A method, device, computer equipment and medium for diagnosing bearing faults under variable working conditions

Abstract

The present invention provides a variable-operating-condition bearing fault diagnosis method, apparatus, computer equipment, and medium, belonging to the field of mechanical equipment fault diagnosis. The method comprises: acquiring a vibration signal of a target bearing; constructing a domain-adaptive network model; the domain-adaptive network model comprising a deep residual semi-shrinkage wavelet convolutional network and a softmax classifier; sequentially optimizing the domain-adaptive network model through momentum contrastive learning, nearest neighbor contrastive learning, a domain discriminator, and a maximum mean difference metric to obtain a bearing fault diagnosis model; inputting the target bearing's vibration signal into the bearing fault diagnosis model to determine the cause of the fault. In this way, cross-domain diagnosis of bearing faults is achieved by aligning the common distribution of contextual representations and the feature space distributions of the same fault under multiple operating conditions. Comprehensive consideration of the fault's contextual representation and feature space distribution facilitates improved diagnostic accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of mechanical equipment fault diagnosis, and in particular relates to a variable operating condition bearing fault diagnosis method, device, computer equipment and medium. Background Art

[0002] Bearings are core components of most machinery and are also extremely vulnerable. Bearing failures can cause production line shutdowns at best, or even lead to casualties. Therefore, bearing fault diagnosis is essential to determine the cause of the failure and promptly repair or replace the faulty component.

[0003] In the problem of variable operating condition fault diagnosis, researchers mostly use transfer learning-based models to learn diagnostic knowledge from different domains. In the past decade, transfer learning has attracted significant attention as a new learning paradigm, and fault diagnosis methods based on transfer learning have achieved extensive development. Transfer learning uses knowledge learned from the source domain to handle the same or similar tasks in the target domain. Due to unavoidable factors such as changing operating conditions and components with degraded performance, the source and target data differ, resulting in some common knowledge (CK) and different knowledge (DK). Therefore, CK can be transferred and reused in the target domain, while DK needs to be adjusted or relearned to handle new target tasks.

[0004] Transfer learning is used in the existing technology to train the fault diagnosis model. However, since the bearing vibration signal data collected in actual industrial scenarios is long-term and non-stationary, especially the distribution of fault signal data varies greatly, this leads to higher requirements for the diagnosis of fault signals. However, the existing cross-domain fault diagnosis technology often fails to meet the requirements, resulting in unsatisfactory results of transfer learning. In this way, there will be problems with inaccurate diagnostic results when performing cross-domain diagnosis. Summary of the Invention

[0005] In order to solve the above-mentioned problem of inaccurate diagnostic results in the case of cross-domain diagnosis, the present invention provides a variable operating condition bearing fault diagnosis method, device, computer equipment and storage medium.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] In a first aspect, a method for diagnosing bearing faults under variable operating conditions is provided, comprising:

[0008] Acquire the vibration signal of the target bearing; construct a domain adaptive network model; the domain adaptive network model includes a deep residual semi-shrinkage wavelet convolutional network and a Softmax classifier; optimize the domain adaptive network model to obtain a bearing fault diagnosis model; specifically, perform gradient update on the deep residual semi-shrinkage wavelet convolutional network through momentum contrastive learning, and optimize the deep residual semi-shrinkage wavelet convolutional network based on the current gradient and the previous gradient, so that the deep residual semi-shrinkage wavelet convolutional network can extract the context representation of the input signal; achieve alignment of the context representations between different data domains through nearest neighbor contrastive learning, and optimize the deep residual semi-shrinkage wavelet convolutional network according to the alignment result, so that the context of the input signal extracted by the deep residual semi-shrinkage wavelet convolutional network remains unchanged. features; the different data domains characterize the bearing vibration signals under different working conditions; the data domain from which the context-invariant features come is identified by a domain discriminator, and the deep residual semi-shrinkage wavelet convolutional network is optimized according to the identification results; the structural alignment loss between different data domains is determined by the maximum mean difference metric, and the deep residual semi-shrinkage wavelet convolutional network is optimized according to the structural alignment loss, so that the deep residual semi-shrinkage wavelet convolutional network can extract the domain-invariant features of the input signal; the vibration signal of the target bearing is input into the bearing fault diagnosis model, and the vibration signal is feature extracted by the optimized deep residual semi-shrinkage wavelet convolutional network to obtain domain-invariant features, the domain-invariant features are classified by a Softmax classifier, and the fault cause of the target bearing is determined according to the classification result.

[0009] Optionally, the deep residual semi-shrinkage wavelet convolution network includes an extraction unit and a filtering unit, and the construction steps of the deep residual semi-shrinkage wavelet convolution network include: constructing a deep residual network, initializing the first layer of convolution of the deep residual network according to the wavelet convolution, and obtaining a wavelet weight initialization layer as an extraction unit; inserting a correction coefficient into the soft thresholding function of the deep residual network to obtain a filtering unit.

[0010] Optionally, initializing the first convolution layer of the deep residual network according to wavelet convolution includes: constructing a wavelet convolution, enhancing the wavelet convolution through Laplace wavelet basis function construction, and initializing the first convolution layer of the deep residual network according to the wavelet weight of the enhanced wavelet convolution to obtain a wavelet weight initialization layer.

[0011] Optionally, the enhancing the wavelet convolution by constructing the Laplace wavelet basis function and initializing the first convolution layer of the deep residual network according to the wavelet weight of the enhanced wavelet convolution includes: enhancing the wavelet convolution by constructing the Laplace wavelet basis function to obtain the enhanced Laplace wavelet dictionary

[0012]

[0013] Where t is time, s is the scale factor, u is the translation factor, f is the sampling frequency, ξ is the viscous damping ratio, A is a wavelet normalization function, τ is the time parameter, e is a natural constant, and L is the enhancement flag of the Laplace wavelet dictionary. The scale smoothing factor ζ is introduced, and the Sigmoid activation function is used to process the exponential component to determine the wavelet weight:

[0014]

[0015] The convolution kernel weights of the first convolution layer of the deep residual network are initialized as wavelet weights.

[0016] Optionally, inserting the correction coefficient into the soft thresholding function of the deep residual network includes: inserting the correction coefficient α into the soft thresholding function of the deep residual network to obtain:

[0017]

[0018] Where x is the vibration signal, η is the adaptive threshold, and y is the wavelet coefficient.

[0019] Optionally, optimizing the deep residual semi-shrinkage wavelet convolutional network in combination with the current gradient and the previous gradient includes: updating parameters of the deep residual semi-shrinkage wavelet convolutional network according to a weighted average of the current gradient and the previous gradient.

[0020] Optionally, the data domain from which the context-invariant features originate is identified through a domain discriminator, and the deep residual semi-shrinkage wavelet convolutional network is updated according to the identification result, including: identifying the features extracted by the deep residual semi-shrinkage wavelet convolutional network through the domain discriminator, iteratively optimizing the deep residual semi-shrinkage wavelet convolutional network according to the identification result, and stopping the iteration when the domain discriminator and the deep residual semi-shrinkage wavelet convolutional network reach a balance.

[0021] In a second aspect, the present invention further provides a variable operating condition bearing fault diagnosis device, comprising:

[0022] An acquisition module, used for acquiring a vibration signal of a target bearing;

[0023] A construction module for constructing a domain adaptive network model; the domain adaptive network model includes a deep residual semi-shrinkage wavelet convolutional network and a Softmax classifier;

[0024] An optimization module is used to optimize the domain adaptive network model to obtain a bearing fault diagnosis model; specifically, the deep residual semi-shrinkage wavelet convolution network is gradient updated through momentum contrastive learning, and the deep residual semi-shrinkage wavelet convolution network is optimized in combination with the current gradient and the previous gradient, so that the deep residual semi-shrinkage wavelet convolution network can extract the context representation of the input signal; the context representation between different data domains is aligned through nearest neighbor contrastive learning, and the deep residual semi-shrinkage wavelet convolution network is optimized according to the alignment result, so that the deep residual semi-shrinkage wavelet convolution network can extract the context invariant features of the input signal; the different data domains represent the bearing vibration signals under different working conditions; the data domain from which the context invariant features come is identified through a domain discriminator, and the deep residual semi-shrinkage wavelet convolution network is optimized according to the identification result; the structural alignment loss between different data domains is determined through the maximum mean difference metric, and the deep residual semi-shrinkage wavelet convolution network is optimized according to the structural alignment loss, so that the deep residual semi-shrinkage wavelet convolution network can extract the domain invariant features of the input signal;

[0025] The diagnosis module is used to input the vibration signal of the target bearing into the bearing fault diagnosis model, extract the features of the vibration signal through the optimized deep residual semi-shrinkage wavelet convolutional network, obtain domain-invariant features, classify the domain-invariant features through the Softmax classifier, and determine the fault cause of the target bearing based on the classification results.

[0026] In a third aspect, the present invention further provides a computer-readable storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above-mentioned variable operating condition bearing fault diagnosis method are implemented.

[0027] In a fourth aspect, the present invention further provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above-mentioned variable operating condition bearing fault diagnosis method when executing the program.

[0028] The variable operating condition bearing fault diagnosis method provided by the present invention has the following beneficial effects:

[0029] The deep residual semi-shrinkage wavelet convolutional network is optimized through momentum contrastive learning and nearest neighbor contrastive learning to achieve alignment of contextual representations of faults under multiple working conditions, and improve the accuracy of fault diagnosis through the common distribution of contextual representations; and on this basis, the deep residual semi-shrinkage wavelet convolutional network is optimized through domain discriminator recognition and maximum mean difference measurement to achieve alignment of feature space distribution of the same fault under multiple working conditions, and realize cross-domain diagnosis of bearing faults under different working conditions; in this way, by comprehensively considering the contextual representation and feature space distribution of the fault, it is beneficial to improve the accuracy of the diagnosis results. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] To more clearly illustrate the embodiments of the present invention and its design, the following briefly introduces the drawings required for this embodiment. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be derived from these drawings without inventive effort.

[0031] Figure 1 The figure is a flow chart of a method for diagnosing bearing faults under variable operating conditions provided by the present invention according to an exemplary embodiment.

[0032] Figure 2 A schematic diagram of the structure of a deep residual half-shrinkage wavelet convolutional network provided by the present invention according to an exemplary embodiment.

[0033] Figure 3 A schematic diagram of the optimization process of a deep residual semi-shrinkage wavelet convolutional network provided by the present invention according to an exemplary embodiment.

[0034] Figure 4 This is a block diagram of a variable operating condition bearing fault diagnosis device provided according to an exemplary embodiment of the present invention. DETAILED DESCRIPTION

[0035] In order to enable those skilled in the art to better understand the technical solution of the present invention and to be able to implement it, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solution of the present invention and are not intended to limit the scope of protection of the present invention.

[0036] Domain adaptation refers to the attempt to reduce the distribution difference between the source domain and the target domain through some technical means, such as maximum mean difference minimization and adversarial training, in order to better transfer knowledge to the target domain.

[0037] This paper proposes a domain adaptation network model based on momentum contrastive learning, combining contrastive learning and transfer learning. First, momentum contrastive learning is used to capture the contextual representation of data samples. Second, the contextual representations of the two domains are aligned using a nearest neighbor contrast algorithm. Finally, domain adversarial learning and the maximum mean difference metric are used to align the feature space distributions between the source and target domains. This approach not only considers the connection between the contextual representations of the two domains but also addresses the commonalities in the feature space distributions, achieving the goal of domain alignment with minimal dissimilarity.

[0038] The technical solutions provided by various embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0039] First, the present invention provides a method for diagnosing bearing faults under variable working conditions. Figure 1 As shown, the following steps are included:

[0040] S101. Acquire a vibration signal of a target bearing.

[0041] In this step, based on the measurement object and measurement environment, the size and type of the sensor range should be considered when selecting a vibration sensor. According to specific needs, the appropriate sensor type should be selected, and specific performance indicators such as its sensitivity, frequency response characteristics, linear range and accuracy should be considered. The vibration sensor is then used to sample the target bearing according to the preset sampling frequency.

[0042] S102. Build a domain adaptive network model.

[0043] Among them, the adaptive network models in this field include deep residual semi-shrinkage wavelet convolutional network and Softmax classifier.

[0044] In this step, the deep residual semi-shrinkage wavelet convolutional network includes an extraction unit and a filtering unit. Constructing the deep residual semi-shrinkage wavelet convolutional network includes the following steps: constructing a deep residual network, initializing the first layer of convolution of the deep residual network according to the wavelet convolution, and obtaining a wavelet weight initialization layer as an extraction unit; inserting a correction coefficient into the soft thresholding function of the deep residual network to obtain a filtering unit.

[0045] In one embodiment, the first convolution layer of the deep residual network can be initialized by first constructing a wavelet convolution, enhancing the wavelet convolution through the Laplace wavelet basis function construction, and initializing the first convolution layer of the deep residual network according to the wavelet weight of the enhanced wavelet convolution to obtain a wavelet weight initialization layer, specifically as follows: Figure 2 As shown in Figure 2, enhancing the wavelet convolution and initializing the first convolution layer of the deep residual network based on the enhanced wavelet convolution is beneficial to improving the accuracy of feature extraction.

[0046] For example, the original convolution output of the first convolution layer of the deep residual network is as follows:

[0047]

[0048] Among them, k w and b w Denote weight and bias respectively, and x is the vibration signal. In the time domain, the basic wavelet dictionary is ψ u,s (t), as follows:

[0049]

[0050] Where ψ(·) is the wavelet basis function, t is time, s is the scale factor, u is the translation factor, s and u are dynamic adaptive parameters and are completely differentiable. The wavelet convolution h operation is as follows:

[0051] h=ψ u,s (t)×x

[0052] Since the original bearing signal is a real signal, the wavelet convolution is enhanced by constructing the Laplace wavelet basis function, as shown in the following formula:

[0053]

[0054] Where f is the sampling frequency, ξ is the viscous damping ratio, A is a wavelet normalization function, and τ represents the time parameter.

[0055] Therefore, the actual enhanced Laplace Wavelet Dictionary (LWD) is:

[0056]

[0057] Because the value ranges of u and s are different, the scale smoothing factor ζ is introduced, and the value range of s is modified to s∈(0,N k ), N k It is expressed as the number of output channels, and L is the enhancement flag of the Laplace wavelet dictionary, from which the following formula can be obtained:

[0058]

[0059] At this time, since the fluctuation range of the signal is large, it is not conducive to optimization, so the Sigmoid activation function is used to process the exponential component. The Sigmoid function formula is as follows:

[0060]

[0061] Then the wavelet weights are determined as follows:

[0062]

[0063] The convolution kernel weights of the first convolution layer of the deep residual network are initialized as wavelet weights.

[0064] Alternatively, in this step, wavelet convolution can be used to replace the convolution of the first layer of the deep residual network to establish a physical prior wavelet kernel. The parameters can then be adaptively adjusted through backpropagation to obtain the optimal solution for the wavelet kernel. The wavelet convolution is then enhanced through the construction of Laplace wavelet basis functions to obtain an enhanced Laplace wavelet dictionary.

[0065] In another embodiment, in order to improve the accuracy of filtering, the wavelet coefficients may be fitted by inserting correction coefficients into the soft thresholding function of the deep residual network.

[0066] Specifically, the residual shrinkage block uses the Squeeze-and-Excitation (SE) module to obtain the adaptive threshold η, and the following formula is used to set the soft threshold denoising:

[0067]

[0068] Where y is the wavelet coefficient. Taking into account the adaptive threshold correction coefficient, the correction coefficient α is inserted into the soft thresholding function of the deep residual network to obtain:

[0069]

[0070] Where α and η are differentiable (α∈(0,1), α≠0,1). When α=0 or 1, the above equation degenerates into hard threshold and soft threshold, respectively. α can be adjusted appropriately to make y closer to the true wavelet coefficients. Based on the above balanced dynamic adaptive threshold and Adam (Adaptive Moment Estimation), the parameter update is as follows:

[0071]

[0072] Here, μ represents the learning rate.

[0073] In addition, the present invention optimizes the domain adaptive network model in many aspects to obtain a bearing fault diagnosis model, so that the bearing fault diagnosis model can perform adaptive working condition fault diagnosis according to the bearing vibration signal.

[0074] First, training samples of bearing vibration signals under different working conditions are obtained. The training samples include a source domain and a target domain. The source domain includes a sample of fault signals to be identified under a certain working condition and its corresponding fault cause. The target domain includes fault signals to be identified under other undetermined working conditions. The domain adaptive network model is trained through the source domain, and the trained domain adaptive network model is subjected to transfer learning training through the target domain to obtain a bearing fault diagnosis model.

[0075] Specifically, the fault signal samples to be identified in the source domain are input into the domain adaptive network model to obtain the recognition result. The domain adaptive network model is trained with the goal of minimizing the deviation between the recognition result and the fault cause. Then, the trained domain adaptive network model is transferred through the target domain to obtain the bearing fault diagnosis model. The specific transfer learning process includes optimizing the deep residual semi-shrinkage wavelet convolutional network through momentum contrast learning, nearest neighbor contrast learning, domain discriminator and maximum mean difference metric in sequence. Figure 3 shown.

[0076] S103. Update the gradient of the deep residual semi-shrinkage wavelet convolutional network through momentum contrast learning, and optimize the deep residual semi-shrinkage wavelet convolutional network by combining the current gradient and the previous gradient.

[0077] In order to enable the deep residual semi-shrinkage wavelet convolutional network to extract the contextual representation of the input signal, this paper introduces momentum contrastive learning. Optimizing the deep residual semi-shrinkage wavelet convolutional network through momentum contrastive learning is beneficial to maximize the mutual information between the two variables and thus capture the contextual representation. Specifically, the bearing vibration signal is first randomly enhanced twice to obtain x q and x k Then, the two enhanced signals are processed by the deep residual semi-shrinkage wavelet convolutional network to obtain their embedded features z q =F(x q )and in It is a deep residual semi-shrinkage wavelet convolutional network with momentum update.

[0078] The parameter update fails Instead of backpropagation, the parameters of the deep residual semi-shrinkage wavelet convolutional network are updated based on the weighted average of the current gradient and the previous gradient. The momentum term combines the current gradient and the previous momentum value at a certain ratio (called the momentum coefficient m). This can smooth the change of gradient, reduce oscillation, and make the parameter update process more stable. Since the information of past gradients is introduced in the parameter update process, the context representation of each domain is also captured. Specifically,

[0079] θ k ←mθ k +(1-m)θ q

[0080] Where m is the momentum coefficient, m∈[0,1), θ q is updated by back-propagation, θ k is based on the accumulated θ q The update direction is updated, so θ k The transformation is smoother. Although the units in the queue are encoded by different encoders for different batches, the differences between them can be small enough. In momentum contrastive learning, the contrast loss function L cl To measure the performance of the model, when the sample z q With z k The more similar, the contrast loss L cl To minimize the difference between different samples expressing the same fault, feature z is projected via the projector network Q(·) q , so that Q(zq ) is closer to the positive sample z k , while storing in queue The negative samples in are farther away, and the negative samples are from the earlier batches z k So this generates a large set of negative pairs with the queue size and batch size being equal. This allows for better context representation.

[0081] After each training step, z k Batches are stored in a queue of size I, so for each domain, there is a contrastive loss as follows to update the network parameters.

[0082]

[0083] Among them, N represents the number of features, exp represents the exponential operation of the natural constant e, τ>0 is the temperature scaling parameter, and z qi and z ki Represent the enhanced signal x qi and x ki The feature representation of all features is normalized. Since there are two domains, the source domain and the target domain, two contrast losses are calculated by two queues respectively. and

[0084] In this way, the contextual representation of the input signal can be extracted through the optimized deep residual semi-shrinkage wavelet convolutional network.

[0085] S104. Align the context representations between different data domains through nearest neighbor contrast learning, and optimize the deep residual semi-shrinkage wavelet convolutional network based on the alignment results.

[0086] The different data domains represent bearing vibration signals under different working conditions.

[0087] Specifically, the deep residual semi-shrinkage wavelet convolutional network can extract the contextual representation of each domain of the bearing time series data after being optimized by momentum contrastive learning. Then, in order to enable the deep residual semi-shrinkage wavelet convolutional network to extract the context-invariant features of the input signal, the nearest neighbor contrastive learning is completed by creating positive pairs between the source domain and the target domain. That is, the samples that are most similar to the target domain samples are retrieved from the source domain as positive samples for contrastive learning, thereby realizing the alignment of domain context-invariant features. In the model framework proposed in the present invention, the nearest neighbor contrastive learning helps the classifier to achieve accurate prediction of the target domain by aligning the domain context representation. The nearest neighbor contrast loss is as follows:

[0088]

[0089] Among them, N tThe number of features that characterize the target domain, and Represent the feature representation of the target domain and the source domain respectively, NN s (·) represents the nearest neighbor embedding retrieved from the features of the source domain augmented samples.

[0090] In this way, the context-invariant features of the input signal can be extracted through the updated deep residual semi-shrinkage wavelet convolutional network.

[0091] In addition, through the above analysis of the cross-domain alignment context representation of bearing signal data, the momentum contrast loss can be obtained and the nearest neighbor contrast loss L NNCL Then each loss is assigned its own weight. Therefore, the total loss of context representation alignment in this step is as follows:

[0092]

[0093] Among them, λ cl ,λ NNCL is a hyperparameter, which represents the weight given to each component.

[0094] In addition, in order to further improve the accuracy of fault diagnosis, the present invention also defines the classification loss function as follows:

[0095]

[0096] Among them, L cel (·) is the cross entropy loss, N S The number of features that characterize the source domain, Characterize the predicted fault of the source domain vibration signal through the Softmax classifier, is the real fault of the source domain vibration signal.

[0097] The Softmax classifier is trained with the goal of minimizing the loss value of the classification loss function. When the loss value is less than or equal to the preset loss value, the training is determined to be complete. By training the Softmax classifier, the accuracy of the diagnosis results of the domain adaptive network model can be improved.

[0098] S105. Identify the data domain from which the context-invariant feature originates through a domain discriminator, and optimize the deep residual semi-shrinkage wavelet convolutional network based on the identification result.

[0099] In this step, due to the domain shift caused by the variable operating conditions of bearings, using only data from the source domain for training results in poor performance of the bearing fault diagnosis model in cross-operating condition fault diagnosis. To address this, the present invention introduces a domain discriminator to determine whether the extracted features are from the source or target domain. The feature extractor is also trained to deceive the domain discriminator.

[0100] The features extracted by the deep residual semi-shrinkage wavelet convolutional network are identified by the domain discriminator, and the deep residual semi-shrinkage wavelet convolutional network is iteratively optimized based on the identification results. When the domain discriminator and the deep residual semi-shrinkage wavelet convolutional network reach a balance, the iteration is stopped. When the two sides reach a balance, the domain-invariant features can be captured. For a given target domain sample x t , use the feature extractor to extract feature z t , and then use the gradient reversal layer to transform the feature z t This process can be performed to train the domain discriminator to minimize the domain classification loss, and the feature extractor to maximize the same loss. Finally, the present invention inputs the gradient-reversed features into the domain discriminator for binary classification and calculates the domain discriminator loss. The present invention defines the discriminator loss function as follows:

[0101]

[0102] Among them, N S The number of features representing the source domain, N t The number of features that characterize the target domain, 0 represents the domain label of the source domain sample, 1 represents the domain label of the target domain sample, L cel (·) is the cross entropy loss, D(R(·)) represents the gradient reversal processing, and Characterize the characteristics of the target domain and the source domain respectively.

[0103] In this way, the deep residual semi-shrinkage wavelet convolutional network is further optimized through the domain discriminator to improve the accuracy of feature extraction of the deep residual semi-shrinkage wavelet convolutional network.

[0104] S106. Determine the structural alignment loss between different data domains through the maximum mean difference metric, and optimize the deep residual semi-shrinkage wavelet convolutional network based on the structural alignment loss.

[0105] In order to align the feature space structures of the source domain and the target domain so that the deep residual semi-shrinkage wavelet convolutional network can extract the domain-invariant features of the input signal, the present invention adopts the maximum mean difference metric as the structural difference alignment loss, as shown in the following formula:

[0106]

[0107] Where φ(·) represents the nonlinear mapping function, ||·|| Ω Characterizing distance computations in reproducing kernel Hilbert spaces.

[0108] In summary, the present invention can obtain the total loss of the part aligned with the feature space distribution as shown in the following formula:

[0109] L2=L c +λ DA L DA +λ MMD L MMD

[0110] Among them, λ DA and λ MMD are the hyperparameters of the discriminator loss function and the maximum mean divergence metric, representing the weights given to each part.

[0111] In this way, the domain-invariant features of the input signal can be extracted through the updated deep residual semi-shrinkage wavelet convolutional network.

[0112] S107 , inputting the vibration signal of the target bearing into a bearing fault diagnosis model, and determining the fault cause of the target bearing through the bearing fault diagnosis model.

[0113] Specifically, the vibration signal of the target bearing is input into the bearing fault diagnosis model. The vibration signal is extracted through the optimized deep residual semi-shrinkage wavelet convolutional network to obtain domain-invariant features. The domain-invariant features are classified through the Softmax classifier, and the fault cause of the target bearing is determined according to the classification results.

[0114] Using the above method, the deep residual semi-shrinkage wavelet convolutional network is optimized through momentum contrastive learning and nearest neighbor contrastive learning to achieve alignment of the contextual representation of faults under multiple working conditions, and improve the accuracy of fault diagnosis through the common distribution of contextual representation; and on this basis, the deep residual semi-shrinkage wavelet convolutional network is optimized through domain discriminator recognition and maximum mean difference measurement to achieve alignment of the feature space distribution of the same fault under multiple working conditions, and realize cross-domain diagnosis of bearing faults under different working conditions; in this way, by comprehensively considering the contextual representation and feature space distribution of the fault, it is beneficial to improve the accuracy of the diagnosis results.

[0115] Secondly, the present invention also provides a variable working condition bearing fault diagnosis device, such as Figure 4 Shown, including:

[0116] The acquisition module 401 is used to acquire the vibration signal of the target bearing.

[0117] The construction module 402 is used to construct a domain adaptive network model; the domain adaptive network model includes a deep residual semi-shrinkage wavelet convolutional network and a Softmax classifier.

[0118] Optimization module 403 is used to optimize the domain adaptive network model to obtain a bearing fault diagnosis model; specifically, the deep residual semi-shrinkage wavelet convolution network is gradient updated through momentum contrastive learning, and the deep residual semi-shrinkage wavelet convolution network is optimized in combination with the current gradient and the previous gradient, so that the deep residual semi-shrinkage wavelet convolution network extracts the context representation of the input signal; the context representation between different data domains is aligned through nearest neighbor contrastive learning, and the deep residual semi-shrinkage wavelet convolution network is optimized according to the alignment result, so that the deep residual semi-shrinkage wavelet convolution network extracts the context invariant features of the input signal; the different data domains represent the bearing vibration signals under different working conditions; the data domain of the source of the context invariant features is identified through a domain discriminator, and the deep residual semi-shrinkage wavelet convolution network is optimized according to the identification result; the structural alignment loss between different data domains is determined through the maximum mean difference metric, and the deep residual semi-shrinkage wavelet convolution network is optimized according to the structural alignment loss, so that the deep residual semi-shrinkage wavelet convolution network extracts the domain invariant features of the input signal.

[0119] Diagnostic module 404 is used to input the vibration signal of the target bearing into the bearing fault diagnosis model, extract features of the vibration signal through the optimized deep residual semi-shrinkage wavelet convolutional network to obtain domain-invariant features, classify the domain-invariant features through the Softmax classifier, and determine the cause of the target bearing fault based on the classification results.

[0120] Optionally, the deep residual semi-shrinkage wavelet convolution network includes an extraction unit and a filtering unit; the construction module 402 is also used to construct a deep residual network, initialize the first layer of convolution of the deep residual network according to the wavelet convolution, and obtain a wavelet weight initialization layer as an extraction unit; insert a correction coefficient into the soft thresholding function of the deep residual network to obtain a filtering unit.

[0121] Among them, the construction module 402 is also used to construct wavelet convolution, enhance the wavelet convolution through Laplace wavelet basis function construction, and initialize the first layer of convolution of the deep residual network according to the wavelet weight of the enhanced wavelet convolution to obtain a wavelet weight initialization layer.

[0122] Specifically, the building module 402 is also used to enhance the wavelet convolution by constructing the Laplace wavelet basis function to obtain the enhanced Laplace wavelet dictionary

[0123]

[0124] Where t is time, s is the scale factor, u is the translation factor, f is the sampling frequency, ξ is the viscous damping ratio, A is a wavelet normalization function, τ is the time parameter, e is a natural constant, and L is the enhancement flag of the Laplace wavelet dictionary;

[0125] The scale smoothing factor ζ is introduced, and the Sigmoid activation function is used to process the exponential component to determine the wavelet weight:

[0126]

[0127] The convolution kernel weights of the first convolution layer of the deep residual network are initialized as wavelet weights.

[0128] Furthermore, the construction module 402 is further configured to insert the correction coefficient α into the soft thresholding function of the deep residual network to obtain:

[0129]

[0130] Where x is the vibration signal, η is the adaptive threshold, and y is the wavelet coefficient.

[0131] The optimization module 403 is also used to update the parameters of the deep residual semi-shrinkage wavelet convolutional network according to the weighted average of the current gradient and the previous gradient.

[0132] The optimization module 403 is also used to identify the features extracted by the deep residual semi-shrinkage wavelet convolutional network through a domain discriminator, iteratively optimize the deep residual semi-shrinkage wavelet convolutional network based on the identification results, and stop the iteration when the domain discriminator and the deep residual semi-shrinkage wavelet convolutional network reach a balance.

[0133] Using the above-mentioned device, the deep residual semi-shrinkage wavelet convolutional network is optimized through momentum contrastive learning and nearest neighbor contrastive learning, so as to realize the alignment of the contextual representation of faults under multiple working conditions, and improve the accuracy of fault diagnosis through the common distribution of contextual representation; and on this basis, the deep residual semi-shrinkage wavelet convolutional network is optimized through domain discriminator recognition and maximum mean difference measurement, so as to realize the alignment of the feature space distribution of the same fault under multiple working conditions, and realize cross-domain diagnosis of bearing faults under different working conditions; in this way, by comprehensively considering the contextual representation and feature space distribution of the fault, it is beneficial to improve the accuracy of the diagnosis results.

[0134] The present invention also provides a computer-readable storage medium, which stores a computer program, which can be used to execute the above Figure 1 The steps of the variable operating condition bearing fault diagnosis method are provided.

[0135] The present invention also provides a computer device. At the hardware level, the computer device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory. Of course, it may also include hardware required for other services. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to achieve the above Figure 1 The steps of the variable operating condition bearing fault diagnosis method are provided.

[0136] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0137] The present invention is described with reference to flowcharts and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0138] These computer program instructions may 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, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0139] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0140] It should be noted that the above specific embodiments can enable those skilled in the art to more fully understand the invention, but do not limit the invention in any way. Therefore, although the present invention and the embodiments have been described in detail, those skilled in the art should understand that the invention can still be modified or replaced with equivalents; and all technical solutions and improvements that do not depart from the spirit and scope of the invention are included in the scope of protection of the patent for the invention. Any reference numerals in the claims should not be construed as limiting the claims involved.

Claims

1. A method for diagnosing bearing faults under variable operating conditions, characterized in that: include: Obtaining the vibration signal of the target bearing; Constructing a domain adaptive network model; the domain adaptive network model includes a deep residual semi-shrinkage wavelet convolutional network and a Softmax classifier; The domain adaptive network model is optimized to obtain a bearing fault diagnosis model; specifically, the deep residual semi-shrinkage wavelet convolution network is gradient updated through momentum contrastive learning, and the deep residual semi-shrinkage wavelet convolution network is optimized in combination with the current gradient and the previous gradient, so that the deep residual semi-shrinkage wavelet convolution network can extract the context representation of the input signal; the context representation between different data domains is aligned through nearest neighbor contrastive learning, and the deep residual semi-shrinkage wavelet convolution network is optimized according to the alignment result, so that the deep residual semi-shrinkage wavelet convolution network can extract the context invariant features of the input signal; the different data domains represent the bearing vibration signals under different working conditions; the data domain from which the context invariant features come is identified through a domain discriminator, and the deep residual semi-shrinkage wavelet convolution network is optimized according to the identification result; the structural alignment loss between different data domains is determined through the maximum mean difference metric, and the deep residual semi-shrinkage wavelet convolution network is optimized according to the structural alignment loss, so that the deep residual semi-shrinkage wavelet convolution network can extract the domain invariant features of the input signal; The vibration signal of the target bearing is input into the bearing fault diagnosis model. The optimized deep residual semi-shrinkage wavelet convolutional network is used to extract the vibration signal features to obtain domain-invariant features. The domain-invariant features are then classified using the Softmax classifier, and the fault cause of the target bearing is determined based on the classification results. The deep residual semi-shrinkage wavelet convolutional network includes an extraction unit and a filtering unit, and the steps of constructing the deep residual semi-shrinkage wavelet convolutional network include: A deep residual network is constructed, and the first convolution layer of the deep residual network is initialized according to the wavelet convolution to obtain the wavelet weight initialization layer as the extraction unit; the correction coefficient is inserted into the soft thresholding function of the deep residual network to obtain the filtering unit; Initializing the first convolution layer of the deep residual network according to wavelet convolution includes: Construct wavelet convolution, enhance the wavelet convolution through Laplace wavelet basis function construction, and initialize the first convolution layer of the deep residual network according to the wavelet weight of the enhanced wavelet convolution to obtain the wavelet weight initialization layer; The method of enhancing the wavelet convolution by constructing a Laplace wavelet basis function and initializing the first convolution layer of the deep residual network according to the wavelet weight of the enhanced wavelet convolution includes: The wavelet convolution is enhanced by constructing the Laplace wavelet basis function to obtain the enhanced Laplace wavelet dictionary : in, For time, is the scale factor, is the translation factor, is the sampling frequency, is the viscous damping ratio, is a wavelet normalization function, is the time parameter, is a natural constant, L is the enhancement flag of the Laplace wavelet dictionary; Introducing scale smoothing factor , and use the Sigmoid activation function to process the exponential component and determine the wavelet weight: The convolution kernel weights of the first convolution layer of the deep residual network are initialized as wavelet weights.

2. The variable operating condition bearing fault diagnosis method according to claim 1, characterized in that: The inserting of a correction coefficient into the soft thresholding function of the deep residual network comprises: The correction factor Inserting it into the soft thresholding function of the deep residual network yields: in, is the vibration signal, is the adaptive threshold, is the wavelet coefficient.

3. The variable operating condition bearing fault diagnosis method according to claim 1, characterized in that: The optimization of the deep residual semi-shrinkage wavelet convolutional network by combining the current gradient and the previous gradient includes: The parameters of the deep residual semi-shrinkage wavelet convolutional network are updated according to the weighted average of the current gradient and the previous gradient.

4. The variable operating condition bearing fault diagnosis method according to claim 1, characterized in that: Identifying the data domain of the context-invariant feature source through a domain discriminator, and updating the deep residual semi-shrinkage wavelet convolutional network according to the identification result includes: The features extracted by the deep residual semi-shrinkage wavelet convolutional network are recognized by a domain discriminator, and the deep residual semi-shrinkage wavelet convolutional network is iteratively optimized according to the recognition results. The iteration is stopped when the domain discriminator and the deep residual semi-shrinkage wavelet convolutional network reach a balance.

5. A variable operating condition bearing fault diagnosis device, characterized in that: A variable operating condition bearing fault diagnosis method as described in claim 1, comprising: An acquisition module, used for acquiring a vibration signal of a target bearing; A construction module for constructing a domain adaptive network model; the domain adaptive network model includes a deep residual semi-shrinkage wavelet convolutional network and a Softmax classifier; An optimization module is used to optimize the domain adaptive network model to obtain a bearing fault diagnosis model; specifically, the deep residual semi-shrinkage wavelet convolution network is gradient updated through momentum contrastive learning, and the deep residual semi-shrinkage wavelet convolution network is optimized in combination with the current gradient and the previous gradient, so that the deep residual semi-shrinkage wavelet convolution network can extract the context representation of the input signal; the context representation between different data domains is aligned through nearest neighbor contrastive learning, and the deep residual semi-shrinkage wavelet convolution network is optimized according to the alignment result, so that the deep residual semi-shrinkage wavelet convolution network can extract the context invariant features of the input signal; the different data domains represent the bearing vibration signals under different working conditions; the data domain from which the context invariant features come is identified through a domain discriminator, and the deep residual semi-shrinkage wavelet convolution network is optimized according to the identification result; the structural alignment loss between different data domains is determined through the maximum mean difference metric, and the deep residual semi-shrinkage wavelet convolution network is optimized according to the structural alignment loss, so that the deep residual semi-shrinkage wavelet convolution network can extract the domain invariant features of the input signal; The diagnosis module is used to input the vibration signal of the target bearing into the bearing fault diagnosis model, extract the features of the vibration signal through the optimized deep residual semi-shrinkage wavelet convolutional network, obtain domain-invariant features, classify the domain-invariant features through the Softmax classifier, and determine the fault cause of the target bearing based on the classification results.

6. A computer-readable storage medium, characterized in that The storage medium stores a computer program, and when the computer program is executed by the processor, the steps of the variable operating condition bearing fault diagnosis method according to any one of claims 1 to 4 are implemented.

7. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps of the variable operating condition bearing fault diagnosis method according to any one of claims 1 to 4 are implemented.

Images