solid foundation is laid for subsequent feature extraction and model training.

[0069] Step S30, multi-mode feature extraction: the vibration signal data filtered in step S20 is decomposed into multiple feature signals by a signal processing method based on variational mode decomposition, and statistical feature extraction is performed on these feature signals. The main extracted features include four key indicators: mean, standard deviation, variance and kurtosis. The extracted features are then integrated to replace the original vibration signal data to generate multi-mode feature signals with higher discrimination and diversity. These newly generated feature data provide a richer, more stable and more representative input basis for subsequent algorithm training.

[0070] Step S40, data normalization processing: normalizing the multi-modal feature signal generated in step S30 to eliminate the dimensional differences between different feature dimensions and ensure the balance and stability of the feature data in algorithm training.

[0071] Step S50, data loader production: In order to verify the universality of the method, two data partitioning strategies are used for experiments; on the one hand, the laboratory simulation bearing vibration signal data set collected at high frequency is used as the training set, and the real sampling bearing vibration signal data set collected at low frequency is used as the test set to verify the performance of the model in mapping from high frequency to low frequency data in cross-domain scenarios; on the other hand, the data partitioning strategy is reversed, and the real sampling bearing vibration signal data set collected at low frequency is used as the training set, and the laboratory simulation bearing vibration signal data set collected at high frequency is used as the test set to verify the performance of the model in mapping from low frequency to high frequency data. This two-way verification strategy can not only evaluate the domain adaptability of the method between different frequency data, but also further verify its robustness and universality when migrating between high-frequency and low-frequency data.

[0072] Step S60, algorithm evaluation and structural design: After comprehensive evaluation of data characteristics and multiple machine learning methods, convolutional neural network is selected as the structural design of the model, with its powerful feature extraction ability to adapt to the pattern analysis of complex signals; at the same time, variance difference representation (VDR) is used as a distribution alignment measurement method to reduce the difference between the source domain and target domain data distribution and enhance the adaptability of the model in the unlabeled target domain. During the model training process, the deep learning ability of CNN is combined with the domain alignment technology of VDR, so that the model can make full use of cross-domain data features and achieve high-precision fault diagnosis performance. This design provides an efficient solution to the challenge of the difference in feature distribution between the source domain and the target domain in actual scenarios.

[0073] Step S70, algorithm training and performance evaluation: Use the data set processed in step S50 to train the model designed in step S60. The present invention sets up multiple rounds of iterative training and designs four loss functions and evaluation indicators, namely VDR loss, classification loss, training set accuracy and test set accuracy, to intuitively judge the training progress and performance of the network. At the end of each round of iteration, the above four indicators are calculated, the optimization effect of the model is dynamically evaluated, and the current network parameters are saved to ensure that the best model state can be recorded. Through this training strategy, the convergence speed and performance stability of the network can be effectively improved, laying the foundation for achieving high-quality fault diagnosis.

[0074] In step S10, data loading includes the following steps:

[0075] Step S11, downloading the specified laboratory simulation bearing vibration signal data set and the real sampling bearing vibration signal data set, as shown in Figure 2(a) and Figure 2(b), to ensure that the data sets are complete and in a standardized format;

[0076] Step S12, classify and load the downloaded laboratory simulation bearing vibration signal data set and the real sampling bearing vibration signal data set according to the vibration signal data and the corresponding label data, so as to provide a clear data structure for subsequent processing.

[0077] In step S10, the loaded bearing vibration signal data has a fixed length, each piece of data contains the same number of sampling points, covers a variety of fault states, each state contains several pieces of data, the data distribution is balanced and the sample size is sufficient. These data provide a diverse basis for subsequent feature extraction and model training.

[0078] In step S20, data preprocessing includes the following steps:

[0079] Step S21, detecting the bearing vibration signal data loaded in step S10, and clearing possible null value data to ensure the validity of each data sample;

[0080] Step S22, eliminating outliers and noise in the bearing vibration signal data to avoid interference of extreme values ​​on model training;

[0081] Step S23, checking the data length of the bearing vibration signal data, confirming that all samples have the same number of sampling points, and removing data samples with abnormal lengths.

[0082] The above cleaning and filtration steps provide high-quality and consistent input data for subsequent steps.

[0083] In step S30, multi-modal feature extraction includes the following steps:

[0084] Step S31, performing frequency domain evaluation on the signals in the healthy state and the faulty state to ensure that the signal strength does not exceed the specified threshold set by the method, so as to ensure the availability of the signal features and the effectiveness of the analysis;

[0085] Step S32, using VMD method to process the original signal, decompose the original signal into several modes based on the threshold, and extract statistical features from each decomposed mode signal, including multiple statistical feature indicators, to form multi-dimensional feature data. The processing effect is as follows: Figure 3(a) to Figure 3(d) As shown, the objective function of VMD is as follows;

[0086]

[0087] Among them, u k (t) represents the kth modal signal obtained by decomposition; w k represents the center frequency of the mode uk(t); α represents the balance parameter, which is used to control the mode smoothness; represents the derivative with respect to time t; δ(t) represents the unit pulse signal; j is an imaginary unit, representing the Hilbert transform of the signal;

[0088] In step S33, the signal features of all modes are integrated to generate a new data vector, the integration effect of which is shown in Figures 4(a) and 4(b), which replaces the original signal and provides a richer and more discriminative feature representation for subsequent modeling and training.

[0089] In step S40, the data normalization process is specifically as follows:

[0090] The newly generated multidimensional data vector is normalized. The processing results are shown in Figure 5(a) and Figure 5(b). The minimum-maximum normalization method with translation term is used to linearly scale the eigenvalues ​​to the interval [0,1]. The normalization formula is as follows:

[0091]

[0092] Among them, x represents the original eigenvalue; min(x) represents the minimum value among the original eigenvalues; max(x) represents the maximum value among the original eigenvalues; and x' represents the normalized eigenvalue.

[0093] In step S50, the data loader production includes the following steps:

[0094] Step S51, in order to verify the performance of data mapping from high frequency to low frequency, the laboratory simulated bearing vibration signal dataset collected at high frequency is used as the training set, and its rich frequency range and detailed information are used to train the model; at the same time, the real sampled bearing vibration signal dataset used at low frequency is used as the test set, and cross-domain prediction is performed on it through the model; this experiment aims to evaluate the performance of the model in mapping data from high frequency to low frequency, including the ability of feature extraction and fault pattern recognition, and simulates the application scenario of the model trained under high sampling frequency conditions in a low sampling frequency environment, so as to verify the domain adaptability of the model when the frequency decreases.

[0095] Step S52: In order to verify the performance of data mapping from low frequency to high frequency, the real sampled fault data set collected at low frequency is used as the training set, and its relatively simplified features are used to train the model; at the same time, the laboratory fault simulation data set collected at high frequency is used as the test set to test the performance of the model under high frequency data conditions. This experiment aims to evaluate the model's ability to map from low frequency to high frequency data, especially the performance of extracting subtle features at higher sampling frequencies, and simulates the application scenario of the model trained under low sampling frequency conditions in a high sampling frequency environment, thereby verifying the model's domain migration ability and robustness under frequency increase conditions.

[0096] In step S60, the algorithm evaluation and structure design includes the following steps:

[0097] Step S61, through comprehensive evaluation of data characteristics and multiple machine learning methods, a convolutional neural network (CNN) is selected as the structural design of the model; CNN, with its powerful feature extraction capability, can effectively capture local pattern features in complex signals and adapt to the multi-mode characteristic requirements of wind turbine gearbox bearing vibration signals;

[0098] In step S62, the variance difference representation method is used as a distribution alignment measurement method. VDR measures the variance distribution difference between the source domain and the target domain to guide the model to continuously reduce the deviation of the feature distribution of the two domains during the training process, as shown in Figure 6(a) and Figure 6(b). It is particularly suitable for scenarios where the target domain has no labels, and enhances the adaptability of the model in cross-domain data. The objective function of VDR is as follows:

[0099]

[0100] in, Represents the variance information of the feature map, which is mapped to the Reproducing Kernel Hilbert Space (RKHS) through the kernel function; represents the tensor product of two Hilbert spaces, which is used to represent variance information; κ(x,·) represents the kernel function, which is used to map the input to a high-dimensional feature space; E x～p(x) represents the expectation of the source domain distribution p(x); E y～q(y) Represents the expectation of the target domain distribution q(y);

[0101] Step S63, during the model training process, a fault diagnosis model that can efficiently utilize cross-domain data features is constructed by combining the deep feature extraction capability of CNN with the domain alignment optimization technology of VDR. By gradually optimizing the training loss function, the model can simultaneously learn the discrimination capability of source domain features and the consistency of inter-domain distribution, and finally achieve high-precision fault diagnosis of target domain data. This design provides an efficient and innovative solution to the problem of feature distribution differences between source and target domains.

[0102] In step S70, algorithm training and performance evaluation includes the following steps:

[0103] Step S71, using the data set processed in step S50 as input data, to train the algorithm network designed in step S60. The present invention sets multiple rounds of iterative training, initializes the parameters and optimizer of the model, and loads the divided source domain training set and target domain test set data at the same time, so as to ensure that the training process can fully utilize the diversity of cross-domain data;

[0104] Step S72, during the training process, four key loss functions and evaluation indicators are designed to comprehensively monitor the training progress and performance of the network, namely VDR loss, classification loss, training set accuracy and test set accuracy. Among them, VDR loss represents the degree of alignment of the difference in feature distribution between the source domain and the target domain; classification loss represents the accuracy of label classification in the source domain training set; training set accuracy represents the learning effect of the model on the source domain training set; test set accuracy represents the fault diagnosis ability of the evaluation model on the unlabeled test set of the target domain. At the end of each round of iteration, these four indicators are dynamically calculated to intuitively judge the optimization effect of the model;

[0105] Step S73, after each round of iteration is completed, the current performance of the model is evaluated based on the calculated loss function value and accuracy, and the parameter status of the network is saved in real time. If the performance of a certain round of iteration reaches the optimal value, the status is recorded as the final model. Through this dynamic training strategy, the convergence speed and performance stability of the network can be effectively improved, laying a solid foundation for achieving high-quality cross-domain fault diagnosis.

[0106] The following is a table showing the comparison results between the method of the present invention and the traditional method, and see FIG. 7(a) and FIG. 7(b):

[0107]

[0108] Compared with the traditional method, the present invention has the following advantages:

[0109] 1. Domain adaptability from high-frequency to low-frequency data

[0110] In the experiment from the high-frequency laboratory simulation bearing data set to the low-frequency real sampling bearing data set, the test accuracy of the MMFE-DA method reached 93.75%, which is significantly higher than the 90.42% of the CNN+VDR method and the 16.27% of the traditional CNN method. This result shows that the MMFE-DA method can effectively adapt to the frequency reduction scenario and capture the key features of the low-frequency domain.

[0111] 2. Mapping capability from low frequency to high frequency data

[0112] In the experiment from low-frequency real sampling bearing dataset to high-frequency laboratory simulation bearing dataset, the test accuracy of the MMFE-DA method reached 83.80%, which is 28.30% higher than the 55.50% of the CNN+VDR method, and much higher than the 14.25% of the traditional CNN method. This result verifies the excellent performance of the MMFE-DA method in feature mapping and distribution alignment, especially in the complex patterns of high-frequency signals.

[0113] 3. Advantages of multimodal feature extraction

[0114] In the experiment, the CNN+VDR method combined the variational mode decomposition (VMD) technology to perform multi-mode decomposition and feature extraction on complex signals, significantly improving the fault diagnosis performance in cross-domain scenarios compared to the traditional CNN method. For example, in the test scenario of high-frequency to low-frequency fault mapping, the test accuracy of the CNN+VDR method reached 90.42%, while the test accuracy of the traditional CNN method was only 16.27%. This result shows that the introduction of VMD can effectively extract the key features of multi-mode signals and enhance the stability and separability of the features.

[0115] It should be pointed out that the traditional CNN method does not apply any cross-domain adaptation strategy, so its performance in cross-domain scenarios is limited. The CNN+VDR method combines VMD and cross-domain technology, reduces the difference between domains through feature distribution alignment, and significantly improves the diagnostic ability of the model in the target domain without labels.

[0116] 4. Improvement of real-time performance and computing efficiency:

[0117] Compared with the CNN+VDR method, which requires 2335s and 2633s of training time in the fault mapping of high-frequency data sets to low-frequency data sets and its opposite scenario, the MMFE-DA method reduces the training time to 636s and 812s respectively, achieving a training efficiency improvement of more than 70%. This optimization lays the foundation for the application of real-time industrial fault diagnosis.

[0118] 5. Engineering application value:

[0119] The high-accuracy and high-efficiency MMFE-DA method can significantly improve the fault detection level of key equipment such as wind turbine gearboxes. By improving diagnostic reliability and efficiency, the MMFE-DA method can effectively reduce equipment operation and maintenance costs, providing strong support for the intelligence and stability of industrial equipment.

[0120] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A complex signal fault diagnosis method based on multi-modal feature extraction and domain adaptation, characterized in that: The following steps are involved: Step S10, data loading: performing data loading processing on the laboratory simulation bearing vibration signal data set and the real sampled bearing vibration signal data set; Step S20, data preprocessing: filtering the vibration signal data loaded in step S10; Step S30, multi-mode feature extraction: Decomposing the vibration signal data filtered in step S20 into multiple feature signals by using a signal processing method based on variational mode decomposition, and performing statistical feature extraction processing on these feature signals, and then integrating the extracted features to replace the original vibration signal data, thereby generating a multi-mode feature signal with higher discrimination and diversity; Step S40, data normalization processing: normalizing the multi-mode feature signal generated in step S30; Step S50, data loader production: on the one hand, the laboratory simulation bearing vibration signal data set collected at high frequency is used as a training set, and the real sampling bearing vibration signal data set collected at low frequency is used as a test set to verify the performance of the model in mapping from high frequency to low frequency data in a cross-domain scenario; on the other hand, the real sampling bearing vibration signal data set collected at low frequency is used as a training set, and the laboratory simulation bearing vibration signal data set collected at high frequency is used as a test set to verify the performance of the model in mapping from low frequency to high frequency data; Step S60, algorithm evaluation and structural design: convolutional neural network is selected as the structural design of the model, and variance difference representation is used as the measurement method of distribution alignment. During the model training process, the deep learning ability of CNN and the domain alignment technology of VDR are combined to enable the model to make full use of cross-domain data features and achieve high-precision fault diagnosis performance; Step S70, algorithm training and performance evaluation: Use the data set processed in step S50 to train the model designed in step S60.

2. The complex signal fault diagnosis method based on multi-mode feature extraction and domain adaptation according to claim 1, characterized in that: In step S10, the data loading includes the following steps: Step S11, downloading the specified laboratory simulation bearing vibration signal data set and the real sampling bearing vibration signal data set, and ensuring that the data sets are complete and in a standardized format; Step S12, classify and load the downloaded laboratory simulation bearing vibration signal data set and the real sampling bearing vibration signal data set according to the vibration signal data and the corresponding label data, so as to provide a clear data structure for subsequent processing.

3. The complex signal fault diagnosis method based on multi-mode feature extraction and domain adaptation as claimed in claim 2, characterized in that: The bearing vibration signal data loaded in step S10 has a fixed length, each piece of data contains the same number of sampling points, covers a variety of fault states, each state contains several pieces of data, the data distribution is balanced and the sample size is sufficient.

4. The complex signal fault diagnosis method based on multi-mode feature extraction and domain adaptation as claimed in claim 1, characterized in that: In step S20, the data preprocessing includes the following steps: Step S21, detecting the bearing vibration signal data loaded in step S10, and clearing possible null value data to ensure the validity of each data sample; Step S22, eliminating outliers and noise in the bearing vibration signal data to avoid interference of extreme values ​​on model training; Step S23, checking the data length of the bearing vibration signal data, confirming that all samples have the same number of sampling points, and removing data samples with abnormal lengths.

5. The complex signal fault diagnosis method based on multi-mode feature extraction and domain adaptation as claimed in claim 1, characterized in that: In step S30, the multi-modal feature extraction includes the following steps: Step S31, performing frequency domain evaluation on the signals in the healthy state and the faulty state to ensure that the signal strength does not exceed the specified threshold set by the method, so as to ensure the availability of the signal features and the effectiveness of the analysis; Step S32, using VMD method to process the original signal, decompose the original signal into several modes based on the threshold, and extract statistical features from each decomposed mode signal, including multiple statistical feature indicators, to form multi-dimensional feature data. The objective function of VMD is as follows; Among them, u k (t) represents the kth modal signal obtained by decomposition; w k represents the center frequency of the mode uk(t); α represents the balance parameter, which is used to control the mode smoothness; represents the derivative with respect to time t; δ(t) represents the unit pulse signal; j is an imaginary unit, representing the Hilbert transform of the signal; Step S33, integrates the signal features of all modes to generate a new data vector to replace the original signal, providing a richer and more discriminative feature representation for subsequent modeling and training.

6. The complex signal fault diagnosis method based on multi-mode feature extraction and domain adaptation as claimed in claim 1, characterized in that: In step S40, the data is normalized, specifically: The newly generated multidimensional data vector is normalized by using the minimum-maximum normalization method with translation term to linearly scale the eigenvalues ​​to the interval [0,1]. The normalization formula is as follows: Where x represents the original eigenvalue; min(x) represents the minimum value among the original eigenvalues; max(x) represents the maximum value among the original eigenvalues; and x' represents the normalized eigenvalue.

7. The complex signal fault diagnosis method based on multi-mode feature extraction and domain adaptation as claimed in claim 1, characterized in that: In step S50, the data loader is produced, including the following steps: Step S51: In order to verify the performance of data mapping from high frequency to low frequency, the laboratory simulation bearing vibration signal data set collected at high frequency is used as a training set, and its rich frequency range and detailed information are used to train the model; at the same time, the real sampled bearing vibration signal data set used at low frequency is used as a test set, and cross-domain prediction is performed on it through the model; Step S52, in order to verify the performance of mapping from low-frequency to high-frequency data, the real sampling fault data set collected at low frequency is used as the training set, and its relatively simplified features are used to train the model; at the same time, the laboratory fault simulation data set collected at high frequency is used as the test set to test the performance of the model under high-frequency data conditions.

8. The complex signal fault diagnosis method based on multi-mode feature extraction and domain adaptation as claimed in claim 1, characterized in that: In step S60, the algorithm evaluation and structure design includes the following steps: Step S61, selecting a convolutional neural network (CNN) as the structural design of the model; Step S62, using variance difference representation as a measurement method for distribution alignment, the objective function of VDR is as follows: in, Represents the variance information of the feature map, which is mapped to the Reproducing Kernel Hilbert Space (RKHS) through the kernel function; represents the tensor product of two Hilbert spaces, which is used to represent variance information; κ(x,·) represents the kernel function, which is used to map the input to a high-dimensional feature space; E x～p(x) represents the expectation of the source domain distribution p(x); E y～q(y) Represents the expectation of the target domain distribution q(y); Step S63, during the model training process, the deep feature extraction capability of CNN and the domain alignment optimization technology of VDR are combined to build a fault diagnosis model that can efficiently utilize cross-domain data features.

9. The complex signal fault diagnosis method based on multi-mode feature extraction and domain adaptation as claimed in claim 1, characterized in that: In step S70, the algorithm training and performance evaluation includes the following steps: Step S71, using the data set processed in step S50 as input data to train the algorithm network designed in step S60; Step S72: During the training process, four key loss functions and evaluation indicators are designed to comprehensively monitor the training progress and performance of the network, namely, VDR loss, classification loss, training set accuracy, and test set accuracy; at the end of each round of iteration, these four indicators are dynamically calculated to intuitively judge the optimization effect of the model; Step S73, after each round of iteration is completed, the current performance of the model is evaluated based on the calculated loss function value and accuracy, and the parameter status of the network is saved in real time. If the performance of a certain round of iteration reaches the optimal value, the status is recorded as the final model.

Images