Cross-working-condition bearing fault diagnosis method and device based on multi-scale convolutional fuzzy neural network migration model

Abstract

The invention discloses a cross-working-condition bearing fault diagnosis method and device based on a multi-scale convolutional fuzzy neural network migration model, and the method comprises the steps: inputting a vibration signal of a to-be-diagnosed bearing into a trained cross-working-condition bearing fault diagnosis model, and outputting a diagnosis result; wherein the trained cross-working-condition bearing fault diagnosis model is obtained by training a multi-scale convolutional fuzzy neural network migration model by adopting a training data set, and a total loss function in the training process comprises classification loss guided by error samples, depth domain adaptive loss and confusion loss of an adversarial network domain; the training data set comprises a vibration signal from the bearing under the source domain working condition and a corresponding fault label, and a vibration signal from the bearing under the target domain working condition; the migration model based on the multi-scale convolutional fuzzy neural network comprises a multi-scale convolutional neural network module, a gated recursive convolutional attention module, an improved adaptive fuzzy inference system and an adversarial network module which are connected in sequence. The objective of the invention is to solve the problem of poor bearing fault diagnosis accuracy under variable working condition migration conditions.

Description

Technical Field

[0001] The present invention belongs to the technical field of mechanical equipment fault diagnosis, and particularly relates to a cross-condition bearing fault diagnosis method and device based on a multi-scale convolutional fuzzy neural network transfer model. Background Art

[0002] In high-speed rotating mechanical equipment such as wind turbines, bearings are an essential component for the mechanical equipment to achieve rotational, swinging and other motion modes, and are widely used in various large mechanical equipment such as high-speed railways, automobiles, large machine tools, aero-engines, etc. Bearing faults will cause wear of equipment components, affect the service life of the equipment, and even lead to equipment damage and scrapping in severe cases, posing potential safety hazards to personnel. Their operating conditions often directly determine the performance of the entire equipment. Therefore, it is very crucial to carry out accurate intelligent diagnosis for such major components. Due to its deep and stackable network structure, the diagnostic method based on deep learning has stronger capabilities in the field of mining and analyzing implicit features of complex signals, and has gradually become a research upsurge in the field of fault diagnosis. Typical deep learning models include convolutional neural network (CNN), residual neural network (ResNet), stacked autoencoder (SAE), etc. Most of the above deep learning models are encoding-decoding architectures based on feature extraction and classification, and adopt the combination of convolutional layers and fully connected classification networks to obtain accurate detection effects for specific online and offline detection tasks. However, the feature recognition and classification capabilities of the fully connected layer are limited, and often require a powerful feature extractor to achieve better performance, which comes at the cost of longer training time and greater parameter storage consumption. More critically, these models are trained for a specific diagnostic scenario, and there will inevitably be a significant reduction in diagnostic performance when facing transfer diagnostic tasks under different working conditions.

[0003] In view of the problems of low transfer diagnosis accuracy and poor model generalization performance of deep learning models under variable working conditions, in recent years, scholars have carried out a large amount of research work to solve the problem of different distributions between source domain data and target domain data. These methods can be roughly divided into (1) methods based on network structure changes and (2) methods based on measuring domain distribution differences. However, these methods often have limited effectiveness. When the span of working conditions increases, resulting in large domain distribution differences and complex boundary distributions, the above methods often cannot achieve very ideal diagnostic results. For example, the patent with the publication number CN110751207A and the invention name of a fault diagnosis method based on deep convolutional domain adversarial transfer learning is suitable for dataset samples with obvious differentiation between different working conditions and very standard fault settings. When the data acquisition working conditions show strong time-varying characteristics while the fault features are not obvious, the above method will inevitably experience a decline in diagnostic performance, thus showing shortcomings in the application level facing the actual industrial environment. Summary of the Invention

[0004] In view of the problems existing in the prior art, the present invention provides a cross-condition bearing fault diagnosis method and device based on a multi-scale convolutional fuzzy neural network transfer model, aiming to solve the problem of poor accuracy of bearing fault diagnosis under variable-condition transfer conditions.

[0005] To solve the above technical problems, the present invention is realized through the following technical solutions:

[0006] According to the first aspect of the present invention, there is provided a cross-condition bearing fault diagnosis method based on a multi-scale convolutional fuzzy neural network transfer model, including:

[0007] Obtain the vibration signal of the bearing to be diagnosed;

[0008] Input the vibration signal of the bearing to be diagnosed into the trained cross-condition bearing fault diagnosis model, and output a diagnosis result; wherein, the trained cross-condition bearing fault diagnosis model is obtained by training a multi-scale convolutional fuzzy neural network transfer model using a training data set, and the total loss function in the training process includes a classification loss guided by wrong samples, a deep domain adaptation loss, and an adversarial network domain confusion loss; the training data set includes the vibration signals and corresponding fault labels of bearings under source domain conditions, and the vibration signals of bearings under target domain conditions; the multi-scale convolutional fuzzy neural network transfer model includes a multi-scale convolutional neural network module, a gated recurrent convolutional attention module, an improved adaptive fuzzy inference system, and an adversarial network module connected in sequence.

[0009] In a possible implementation manner of the first aspect, the multi-scale convolutional neural network module uses convolutional-pooling layers with multiple different convolutional kernel sizes, and the different convolutional-pooling layers are connected in a sequential connection and a parallel connection manner for multi-scale feature extraction of vibration signals, specifically as follows:

[0010]

[0011] Among them, x is the training sample input in batches; B×K×N respectively represent the batch size of the input sample, the number of feature maps input

[0012] And the input feature length; ConvPooling i (·) and ConvPooling j 3 (·) respectively represent the 1st to 3rd and 4th convolutional-pooling operations; is the output of the first 3 convolutional-pooling operations, where i indicates the operation round; represents the input of the 4th convolutional-pooling operation, which is the output after the previous convolutional-pooling The number of representations at different scales obtained after segmentation by the power series length of 2 is m; It is the different-length signal feature segments obtained by convolving with convolutional kernels of different scales in the 4th convolution-pooling operation, and finally the final output y of the multi-scale convolutional neural network module obtained through the feature concatenation operation Concat(·) MC .

[0013] In a possible implementation manner of the first aspect, the gated recurrent convolutional attention module is a recurrent convolutional module with residual connections, including a convolution-pooling projection layer for feature projection, an adaptive gated recurrent convolutional layer, and a fully connected layer for signal scale restoration;

[0014] For the input feature with a batch size of B, the number of feature maps of K, and the feature length of N First, a convolution-pooling operation of different scales is performed to establish the mapping relationship of the input feature in different spatial dimensions:

[0015]

[0016] In the formula, MConv(·) represents the convolution operation of different scales; is the output corresponding to the convolution operation of different scales, and its size is a geometric sequence arranged in a series of 2 AvgPooling(·) represents the average pooling operation; is the feature output of different scales obtained through this convolution-pooling layer;

[0017] Then, each feature segment continuously sums with the recurrent convolution result of a longer size in the way of element-wise multiplication, so as to realize the recurrent convolution operation:

[0018]

[0019] In the formula, ⊙ represents the element-wise multiplication operation; {f t (·)} t＝0,1,...n is a set of recurrent convolution operations with continuously increasing depth;

[0020] {g t (·)} t＝0,1,...n is the corresponding feature size matching function; α is a scalar parameter used to stabilize the training process;

[0021] Finally, all the output results obtained by the recurrent convolution are connected through the Concat(·) operation, and an output y with the same feature dimension as the input gated recurrent convolutional attention module is obtained through a fully connected layer Proj_out , as follows:

[0022]

[0023] In the formula, p 0 , p 1 ,..., p n represents the multi-level output of the recursive convolution; Linear(·) represents the linear connection layer that finally obtains the output of the module.

[0024] In a possible implementation of the first aspect, the improved adaptive fuzzy inference system includes a plurality of parallel adaptive pruning fuzzy inference structures. The adaptive pruning fuzzy inference structure is based on fuzzy logic and includes a fuzzification layer, a fuzzy rule layer, a normalization layer, and a defuzzification layer;

[0025] The fuzzification layer is used to perform fuzzy mapping on the input features, mapping the input features to fuzzy values on different fuzzy sets, which are called membership degrees:

[0026]

[0027] Among them, is the membership degree of the jth fuzzy set belonging to the input of the ith channel, and its value is between 0 and 1; MF j (·) is the membership function of the jth fuzzy set, and a Gaussian function is selected as the measurement standard for the membership degree; μ j and σ j are the parameters to be trained in the membership function of the jth fuzzy set; is the data on the ith channel of the fuzzification layer input x F ,

[0028] The fuzzy rule layer calculates the triggering intensity of each rule from the membership degree; for the kth rule R k in the fuzzy rule base, its expression is as follows:

[0029] R k : IF x 1 is A 1j , and x 2 is A 2j ,... and x N is A Nj , THEN Y k = f k (x 1 , x 2 , …, x N )

[0030] Among them, A Nj is the jth fuzzy set in the universe of discourse, which is a fuzzy clustering determined by the fuzzy input distribution. The universe of discourse is defined as two fuzzy sets with positive and negative dual properties: and where Ξ pos represents a positive fuzzy set that contains the membership degrees of the data of each channel after fuzzification corresponds to a relatively large activation degree of the neuron, while Ξ neg represents a negative fuzzy set that contains the membership degrees of the data of each channel after fuzzification corresponds to an activation degree smaller than the former; f k (·) is the output form of the rule, taking f k (·) as a first-order polynomial about the input x F of the fuzzification layer; the triggering strength of each fuzzy rule is calculated by the S-norm regularization method, and the output of the fuzzy rule layer can be expressed as the triggering strength values of each fuzzy rule, as shown in the following formula:

[0031]

[0032] where, r k is the triggering strength of the k-th fuzzy rule, represents the membership degrees of the input of the i-th channel on the positive and negative fuzzy sets, and its different permutation orders correspond to different fuzzy rules;

[0033] The normalization layer is used to calculate the normalized values of the triggering strengths of the rules derived from the fuzzy rule layer; for the normalized triggering strength corresponding to a certain fuzzy rule, its calculation method is as follows:

[0034]

[0035] where, S k is the normalized triggering strength of the k-th fuzzy rule after the normalization operation, and z is the number of fuzzy rules; the defuzzification layer finally obtains the final output of the adaptive pruning fuzzy inference structure by calculating the product of the normalized triggering strength of each rule and the rule output value, as shown in the following formula:

[0036]

[0037] where, y out is the final output result, Y k = f k (x 1 , x 2 , …, x N ) is the output value of the k-th rule, and Linear(·) is a linear layer used to map the obtained result to the final output dimension.

[0038] In a possible implementation of the first aspect, the adversarial network module is a multi-layer perceptron structure composed of three fully-connected layers, which is used to confuse the source domain and the target domain from which the features are extracted, so as to extract domain-invariant features of the signal.

[0039] In a possible implementation of the first aspect, the total loss function of the training process is:

[0040] L ttl = L JMMD (θ f , θ y ) + L d (θ f , θ d ) + L Focal (θ f , θ y )

[0041]

[0042] In the formula, L ttl is the total loss of the training process; L Focal (θ f , θ y ) is the classification loss guided by misclassified samples; L JMMD (θ f , θ y ) is the deep domain adaptation loss; L d (θ f , θ d ) is the adversarial network domain confusion loss; θ f is the trainable parameter of the feature extractor G f composed of a multi-scale convolutional neural network module and a gated recurrent convolutional attention module; θ y is the trainable parameter of the feature classifier G y composed of an improved adaptive fuzzy inference system; θ d is the trainable parameter in the domain discriminator G d composed of an adversarial network module; n s and n t represent the number of source domain and target domain samples respectively; K s represents the number of classification labels; (1 - y) γ is a modulation factor controlled by the adjustable focusing parameter γ ≥ 0; k(·) is a Gaussian kernel function; and represent the samples on the source domain and the target domain respectively; and are the intermediate features output by the feature extractor G f on the source domain and the target domain respectively; and They are the final output results in the source domain and the target domain respectively.

[0043] According to a second aspect of the present invention, there is provided a cross-condition bearing fault diagnosis device based on a multi-scale convolutional fuzzy neural network transfer model, comprising:

[0044] An acquisition module, configured to acquire the vibration signal of the bearing to be diagnosed;

[0045] A diagnosis module, configured to input the vibration signal of the bearing to be diagnosed into the trained cross-condition bearing fault diagnosis model, and output a diagnosis result; wherein, the trained cross-condition bearing fault diagnosis model is obtained by training a multi-scale convolutional fuzzy neural network transfer model using a training data set, and the total loss function of the training process includes a classification loss guided by wrong samples, a deep domain adaptation loss, and an adversarial network domain confusion loss; the training data set includes the vibration signals of the bearings under the source domain conditions and the corresponding fault labels, as well as the vibration signals of the bearings under the target domain conditions; the multi-scale convolutional fuzzy neural network transfer model includes a multi-scale convolutional neural network module, a gated recurrent convolutional attention module, an improved adaptive fuzzy inference system, and an adversarial network module connected in sequence.

[0046] According to a third aspect of the present invention, there is provided a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the cross-condition bearing fault diagnosis method based on a multi-scale convolutional fuzzy neural network transfer model as described above.

[0047] According to a fourth aspect of the present invention, there is provided a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, it implements the cross-condition bearing fault diagnosis method based on a multi-scale convolutional fuzzy neural network transfer model as described above.

[0048] According to a fifth aspect of the present invention, there is provided a computer program product, and when the computer program product is executed by a processor, it implements the cross-condition bearing fault diagnosis method based on a multi-scale convolutional fuzzy neural network transfer model as described above.

[0049] Compared with the prior art, the present invention has at least the following beneficial effects:

[0050] A cross-condition bearing fault diagnosis method based on a multi-scale convolutional fuzzy neural network transfer model provided by the present invention combines the advantages of multi-scale, recursive convolution, and fuzzy inference to improve the accuracy and adaptability of fault diagnosis for the problems of high accuracy and cross-condition transfer diagnosis in the bearing fault diagnosis of rotating machinery. Specifically, first, a multi-scale convolutional neural network module is used to extract features of the input signal at different levels; then, based on the gated recursive convolutional attention module, the expressions of the obtained features at different order levels are used to fully express the potential information of the fault; subsequently, based on an improved adaptive fuzzy inference system, the knowledge representation and decision inference in the fuzzy inference process are optimized, and the fault classification ability of the model in a complex and uncertain environment is enhanced. By constructing a hybrid model combining a multi-scale convolutional network and the fuzzy inference system, the present invention can achieve the adaptive fusion of deep features and fuzzy knowledge, thereby improving the transfer ability of the model under variable working conditions. Through adversarial training of the data in the source domain and the target domain, the cross-domain adaptability of the features is further enhanced, enabling the model to achieve accurate fault diagnosis under different working conditions. Compared with traditional deep learning methods, the present invention can effectively improve the diagnosis accuracy under variable working conditions and has important industrial application value.

[0051] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following specifically enumerates preferred embodiments and, in conjunction with the accompanying drawings, makes a detailed description as follows. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the specific embodiments of the present invention, the following will briefly introduce the drawings required for use in the description of the specific embodiments. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0053] Figure 1 It is a flowchart of a cross-condition bearing fault diagnosis method based on a multi-scale convolutional fuzzy neural network transfer model of the present invention;

[0054] Figure 2 It is the overall structure diagram of a multi-scale convolutional fuzzy neural network transfer model in a cross-condition bearing fault diagnosis method based on a multi-scale convolutional fuzzy neural network transfer model of the present invention;

[0055] Figure 3 It is the specific structure of a multi-scale convolutional neural network module in a multi-scale convolutional fuzzy neural network transfer model of the present invention;

[0056] Figure 4 It is the specific structure of a gated recursive convolutional attention module in a multi-scale convolutional fuzzy neural network transfer model of the present invention;

[0057] Figure 5 This is the specific structure of the improved adaptive fuzzy inference system in the multi-scale convolutional fuzzy neural network transfer model of the present invention;

[0058] Figure 6 This is the inference process of the adaptive pruning fuzzy inference structure proposed in the multi-scale convolutional fuzzy neural network transfer model of the present invention;

[0059] Figure 7 This is the classification result of the time-varying speed fan gearbox bearing dataset of Xi'an Jiaotong University in the embodiments of the present invention;

[0060] Figure 8 This is the confusion matrix analysis of the time-varying speed fan gearbox bearing dataset of Xi'an Jiaotong University in the embodiments of the present invention;

[0061] Figure 9 This is the t-SNE visualization result of the time-varying speed fan gearbox bearing dataset of Xi'an Jiaotong University in the embodiments of the present invention. Detailed implementation manners

[0062] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Apparently, the described embodiments are some but not all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0063] As Figure 1 shown, the embodiments of the present invention provide a cross-condition bearing fault diagnosis method based on a multi-scale convolutional fuzzy neural network transfer model, which specifically includes the following steps:

[0064] Step 1, obtain the vibration signal of the bearing to be diagnosed. Specifically, based on sensors, the vibration signals of the bearings of rotating mechanical equipment are collected.

[0065] Step 2, input the vibration signal of the bearing to be diagnosed into the trained cross-condition bearing fault diagnosis model, and output a diagnosis result. Among them, the trained cross-condition bearing fault diagnosis model is obtained by training a multi-scale convolutional fuzzy neural network transfer model using a training dataset, and the details are as follows.

[0066] The total loss function in the training process includes a classification loss guided by wrong samples, a deep domain adaptation loss, and an adversarial network domain confusion loss. Specifically, the total loss function in the training process is:

[0067] L ttl = L JMMD (θf , θ y ) + L d (θ f , θ d ) + L Focal (θ f , θ y )

[0068] Specifically, the classification loss L Focal (θ f , θ y ) is obtained by adding a penalty coefficient term guided by misclassified samples to the cross-entropy loss function, which is used to strengthen the learning of misclassified samples, so as to achieve better learning results. For the cross-entropy loss function:

[0069]

[0070] Among them, is a 0-1 discriminant function used to judge whether the classification is correct;

[0071] For the classification loss function guided by misclassified samples:

[0072]

[0073] Specifically, the deep domain adaptation loss L JMMD (θ f , θ y ) is the total loss of the joint maximum mean discrepancy domain adaptation distribution based on features and classification results, which can be expressed by the following formula:

[0074]

[0075] Specifically, the adversarial network domain confusion loss L d (θ f , θ d ) is the cross-entropy loss function based on domain classification, and its expression is as follows:

[0076]

[0077] In the formula, L ttl is the total loss during the training process; L Focal (θ f , θ y ) is the classification loss guided by misclassified samples; L JMMD (θ f , θ y ) is the deep domain adaptation loss; L d (θ f , θ d ) is the adversarial network domain confusion loss; θ fThe feature extractor G composed of a multi-scale convolutional neural network module and a gated recurrent convolutional attention module f The training parameters of; θ y The feature classifier G composed of an improved adaptive fuzzy inference system y The training parameters of; θ d The domain discriminator G composed of an adversarial network module d The training parameters in; n s And n t Respectively represent the number of source domain and target domain samples; K s Represents the number of classification labels; (1 - y) γ The modulation factor controlled by the adjustable focusing parameter γ ≥ 0. Exemplarily, γ is set to 2; k(·) is the Gaussian kernel function; And Respectively represent the samples on the source domain and the target domain; And Are respectively the intermediate features output by the feature extractor G on the source domain and the target domain f Output intermediate features; And Are respectively the final output results on the source domain and the target domain.

[0078] Specifically, the training data set includes the vibration signals of bearings under source domain working conditions and the corresponding fault labels, as well as the vibration signals of bearings under target domain working conditions. Specifically, based on the sensors and data acquisition system, the vibration signals of the bearings of rotating machinery are collected under different working conditions. Among them, the collected signal samples with labels can be regarded as source domain data because they are completely labeled and easy to obtain; while the collected signal samples without labels can be regarded as target domain data because of their lack of labeling characteristics, which often conform to the characteristics of the signal acquisition process under actual working conditions.

[0079] It should also be noted that data preprocessing needs to be performed on the source domain and target domain signal samples. By means of random sliding window sampling or equidistant sliding window sampling, a sample set that meets the length of the input cross-condition bearing fault diagnosis model is obtained while realizing sampling data enhancement, ensuring the feature adaptability and transferability under different working conditions.

[0080] Specifically, as shown in combination with Figures 2 to 6 The multi-scale convolutional fuzzy neural network transfer model includes a multi-scale convolutional neural network module, a gated recurrent convolutional attention module, an improved adaptive fuzzy inference system, and an adversarial network module connected in sequence. Among them:

[0081] The multi-scale convolutional neural network module adopts convolutional-pooling layers with multiple different convolutional kernel sizes, and sequential connection and parallel connection are used between different convolutional-pooling layers for multi-scale feature extraction of vibration signals, specifically as follows:

[0082]

[0083] Among them, x is the training sample input in batches; B×K×N respectively represent the batch size of the input sample, the number of feature maps input

[0084] and the input feature length; ConvPooling i (·) and ConvPooling j 3 (·) respectively represent the 1st to 3rd and the 4th convolutional-pooling operations; is the output of the first 3 convolutional-pooling operations, where i indicates the operation round; represents the input of the 4th convolutional-pooling operation, which is the output after the previous convolutional-pooling The representations at different scales obtained after being segmented by the length of the power series of 2, and the number of them is m; are the signal feature segments of different lengths obtained by convolving with convolutional kernels of different scales in the 4th convolutional-pooling operation, and finally the final output y of the multi-scale convolutional neural network module is obtained through the feature concatenation operation Concat(·) MC .

[0085] Exemplarily, as shown by Figure 3 , the multi-scale convolutional neural network module is a multi-scale convolutional feature extractor, which consists of 3 convolutional-pooling block structures for feature extraction of input signals. The design of the convolutional kernel size follows the design criterion of gradually decreasing from large to small, and feature fusion and concatenation are added to comprehensively extract and express the potential features of the sample from the perspectives of multi-scale and multi-resolution.

[0086] The gated recurrent convolutional attention module is a recurrent convolutional module with residual connections, including a convolutional-pooling projection layer for feature projection, an adaptive gated recurrent convolutional layer, and a fully connected layer for signal scale restoration;

[0087] For the input feature with a batch size of B, the number of feature maps of K, and the feature length of N First, a convolutional-pooling operation with different scales is performed to establish the mapping relationship of the input feature in different spatial dimensions:

[0088]

[0089] In the formula, MConv(·) represents the convolutional operation with different scales; The output corresponding to the convolution operation of different scales is a geometric progression with a size of 2. AvgPooling(·) represents the average pooling operation; are the feature outputs of different scales obtained through the convolution-pooling layer;

[0090] Then, each feature fragment is continuously summed with the recursive convolution result of a longer size by element-by-element multiplication, thereby implementing the recursive convolution operation:

[0091]

[0092] Where ⊙ represents the element-by-element multiplication operation; {f t (·)} t＝0,1,...n is a set of recursive convolution operations with increasing depth;

[0093] {g t (·)} t＝0,1,...n is the corresponding feature size matching function; α is a scalar parameter used to stabilize the training process;

[0094] Finally, all the output results obtained by the recursive convolution are connected through the Concat(·) operation and passed through a fully connected layer to obtain the output y with the same feature dimension as the input gated recursive convolution attention module. Proj_out , as shown below:

[0095]

[0096] In the formula, p 0 ,p 1 ,...,p n Represents the multi-level output of recursive convolution; Linear(·) represents the linear connection layer that finally obtains the module output.

[0097] Through gated convolution and recursive design, the spatial dimension of the signal is linked to the feature order, so that the potential features of multiple spatial scales and high-order convolution scales of the input signal can be deeply mined without introducing a lot of additional calculations.

[0098] The improved adaptive fuzzy inference system includes a plurality of parallel adaptive pruning fuzzy inference structures, wherein the adaptive pruning fuzzy inference structure is based on the fuzzy logic concept and includes a fuzzification layer, a fuzzy rule layer, a normalization layer and a defuzzification layer;

[0099] The fuzzification layer is used to perform fuzzy mapping on the input features, mapping the input features to fuzzy values ​​on different fuzzy sets, which are called membership degrees:

[0100]

[0101] Among them, is the membership degree of the j-th fuzzy set belonging to the input of the i-th channel, and its value is between 0 and 1; MF j (·) is the membership function of the j-th fuzzy set, and a Gaussian function is selected as the measurement standard for the membership degree; μ j and σ j are the parameters to be trained in the membership function of the j-th fuzzy set; is the data on the i-th channel of the input x F to the fuzzification layer,

[0102] The fuzzy rule layer calculates the triggering strength of each rule from the membership degree (the fuzzy output completes the inference operation of the membership degree of the original input through a layer of fuzzy rule layer, and the basis for the inference is the pre-initialized fuzzy rule base, that is, the triggering strength of each rule is calculated from the membership degree); for the k-th rule R k in the fuzzy rule base, its expression is as follows:

[0103] R k : IF x 1 is A 1j , and x 2 is A 2j ,... and x N is A Nj , THEN Y k = f k (x 1 , x 2 , …, x N )

[0104] Among them, A Nj is the j-th fuzzy set in the universe of discourse, which is a fuzzy clustering determined by the fuzzy input distribution. The universe of discourse is defined as two fuzzy sets with positive and negative dual properties: and Among them, Ξ pos represents the positive fuzzy set, and this set contains the membership degrees of the data of each channel after fuzzification corresponding to a relatively large neuron activation degree, while Ξ neg represents the negative fuzzy set, and this set contains the membership degrees of the data of each channel after fuzzification corresponding to an activation degree less than the former; f k (·) is the output form of the rule, and take f k (·) as a first-order polynomial about the input x F to the fuzzification layer; the triggering strength of each fuzzy rule is calculated by the S-norm regularization method, then the output of the fuzzy rule layer can be expressed as the triggering strength values of each fuzzy rule, as shown in the following formula:

[0105]

[0106] Among them, r k is the triggering intensity of the k-th fuzzy rule, represents the membership degrees of the input of the i-th channel on the positive and negative fuzzy sets, and different permutation orders correspond to different fuzzy rules;

[0107] The normalization layer is used to calculate the normalized values of the triggering intensities of the rules derived by the fuzzy rule layer; for the normalized triggering intensity corresponding to a certain fuzzy rule, its calculation method is as follows:

[0108]

[0109] Among them, S k is the normalized triggering intensity of the k-th fuzzy rule after the normalization operation, and z is the number of fuzzy rules; the defuzzification layer finally obtains the final output of the adaptive pruning fuzzy inference structure by calculating the product of the normalized triggering intensity of each rule and the rule output value, as shown in the following formula:

[0110]

[0111] Among them, y out is the final output result, Y k = f k (x 1 , x 2 ,..., x N ) is the output value of the k-th rule, and Linear(·) is a linear layer used to map the obtained result to the final output dimension.

[0112] That is to say, as shown by Figure 5 and Figure 6 , the improved adaptive fuzzy inference system includes multiple parallel adaptive pruning fuzzy inference structures, which are used as feature classifiers to establish the final mapping of multi-scale features. The adaptive pruning fuzzy inference structure is based on the fuzzy logic idea and mainly consists of four network layers, including the fuzzification layer, the fuzzy rule layer, the normalization layer and the defuzzification layer, and finally obtains the output of the fuzzy inference structure.

[0113] The adversarial network module is a multi-layer perceptron structure composed of 3 fully connected layers, which is used to confuse the source domain and the target domain sources of the extracted features in order to extract the domain-invariant features of the signals. Specifically, each fully connected layer uses the ReLU function as the activation function for non-linear mapping. By setting a gradient reversal layer during the gradient propagation process, it is used to confuse the source domain and the target domain sources of the extracted features, and at the same time promote the learning of the domain-invariant features by the aforementioned multi-scale convolutional neural network.

[0114] The training process of the cross-condition bearing fault diagnosis model is as follows:

[0115]

[0116] Among them, the feature extractor G f , the feature classifier G y and the domain discriminator G d respectively refer to the multi-scale convolutional neural network module, the gated recurrent convolutional attention module, the adaptive fuzzy inference system and the adversarial network module introduced above.

[0117] In the following specific implementation cases, the experimental results for the time-varying speed condition scenarios show that a multi-scale convolutional fuzzy neural network transfer fault diagnosis method for cross-condition conditions provided by the present invention can achieve the best diagnosis effect based on the variable-condition scenarios under different time-varying speed schemes, providing an effective solution for the application of intelligent diagnosis methods in actual transfer industrial environments.

[0118] Specifically, it is verified using the fan gearbox bearing dataset obtained from the variable-speed fan drive system fault experiment platform of Xi'an Jiaotong University.

[0119] Step 1: Data acquisition. The vibration signals collected by the vibration acceleration sensors installed on the gearbox are used for the verification experiment of the method. A total of five types of vibration signals of bearings in different health states, namely ball fault (BF), cage fault (CF), inner ring fault (IF), normal (NS) and outer ring fault (OF), are collected in the experiment. The experimental speed schemes include a linearly increasing type (S 1 ), a long-period sinusoidal change type (S 2 ) and a short-period sinusoidal change type (S 3 ), and the speed magnitude change range is between 10 Hz and 34 Hz.

[0120] Step 2: Data preprocessing. The dataset is divided and enhanced using the sliding window sampling method. Each type of fault contains 100 samples, of which 70 samples are divided into the training set and 30 samples are the verification set. Each sample uses the time-domain signal of the drive end channel as the input, which contains 2048 sampling points. Six types of transfer tasks are set based on different variable-speed schemes, as shown in the following table:

[0121] Table 1 Dataset division and transfer task settings

[0122]

[0123]

[0124] Step 3: Build the model, and its specific parameters are shown in the following table:

[0125] Table 2 Specific Parameter Table of the Model

[0126]

[0127] Step 4: Train the established model based on the divided dataset. During the training process described in the example, the Adaptive Gradient Descent Algorithm (Adam) is used to optimize the model parameters. The number of training rounds is set to 100 times, and the performance on the test set after training is taken as the measurement standard for the final diagnostic accuracy of the model.

[0128] To prove the superiority of the method proposed in the present invention, five types of models widely used in the field of transfer learning are introduced for comparison, including: WDCNN, DANN, DDC, D-CORAL, and ResNet-18 models. At the same time, in addition to the diagnostic accuracy rate, recall rate (Recall), precision rate (Precision), and F1-score are introduced as evaluation indicators to measure the effectiveness of the model from multiple perspectives. The comparative analysis of the six types of models is as follows:

[0129] (1) As Figure 7 shown in Table 3, it is obvious that the model can achieve the highest average accuracy rate on different transfer tasks, with an average of 76.62%, and the diagnostic accuracy is superior to other models in the same field in most cases.

[0130] Table 3 Diagnostic Results of Different Transfer Tasks

[0131]

[0132]

[0133] (2) Taking the first type of transfer task as an example, analyze the domain adaptation ability of each method. Perform t-distributed Stochastic Neighbor Embedding (t-SNE) visualization processing. The feature distribution is as Figure 8 shown. It can be seen that the model well demonstrates the final classification of the vibration signals of bearings in 5 different health conditions, and good classification effects can be achieved for each type of fault. In contrast, as can be seen from Figure 8 (a)-(e), although other methods can distinguish the fault features of some categories to a certain extent, the discrimination degree is still limited, resulting in many misclassifications. Therefore, the above results prove the strong classification recognition and domain adaptation ability of the method of the present invention.

[0134] (3) Based on the same task, use the confusion matrix to further observe the classification results as Figure 9As shown. It can be seen that the comparison methods have different degrees of classification errors when identifying different fault categories. These methods fluctuate greatly and there is a certain gap from the migration results of the model. In contrast, the present invention has obvious advantages, achieving higher classification accuracy and better robustness.

[0135] In summary, the above results indicate that the model can achieve the maximum diagnostic accuracy under various evaluation criteria, verifying the good cross-condition diagnostic performance of the model.

[0136] The embodiment of the present invention provides a cross-condition bearing fault diagnosis device based on a multi-scale convolutional fuzzy neural network migration model, which is used to implement the foregoing cross-condition bearing fault diagnosis method based on a multi-scale convolutional fuzzy neural network migration model, and specifically includes the following modules:

[0137] An acquisition module, which is used to acquire the vibration signal of the bearing to be diagnosed.

[0138] A diagnosis module, which is used to input the vibration signal of the bearing to be diagnosed into the trained cross-condition bearing fault diagnosis model and output a diagnosis result; wherein, the trained cross-condition bearing fault diagnosis model is obtained by training the multi-scale convolutional fuzzy neural network migration model with a training data set, and the total loss function of the training process includes a classification loss guided by wrong samples, a deep domain adaptive loss, and an adversarial network domain confusion loss; the training data set includes the vibration signals of bearings under the source domain conditions and the corresponding fault labels, as well as the vibration signals of bearings under the target domain conditions; the multi-scale convolutional fuzzy neural network migration model includes a multi-scale convolutional neural network module, a gated recurrent convolutional attention module, an improved adaptive fuzzy inference system, and an adversarial network module connected in sequence.

[0139] All relevant contents of each step involved in the embodiment of the foregoing cross-condition bearing fault diagnosis method based on a multi-scale convolutional fuzzy neural network migration model can be cited in the function description of the corresponding functional modules of the cross-condition bearing fault diagnosis device based on a multi-scale convolutional fuzzy neural network migration model in the embodiment of the present invention, and will not be elaborated here. The division of modules in the embodiment of the present invention is illustrative, only a logical function division, and there may be other division methods in actual implementation. In addition, in each embodiment of the present invention, each functional module can be integrated in a processor, can also exist separately physically, or two or more modules can be integrated in one module. The above integrated modules can be implemented in the form of hardware or in the form of software functional modules.

[0140] In another embodiment of the present invention, a computer device is provided. The computer device includes a processor and a memory. The memory is used to store a computer program, and the computer program includes program instructions. The processor is used to execute the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or may also be other general-purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field-Programmable Gate Arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, and is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions in the computer storage medium to implement the corresponding method flow or corresponding function. The processor described in the embodiment of the present invention can be used for the operation of a cross-condition bearing fault diagnosis method based on a multi-scale convolutional fuzzy neural network migration model.

[0141] In another embodiment of the present invention, a storage medium is also provided, specifically a computer-readable storage medium (Memory). The computer-readable storage medium is a memory device in a computer device and is used to store programs and data. It can be understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and, of course, the extended storage medium supported by the computer device. The computer-readable storage medium provides a storage space, and the operating system of the terminal is stored in this storage space. And, one or more instructions suitable for being loaded and executed by the processor are also stored in this storage space. These instructions can be one or more computer programs (including program codes). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk memory. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the cross-condition bearing fault diagnosis method based on a multi-scale convolutional fuzzy neural network migration model in the above embodiment.

[0142] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of an all-hardware embodiment, an all-software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0143] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present invention. It should be understood that each flow and / or block in the flowchart and / or block diagram can be implemented by computer program instructions, and the combination of the flows and / or blocks in the flowchart and / or block diagram can also be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in Figure 1 one or more of the flows Figure 1 or blocks or the combination of blocks.

[0144] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including instruction means for implementing the functions specified in Figure 1 one or more of the flows Figure 1 or blocks or the combination of blocks.

[0145] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in Figure 1 one or more of the flows Figure 1 or blocks or the combination of blocks.

[0146] The present invention also provides a computer program product, which is used to execute any one of the above-mentioned cross-condition bearing fault diagnosis methods based on a multi-scale convolutional fuzzy neural network transfer model. Since the computer program product provided by the present invention and the above-mentioned cross-condition bearing fault diagnosis method based on a multi-scale convolutional fuzzy neural network transfer model belong to the same inventive concept, the computer program product provided by the present invention has all the advantages of the above-mentioned cross-condition bearing fault diagnosis method based on a multi-scale convolutional fuzzy neural network transfer model. Therefore, the beneficial effects of the computer program product provided by the present invention will not be elaborated one by one herein.

[0147] In the present invention, the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples", etc. mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, without contradiction, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0148] Finally, it should be noted that the above-mentioned embodiments are only specific embodiments of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: any person skilled in the art within the technical scope disclosed by the present invention can still modify the technical solutions recorded in the foregoing embodiments or can easily think of changes, or make equivalent replacements for some of the technical features; and these modifications, changes, or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A cross-operating bearing fault diagnosis method based on a multi-scale convolutional fuzzy neural network migration model, characterized in that: include: Obtaining the vibration signal of the bearing to be diagnosed; The vibration signal of the bearing to be diagnosed is input into a trained cross-operating condition bearing fault diagnosis model, and a diagnosis result is output; wherein the trained cross-operating condition bearing fault diagnosis model is obtained by training a multi-scale convolutional fuzzy neural network migration model using a training data set, and the total loss function of the training process includes classification loss guided by error samples, deep domain adaptive loss and adversarial network domain confusion loss; the training data set includes vibration signals from bearings under source domain conditions and corresponding fault labels, as well as vibration signals from bearings under target domain conditions; the multi-scale convolutional fuzzy neural network migration model includes a multi-scale convolutional neural network module, a gated recursive convolutional attention module, an improved adaptive fuzzy reasoning system and an adversarial network module connected in sequence.

2. According to claim 1, a cross-operating bearing fault diagnosis method based on a multi-scale convolutional fuzzy neural network migration model is characterized in that: The multi-scale convolutional neural network module uses multiple convolution-pooling layers with different convolution kernel sizes, and different convolution-pooling layers are connected sequentially and in parallel to extract multi-scale features of vibration signals, as follows: Where x is the batch input training sample; B×K×N represents the batch size of the input sample and the number of feature maps input and input feature length; ConvPooling i (·) and ConvPooling j 3(·) represents the 1st to 3rd and 4th convolution-pooling operations respectively; is the output of the first three convolution-pooling operations, where i indicates the operation round; Represents the input of the fourth convolution-pooling operation, which is the output of the previous convolution-pooling operation The number of representations at different scales obtained after segmentation by the length of the power series of 2 is m; The final output y of the multi-scale convolutional neural network module is obtained by convolution of different scales and kernel sizes in the fourth convolution-pooling operation. MC .

3. According to claim 1, a cross-operating bearing fault diagnosis method based on a multi-scale convolutional fuzzy neural network migration model is characterized in that: The gated recursive convolutional attention module is a recursive convolutional module with residual connections, including a convolution-pooling projection layer for feature projection, an adaptive gated recursive convolution layer, and a fully connected layer for signal scale restoration; For input features with batch size B, number of feature maps K, and feature length N First, a convolution-pooling operation of different scales is performed to establish the mapping relationship between input features in different spatial dimensions: Where MConv(·) represents convolution operations of different scales; The output corresponding to the convolution operation of different scales is a geometric progression with a size of 2. AvgPooling(·) represents the average pooling operation; are the feature outputs of different scales obtained through the convolution-pooling layer; Then, each feature fragment is summed with the recursive convolution result of a longer size by element-by-element multiplication, thereby implementing the recursive convolution operation: Where ⊙ represents the element-by-element multiplication operation; {f t (·)} t＝0,1,...n is a set of recursive convolution operations with increasing depth; {g t (·)} t＝0,1,...n is the corresponding feature size matching function; α is a scalar parameter used to stabilize the training process; Finally, all the output results obtained by the recursive convolution are connected through the Concat(·) operation and passed through a fully connected layer to obtain the output y with the same feature dimension as the input gated recursive convolution attention module. Proj_out , as shown below: Where p0, p1, ..., p n Represents the multi-level output of recursive convolution; Linear(·) represents the linear connection layer that finally obtains the module output.

4. The cross-operating bearing fault diagnosis method based on a multi-scale convolutional fuzzy neural network migration model according to claim 1 is characterized in that: The improved adaptive fuzzy inference system includes a plurality of parallel adaptive pruning fuzzy inference structures, wherein the adaptive pruning fuzzy inference structure is based on the fuzzy logic concept and includes a fuzzification layer, a fuzzy rule layer, a normalization layer and a defuzzification layer; The fuzzification layer is used to perform fuzzy mapping on the input features, mapping the input features to fuzzy values ​​on different fuzzy sets, which are called membership degrees: in, is the membership degree of the jth fuzzy set belonging to the i-th channel input, and its value is between 0 and 1; MF j (·) is the membership function of the jth fuzzy set, and Gaussian function is selected as the metric of membership; μ j With σ j is the parameter to be trained in the membership function of the jth fuzzy set; Input x to the fuzzification layer F The data on the i-th channel, The fuzzy rule layer calculates the trigger strength of each rule by the membership degree; for the kth rule R in the fuzzy rule base k , which is expressed as follows: R k :IF x1 is A 1j ,and x2 is A 2j ,...and x N is A Nj ,THEN Y k =f k (x1,x2,…,x N ) Among them, A Nj is the jth fuzzy set in the domain, which is a fuzzy cluster determined by the fuzzy input distribution. The domain is defined as two fuzzy sets with positive and negative duality: and Among them pos Represents a positive fuzzy set, which contains the membership degree of each channel data after fuzzification The corresponding neuron activation degree is greater, while Ξ neg Represents a negative fuzzy set, which contains the membership degree of each channel data after fuzzification The corresponding activation level is less than the former; f k (·) is the output form of the rule, take f k (·) is the input x of the fuzzy layer F The first-order polynomial of ; the trigger strength of each fuzzy rule is calculated by the S-norm regularization method, then the output of the fuzzy rule layer can be expressed as the trigger strength value of each fuzzy rule, as shown in the following formula: Among them, r k is the trigger strength of the kth fuzzy rule, It represents the membership of the i-th channel input on the positive and negative fuzzy sets, and its different arrangement orders correspond to different fuzzy rules; The normalization layer is used to calculate the normalized value of the trigger strength of each rule derived from the fuzzy rule layer; the normalized trigger strength corresponding to a fuzzy rule is calculated as follows: Among them, S k is the normalized trigger strength of the kth fuzzy rule after normalization, and z is the number of fuzzy rules; the defuzzification layer calculates the product of the normalized trigger strength of each rule and the rule output value, and finally obtains the final output of the adaptive pruning fuzzy reasoning structure, as shown in the following formula: Among them, y out is the final output result, Y k =f k (x1,x2,…,x N ) is the output value of the kth rule, and Linear(·) is a linear layer used to map the obtained results to the final output dimension.

5. The cross-operating bearing fault diagnosis method based on a multi-scale convolutional fuzzy neural network migration model according to claim 1 is characterized in that: The adversarial network module is a multi-layer perceptron structure composed of three fully connected layers, which is used to confuse the source domain and the target domain of the extracted features to extract the domain-invariant features of the signal.

6. The cross-operating bearing fault diagnosis method based on a multi-scale convolutional fuzzy neural network migration model according to claim 1 is characterized in that: The total loss function of the training process is: L ttl =L JMMD (i f ,i y )+L d (i f ,i d )+L Focal (i f ,i y ) Where, L ttl is the total loss of the training process; L Focal (θ f ,θ y ) is the classification loss guided by the wrong sample; L JMMD (θ f ,θ y ) is the depth domain adaptive loss; L d (θ f ,θ d ) is the adversarial network domain confusion loss; θ f is a feature extractor G consisting of a multi-scale convolutional neural network module and a gated recursive convolutional attention module. f The parameters to be trained; θ y is a feature classifier G composed of an improved adaptive fuzzy inference system y The parameters to be trained; θ d is the domain discriminator G composed of adversarial network modules d The parameters to be trained in n s With n t Represents the number of samples in the source domain and the target domain respectively; K s Represents the number of classification labels; (1-y) γ is the modulation factor controlled by the adjustable focusing parameter γ≥0; k(·) is the Gaussian kernel function; and Represent samples in the source domain and target domain respectively; and The feature extractor G in the source domain and the target domain are f Output intermediate features; and They are the final output results in the source domain and the target domain respectively.

7. A cross-operating bearing fault diagnosis device based on a multi-scale convolutional fuzzy neural network migration model, characterized in that: include: An acquisition module, used for acquiring a vibration signal of a bearing to be diagnosed; A diagnosis module is used to input the vibration signal of the bearing to be diagnosed into a trained cross-operating condition bearing fault diagnosis model and output a diagnosis result; wherein the trained cross-operating condition bearing fault diagnosis model is obtained by training a multi-scale convolutional fuzzy neural network migration model using a training data set, and the total loss function of the training process includes classification loss guided by error samples, deep domain adaptive loss and adversarial network domain confusion loss; the training data set includes vibration signals from bearings under source domain conditions and corresponding fault labels, as well as vibration signals from bearings under target domain conditions; the multi-scale convolutional fuzzy neural network migration model includes a multi-scale convolutional neural network module, a gated recursive convolutional attention module, an improved adaptive fuzzy reasoning system and an adversarial network module connected in sequence.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, it implements a cross-operating-condition bearing fault diagnosis method based on a multi-scale convolutional fuzzy neural network migration model as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by the processor, it implements a cross-operating-condition bearing fault diagnosis method based on a multi-scale convolutional fuzzy neural network migration model as described in any one of claims 1 to 6.

10. A computer program product, characterized in that When the computer program product is executed by a processor, it implements a cross-operating-condition bearing fault diagnosis method based on a multi-scale convolutional fuzzy neural network migration model as described in any one of claims 1 to 6.

Images