A bearing fault diagnosis method based on a bidirectional Mamba network

Abstract

The application is suitable for the field of high-speed train bearing fault diagnosis technology, and provides a bearing fault diagnosis method based on a bidirectional Mamba network, comprising: collecting a fault vibration signal of a high-speed train bearing; inputting the fault vibration signal into a fault diagnosis model for analysis and processing to obtain a fault diagnosis result of the high-speed train bearing; the fault diagnosis model comprises a two-dimensional CNN network, a time attention network, a bidirectional Transformer network, a one-dimensional time sequence processing network, a first fusion network based on channel attention, a data compression network, a bidirectional Mamba network, a second fusion network based on channel attention, and an output network which are connected in sequence. The application can improve the accuracy of high-speed train bearing fault diagnosis.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of high-speed train bearing fault diagnosis, and particularly relates to a bearing fault diagnosis method based on a bidirectional Mamba network. BACKGROUND

[0002] In industrial production, intelligent fault diagnosis technology of rotating machinery plays a key role. The bearing of a high-speed train, as one of the core rotating components in a bogie, not only bears the weight of the train, but also directly relates to the stability and safety of the train in operation. Therefore, timely and effective state monitoring and fault diagnosis of the bearing are of great significance to ensure the safe operation of the train. However, the current state detection and fault diagnosis method still faces many challenges when facing massive monitoring data.

[0003] For a high-speed train: 1. A large amount of monitoring data is generated during the operation of the train, which contains a variety of working condition information. However, the existing diagnosis model often faces the problem of insufficient generalization ability when dealing with these complex data, and it is difficult to maintain high-precision diagnosis effect under different working conditions, thereby affecting the actual application effect and reliability of the fault diagnosis system. 2. The dependence relationship of massive data in time and channel is often ignored, which further leads to problems such as insufficient feature extraction, increasing the difficulty of fault diagnosis. These characteristics make it difficult for intelligent fault diagnosis models to generalize to different working condition data of the same type of equipment. Therefore, how to complete the deep mining of the relationship between time and channel of existing data under limited available data is a problem to be solved in intelligent fault diagnosis.

[0004] In view of the deep extraction of the relationship between time and channel in the existing sample data, a direct and effective method is to introduce an attention mechanism to automatically focus on important features and improve the accuracy of fault diagnosis. However, the current mainstream attention mechanism method is usually a single channel attention mechanism or a time attention mechanism, which can only capture the single dependence relationship of channel sequence and time sequence, resulting in that the model does not fully understand the channel and time sequence, and the accuracy of bearing fault diagnosis is low. SUMMARY

[0005] The embodiment of the application provides a bearing fault diagnosis method based on a bidirectional Mamba network, which can solve the problem of low accuracy of high-speed train bearing fault diagnosis.

[0006] The embodiment of the application provides a bearing fault diagnosis method based on a bidirectional Mamba network, which can solve the problem of low accuracy of high-speed train bearing fault diagnosis.

[0007] Collecting a fault vibration signal of a high-speed train bearing;

[0008] The fault vibration signal is input into a fault diagnosis model for analysis and processing, and a fault diagnosis result of the bearing of the high-speed train is obtained.

[0009] The fault diagnosis model comprises a two-dimensional CNN network, a time attention network, a bidirectional Transformer network, a one-dimensional time sequence processing network, a first fusion network based on channel attention, a data compression network, a bidirectional Mamba network, a second fusion network based on channel attention, and an output network connected in sequence.

[0010] The two-dimensional CNN network is configured to extract feature information of the fault vibration signal, the time attention network is configured to process the feature information output by the two-dimensional CNN network to obtain attention feature information, the bidirectional Transformer network is configured to extract features of the attention feature information from both time and sequence directions, the one-dimensional time sequence processing network is configured to extract one-dimensional feature information of the features output by the bidirectional Transformer network, the first fusion network is configured to fuse a plurality of one-dimensional feature information output by the one-dimensional time sequence processing network, the data compression network is configured to compress a fusion result output by the first fusion network, the bidirectional Mamba network is configured to extract features of the compressed data output by the data compression network, the second fusion network is configured to fuse a plurality of features output by the bidirectional Mamba network, and the output network is configured to map a fusion result output by the second fusion network to output the fault diagnosis result.

[0011] Optionally, the two-dimensional CNN network comprises a plurality of pseudo two-dimensional convolution kernels and a first addition module, the input end of each pseudo two-dimensional convolution kernel receives the fault vibration signal, the output end of each pseudo two-dimensional convolution kernel is connected with the input end of the first addition module, the output end of the first addition module is connected with the input end of the time attention network, and the sizes of the plurality of pseudo two-dimensional convolution kernels are different from each other.

[0012] Optionally, the time attention network comprises a first branch network configured to extract attention information of different sequences at the same time, a second branch network configured to extract attention information of different times at the same sequence, a first multiplication module, a second multiplication module, and a second addition module.

[0013] The input end of the first branch network and the input end of the second branch network both receive the feature information output by the two-dimensional CNN network, the output end of the first branch network and the output end of the second branch network are both connected with the input end of the first multiplication module, the output end of the first multiplication module is connected with the input end of the second multiplication module, the output end of the two-dimensional CNN network is connected with the input end of the second multiplication module, the output end of the second multiplication module and the output end of the two-dimensional CNN network are both connected with the input end of the second addition module, and the output end of the second addition module is connected with the input end of the bidirectional Transformer network.

[0014] Optionally, the first branch network comprises a first pooling layer and a first attention information extraction module connected in sequence, and the second branch network comprises a second pooling layer and a second attention information extraction module connected in sequence.

[0015] The first attention information extraction module is configured to obtain attention information A1 of different sequences at the same time by formula A1=Sigmoid(Relu(W1*Data t ))

[0016] The second attention information extraction module is configured to obtain attention information A2 of different times of the same sequence by formula A2=Sigmoid(Relu(W2*Data c ))

[0017] Wherein, Sigmoid represents a sigmoid function, Relu represents a Relu function, Data t is data output by the first pooling layer, W1 represents a weight of Data t , Data c is data output by the second pooling layer, W2 represents a weight of Data c , a pooling window size of the first pooling layer is 1*none, a pooling window size of the second pooling layer is none*1, and none is a placeholder.

[0018] Optionally, the one-dimensional time sequence processing network comprises a plurality of one-dimensional convolution kernels, an input end of each one-dimensional convolution kernel receives features output by the bidirectional Transformer network, an output end of each one-dimensional convolution kernel is connected with an input end of the first fusion network, and sizes of the plurality of one-dimensional convolution kernels are different from each other.

[0019] Optionally, the first fusion network and the second fusion network fuse the received data by the following formula and output a fusion result:

[0020] s=Stack(input)

[0021] A c =Sigmoid(Relu(Maxpool(s)+Avgpool(s)))

[0022] stack out =A c *s

[0023] Out=sum(stack out ,dim=1)

[0024] wherein, input represents received data, Stack represents a stacking operation, s represents data obtained by stacking input along a channel dimension, Sigmoid represents a sigmoid function, Relu represents a Relu function, Maxpool represents a max-pooling function, Avgpool represents an average-pooling function, A c represents a channel attention coefficient, dim represents a dimension, sum represents a summation function, and Out represents a fusion result.

[0025] Optionally, the data compression network is configured to compress the fusion result output by the first fusion network by a plurality of multiples to obtain compressed data under the plurality of multiples.

[0026] Optionally, the bidirectional Mamba network comprises a plurality of bidirectional Mamba layers, and an output end of each bidirectional Mamba layer is connected to an input end of the second fusion network.

[0027] The plurality of bidirectional Mamba layers are configured to receive the compressed data under the plurality of multiples, and the plurality of bidirectional Mamba layers correspond to the compressed data under the plurality of multiples one by one.

[0028] Optionally, the output network comprises an average-pooling layer, a reshaping layer and a fully connected layer connected in sequence.

[0029] Optionally, the fault vibration signal is collected by a plurality of vibration sensors arranged on a bogie of a high-speed train.

[0030] The above-mentioned scheme of the present application has the following advantages:

[0031] In the embodiments of the present application, the one-dimensional time sequence processing network and the two-dimensional CNN network are combined to extract fault features, which can fully capture the channel sequence and time sequence information; the time attention network is introduced to further strengthen the capture of channel sequence and time sequence information in the feature extraction between different channels, ensuring sufficient feature extraction; the bidirectional Transformer network is introduced to use the channel-time double-direction convolution operation to extract the KQV of the multi-head attention mechanism, fully extracting the features of different channels and different times while ensuring that the channel information is not chaotic; the bidirectional Mamba network is introduced to ensure the unity of information loss filling and information precision when the SSM is long-distance modeling through parallel and different compression degree input data, ensuring that the data in the original fault vibration signal can be fully extracted, and then the output network based on the fully extracted data can greatly improve the accuracy of fault diagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0032] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0033] Figure 1 The flow chart of the bearing fault diagnosis method based on the bidirectional Mamba network provided by an embodiment of the present application;

[0034] Figure 2 The structural schematic diagram of the fault diagnosis model provided by an embodiment of the present application;

[0035] Figure 3 The structural schematic diagram of the test bench model in the example;

[0036] Figure 4a The schematic diagram of the sensor arrangement position in the example Figure 1 ;

[0037] Figure 4b The schematic diagram of the sensor arrangement position in the example Figure 2 ;

[0038] Fig. 5a-5e are the result schematic diagrams of the confusion matrix and accuracy of the bearing fault diagnosis method in the example, in which a, b, c, d, e respectively represent a one-dimensional basic model, a bidirectional multi-granularity Transformer, a two-dimensional preprocessing + channel-time attention, a bidirectional multi-granularity Mamba under channel fusion, and the final model of the experiment. DETAILED DESCRIPTION

[0039] In the following description, for the purpose of explanation and not limitation, specific details are set forth, such as specific system structures, techniques, etc., in order to thoroughly disclose the embodiments of the present application. However, it should be clear to those skilled in the art that the present application can also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted in order not to obscure the description of the present application with unnecessary details.

[0040] It should be understood that when used in the specification and the appended claims of the present application, the term "comprising" indicates the presence of the described features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0041] It should also be understood that, in the description of the application and in the appended claims, the term "and / or" is used to mean one or more of the associated listed items, as well as the sum of all possible combinations of the associated list items.

[0042] As used in the description of the application and the appended claims, the term "if' can be interpreted to mean "when" or "upon" or "in response to determining" or "in response to detecting." Similarly, the phrase "if determined" or "if detected [a described condition or event]" can be interpreted to mean "upon determining" or "in response to determining" or "upon detecting [a described condition or event]" or "in response to detecting [a described condition or event]," as appropriate.

[0043] In addition, the description in the specification of the application and the appended claims, the terms "first," "second," "third," etc. are merely used to distinguish one element from another, and are not intended to imply or suggest relative importance.

[0044] Reference throughout this specification to "one embodiment" or "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the application. Thus, appearances of the phrases "in one embodiment" or "in an embodiment" or "in some embodiments" or "in other embodiments" or "in still other embodiments" in various places throughout this specification are not necessarily all referring to the same embodiment, unless otherwise indicated. Furthermore, the terms "comprises," "comprising," "includes," "including," "has," "having" and the like are intended to be open-ended terms that do not exclude additional, unrecited elements or methods. Thus, the term "comprising" or "comprises" as used in the specification and claims does not exclude the presence of unrecited elements or steps. Furthermore, as used herein, the common abbreviation "e.g." will be understood to mean "for example."

[0045] In view of the low accuracy of the current high-speed train bearing fault diagnosis, the embodiment of the application provides a bearing fault diagnosis method based on a bidirectional Mamba network. The method extracts fault features by introducing a one-dimensional time sequence processing network and a two-dimensional CNN network, can fully capture the information of channel sequences and time sequences, introduces a time attention network between different channels to further strengthen the capture of channel sequence and time sequence information, ensures sufficient feature extraction, introduces a bidirectional Transformer network, uses channel-time double-direction convolution operation to extract the KQV of the multi-head attention mechanism, fully extracts the features of different channels and different times under the condition of ensuring that the channel information is not chaotic, introduces the bidirectional Mamba network, fills the information loss and ensures the unity of the information accuracy when the SSM models the long distance, ensures that the data in the original fault vibration signal can be fully extracted, and then the output network based on the fully extracted data can greatly improve the accuracy of fault diagnosis when performing fault diagnosis.

[0046] The bearing fault diagnosis method based on the bidirectional Mamba network provided by the application will be exemplarily described below in combination with specific embodiments.

[0047] As shown in Figure 1 The bearing fault diagnosis method based on the bidirectional Mamba network provided by the embodiment of the application comprises the following steps:

[0048] Step 11, collecting the fault vibration signal of the high-speed train bearing.

[0049] In some embodiments of the application, the above-mentioned fault vibration signal can be collected by a plurality of vibration sensors arranged at different positions of the high-speed train bogie. That is, the above-mentioned fault vibration signal is collected by a plurality of vibration sensors arranged on the high-speed train bogie.

[0050] It can be understood that, for the convenience of data processing, the fault vibration signal can be intercepted by using a fixed-length sliding window, and the subsequent processing is performed by using the intercepted data.

[0051] Step 12, inputting the fault vibration signal into a fault diagnosis model for analysis and processing to obtain a fault diagnosis result of the high-speed train bearing.

[0052] The above-mentioned fault diagnosis result can be that the high-speed train bearing has an outer ring crack, an outer ring pitting, a roller crack, etc.

[0053] It can be understood that the above fault diagnosis model is a trained model. That is, before using the fault diagnosis model for fault diagnosis, the model needs to be trained using training data. The training data contains sample data of different fault categories, which includes fault vibration signals of high-speed train bearings under different fault categories, and the fault category corresponding to each fault vibration signal. Specifically, the fault vibration signal can be non-repeatedly intercepted using a sliding window with a length of 1024, 80 samples are taken from each fault category, the total number of samples is 80*19, and the training set, test set and validation set are divided according to the ratio of 8:1:1. In some optional examples, the training of the fault diagnosis model can be completed in a supervised learning manner.

[0054] The specific structure of the above fault diagnosis model will be described below.

[0055] As shown in Figure 2 , the fault diagnosis model includes a two-dimensional CNN network, a time attention network, a bidirectional Transformer network, a one-dimensional time sequence processing network, a first fusion network based on channel attention, a data compression network, a bidirectional Mamba network, a second fusion network based on channel attention, and an output network connected in sequence.

[0056] Among them, the two-dimensional CNN network is used to extract the feature information of the fault vibration signal; the time attention network is used to process the feature information output by the two-dimensional CNN network to obtain attention feature information; the bidirectional Transformer network is used to extract features from the attention feature information in the time and sequence directions; the one-dimensional time sequence processing network is used to extract one-dimensional feature information of the features output by the bidirectional Transformer network; the first fusion network is used to fuse multiple one-dimensional feature information output by the one-dimensional time sequence processing network; the data compression network is used to compress the fusion result output by the first fusion network; the bidirectional Mamba network is used to extract features from the compressed data output by the data compression network; the second fusion network is used to fuse multiple features output by the bidirectional Mamba network; and the output network is used to map the fusion result output by the second fusion network to output the fault diagnosis result.

[0057] The above two-dimensional CNN network is mainly used for channel processing and preliminary information exchange of data (i.e. the fault vibration signal in step 11), and a multi-granularity two-dimensional CNN is used to extract features of each channel to output the feature information of the fault vibration signal. Specifically, the two-dimensional CNN network includes a plurality of pseudo two-dimensional convolution kernels and a first addition module, the input end of each pseudo two-dimensional convolution kernel receives the fault vibration signal, the output end of each pseudo two-dimensional convolution kernel is connected with the input end of the first addition module, the output end of the first addition module is connected with the input end of the time attention network, and the sizes of the plurality of pseudo two-dimensional convolution kernels are different from each other.

[0058] Pseudo two-dimensional convolution kernel: compared with a two-dimensional convolution kernel of 2*2, a two-dimensional feature can also be extracted by using 1*3, 3*1, 1*5, 5*1, 1*7, 7*1, 1*9, 9*1 convolution kernels of different sizes in pairs, but it is essentially the fusion of two one-dimensional feature extraction results. In order to prevent the mixing of features between channels, 1*3, 1*5, 1*7, 1*9 pseudo two-dimensional convolution kernels of four sizes can be used for extraction.

[0059] In some embodiments of the present application, the above plurality of pseudo two-dimensional convolution kernels receive the fault vibration signals of the high-speed train bearing in parallel, and respectively perform preliminary feature extraction on the fault vibration signals of the high-speed train bearing, so as to integrate multi-scale information through convolution kernels of different sizes. As an optional example, the number of pseudo two-dimensional convolution kernels can be four. The above first adding module is mainly used for adding the features extracted by the plurality of pseudo two-dimensional convolution kernels to obtain the feature information of the fault vibration signals.

[0060] The above time attention network is mainly used for extracting and uniformly processing the attention of different sequences at the same time and the attention of the same sequence at different times, so as to effectively capture the time sequence relationship of the same sequence and the dependency relationship between different sequences, thereby avoiding the confusion of sequence information. Specifically, the time attention network includes a first branch network for extracting attention information of different sequences at the same time, a second branch network for extracting attention information of the same sequence at different times, a first multiplication module, a second multiplication module and a second adding module. The input end of the first branch network and the input end of the second branch network both receive the feature information output by the two-dimensional CNN network, the output end of the first branch network and the output end of the second branch network are both connected with the input end of the first multiplication module, the output end of the first multiplication module is connected with the input end of the second multiplication module, the output end of the two-dimensional CNN network is connected with the input end of the second multiplication module, the output end of the second multiplication module and the output end of the two-dimensional CNN network are both connected with the input end of the second adding module, and the output end of the second adding module is connected with the input end of the bidirectional Transformer network.

[0061] The first branch network includes a first pooling layer and a first attention information extraction module connected in sequence, and the second branch network includes a second pooling layer and a second attention information extraction module connected in sequence. The functions of the first pooling layer and the second pooling layer are both to perform rotation operation on the input data.

[0062] Specifically, the feature information output by the two-dimensional CNN network is input into the first branch network and the second branch network in parallel, so that the first branch network extracts attention information of different sequences at the same time, and the second branch network extracts attention information of the same sequence at different times.

[0063] wherein the first attention information extraction module is configured to obtain the attention information A1 of the same time and different sequences by formula A1 = Sigmoid(Relu(W1*Data t )) and the second attention information extraction module is configured to obtain the attention information A2 of the same sequence and different times by formula A2 = Sigmoid(Relu(W2*Data c )). Sigmoid represents a sigmoid function, which compresses the dynamic range of the input activation vector to [0, 1], Relu represents a Relu function, Data t is the data output by the first pooling layer, W1 represents the weight of Data t , Data c is the data output by the second pooling layer, and W2 represents the weight of Data c . The pooling window size of the first pooling layer is 1*none, and the pooling window size of the second pooling layer is none*1. None is a placeholder, and usually represents that the size of this dimension is determined by the actual size of the input data.

[0064] Here, none represents that the size of the pooling window in this dimension is determined according to the actual size of the input data. Usually, none is used to represent an unfixed dimension, and the size of the other dimension is fixed or determined by design. For example, 100*100 enters (none, 1) and becomes 100*1.

[0065] The first multiplication module is mainly used for multiplying the attention information A1 of the same time and different sequences and the attention information A2 of the same sequence and different times, the second multiplication module is mainly used for multiplying the data output by the first multiplication module and the feature information output by the two-dimensional CNN network, and the second addition module is mainly used for adding the data output by the second multiplication module and the feature information output by the two-dimensional CNN network. Specifically, the expression of the attention feature information Out output by the time attention network is as follows:

[0066] Out = Data in +(A1*A2)*Data in

[0067] In the above formula, Data in is the feature information output by the two-dimensional CNN network.

[0068] To better capture the relationship between sequences in two-dimensional data, the multi-head attention mechanism of the above bidirectional Transformer network uses single-direction 2D convolution kernels of multiple sizes to extract time and sequence information respectively. Specifically, when extracting time information horizontally, a 1x3 convolution kernel is selected; when extracting sequence information vertically, a 3x1 convolution kernel is selected. Such a design ensures that time information and sequence information can be effectively extracted at multiple granularities, avoiding information mixing.

[0069] Among them, it needs to be pointed out that in the vertical information extraction, due to the close size of the convolution kernel and the number of channels, it will lead to the difference in the number of channel information contained in the data extracted by the convolution operation. For example, when applying a 5x1 convolution kernel to 8-channel data, part of the data contains 5-channel information, and another part contains only 3-channel information. In order to avoid this imbalance, a copy of the data is made and spliced in the channel dimension to ensure the consistency of vertical channel feature extraction.

[0070] Specifically, the feature extraction process of the bidirectional Transformer network is as follows:

[0071] Step 1.1, extract the KQV of the Transformer multi-head attention from the same sequence and different time (3*1, 5*1, 7*1) convolution group of the attention feature information output by the time attention network.

[0072] Step 1.2, extract the KQV of the Transformer multi-head attention from the same time and different sequence (1*3, 1*5, 1*7) convolution group of the attention feature information output by the time attention network.

[0073] Step 1.3, due to the fact that some data extracted by the 7*1 convolution in step 1.1 only contains 1 channel of data, and some contains 7 channels of data, the input data is spliced along the channel using the splicing method, and the data is changed from 9*1024 to 18*1024. Then use the convolution kernel to extract, and the first 9 rows of the obtained results are used as KQV.

[0074] Step 1.4: finally, all the results are summarized to obtain KQV with different channel combinations but the same number of channels:

[0075] K=K t1 +K t2 +K t3 +K c1 +K c2 +K c3

[0076] Q=Q t1 +Q t2 +Q t3 +Qc1 +Q c2 +Q c3

[0077] V = V t1 +V t2 +V t3 +V c1 +V c2 +V c3

[0078] ATT = softmax(K*Q)*V

[0079] In the above formula, K t1 , K t2 , K t3 , Q t1 , Q t2 , Q t3 , V t1 , V t2 , V t3 are extracted from the time direction (1*3, 1*5, 1*7) convolution kernel respectively, K c1 , K c2 , K c3 , Q c1 , Q c2 , Q c3 , V c1 , V c2 , V c3 are extracted from the channel (i.e. sequence) direction (3*1, 5*1, 7*1) convolution kernel respectively, softmax is a softmax function, and ATT is an “attention” matrix (Attention) which is used to determine how to weight the value (Value) matrix by calculating the similarity between the query (Query) and the key (Key). Here, ATT is the weighted value matrix after being processed by the softmax function.

[0080] The one-dimensional time sequence processing network is mainly used for extracting one-dimensional feature information. The two-dimensional one-dimensional fusion network is adopted to fully exert the advantages of the two-dimensional network in information interaction between channels and the powerful ability of the one-dimensional network in extracting single sequence time sequence relationship, so as to construct an efficient fault diagnosis network. The network first performs preliminary information preprocessing through the two-dimensional network to ensure sufficient exchange of information between channels and perform preliminary time sequence relationship analysis. Specifically, the one-dimensional time sequence processing network includes a plurality of one-dimensional convolution kernels, the input end of each one-dimensional convolution kernel receives the features output by the bidirectional Transformer network, the plurality of one-dimensional convolution kernels process the features output by the bidirectional Transformer network in parallel to obtain a plurality of one-dimensional feature information, the output end of each one-dimensional convolution kernel is connected with the input end of the first fusion network, and the sizes of the plurality of one-dimensional convolution kernels are different from each other. As an optional example, the number of one-dimensional convolution kernels can be 4.

[0081] In some embodiments of the present application, the first fusion network is mainly used to realize intelligent fusion of features. Specifically, the first fusion network fuses the received data (i.e., the plurality of one-dimensional feature information) by the following formula and outputs the fusion result:

[0082] s=Stack(input)

[0083] A c =Sigmoid(Relu(Maxpool(s)+Avgpool(s)))

[0084] stack out =A c *s

[0085] Out=sum(stack out ,dim=1)

[0086] Wherein, input represents the received data (i.e., the plurality of one-dimensional feature information), Stack represents the stacking operation, s represents the data obtained by stacking input along the channel dimension, Sigmoid represents the sigmoid function, Relu represents the Relu function, Maxpool represents the maximum pooling function, Avgpool represents the average pooling function, A c represents the channel attention coefficient, dim represents the dimension, sum represents the summation function, and Out represents the fusion result.

[0087] In some embodiments of the present application, the data compression network is mainly used to compress the fusion result output by the first fusion network by multiple times to obtain compressed data under multiple times.

[0088] The bidirectional Mamba network includes a plurality of bidirectional Mamba layers, and an output end of each bidirectional Mamba layer is connected with an input end of the second fusion network; the plurality of bidirectional Mamba layers are used for receiving compressed data under a plurality of multiples, and the plurality of bidirectional Mamba layers correspond to the compressed data under the plurality of multiples one by one, that is, the data compression network inputs the compressed data under different multiples into different bidirectional Mamba layers. As a preferred example, the number of bidirectional Mamba layers is 4, the data compression network respectively compresses the fusion results output by the first fusion network by 2 times, 4 times, 8 times and 16 times, and inputs the compressed data after 2 times, 4 times, 8 times and 16 times compression processing into the 4 bidirectional Mamba layers (each bidirectional Mamba layer only receives the compressed data after 1 type of multiple compression processing) respectively, so as to simultaneously obtain relatively complete and real data, and finally the second fusion network integrates the results of the 4 bidirectional Mamba layers as output. It is worth mentioning that using the bidirectional Mamba model (i.e. bidirectional Mamba layer) can comprehensively capture the time sequence characteristics in the forward and reverse directions, fuse and add the output data of the forward and reverse input sequences through parallel processing, so as to more accurately extract and output the comprehensive characteristics of the time sequence relationship.

[0089] The calculation formula for processing data by the bidirectional Mamba layer is:

[0090]

[0091] Wherein, Bm out is the output of the bidirectional Mamba layer, Mamba is the Mamba network model, X in is the data output by the data compression network to the bidirectional Mamba layer (such as compressed data after 2 times compression processing).

[0092] The second fusion network is mainly used for realizing intelligent fusion of features. Specifically, the second fusion network fuses the received data (i.e. the data output by the plurality of bidirectional Mamba layers) by the following formula and outputs the fusion result:

[0093] s=Stack(input)

[0094] A c =Sigmoid(Relu(Maxpool(s)+Avgpool(s)))

[0095] stack out =A c *s

[0096] Out=sum(stack out ,dim=1)

[0097] wherein, input represents the received data (i.e. the data output by the plurality of bidirectional Mamba layers), Stack represents a stacking operation, s represents the data obtained by stacking the input along the channel dimension, Sigmoid represents a sigmoid function, Relu represents a Relu function, Maxpool represents a max-pooling function, Avgpool represents an average-pooling function, A c represents a channel attention coefficient, dim represents a dimension, sum represents a summation function, and Out represents the fusion result.

[0098] Specifically, the process of processing data by the one-dimensional time series processing network, the first fusion network based on channel attention, the data compression network, the bidirectional Mamba network, and the second fusion network based on channel attention is as follows:

[0099] Step 2.1, dimension reduction is performed by multi-granularity one-dimensional convolution (3, 5, 7, 9), and then the information obtained by multi-granularity is merged by using a channel-based attention mechanism;

[0100] Step 2.2, the first fusion network splices the obtained four groups of one-dimensional data, then extracts attention for the stacked input according to different times of the same channel, multiplies the obtained result with the stacked input, sums the stacked input according to the channel dimension to obtain the final output, and uses residual connection to ensure the effect, and the sum of the four groups of data is input into the output as the residual connection;

[0101] Step 2.3, the data is processed by using a compression module, and the data is preliminarily compressed to 1 / 2 by using a convolution layer and a max-pooling layer, and then input into a multi-granularity bidirectional Mamba;

[0102] Step 2.4, the data is compressed by 2 times, 4 times, 8 times and 16 times respectively, and then input into four bidirectional Mambas, and the obtained four groups of one-dimensional data are input into the fusion network based on channel attention again to obtain the fusion result.

[0103] It is worth mentioning that the data is compressed by 2 times, 4 times, 8 times and 16 times respectively in a parallel manner, and input into four different bidirectional Mamba layers to simultaneously obtain relatively complete and real data, and finally the results are integrated as output. In order to comprehensively capture the time series features of the positive and negative directions, the bidirectional Mamba model is used, the output data of the positive and negative input sequences are fused and added by parallel processing, so as to more accurately extract and output the comprehensive features of the time series relationship, balance the accuracy of sequence analysis and the integrity of data, and enhance the adaptability and robustness of the model in practical application.

[0104] In some embodiments of the present application, the output network comprises, in sequence, an average pooling layer, a reshaping layer and a fully connected layer. The average pooling layer is used to perform a rotation operation on the fusion result output by the second fusion network, the reshaping layer is used to adjust the dimensions of the result output by the average pooling layer so as to be processed by the fully connected layer, and the fully connected layer is mainly used to map the data output by the reshaping layer to output the fault diagnosis result.

[0105] It is worth mentioning that, in the present application, the one-dimensional time sequence processing network and the two-dimensional CNN network are combined to extract fault features, which can fully capture the channel sequence and time sequence information; the time attention network is introduced to further strengthen the capture of channel sequence and time sequence information in the aspect of feature extraction between different channels, ensuring sufficient feature extraction; the bidirectional Transformer network is introduced to extract the KQV of the multi-head attention mechanism through channel-time double-direction convolution operation, fully extracting the features of different channels and different times while ensuring that the channel information is not chaotic; the bidirectional Mamba network is introduced to ensure the unity of information loss filling and information precision in long-distance modeling of the selective structured state space model (SSM) through parallel input data with different compression degrees, ensuring that the data in the original fault vibration signal can be fully extracted, and then the output network based on the fully extracted data can greatly improve the accuracy of fault diagnosis.

[0106] The bearing fault diagnosis method based on the bidirectional Mamba network of the present application will be described below in conjunction with specific examples.

[0107] The present application is applied to the bearing fault diagnosis experiment platform of the comprehensive test bed of the high-speed train bogie. Specifically applied to rolling bearing fault diagnosis, the present application uses the bearing data collected by the high-speed train bogie comprehensive test bed to verify the embodiments, takes NTN CRI-2692 double-row tapered roller bearing for high-speed trains as the research object, selects the outer ring, roller and retainer test pieces containing normal, crack and pitting and installs them on the entire rolling test bed. The test bed model is shown in Figure 3 , the sensor arrangement position is shown in Figure 4a and Figure 4b . Different types of bearing faults are installed in the test bed with sensors to collect data, 3 measuring points are arranged on each bearing, a total of 3 sensors, the motor speed is spaced at 500 RPM, the brake output is spaced at 20%, the vibration data of the test bed single drive system under the conditions of 1000-3000 RPM speed and 0-40% torque load are collected, all experiments use 12k sampling frequency, 2s sampling time, and three groups of samples are collected for each working condition.

[0108] Based on the collected bearing fault vibration signal dataset, the method is used for bearing fault diagnosis. First, the original fault signal of the bearing is obtained and pretreated (such as using fixed-length sliding window to intercept). Second, the two-dimensional CNN based on channel-time attention mechanism is used for feature extraction to capture the relationship of the signal in channel sequence and time sequence. Then, the bidirectional multi-granularity Transformer is used for channel-time sequence feature re-extraction to further strengthen the information of channel sequence and time sequence captured, ensuring sufficient feature extraction. Finally, the signal is input into the fusion channel attention bidirectional multi-granularity Mamba network for fault diagnosis under different compression multiples. Then, the ablation experiment is carried out, and the performance of each module is verified by gradually adding each module to the basic model.

[0109] In this experiment, a one-dimensional basic model is selected as the starting point, which mainly consists of a multi-modal convolution layer, a single compression process, and subsequent pooling and linear connection layers. On this basis, the performance of the model is expanded and verified by gradually introducing different modules to highlight the superiority of each module. These new modules include: bidirectional multi-granularity Transformer module (i.e. bidirectional Transformer network in the previous text), two-dimensional preprocessing (i.e. two-dimensional CNN network in the previous text) and channel-time attention mechanism module (i.e. time attention network in the previous text), and bidirectional multi-granularity Mamba module under channel fusion (i.e. bidirectional Mamba network in the previous text). The experiment is set to 100 iterations, with a batch size of 32 and a learning rate of 1e-5.

[0110] The final output is the accuracy of fault diagnosis, where a higher bearing fault diagnosis accuracy represents a more accurate diagnosis. The comparison results are shown in Table 1 and Figures 5a1 to 5e2 As shown in Table 1, ACC is the accuracy, and F1 is a comprehensive evaluation index of precision and recall.

[0111] Table 1 Ablation experiment test results of bearing fault diagnosis

[0112]

[0113] In summary, the bearing fault diagnosis method based on channel-time attention, bidirectional multi-granularity Transformer and combined with bidirectional multi-granularity Mamba model optimization provided in the application can sufficiently mine the existing data by introducing the channel-time attention mechanism and bidirectional multi-granularity Transformer under the condition of time-varying working conditions and limited data, and can optimize the time sequence and extraction ability through the bidirectional multi-granularity Mamba model, so as to solve the problem of low fault diagnosis accuracy caused by insufficient data extraction of the field generalization bearing fault diagnosis method.

[0114] The above is the preferred embodiment of the application. It should be pointed out that, for ordinary skilled persons in the technical field, a number of improvements and refinements can be made without departing from the principles described in the application, and these improvements and refinements should also be considered as the protection scope of the application.

Claims

1. A bearing fault diagnosis method based on a bidirectional Mamba network, characterized in that, include: Collect fault vibration signals of high-speed train bearings; The fault vibration signal is input into the fault diagnosis model for analysis and processing to obtain the fault diagnosis result of the high-speed train bearing. The fault diagnosis model includes a two-dimensional CNN network, a temporal attention network, a bidirectional Transformer network, a one-dimensional temporal processing network, a first fusion network based on channel attention, a data compression network, a bidirectional Mamba network, a second fusion network based on channel attention, and an output network, connected in sequence. The two-dimensional CNN network is used to extract feature information of the fault vibration signal. The temporal attention network is used to process the feature information output by the two-dimensional CNN network to obtain attention feature information. The bidirectional Transformer network is used to extract features from the attention feature information from both temporal and sequential directions. The one-dimensional temporal processing network is used to extract one-dimensional feature information of the features output by the bidirectional Transformer network. The first fusion network is used to fuse multiple one-dimensional feature information output by the one-dimensional temporal processing network. The data compression network is used to compress the fusion result output by the first fusion network. The bidirectional Mamba network is used to extract features from the compressed data output by the data compression network. The second fusion network is used to fuse multiple features output by the bidirectional Mamba network. The output network is used to map the fusion result output by the second fusion network and output the fault diagnosis result.

2. The bearing fault diagnosis method according to claim 1, characterized in that, The two-dimensional CNN network includes multiple pseudo-two-dimensional convolutional kernels and a first summing module. The input of each pseudo-two-dimensional convolutional kernel receives the fault vibration signal, and the output of each pseudo-two-dimensional convolutional kernel is connected to the input of the first summing module. The output of the first summing module is connected to the input of the temporal attention network. The sizes of the multiple pseudo-two-dimensional convolutional kernels are different from each other.

3. The bearing fault diagnosis method according to claim 1, characterized in that, The temporal attention network includes: a first branch network for extracting attention information from different sequences at the same time, a second branch network for extracting attention information from different times within the same sequence, a first multiplication module, a second multiplication module, and a second addition module; In this configuration, the input terminals of the first branch network and the second branch network both receive feature information output by the 2D CNN network. The output terminals of the first branch network and the second branch network are both connected to the input terminal of the first multiplication module. The output terminal of the first multiplication module is connected to the input terminal of the second multiplication module. The output terminal of the 2D CNN network is connected to the input terminal of the second multiplication module. The output terminals of the second multiplication module and the 2D CNN network are both connected to the input terminal of the second addition module. The output terminal of the second addition module is connected to the input terminal of the bidirectional Transformer network.

4. The bearing fault diagnosis method according to claim 3, characterized in that, The first branch network includes a first pooling layer and a first attention information extraction module connected in sequence, and the second branch network includes a second pooling layer and a second attention information extraction module connected in sequence. The first attention information extraction module is used to extract information through the formula A1 = Sigmoid(Relu(W1*Data). t We obtain attention information A1 for different sequences at the same time. The second attention information extraction module is used to extract information using the formula A2 = Sigmoid(Relu(W2*Data). c We obtain attention information A2 from different times within the same sequence. Where Sigmoid represents the sigmoid function, Relu represents the Relu function, and Data t The data output from the first pooling layer, W1 represents Data. t Weights, Data c The data output from the second pooling layer, W2 represents Data. c The weights are determined by the pooling window size of the first pooling layer (1*none) and the pooling window size of the second pooling layer (none*1), where none is a placeholder.

5. The bearing fault diagnosis method according to claim 1, characterized in that, The one-dimensional temporal processing network includes multiple one-dimensional convolutional kernels. The input of each one-dimensional convolutional kernel receives features output by the bidirectional Transformer network, and the output of each one-dimensional convolutional kernel is connected to the input of the first fusion network. The sizes of the multiple one-dimensional convolutional kernels are different from each other.

6. The bearing fault diagnosis method according to claim 1, characterized in that, Both the first fusion network and the second fusion network fuse the received data using the following formula and output the fusion result: s = Stack(input) A c =Sigmoid(Relu(Maxpool(s)+Avgpool(s))) stack out =A c *s Out=sum(stack out ,dim=1) Where, input represents the received data, Stack represents the stacking operation, s represents the data obtained by stacking input along the channel dimension, Sigmoid represents the sigmoid function, ReLU represents the ReLU function, Maxpool represents the max pooling function, and Avgpool represents the average pooling function. c represents the channel attention coefficient, dim represents the dimension, sum represents the summation function, and Out represents the fusion result.

7. The bearing fault diagnosis method according to claim 1, characterized in that, The data compression network is used to compress the fusion result output by the first fusion network by multiple factors to obtain compressed data at multiple factors.

8. The bearing fault diagnosis method according to claim 7, characterized in that, The bidirectional Mamba network includes multiple bidirectional Mamba layers, and the output of each bidirectional Mamba layer is connected to the input of the second fusion network. The plurality of bidirectional Mamba layers are used to receive compressed data at the various compression ratios, and the plurality of bidirectional Mamba layers correspond one-to-one with the compressed data at the various compression ratios.

9. The bearing fault diagnosis method according to claim 1, characterized in that, The output network comprises an average pooling layer, a reshaping layer, and a fully connected layer connected in sequence.

10. The bearing fault diagnosis method according to claim 1, characterized in that, The fault vibration signal was collected by multiple vibration sensors installed on the bogie of the high-speed train.

Images