Gearbox composite fault diagnosis method and storage medium

Abstract

The invention discloses a gearbox composite fault diagnosis method and a storage medium, and relates to the technical field of fault diagnosis. The gearbox composite fault diagnosis method comprises the steps that mode decomposition is carried out on gearbox vibration signals under multiple working conditions through an improved characteristic mode decomposition algorithm, data samples are obtained, and the data samples comprise source domain data with labels and target domain data without labels; the source domain data and the corresponding labels are used for pre-training a fault diagnosis model, and the fault diagnosis model comprises a multi-scale attention enhancement network and a Transform network; dual transfer learning is carried out on the target domain data based on a feature library and a pre-trained fault diagnosis model, fault diagnosis of the target domain data is achieved, and the feature library comprises source domain features obtained by carrying out feature extraction on the source domain data through the pre-trained fault diagnosis model. The method can realize accurate diagnosis of the composite fault of the gearbox.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault diagnosis, and in particular to a gearbox compound fault diagnosis method and a storage medium. BACKGROUND

[0002] As a core component of mechanical transmission system, gearbox is widely used in wind power, rail transportation, aerospace and many other fields, and its running state is directly related to the safety and reliability of the equipment. Under complex working conditions such as variable load, high speed rotation and high temperature, gearbox is prone to compound faults such as gear wear and bearing crack. With the advancement of equipment intelligence, it has become a research focus in industry and academia to realize the diagnosis of new or rare faults of gearbox.

[0003] Currently, traditional fault diagnosis methods face many difficulties: on the one hand, the gearbox vibration signal has nonlinear and non-stationary characteristics, and traditional signal processing methods such as wavelet transform and Fourier transform cannot effectively extract the weak fault features. On the other hand, deep learning-based methods highly depend on large-scale labeled data for training, and their diagnosis ability is obviously insufficient when facing unseen fault types. Existing zero-shot diagnosis techniques, such as attribute-based transfer learning, rely on expert-defined fault attributes, and their feature extraction is not optimized for gearbox compound faults, which cannot meet the actual engineering requirements. In addition, most existing methods are based on single working condition data, and the utilization of multi-source information is not sufficient, which limits the generalization ability of the model. SUMMARY

[0004] The purpose of the present application is to provide a gearbox compound fault diagnosis method and a storage medium to realize accurate diagnosis of gearbox compound faults.

[0005] In a first aspect, the present application provides a gearbox compound fault diagnosis method, which comprises: using an improved feature mode decomposition algorithm to decompose gearbox vibration signals under multiple working conditions to obtain data samples, wherein the data samples include source domain data with labels and target domain data without labels; pre-training a fault diagnosis model using the source domain data and its corresponding labels, wherein the fault diagnosis model includes a multi-scale attention enhancement network and a Transformer network; performing double transfer learning on the target domain data based on a feature library and the pre-trained fault diagnosis model to realize fault diagnosis of the target domain data, wherein the feature library includes source domain features obtained by the pre-trained fault diagnosis model performing feature extraction on the source domain data.

[0006] In some embodiments of the present application, the method further comprises: training the pre-trained fault diagnosis model using the target domain data and its corresponding fault diagnosis result to obtain a trained fault diagnosis model; and using the trained fault diagnosis model to diagnose the compound fault of the gearbox.

[0007] In some embodiments of the present application, the number of decomposition modes and the number of frequency band segments used in the improved characteristic mode decomposition algorithm are obtained based on genetic algorithm optimization.

[0008] In some embodiments of the present application, when the number of decomposition modes and the number of frequency band segments are optimized based on genetic algorithm, chromosomes are formed by binary coding parameters, the correlation kurtosis is used as the fitness function, and the population is iteratively updated by roulette selection, crossover and mutation until the fitness converges, to obtain the optimized number of decomposition modes and the number of frequency band segments.

[0009] In some embodiments of the present application, the multi-scale attention enhancement network first strengthens the key channel features through the channel attention module, then fuses the spatial position correlation information through the position attention module, and fuses the output features of the channel attention module and the position attention module to realize multi-scale feature enhancement.

[0010] In some embodiments of the present application, the output features of the channel attention module and the position attention module are fused by the following formula: ; Wherein, represents the fusion result, represents the output feature of the channel attention module, represents the output feature of the position attention module, represents a scaling factor.

[0011] In some embodiments of the present application, the feature library is further optimized by the following objective function: ; Wherein, represents the source domain feature, D represents the feature library, represents a reconstruction error term, represents a sparse coding coefficient matrix and represents a regularization term.

[0012] In some embodiments of the present application, the double transfer learning of the target domain data based on the feature library and the pre-trained fault diagnosis model comprises: generating pseudo labels by sequentially performing DBSCAN clustering and k-means optimization on the target domain data; performing feature extraction on the target domain data by using the pre-trained fault diagnosis model to obtain target domain features; performing transfer learning on the target domain features based on the features in the feature library; taking the pseudo labels and the target domain features after the transfer learning as feedback, and performing k-means optimization again to update the pseudo labels and the feature library, and returning to the step of performing transfer learning on the target domain features based on the features in the feature library until the fault diagnosis result of the target domain data is obtained.

[0013] In some embodiments of the present application, the k-means optimization is performed by the following formula: ; Wherein, represents the center vector of the jth cluster, represents the number of data points in the jth cluster, represents the ith data point in the jth cluster, represents the ith data point in the jth cluster, represents the sum operation on all data points in the jth cluster.

[0014] In a second aspect, the embodiments of the present application propose a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the gearbox compound fault diagnosis method of the first aspect.

[0015] The gearbox compound fault diagnosis method and the storage medium of the embodiments of the present application respectively perform pattern decomposition on the gearbox vibration signals under multiple working conditions by using the improved feature pattern decomposition algorithm, pre-train the fault diagnosis model including the multi-scale attention enhancement network and the Transformer network by using the source domain data and the corresponding labels, and then perform double transfer learning on the target domain data based on the feature library and the pre-trained fault diagnosis model, so that the unknown faults of the target domain data can be accurately diagnosed. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a flowchart of the gearbox compound fault diagnosis method of the embodiments of the present application; Figure 2 is a flowchart of the gearbox compound fault diagnosis method of an example of the present application; Figure 3 is a flowchart of the improved feature pattern decomposition algorithm of an embodiment of the present application; Figure 4is a structural schematic diagram of a channel attention module of one example of the present application. Figure 5 is a structural schematic diagram of a position attention module of one example of the present application. Figure 6 is a structural schematic diagram of a double migration clustering diagnosis module of one embodiment of the present application. Figure 7 is a structural schematic diagram of a Transformer network of one example of the present application. Figure 8 is a schematic diagram of clustering of one example of the present application. DETAILED DESCRIPTION

[0017] Embodiments of the present application are described in detail below with reference to the accompanying drawings, examples of which are shown in the drawings, wherein the same or similar notations represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to explain the present application, and cannot be understood as a limitation of the present application.

[0018] A gearbox composite fault diagnosis method and a storage medium of an embodiment of the present application are described below with reference to the accompanying drawings.

[0019] The conventional gearbox fault diagnosis method has difficulty in effectively extracting weak fault features when processing nonlinear and non-stationary vibration signals, and is highly dependent on a large amount of labeled data, and has insufficient diagnosis capability when facing unobserved fault types. At the same time, the existing zero-shot diagnosis technology relies on expert-defined fault attributes, and the feature extraction is not optimized for gearbox composite faults, which cannot meet the actual engineering requirements.

[0020] Therefore, the present application proposes a gearbox composite fault diagnosis method to break through the dependence of the conventional technology on labeled data and significantly improve the diagnosis accuracy and robustness of gearbox composite faults under complex working conditions. The method uses an IFMD (Improved Feature Mode Decomposition) algorithm to preprocess the vibration signal, suppresses noise interference and enhances fault features; with the multi-scale attention mechanism of the multi-scale attention enhancement network, it adaptively captures key frequency band features and realizes feature enhancement; through a double migration clustering diagnosis method, it completes knowledge transfer and pseudo-label mining, and then realizes effective diagnosis of unobserved fault types of the gearbox.

[0021] Figure 1 is a flowchart of a gearbox composite fault diagnosis method of an embodiment of the present application.

[0022] As shown in Figure 1 , the gearbox composite fault diagnosis method comprises: S11, the improved feature mode decomposition algorithm is used to respectively decompose the gear box vibration signals under multiple working conditions to obtain data samples, wherein the data samples include source domain data with labels and target domain data without labels.

[0023] The gear box vibration signals under multiple working conditions can be gear box vibration signals collected by vibration displacement sensors under different rotating speeds, different loads, etc. As shown in FIG. 1, the vibration displacement sensors x1, x2, …, xm deployed at the input end of the gear box and the vibration displacement sensors y1, y2, …, ym deployed at the output end of the gear box can be used to collect the gear box vibration signals under multiple working conditions. Figure 2

[0024] Due to the large amount of labeling work, in order to reduce the workload, part of the gear box vibration signals can be labeled for fault and fault type to obtain labeled gear box vibration signals, and part of the gear box vibration signals can be unlabeled to obtain unlabeled gear box vibration signals. Then, the labeled gear box vibration signals and the unlabeled gear box vibration signals can be randomly divided into training sets and test sets according to a certain proportion. For example, the number of labeled gear box vibration signals is N1, the number of unlabeled gear box vibration signals is N2, N1x7 / 10 labeled gear box vibration signals and N2x7 / 10 unlabeled gear box vibration signals are used as the training set, and the rest is used as the test set.

[0025] In order to suppress noise and highlight fault features, the IFMD algorithm can be used to preprocess the gear box vibration signals under multiple working conditions to remove interference information, and the processing result is converted into a frequency domain signal. As shown in FIG. 2, the IFMD algorithm can decompose each gear box vibration signal into n groups of signals to provide input for the multi-scale attention enhancement network in the subsequent fault diagnosis model. Figure 2

[0026] As an embodiment, the number of decomposition modes and the number of frequency band segments used in the improved feature mode decomposition algorithm are obtained based on genetic algorithm optimization.

[0027] Exemplarily, when the number of decomposition modes and the number of frequency band segments are optimized based on the genetic algorithm, a chromosome is formed by binary coding parameters, the correlation kurtosis is used as the fitness function, the roulette selection, crossover and mutation are iteratively updated to update the population until the fitness converges, and the optimized number of decomposition modes and the number of frequency band segments are obtained.

[0028] ​​Specifically, the traditional FMD algorithm is of great significance in the field of rotating machinery fault diagnosis, but it has certain limitations in parameter selection. The FMD algorithm mainly uses an adaptive FIR (Finite Impulse Response) filter to decompose the signal. In order to more accurately evaluate the quality of the decomposition pattern and effectively extract the fault characteristic signal, CK (Cumulant Kurtosis) is used as the objective function. Its calculation formula is as follows (1): (1) in, Indicates the The first mode component The value of the point, Indicates the The first mode component The value of the point, , N represents the signal length, M represents the pre-set correlation order, This method initializes the FIR filter bank through the Hanning window, divides the signal frequency band into K segments, and the cutoff frequency of the kth filter is as follows (2): (2) Where fs represents the sampling frequency, 、 They represent the lower cutoff frequency and upper cutoff frequency of the kth FIR filter respectively.

[0029] However, the decomposition performance of the traditional FMD algorithm is highly dependent on the manual setting of the mode number n (i.e., the number of mode components) and the number of frequency band segments K, and has problems such as strong parameter sensitivity and insufficient modal aliasing suppression.

[0030] In order to solve the parameter dependence problem in the traditional FMD algorithm, the IFMD algorithm in this invention introduces a genetic algorithm to optimize the decomposition mode number n and the frequency band segment number K. The specific optimization process is as follows: Figure 3 As shown, the chromosome is formed by binary coding parameter combination (n, K), and the related kurtosis is used as the fitness function to perform roulette wheel selection, crossover and mutation iteration to output the optimal parameter combination ( , ).

[0031] See also Figure 3 , after getting ( , ) and then ( , ) for feature mode decomposition, so that the IMF (Intrinsic Mode Function) component reflecting the fault feature more accurately can be obtained, thereby improving the quality of the input signal of the subsequent diagnosis process.

[0032] By introducing a genetic algorithm based on a traditional feature mode decomposition algorithm, the number of decomposition layers and filter parameters can be intelligently screened, thereby more efficiently processing the nonlinear and non-stationary vibration signals of the gearbox. The algorithm decomposes the original signal (i.e., the gearbox vibration signal) into multiple feature mode components, each component carrying different scale fault feature information, forming a hierarchical feature mode set, and realizing the transformation from the original signal to the fault feature signal, thereby providing a more valuable data basis for subsequent fault diagnosis.

[0033] As another implementation, the number of decomposition modes and the number of frequency band segments used in the improved feature mode decomposition algorithm are obtained based on an artificial swallow algorithm.

[0034] S12, pre-training a fault diagnosis model using source domain data and its corresponding labels, wherein the fault diagnosis model comprises a multi-scale attention enhancement network and a Transformer network.

[0035] In some embodiments of the present application, the multi-scale attention enhancement network first strengthens the key channel features through the channel attention module, then fuses the spatial position correlation information through the position attention module, and fuses the output features of the channel attention module and the position attention module, thereby realizing multi-scale feature enhancement.

[0036] For example, the output features of the channel attention module and the position attention module are fused by the following formula (3): (3) Wherein, denotes the fusion result, denotes the output feature of the channel attention module, denotes the output feature of the position attention module, denotes the scaling factor.

[0037] Specifically, in the self-attention mechanism, after the linear transformation of the input sequence, the query vector Q, the key vector K and the value vector V are generated respectively. Here, is a scalar, which is usually equal to the dimension number of the key vector K, and mainly serves to scale the attention score to reduce the variance. The specific calculation process is as follows: first, the dot product operation of the query vector Q and the key vector K is performed, so that the correlation degree score between all elements can be obtained. Then, divide these scores by After scaling, the softmax function is used to convert it into probability weights. Finally, these probability weights are weighted and summed with the value vector V to obtain the final output result. The self-attention formula is shown in formula (4).

[0038] (4) Next, we will introduce in detail the position attention module and channel attention module used in the multi-scale attention enhancement network. The channel attention module takes screening key channel features as its core goal and realizes feature enhancement in the channel dimension through a systematic process. The specific flow chart is as follows: Figure 4 As shown. First, the input features Perform average pooling (AvgPool) and maximum pooling (MaxPool) processing respectively to obtain and Subsequently, the two pooling results are fused through a shared multi-layer perceptron (MLP): ,in, is the Sigmoid activation function, is the dimensionality reduction weight matrix (r is the compression ratio, such as r=16), b1 is the bias term. Finally, the generated channel attention weight With input features Multiply them together to achieve weighted reinforcement of key channels, as shown in the following formula 5): (5) in, Represents channel-by-channel multiplication operation, output feature The channel information which is more discriminative for fault diagnosis is retained.

[0039] After the channel attention module completes the channel key information screening of the input features, the output feature A immediately enters the position attention module to start the feature optimization of the spatial dimension. The position attention module is centered on mining spatial position dependencies. Its specific flow chart is as follows: Figure 5 See Figure 5 , for the output feature A of the channel attention module, three sets of convolutions generate features B, C, and D, which are reshaped into 、 、 After that, calculate the spatial attention matrix ( is the feature dimension), and then through Fusion of spatial correlation features ( is the learnable coefficient, That is Figure 5 E in the figure) to achieve weighted aggregation of features at each location and enhance the discriminability of fault features in the spatial dimension.

[0040] After the processing of the channel attention module and the position attention module, a parallel fusion strategy is adopted to realize multi-dimensional feature enhancement. Specifically, the features A output by the channel attention are added to the features E output by the position attention element by element, that is, , so as to retain the key band information of the channel dimension and the position dependence of the spatial dimension. The fusion process can further adjust the weights of the two by a scaling factor , as shown in the above formula (3), to finally generate a dual attention feature containing both channel screening and spatial correlation, obtaining a multi-scale discriminative feature, which provides a more discriminative input representation for subsequent feature classification or clustering.

[0041] By integrating the related features between all channel mappings through the channel attention module, the channel mappings that exist in mutual dependence can be selectively emphasized, effectively highlighting the key channel information; by weighting and aggregating the features at all positions through the position attention module, the features at each position are selectively aggregated to mine the spatial position dependence; by combining the two to realize multi-scale feature enhancement, the obtained fault features are more comprehensive and accurate, and the discriminative ability of the fault diagnosis model is improved.

[0042] As shown in Figure 6 , the fault labels of the source domain data are input into the Transformer network after classification for feature extraction, and the specific process is as shown in Figure 7 . The source sequence (i.e., the sequence output by the source domain data after processing by the multi-scale attention enhancement network) is first mapped to a high dimension through an embedding layer and position encoding . The encoder layer calculates: , , wherein the multi-head self-attention mechanism captures the long-range dependence within the sequence, and the feed-forward network (FFN) enhances the nonlinear expression ability. The decoder layer processes the target sequence (i.e., the label corresponding to the source domain data) , , , wherein the masked multi-head self-attention prevents information leakage and fuses the encoder and decoder features through cross-attention. Finally, the output is obtained as the prediction result.

[0043] When pre-training the fault diagnosis model, the loss function can be calculated based on the prediction result and the corresponding true label, and the parameters of the fault diagnosis model can be updated based on the loss function.

[0044] S13, based on the feature library and the pre-trained fault diagnosis model, double transfer learning is performed on the target domain data to realize fault diagnosis of the target domain data, wherein the feature library includes the source domain features obtained by the pre-trained fault diagnosis model for feature extraction on the source domain data.

[0045] In some embodiments of the present application, the feature library is further optimized by an objective function as follows (6): (6) wherein, represents the source domain feature, D represents the feature library, represents a reconstruction error term, represents a sparse coding coefficient matrix and represents a regularization term.

[0046] Specifically, the source domain feature extracted through the Transformer network will enter the feature library. In order to make the feature library store typical features of various faults in the source domain data, the source domain features in the feature library can be further optimized and processed, including: screening, integrating and learning the extracted source domain features, removing possible noise and redundant information in them, so that the source domain features in the feature library are more refined and representative. Finally, these optimized source domain features will be stored in the feature library. The objective function for optimization processing is as follows (6), The sparse coding coefficient matrix constrains the linear representation ability of D to the source domain feature, ensuring the integrity of the fault feature reservation; and respectively constrain the redundancy of D and the sparsity of the coding to extract the most discriminative fault mode basis vector.

[0047] In some embodiments of the present application, double transfer learning is performed on the target domain data based on the feature library and the pre-trained fault diagnosis model, including: generating pseudo labels by sequentially performing DBSCAN clustering and k-means optimization on the target domain data; extracting target domain features by using the pre-trained fault diagnosis model to extract features of the target domain data; performing transfer learning on the target domain features based on the features in the feature library; taking the transfer-learned target domain features and pseudo labels as feedback, and performing k-means optimization again to update the pseudo labels and the feature library, and returning to the step of performing transfer learning on the target domain features based on the features in the feature library, until the fault diagnosis result of the target domain data is obtained.

[0048] Exemplarily, k-means optimization is performed by the following formula (7): (7) wherein, represents the center vector of the jth cluster, represents the number of data points in the jth cluster, denotes the i-th data point in the j-th cluster, denotes the summation operation on all data points in the j-th cluster.

[0049] In this embodiment, Figure 2 The double migration cluster diagnosis module in the source domain knowledge migration and target domain unsupervised learning cooperatively constructs zero-shot diagnosis, and a specific flowchart is shown in Figure 6 The labeled source domain data are extracted by the encoder in the Transformer network of the pre-trained fault diagnosis model to obtain deep fault features (i.e., source domain features) and stored in a feature library; the unlabeled target domain data are filtered by the DBSCAN algorithm and initially clustered, and then the k-means algorithm is used for iterative optimization to generate pseudo labels. Subsequently, the target domain data are subjected to double migration to realize cross-domain knowledge alignment, the clustering results are optimized in combination with the pseudo label feedback, and deep adjustment is performed through the feature library to form a closed loop of “feature extraction-knowledge migration-cluster optimization-iterative learning”, thereby realizing zero-shot fault recognition.

[0050] Specifically, the target domain data are different from the source domain data in that they lack explicit fault labels. To solve this problem, the pre-trained fault diagnosis model of the source domain data is first used, and the target domain data are preprocessed by using the DBSCAN algorithm. The DBSCAN is an unsupervised clustering algorithm based on data density, which can divide data into different clusters according to the density relationship between data points, and a DBSCAN point classification diagram is shown in Figure 8 In the target domain data, different fault modes may exhibit different data distribution characteristics, and the DBSCAN algorithm can discover the rules of these data distribution to preliminarily divide the cluster structure of the data. However, the clustering results obtained by the DBSCAN algorithm may not be compact and accurate enough, and therefore the k-means algorithm is used to optimize the clustering results, and the calculation formula is as follows (7).

[0051] The k-means algorithm iteratively assigns data points to different clusters so that the points within a cluster are as close to the cluster center as possible to improve the clustering quality. The pseudo labels are generated by the DBSCAN and k-means processing to construct a weak supervision signal for the target domain data. Subsequently, the target domain data are subjected to double migration learning: on the one hand, the pre-trained fault diagnosis model of the source domain data is reused to utilize its rich fault feature extraction capability to quickly extract cross-domain general features, thereby avoiding the tediousness of training from scratch and the lack of data; on the other hand, the source domain features in the feature library are called to associate the typical fault features of the source domain with the target domain features, so that the target domain features learn to move closer to the source domain knowledge system, thereby making up for the label missing problem.

[0052] ​The target domain features after transfer learning are combined with pseudo label feedback and input into k-means again, and the clustering distribution is further optimized to accurately reflect the fault mode. The optimized features enter the feature library, and deep learning is performed based on the source domain knowledge and the clustering results, and the feature representation is continuously adjusted to adapt to the fault characteristics of the target domain. Through the iterative closed loop of "feature extraction-knowledge transfer-feature optimization", the source domain knowledge is finally used to effectively diagnose the unknown faults of the target domain in the zero sample scene.

[0053] Through the double transfer clustering diagnosis method, the double collaborative mechanism of the transfer of the fault diagnosis model parameters and the feature knowledge of the pre-trained source domain data is realized, and the diagnosis difficulty of the unknown faults of the gearbox in the target domain data without labels can be effectively solved.

[0054] In some embodiments of the present application, the gearbox compound fault diagnosis method further comprises: training the pre-trained fault diagnosis model using the target domain data and the corresponding fault diagnosis results thereof to obtain a trained fault diagnosis model; and diagnosing the compound faults of the gearbox by using the trained fault diagnosis model.

[0055] Specifically, after obtaining the fault diagnosis results of the target domain data, the target domain data can obtain labels, and at this time, the pre-trained fault diagnosis model can be continuously trained by using the target domain data and the corresponding labels thereof, so that the trained fault diagnosis model can adapt to the diagnosis of more fault types of the gearbox. In actual use, the trained fault diagnosis model can be directly used to diagnose the compound faults of the gearbox, which is simple and efficient.

[0056] For example, when the fault diagnosis model pre-training and / or training is completed, the test set data can be sequentially passed through the IFMD, the multi-scale attention enhancement network and the double transfer clustering diagnosis module, and the output results of the predicted classification are used to complete the diagnosis of the compound faults of the gearbox.

[0057] For example, after the fault diagnosis model pre-training is completed, the pre-trained fault diagnosis model is tested by using the test set corresponding to the labeled source domain data. For another example, after the fault diagnosis model training is completed, the pre-trained fault diagnosis model is tested by using the test set corresponding to the labeled source domain data. For another example, after the fault diagnosis model pre-training is completed, the first prediction result is obtained by using the test set corresponding to the unlabeled source domain data through the pre-trained fault diagnosis model, and after the fault diagnosis model training is completed, the second prediction result is obtained by using the test set corresponding to the unlabeled source domain data to test the trained fault diagnosis model, and the test is realized by comparing the two prediction results.

[0058] Based on the gearbox compound fault diagnosis method of the above-mentioned embodiments, the present application proposes a computer readable storage medium.

[0059] In the embodiment of the present application, a computer program is stored on a computer readable storage medium, and the computer program is executed by a processor to implement the gearbox compound fault diagnosis method of the above-mentioned embodiment.

[0060] In summary, the gearbox compound fault diagnosis method and storage medium of the embodiment of the present application can efficiently diagnose the compound fault of the gearbox, have good stability, are strong in adaptability to complex working conditions, and have excellent generalization ability. Specifically, first, with the help of the improved feature mode decomposition algorithm, the decomposition parameters are optimized by using the genetic algorithm, the nonlinear non-stationary vibration signal is adaptively processed, the fault mode is separated, the noise interference is suppressed, and the decomposition result containing the fault feature is obtained, greatly improving the quality and efficiency of fault feature extraction. Then, the decomposition result is converted into a frequency domain feature matrix, which is input into a multi-scale attention enhancement network, the key frequency band features are dynamically captured through the parallel structure of channel and position attention, multi-scale feature enhancement is realized, and the fault feature is more comprehensive and accurate. Finally, the enhanced features are input into a double transfer clustering diagnosis module, the source domain knowledge transfer and target domain unsupervised learning are cooperated, the feature library transfer and pseudo-label mining mechanism are used, the dependence on labeled data is broken through, the diagnosis of the gearbox fault type that has not been seen is realized, the diagnosis precision and robustness of the gearbox compound fault under complex working conditions are effectively improved, the fault diagnosis demand of actual engineering is met, and under the multiple actions of fault feature extraction, enhancement and knowledge transfer, the compound fault of the gearbox can be more accurately diagnosed.

[0061] It is to be appreciated that the above description and the examples that follow are intended to be illustrative only and that changes can be made to the description, as represented by the above listed elements, by the steps recited in the flow charts, and by the examples that follow, without departing from the spirit of the application. Accordingly, the scope of the present application is intended to be defined only by the appended claims.

[0062] It should be understood that aspects of the application can be implemented in hardware, software, firmware or a combination thereof. In the above embodiments, various steps or methods can be implemented in software or firmware that is stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, any of the following technologies, known in the art, can be used: a hybrid of the above technologies, a combination of any of the above technologies, etc.

[0063] In the description of the present application, the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" are intended to mean that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the present application. The illustrative appearances of the above-mentioned terms in various places in the specification are not necessarily intended to refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0064] In the description of the application, it should be understood that the orientation or positional relationship indicated by the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "upper", "lower", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise", "axial", "radial", "circumferential" and the like is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the application and simplifying the description, and does not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the application.

[0065] In addition, the terms "first", "second" are only for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined with "first", "second" can explicitly or implicitly include at least one of the features. In the description of the application, the meaning of "a plurality of" is at least two, such as two, three, etc., unless otherwise explicitly specified and limited.

[0066] In the present application, unless otherwise explicitly specified and limited, the terms "mounting", "connecting", "connecting", "fixing" and the like should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or it can be integrated; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium; it can be the internal communication of two elements or the interaction relationship between two elements, unless otherwise explicitly limited. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0067] In the present application, unless otherwise explicitly specified and limited, the first feature is "on" or "under" the second feature, which can be direct contact between the first and second features, or indirect contact between the first and second features through an intermediate medium. Moreover, the first feature "above", "above" and "above" the second feature can be directly above or obliquely above the first feature, or only indicate that the horizontal height of the first feature is higher than that of the second feature. The first feature "below", "below" and "below" the second feature can be directly below or obliquely below the first feature, or only indicate that the horizontal height of the first feature is less than that of the second feature.

[0068] Although the embodiments of the present application have been shown and described above, it should be understood that the above embodiments are exemplary and cannot be understood as a limitation on the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the present application.

Claims

1. A gearbox composite fault diagnosis method, characterized in that: include: An improved eigenmode decomposition algorithm is used to perform mode decomposition on the gearbox vibration signals under multiple working conditions to obtain data samples, wherein the data samples include labeled source domain data and unlabeled target domain data; Pre-training a fault diagnosis model using the source domain data and its corresponding labels, wherein the fault diagnosis model includes a multi-scale attention enhancement network and a Transformer network; Dual transfer learning is performed on the target domain data based on a feature library and a pre-trained fault diagnosis model to achieve fault diagnosis of the target domain data, wherein the feature library includes source domain features obtained by extracting features from the source domain data by the pre-trained fault diagnosis model.

2. The gearbox composite fault diagnosis method according to claim 1, characterized in that: The method further comprises: Using the target domain data and its corresponding fault diagnosis results to train the pre-trained fault diagnosis model to obtain a trained fault diagnosis model; The trained fault diagnosis model is used to diagnose compound faults of the gearbox.

3. The gearbox composite fault diagnosis method according to claim 1, characterized in that: The number of decomposition modes and the number of frequency band segments used in the improved characteristic mode decomposition algorithm are obtained based on genetic algorithm optimization.

4. The gearbox composite fault diagnosis method according to claim 3, characterized in that: When optimizing the number of decomposition modes and the number of frequency band segments based on a genetic algorithm, chromosomes are formed by binary coding parameters, and the relevant kurtosis is used as the fitness function. Roulette wheel selection, crossover and mutation are performed to iteratively update the population until the fitness converges, thereby obtaining the optimized number of decomposition modes and the number of frequency band segments.

5. The gearbox composite fault diagnosis method according to claim 1, characterized in that: The multi-scale attention enhancement network first strengthens the key channel features through the channel attention module, then fuses the spatial position association information through the position attention module, and fuses the output features of the channel attention module and the position attention module to achieve multi-scale feature enhancement.

6. The gearbox composite fault diagnosis method according to claim 5, characterized in that: The output features of the channel attention module and the position attention module are fused by the following formula: ； in, represents the fusion result, represents the output features of the channel attention module, represents the output features of the position attention module, Represents the scaling factor.

7. The gearbox composite fault diagnosis method according to claim 1, characterized in that: The feature library is also optimized by the following objective function: ； in, represents the source domain feature, D represents the feature library, represents the reconstruction error term, Represents the sparse coding coefficient matrix and represents the regularization term.

8. The gearbox composite fault diagnosis method according to claim 1, characterized in that: The dual transfer learning of the target domain data based on the feature library and the pre-trained fault diagnosis model includes: Performing DBSCAN clustering and k-means optimization on the target domain data in turn to generate pseudo labels; Extracting features from the target domain data using the pre-trained fault diagnosis model to obtain target domain features; Performing transfer learning on the target domain features based on the features in the feature library; The target domain features after transfer learning and the pseudo labels are used as feedback, and k-means optimization is performed again to update the pseudo labels and the feature library, and the process returns to the step of performing transfer learning on the target domain features based on the features in the feature library until a fault diagnosis result of the target domain data is obtained.

9. The gearbox composite fault diagnosis method according to claim 8, characterized in that: K-means optimization is performed using the following formula: ； in, represents the center vector of the j-th cluster, represents the number of data points in the jth cluster, Indicates belonging to the jth cluster The i-th data point in Indicates the summation operation of all data points in the jth cluster.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the gearbox composite fault diagnosis method according to any one of claims 1 to 9 is implemented.

Images