A bearing fault diagnosis method based on deep domain generalization network

Abstract

The application discloses a bearing fault diagnosis method based on a deep field generalization network, and comprises the following steps: S1, historical working condition data, new working condition data and random noise data are acquired, and after preprocessing, source domains, target domains and reference domains are obtained respectively; S2, a fault classification deep neural network of a bearing fault diagnosis model is established, three submodules in the fault classification deep neural network, namely, a feature encoder, a fault classifier and a domain scale calculator, are constructed; S3, a learning task corresponding to the bearing fault diagnosis model is established, and a corresponding learning strategy is formulated; S4, hyperparameters in the bearing fault diagnosis model are optimized, and optimal hyperparameters are determined; S5, the optimal hyperparameters are substituted into the fault classification deep neural network, the network is trained, and an optimal bearing fault diagnosis model is obtained; and S6, bearing monitoring target domain data that needs to be subjected to fault diagnosis is created, the optimal bearing fault diagnosis model is adopted to complete fault classification on the target domain data, and a fault type is determined.

Description

Technical Field

[0001] This invention relates to the field of bearing fault diagnosis technology, and specifically to a bearing fault diagnosis method based on deep domain generalization networks. Background Technology

[0002] Rolling bearings, as precision components, are widely used in the transmission mechanisms of various mechanical equipment, serving to support moving shafts and bear the loads on them. During long-term use, external environmental factors and inherent defects in the bearings themselves can cause various bearing failures, such as discoloration, cracks, corrosion, damage to balls and rollers, and damage to raceways. These issues can affect the normal operation of equipment and even cause irreversible damage. Therefore, conducting regular bearing health checks on critical mechanical equipment to detect and replace faulty bearings in advance is of great importance.

[0003] With advancements in computer and sensor technologies, bearing fault diagnosis technology has made significant progress in digitalization and intelligence. Bearing fault diagnosis technology based on digital signal analysis uses sensor systems to sense and collect various bearing-related signals emitted during equipment operation. It then processes and analyzes the collected digital signals, extracting bearing-related features and finally determining the bearing's health status based on these features. In the aforementioned digital signal processing and analysis, many researchers use feature engineering methods to process signals in the time or frequency domain, analyze their components, and then extract bearing fault features.

[0004] Traditional feature engineering methods typically require researchers to possess extensive expertise and experience to obtain satisfactory fault feature representations. Machine learning techniques, based on learning algorithms, can fit complex random data distributions. Some researchers have used machine learning methods to automatically extract bearing fault features and classify faults after performing simple preprocessing on signals. However, existing bearing fault diagnosis networks have poor generalization performance and are difficult to widely meet practical bearing fault diagnosis needs. Summary of the Invention

[0005] The purpose of this invention is to provide a bearing fault diagnosis method based on a deep domain generalization network. This method solves the problem of poor generalization performance of existing bearing fault diagnosis networks. It combines deep learning and domain generalization methods to achieve bearing fault feature alignment across multiple data domains and improves the generalization ability of the fault diagnosis model.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A bearing fault diagnosis method based on deep domain generalization networks includes the following steps:

[0008] S1. Acquire historical operating condition data, new operating condition data, and random noise data. After preprocessing, obtain multiple source domains, target domains, and reference domains respectively.

[0009] S2. Establish a deep neural network for bearing fault diagnosis model, and construct three sub-modules in the deep neural network for fault classification: feature encoder, fault classifier and domain scale calculator.

[0010] S3. Establish the learning task corresponding to the bearing fault diagnosis model and formulate the corresponding learning strategy;

[0011] S4. Optimize the hyperparameters in the bearing fault diagnosis model and determine the optimal hyperparameters;

[0012] S5. Substitute the optimal hyperparameters into the fault classification deep neural network, train the network, and obtain the optimal bearing fault diagnosis model.

[0013] S6. Create bearing monitoring target domain data that requires fault diagnosis, and use the best bearing fault diagnosis model to classify the target domain data for faults and determine the fault type.

[0014] Preferably, in step S1,

[0015] The source domain is represented as: D s,i ={X s,i ,Y s,i ,P s,i (x,y)}, i=1,2,3,...,N d ;

[0016] The target domain is represented as: D t ={X t ,P t (x)};

[0017] The reference domain is represented as: D r ={X r ,P r (x)};

[0018] Where s is the source domain identifier; t is the target domain identifier; r is the reference domain identifier; i represents the operating condition number; N d X represents the total number of historical operating conditions and also the number of source domains; X represents the sample space, where samples are three-dimensional vectors; Y represents the label space, where samples are scalars; P(x) represents the probability distribution of the samples; D s,i This represents the i-th source domain, which consists of samples under the i-th historical working condition;

[0019] The historical operating condition data is greater than or equal to three types, i.e., N d≥3; There are at least three types of faults under each historical operating condition, and the types of faults are the same under all historical operating conditions; The target domain is the data domain for bearing fault diagnosis; The target domain consists of a new operating condition sample that is different from all historical operating conditions, and the fault types under the new operating condition are within the range of fault types under the historical conditions; The reference domain consists of noise samples that do not contain any fault features, and the reference domain is used to provide a feature zero point.

[0020] Preferably, the preprocessing methods for historical and new operating condition data in step S1 are the same, both involving envelope processing of the acquired vibration acceleration signals and extraction of the envelope spectrum. The specific preprocessing procedures for historical and new operating condition data are as follows:

[0021] S11. The vibration acceleration signal acquired over a certain period of time is divided into multiple sample signal segments according to the set sample duration. Then, each signal segment is filtered by multiple frequency bands to obtain the data of the segmented signal after filtering in each frequency band. Among them, the multiple frequency bands are obtained by first determining the frequency range of the bearing's fault characteristic frequency and then dividing the frequency band into segments.

[0022] S12. Perform envelope processing on the filtered signal obtained from multi-band filtering to obtain the envelope signal of each frequency band;

[0023] S13. Perform Fourier transform on the envelope signal of each frequency band to obtain the envelope spectrum;

[0024] S14. Fold the envelope spectrum of each frequency band into two-dimensional data to form the data of one channel; multiple channels of data of a sample signal segment constitute a three-dimensional sample for the final training of the neural network.

[0025] Preferably, obtaining random noise data in step S1 refers to randomly generating a random number that is uniformly distributed within the interval (0,1), the length of which is equal to the length of the vibration acceleration signal to be preprocessed; the preprocessing of the random noise data is the same as the preprocessing of historical operating condition data.

[0026] Preferably, in step S2,

[0027] The feature encoder is represented as:

[0028] The fault classifier is represented as:

[0029] The domain scale calculator is represented as follows:

[0030] Where Z represents the deep feature vector space; P represents the predicted sample class probability space; and W represents the domain scale space.

[0031] The feature encoder is composed of multiple layers of convolutional units connected one after another. Each convolutional unit includes a convolutional operation unit for feature extraction and a downsampling unit. The feature encoder is used to map the sample vector into a depth feature vector.

[0032] The fault classifier is composed of multilayer perceptron neural network units connected one after another. Each layer of the perceptron neural network consists of multiple perceptrons connected in parallel. The fault classifier is used to map the deep feature vector obtained by the feature encoder into a sample class probability vector. The elements of the sample class probability vector are the probability that the sample belongs to each type of fault.

[0033] The domain scaling calculator is composed of multilayer perceptron neural network units connected one after another. Each layer of the perceptron neural network consists of multiple perceptrons connected in parallel. The domain scaling calculator is used to map the deep feature vector obtained by the feature encoder into a domain scaling factor. This domain scaling factor is used to scale the distribution difference between the source domain and the reference domain.

[0034] Preferably, the learning task of the bearing fault diagnosis model in step S3 includes three sub-tasks: source domain classification task, global distribution difference minimization task, and global regularization task.

[0035] The source domain classification task consists of classification tasks for all source domains in step S1. Multiple classification tasks share a feature encoder, a fault classifier, and a domain scale calculator. The source domain classification task first uses the feature encoder to map the sample vectors to deep feature vectors, and then uses the fault classifier to map the deep feature vectors to sample class probability vectors. The optimization objective of the source domain classification task is expressed as:

[0036]

[0037] Where, Θ e and Θ c Let N represent the parameter sets of the feature encoder and the fault classifier, respectively. b This indicates the number of samples in a batch. This represents the j-th sample in the i-th source domain. L represents the one-hot encoded label vector corresponding to the j-th sample in the i-th source domain. ce (·,·) denotes the multivariate cross-entropy loss function, f c (·) represents the fault classifier, g e (·) represents the feature encoder, and min represents minimization;

[0038] The task of minimizing global distribution differences aims to align the distributions of multiple source domains and eliminate the influence of operating conditions on fault characteristics. The task first calculates the maximum mean difference between each source domain distribution and the reference domain distribution, then calculates the domain scaling factor using a domain scaling calculator, and finally scales the maximum mean difference based on the scaling factor, and calculates the variance of all scaled mean differences. The optimization objective of the task is L. myfn Represented as:

[0039]

[0040]

[0041]

[0042] Where, Θ m The parameter set of the domain scaling calculator. This represents the j-th sample in the reference domain. h represents the maximum mean difference function of the joint distribution. m (·) indicates the domain scaling calculator, f c ′(·) denotes a subnetwork of the fault classifier;

[0043] The joint distribution maximum mean difference function The calculation of the joint distribution maximum mean difference function, used to measure the distance between two different distributions, consists of the following steps:

[0044] S31. Randomly and without repetition, extract deep feature vector pairs from the source and reference domains of the bearing fault diagnosis model: and Estimate the maximum mean difference between the marginal distributions of the source and reference domains using the following formula:

[0045]

[0046]

[0047] Where K(·,·) is the kernel function, which can be specifically expressed as: In the formula, σ is the range control parameter;

[0048] S32. Extract the sample class probability vector output by the fault classifier in the bearing fault diagnosis model. in The probability of a sample being diagnosed as a type k fault is predicted. Then, using the deep feature vector from step S31, the maximum mean difference between the source and reference domain conditional distributions of the type k sample is estimated using the following formula:

[0049]

[0050] S33. Based on the maximum mean difference of the marginal distribution and the maximum mean difference of the conditional distribution, calculate the maximum mean difference of the joint distribution of the source domain and the reference domain, as shown in the following formula:

[0051]

[0052]

[0053] Where, λ j is the proportion coefficient of each type of sample, and μ is the weight coefficient of the maximum mean difference of the joint distribution;

[0054] The global regularization task is used to apply global anti-overfitting constraints to the parameter sets of the feature encoder, fault classifier, and domain scaling calculator. The optimization objective of the global regularization task is L. re Represented as:

[0055]

[0056] Where λ represents the regularization parameter, which controls the degree of regularization, ω represents the parameter in the parameter set, and n represents the parameter index.

[0057] Preferably, the learning strategy in step S3 consists of two parts: model pre-training and model generalization training.

[0058] The model pre-training uses labeled source domain samples to pre-train the bearing fault diagnosis model, enabling it to possess basic source domain classification capabilities. The optimization objective is expressed as:

[0059]

[0060] The model generalization training is used to bridge the distribution differences of multiple source domains in the deep feature space, while maintaining the classification ability obtained by the bearing fault diagnosis model during pre-training. The optimization objective is expressed as:

[0061]

[0062] Here, α and β are weight parameters used to regulate the learning process and balance the attention of the bearing fault diagnosis model to the three sub-tasks.

[0063] Preferably, the hyperparameters in step S4 refer to the weight parameters α and β in the model generalization training objective function in step S3, and the optimization process of the hyperparameters is as follows:

[0064] S41. Specify the set T for the grid search for hyperparameters α and β respectively. α and T β ,

[0065] S42. Following the method in step S1, create a set of source domain data, target domain data, and reference domain data for optimizing hyperparameters;

[0066] S43. Construct the bearing fault diagnosis model on a computer using a computer programming language;

[0067] S44. Select the learning rate optimizer, training batch size, and number of training epochs. For each combination of hyperparameters α and β, train the model using source domain data and reference domain data respectively, following the learning strategy in step S3.

[0068] S45. After the model training is completed for each hyperparameter combination, the trained bearing fault diagnosis model is used to predict the faults of the sample data in the target domain. The prediction accuracy is statistically analyzed, and the hyperparameter combination with the highest accuracy is selected as the optimal hyperparameter.

[0069] Preferably, the specific process of step S5 is as follows:

[0070] S51. Following the method in step S1, create a new set of source domain data and reference domain data;

[0071] S52. Substitute the obtained optimal hyperparameters into the bearing fault diagnosis model created in step S4.

[0072] S53. Select the learning rate optimizer, batch size, and number of training rounds. Use the newly created data to train the bearing fault diagnosis model and obtain the optimal bearing fault diagnosis model.

[0073] By adopting the above technical solution, the present invention has the following beneficial effects:

[0074] 1. This invention solves the problem of poor generalization performance of existing bearing fault diagnosis networks. It combines deep learning and domain generalization methods to achieve bearing fault feature alignment across multiple data domains, thereby improving the generalization ability of the fault diagnosis model. This method proposes a new bearing fault diagnosis method for dynamic operating conditions, which can achieve indiscriminate generalization to new operating condition data based on historical bearing operating condition data.

[0075] 2. This invention proposes an end-to-end bearing fault diagnosis model structure that combines bearing fault classification and domain alignment to achieve bearing fault diagnosis under dynamic operating conditions.

[0076] 3. This invention proposes a new domain generalization model training objective function, which enables the model to extract invariant features across the entire operating range of a given device. Attached Figure Description

[0077] Figure 1 This is a flowchart of the present invention;

[0078] Figure 2 This is a structural block diagram of the bearing fault diagnosis method of the present invention;

[0079] Figure 3 This is a three-dimensional training sample image of the bearing fault diagnosis neural network of the present invention;

[0080] Figure 4 This is a schematic diagram of the deep neural network for bearing fault classification of the present invention;

[0081] Figure 5 The graph shows the target domain test accuracy results under different objective function hyperparameters of this invention. Detailed Implementation

[0082] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0083] Example

[0084] like Figures 1 to 5 As shown, a bearing fault diagnosis method based on deep domain generalization networks includes the following steps:

[0085] S1. Acquire historical operating condition data, new operating condition data, and random noise data. After preprocessing, obtain multiple source domains, target domains, and reference domains respectively.

[0086] In step S1,

[0087] The source domain is represented as: D s,i ={X s,i ,Y s,i ,P s,i (x,y)}, i=1,2,3,...,N d ;

[0088] The target domain is represented as: D t ={X t ,P t (x)};

[0089] The reference domain is represented as: D r ={X r ,P r (x)};

[0090] Where s is the source domain identifier; t is the target domain identifier; r is the reference domain identifier; i represents the operating condition number; N dX represents the total number of historical operating conditions and also the number of source domains; X represents the sample space, where samples are three-dimensional vectors; Y represents the label space, where samples are scalars; P(x) represents the probability distribution of the samples; D s,i This represents the i-th source domain, which consists of samples under the i-th historical working condition;

[0091] The number of source domains is 3, which also indicates that there are 3 types of historical operating conditions. The number of fault types under each historical operating condition is 3, and the fault types are the same under all historical operating conditions.

[0092] The three historical operating conditions are as follows: Condition 1: shaft speed 1500 rpm, shaft load torque 0.7 Nm, bearing radial force 400 N; Condition 2: shaft speed 1500 rpm, shaft load torque 0.1 Nm, bearing radial force 1000 N; Condition 3: shaft speed 900 rpm, shaft load torque 0.7 Nm, bearing radial force 1000 N.

[0093] The target domain is the data domain for which bearing fault diagnosis is required; the target domain consists of a new operating condition sample that is different from all historical operating conditions, and the fault types under the new operating condition are within the range of fault types under the historical conditions.

[0094] The target operating conditions are: shaft speed 1500 rpm, shaft load torque 0.7 Nm, and bearing radial force 1000 N.

[0095] The reference domain is composed of noise samples that do not contain any fault features, and this reference domain is used to provide a feature zero point;

[0096] In step S1, acquiring operating condition data refers to, for a bearing life test platform, when it is in service under certain operating conditions, using an acceleration sensor acquisition system to collect vibration acceleration signals that can reflect the bearing condition at the twelve o'clock position of the bearing support housing.

[0097] The bearing life test platform consists of an electric motor, a torque measurement unit, a ball bearing module, a flywheel, and a load motor.

[0098] The preprocessing methods for historical and new operating condition data in step S1 are the same; both involve envelope processing of the acquired vibration acceleration signals to extract the envelope spectrum. The specific preprocessing steps for historical and new operating condition data are as follows:

[0099] S11. The acquired vibration acceleration signal with a duration of 4 seconds is divided into 9 sample signal segments by sliding the sampling duration of 2 seconds and the window of 0.25 seconds. Then, each signal segment is filtered by 3 frequency bands to obtain the data of the segmented signal after filtering in each frequency band. Among them, multiple frequency bands are obtained by first determining the frequency range of the bearing's fault characteristic frequency, and then dividing the frequency band into segments. The 3 frequency bands are 0~3000Hz, 3000~6000Hz, and 6000~10000Hz.

[0100] S12. Perform envelope processing on the filtered signal obtained from multi-band filtering to obtain the envelope signal of each frequency band;

[0101] S13. Perform Fourier transform on the envelope signal of each frequency band to obtain the envelope spectrum;

[0102] S14. Fold the envelope spectrum of each frequency band into two-dimensional data with a shape of (64,64) to form the data of one channel; the three channels of data of a sample signal segment constitute a three-dimensional sample for the final training of the neural network.

[0103] In step S1, obtaining random noise data means randomly generating a random number that is uniformly distributed within the interval (0,1). The length of this random number is equal to the length of the vibration acceleration signal to be preprocessed. The preprocessing of the random noise data is the same as the preprocessing of historical working condition data.

[0104] S2. Establish a deep neural network for bearing fault diagnosis model, and construct three sub-modules in the deep neural network for fault classification: feature encoder, fault classifier and domain scale calculator.

[0105] In step S2,

[0106] The feature encoder is represented as:

[0107] The fault classifier is represented as:

[0108] The domain scale calculator is represented as follows:

[0109] Where Z represents the deep feature vector space; P represents the predicted sample class probability space; and W represents the domain scale space.

[0110] The feature encoder consists of four convolutional units connected sequentially. Each convolutional unit includes a convolutional operation unit for feature extraction and a downsampling unit. The feature encoder maps sample vectors to depth feature vectors. From front to back, the number of convolutional kernels in each two-dimensional convolutional operation unit is 12, 16, 20, and 24, respectively. The kernel size of each two-dimensional convolutional operation unit is (5,5).

[0111] The fault classifier is composed of four layers of perceptron neural network units connected one after the other. Each layer of perceptron neural network consists of multiple perceptrons connected in parallel. From front to back, the number of perceptrons in each layer of perceptron neural network are 96, 32, 8, and 3, respectively. The fault classifier is used to map the deep feature vector obtained by the feature encoder into a sample class probability vector. The elements of the sample class probability vector are the probability that the sample belongs to each type of fault.

[0112] The domain scaling calculator is composed of four layers of perceptron neural network units connected one after the other. Each layer of perceptron neural network consists of multiple perceptrons connected in parallel. From front to back, the number of perceptrons in each layer of perceptron neural network is 96, 32, 8, and 1, respectively. The domain scaling calculator is used to map the deep feature vector obtained by the feature encoder to a domain scaling factor. This domain scaling factor is used to scale the distribution difference between the source domain and the reference domain.

[0113] S3. Establish the learning task corresponding to the bearing fault diagnosis model and formulate the corresponding learning strategy;

[0114] The learning task of the bearing fault diagnosis model in step S3 includes three sub-tasks: source domain classification task, global distribution difference minimization task, and global regularization task.

[0115] The source domain classification task consists of classification tasks for the three source domains in step S1. These three tasks share a feature encoder, a fault classifier, and a domain scale calculator. The source domain classification task first uses the feature encoder to map the sample vectors into deep feature vectors, and then uses the fault classifier to map the deep feature vectors into sample class probability vectors. The optimization objective of the source domain classification task is expressed as:

[0116]

[0117] Where, Θ e and Θ c N represents the parameter sets of the feature encoder and the fault classifier, respectively; b N represents the number of samples in a batch, equal to 300; d This indicates the number of source domains, which is 3. This represents the j-th sample in the i-th source domain. L represents the one-hot encoded label vector corresponding to the j-th sample in the i-th source domain. ce (·,·) denotes the multivariate cross-entropy loss function, f c (·) represents the fault classifier, g e (·) represents the feature encoder, and min represents minimization;

[0118] The task of minimizing the global distribution difference aims to align the distributions of the three source domains and eliminate the influence of operating conditions on fault characteristics. The task first calculates the maximum mean difference between the distributions of each source domain and the reference domain, then calculates the domain scaling factor using a domain scaling calculator, and finally scales the maximum mean difference based on the scaling factor, and calculates the variance of all scaled mean differences. The optimization objective of the task is L. myfn Represented as:

[0119]

[0120]

[0121]

[0122] Where, Θ m The parameter set of the domain scaling calculator. This represents the j-th sample in the reference domain. h represents the maximum mean difference function of the joint distribution. m (·) indicates the domain scaling calculator, f c ′(·) denotes a subnetwork of the fault classifier;

[0123] The joint distribution maximum mean difference function The calculation of the joint distribution maximum mean difference function, used to measure the distance between two different distributions, consists of the following steps:

[0124] S31. Randomly and without repetition, extract deep feature vector pairs from the source and reference domains of the bearing fault diagnosis model: and Estimate the maximum mean difference between the marginal distributions of the source and reference domains using the following formula:

[0125]

[0126] Where K(·,·) is the kernel function, which can be specifically expressed as: In the formula, σ is the control parameter for the effective range, which is equal to 10;

[0127] S32. Extract the sample class probability vector output by the fault classifier in the bearing fault diagnosis model. in The probability of a sample being diagnosed as a type k fault is predicted. Then, using the deep feature vector from step S31, the maximum mean difference between the source and reference domain conditional distributions of the type k sample is estimated using the following formula:

[0128]

[0129] S33. Based on the maximum mean difference of the marginal distribution and the maximum mean difference of the conditional distribution, calculate the maximum mean difference of the joint distribution of the source domain and the reference domain, as shown in the following formula:

[0130]

[0131]

[0132] Where, λ j is the proportion coefficient of each type of sample, and μ is the weight coefficient of the maximum mean difference of the joint distribution;

[0133] The global regularization task is used to apply global anti-overfitting constraints to the parameter sets of the feature encoder, fault classifier, and domain scaling calculator. The optimization objective of the global regularization task is L. re Represented as:

[0134]

[0135] Where λ represents the regularization parameter, which controls the degree of regularization, ω represents the parameter in the parameter set, and n represents the parameter index;

[0136] The learning strategy described in step S3 consists of two parts: model pre-training and model generalization training.

[0137] The model pre-training uses labeled source domain samples to pre-train the bearing fault diagnosis model, enabling it to possess basic source domain classification capabilities. The optimization objective is expressed as:

[0138]

[0139] The model generalization training is used to bridge the distribution differences of multiple source domains in the deep feature space, while maintaining the classification ability obtained by the bearing fault diagnosis model during pre-training. The optimization objective is expressed as:

[0140]

[0141] Here, α and β are weight parameters used to regulate the learning process and to balance the attention of the bearing fault diagnosis model to the three sub-tasks;

[0142] S4. Optimize the hyperparameters in the bearing fault diagnosis model and determine the optimal hyperparameters;

[0143] In step S4, the hyperparameters refer to the weight parameters α and β in the model generalization training objective function in step S3. The optimization process of the hyperparameters is as follows:

[0144] S41. Specify the set T for the grid search for hyperparameters α and β respectively. α and T β , among which, T α ={0,0.05,0.1,0.2,0.4,0.8,1,2,4,8,10} and T β ={0,0.0001,0.0005,0.001,0.005,0.01,0.05,0.1,0.15,0.2};

[0145] S42. Following the method in step S1, create a set of source domain data, target domain data, and reference domain data for optimizing hyperparameters;

[0146] S43. Construct the bearing fault diagnosis model on a computer using the Python programming language;

[0147] S44. Select Adam optimizer as the learning rate optimizer, batch size of training is 300 and number of training epochs is 200. For each combination of hyperparameter α and hyperparameter β, train the model using source domain data and reference domain data respectively according to the learning strategy in step S3.

[0148] S45. After training the model for each hyperparameter combination, use the trained bearing fault diagnosis model to predict faults in the sample data of the target domain, calculate the prediction accuracy, and select the hyperparameter combination with the highest accuracy as the optimal hyperparameters; Figure 5 As shown, the optimal hyperparameter combination is: α = 0.8, β = 0.01, and the target domain test accuracy under this hyperparameter combination is 92%.

[0149] S5. Substitute the optimal hyperparameters into the fault classification deep neural network, train the network, and obtain the optimal bearing fault diagnosis model.

[0150] The specific process of step S5 is as follows:

[0151] S51. Following the method in step S1, create a new set of source domain data and reference domain data;

[0152] S52. Substitute the obtained optimal hyperparameters into the bearing fault diagnosis model created in step S4.

[0153] S53. Select the learning rate optimizer, training batch size, and number of training rounds. Use the newly created data to train the bearing fault diagnosis model and obtain the optimal bearing fault diagnosis model.

[0154] S6. Create bearing monitoring target domain data that requires fault diagnosis, and use the best bearing fault diagnosis model to classify the target domain data for faults and determine the fault type.

[0155] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A bearing fault diagnosis method based on a deep domain generalization network, characterized in that, The method comprises the following steps: S1, obtaining historical working condition data, new working condition data and random noise data, and after preprocessing, obtaining multiple source domains, target domains and reference domains; In step S1, The source domain is represented as: D s,i = {X s,i , Y s,i , P s,i (x, y)}, i = 1, 2, 3,..., N d ; The target domain is represented as: D t = {X t , P t (x)}; The reference domain is represented as: D r = {X r , P r (x)}; where s is the source domain identification letter; t is the target domain identification letter; r is the reference domain identification letter; i represents the working condition serial number; N d represents the total number of historical working conditions, and also represents the number of source domains; X represents the sample space, and the sample in the domain is a three-dimensional vector; Y represents the label space, and the sample in the domain is a scalar; P(x) represents the probability distribution of the sample; D s,i represents the i-th source domain, which is composed of samples under the i-th historical working condition; The historical working condition data is greater than or equal to three, that is, N d ≥ 3; the fault type under each historical working condition is greater than or equal to three, and the fault types under all historical working conditions are the same; the target domain is a data domain that needs to be subjected to bearing fault diagnosis; the target domain is composed of a new working condition sample different from all historical working conditions, and the fault type under the new working condition is within the range of fault types under the historical working conditions; the reference domain is composed of noise samples that do not contain any fault features, and the reference domain is used to provide a feature zero point; S2, establishing a fault classification deep neural network of the bearing fault diagnosis model, and constructing three submodules in the fault classification deep neural network: a feature encoder, a fault classifier and a domain scale calculator; S3, establishing a learning task corresponding to the bearing fault diagnosis model, and formulating a corresponding learning strategy; S4, optimizing the hyperparameters in the bearing fault diagnosis model to determine the best hyperparameters; S5, substituting the best hyperparameters into the fault classification deep neural network, training the network, and obtaining the best bearing fault diagnosis model; S6, creating a bearing monitoring target domain data that needs to be diagnosed, and using the best bearing fault diagnosis model to complete fault classification on the target domain data to determine the fault type.

2. The bearing fault diagnosis method based on the deep domain generalization network according to claim 1, wherein, The preprocessing methods of the historical working condition data and the new working condition data in step S1 are the same, which are both envelope processing of the collected vibration acceleration signals, and extracting envelope spectrum. The specific process of preprocessing the historical working condition data and the new working condition data is as follows: S11, the vibration acceleration signals collected in a certain time period are divided into multiple sample signal segments according to the set sample length, and then each signal segment is filtered in multiple frequency bands to obtain the data of the segmented signal after individual filtering in each frequency band; wherein the multiple frequency bands are obtained by first determining the frequency band range of the bearing fault characteristic frequency, and then segmenting the frequency band; S12, envelope processing is performed on the filtered signals obtained by multiple frequency band filtering to obtain envelope signals of each frequency band; S13, Fourier transform is performed on the envelope signals of each frequency band to obtain envelope spectrum; S14, fold each frequency band envelope spectrum into two-dimensional data to form a channel of data; multiple channel data of a sample signal segment form a three-dimensional sample for training a neural network.

3. The bearing fault diagnosis method based on the deep domain generalization network according to claim 2, characterized in that: In step S1, the random noise data refers to a random number uniformly distributed in the interval (0, 1), and the length of the random number is equal to the length of the vibration acceleration signal to be preprocessed; the preprocessing of the random noise data is the same as the preprocessing of the historical working condition data.

4. The bearing fault diagnosis method based on deep domain generalization network according to claim 1, characterized in that: In step S2, The feature encoder is represented as: The fault classifier is represented as: The domain scale calculator is represented as: Wherein, Z represents a deep feature vector space; P represents a predicted sample class probability space; W represents a domain scale space; The feature encoder is connected by multiple layers of convolution units, each layer of convolution unit includes a convolution operation unit for feature extraction and a down-sampling unit, and the feature encoder is used to map the sample vector to the deep feature vector; The fault classifier is connected by multiple layers of perceptron neural network units, each layer of perceptron neural network is composed of multiple perceptrons in parallel, and the fault classifier is used to map the deep feature vector obtained by the feature encoder to a sample class probability vector; the elements of the sample class probability vector are the probabilities of the sample belonging to each fault; The domain scale calculator is connected by multilayer perceptron neural network units in front and back, each layer of perceptron neural network is composed of multiple perceptrons in parallel, and the domain scale calculator is used for mapping the deep feature vector obtained by the feature encoder into a domain scale factor, and the domain scale factor is used for scaling the distribution difference between the source domain and the reference domain.

5. The bearing fault diagnosis method based on deep domain generalization network according to claim 1, characterized in that: The learning task of the bearing fault diagnosis model in step S3 includes three subtasks, which are respectively: a source domain classification task, a global distribution difference minimization task, and a global regularization task; The source domain classification task is composed of the classification tasks of all source domains in step S1, and the multiple classification tasks share the feature encoder, the fault task classifier, and the domain scale calculator; the source domain classification task first maps the sample vector into a deep feature vector using the feature encoder, and then maps the deep feature vector into a sample class probability vector using the fault classifier, and the optimization objective of the source domain classification task is represented as: where Θ e and Θ c denote the parameter sets of the feature encoder and the fault classifier, respectively, N b denotes the number of samples in a batch, denotes the jth sample in the ith source domain, denotes the one-hot encoded label vector corresponding to the jth sample in the ith source domain, L ce (·,·) denotes the multi-class cross-entropy loss function, f c (·) denotes the fault classifier, g e (·) denotes the feature encoder, and min denotes minimization. The global distribution difference minimization task is used to align the distributions of the plurality of source domains and eliminate the influence of working conditions on fault features; the global distribution difference minimization task first calculates the maximum mean difference between each source domain distribution and the reference domain distribution, then calculates a domain scale factor by a domain scale calculator, finally scales the maximum mean difference according to the scale factor, and calculates the variance of all scaled mean differences; and the optimization objective L of the global distribution difference minimization task is myfn is represented as: where Θ m denotes a parameter set of the domain scale calculator, denotes the j-th sample in the reference domain, denotes the joint distribution maximum mean discrepancy function, h m (·) denotes the domain scale calculator, f c ′(·) denotes a subnetwork of the fault classifier; The joint distribution maximum mean discrepancy function For measuring the distance between two different distributions, the calculation process of the joint distribution maximum mean discrepancy function consists of the following steps: S31, randomly and without repetition, extract a pair of deep feature vectors of the source domain and the reference domain from the bearing fault diagnosis model: and The maximum mean difference of the source domain and the reference domain edge distribution is estimated as follows: Wherein, K(·,·) is a kernel function, which can be specifically represented as In the formula, σ is an action range control parameter. S32, extract the sample class probability vector output by the fault classifier in the bearing fault diagnosis model wherein The probability of the sample being diagnosed as the kth fault is predicted, and then the maximum mean difference of the source domain and the reference domain conditional distribution of the kth sample is estimated according to the following formula using the deep feature vector in step S31: S33, according to the maximum mean difference of the marginal distribution and the maximum mean difference of the conditional distribution, the maximum mean difference of the joint distribution of the source domain and the reference domain is calculated, as shown in the following formula: where λ j is the proportionality coefficient of each type of sample, and μ is the weight coefficient of the maximum mean difference of the joint distribution. The global regularization task is configured to impose a global overfitting prevention constraint on the parameter set of the feature encoder, the failure classifier, and the domain scale calculator, and an optimization objective L of the global regularization task is represented as: re is represented as: Wherein, λ represents a regularization parameter, used to control the regularization degree, ω represents a parameter in the parameter set, and n represents the serial number of the parameter.

6. The bearing fault diagnosis method based on deep domain generalization network according to claim 5, characterized in that: The learning strategy in step S3 is composed of model pre-training and model generalization training; The model pre-training pre-trains the bearing fault diagnosis model using the labeled source domain samples, so that the bearing fault diagnosis model has the source domain classification ability, and the optimization objective is represented as: The model generalization training is used to bridge the distribution difference of multiple source domains in the deep feature space, while maintaining the classification ability obtained by the bearing fault diagnosis model in the pre-training, and the optimization objective is represented as: Wherein, α and β are weight parameters used to regulate the learning process, and are used to balance the attention of the bearing fault diagnosis model on the three subtasks.

7. The bearing fault diagnosis method based on deep domain generalization network according to claim 6, characterized in that: The hyperparameters in step S4 are the weight parameters α and β in the model generalization training objective function in step S3, and the optimization process of the hyperparameters is specifically: S41. Specify a set T of grid searches for hyperparameter a and hyperparameter b, respectively α and T β , S42, according to the method in step S1, a set of optimized hyperparameters is created using source domain data, target domain data and reference domain data; S43, the bearing fault diagnosis model is built on the computer using a computer programming language; S44, select the learning rate optimizer, the batch size of training and the number of training rounds, for each combination of hyperparameters α and hyperparameters β, use the source domain data and the reference domain data to train the model according to the learning strategy in step S3; S45, after the model training under each hyperparameter combination is completed, the trained bearing fault diagnosis model is used to make fault prediction on the sample data in the target domain, the prediction accuracy is counted, and the hyperparameter combination with the highest accuracy is selected as the best hyperparameter.

8. The bearing fault diagnosis method based on deep domain generalization network according to claim 7, characterized in that: The specific process of step S5 is: S51, according to the method in step S1, a new set of source domain data and reference domain data is created; S52, the best hyperparameters obtained are substituted into the bearing fault diagnosis model created in step S4; S53, select the learning rate optimizer, the batch size of training and the number of training rounds, and use the newly created data to train the bearing fault diagnosis model to obtain the best bearing fault diagnosis model.

Images