Intelligent fault diagnosis method and system for rotating machinery

Abstract

The invention belongs to the technical field related to intelligent operation and maintenance of industrial equipment, and discloses an intelligent fault diagnosis method and system for rotating machinery, and the method comprises the steps: S1, constructing a training sample set which comprises sample signals of a plurality of source domains, and each sample signal comprises a vibration signal, an acoustic signal and a fault label; s2, constructing a fault diagnosis model: a vibration feature extractor used for extracting vibration features, an acoustic feature extractor used for extracting acoustic features, a feature fusion module used for fusing the vibration features and the acoustic features to obtain connection features, and a feature mapping module used for extracting unified features from the connection features, the classification module is used for carrying out fault mode classification according to the extracted features; s3, training the fault diagnosis model based on the training sample set; and S4, performing online fault diagnosis by using the trained fault diagnosis model. According to the method, multi-source domain vibration and acoustic signals are utilized, invariant unified fault representation is learned from the multi-source domain vibration and acoustic signals, and the cross-domain diagnosis precision of the model is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field related to intelligent operation and maintenance of industrial equipment, and more specifically, relates to an intelligent fault diagnosis method and system for rotating machinery. Background Art

[0002] Rotating machinery is widely used in modern industry. Rotating machinery failures can cause serious safety hazards and lead to huge maintenance costs. In order to ensure the normal operation of equipment, equipment operation and maintenance departments are often set up in industrial production to perform equipment fault diagnosis and operation and maintenance. Equipment fault diagnosis is also gradually shifting from traditional signal analysis methods to intelligent diagnosis methods based on machine learning.

[0003] In recent years, with the development of sensors and artificial intelligence technology, deep learning has been increasingly applied to mechanical fault diagnosis and has achieved good results. Common diagnostic models include convolutional neural networks, recursive neural networks, and autoencoders. These deep learning models with good results are inseparable from sufficient labeled training samples. That is, the existing intelligent fault diagnosis technology relies on sufficient labeled training data, and the collected samples need to meet the same distribution condition. In order to accurately realize fault diagnosis, it is usually necessary to collect samples of the target working conditions for training for the target working conditions. This also makes the trained diagnostic model only able to make accurate judgments for specific target working conditions. Once the equipment working conditions change, it is difficult for the model to make accurate judgments.

[0004] Due to the diverse operating conditions of machines and equipment and the different physical characteristics of different machines and equipment, the difference in data distribution during the training stage and the application stage will reduce the effectiveness of the diagnosis model. That is, most of the existing intelligent fault diagnosis methods have poor generalization ability and have the problem of poor diagnostic accuracy when the data distribution is different during the training stage and the application stage. Summary of the invention

[0005] In view of the above defects or improvement needs of the prior art, the present invention provides an intelligent fault diagnosis method and system for rotating machinery, which is used to solve the problem that most of the existing intelligent fault diagnosis methods have poor generalization ability and poor diagnostic accuracy when the data distribution is different in the training stage and the application stage.

[0006] To achieve the above object, according to one aspect of the present invention, there is provided an intelligent fault diagnosis method for rotating machinery, comprising:

[0007] Offline data processing and training phase:

[0008] S1, constructing a training sample set, wherein the training sample set includes sample signals of multiple source domains, each sample signal includes a vibration signal, an acoustic signal and a fault label;

[0009] S2, constructing a fault diagnosis model including: a vibration feature extractor, an acoustic feature extractor, a feature fusion module, a feature mapping module and a classification module, wherein the vibration feature extractor is used to extract vibration features from the vibration signal, the acoustic feature extractor is used to extract acoustic features from the acoustic signal, the feature fusion module is used to fuse the vibration features and the acoustic features to obtain connection features, the feature mapping module is used to extract unified features from the connection features, and the classification module is used to classify fault modes according to the vibration features, the acoustic features and the unified features;

[0010] S3, training the fault diagnosis model based on the training sample set to obtain a trained fault diagnosis model;

[0011] Online fault diagnosis stage:

[0012] S4, using the trained fault diagnosis model to perform online fault diagnosis on rotating machinery.

[0013] According to the intelligent fault diagnosis method for rotating machinery provided by the present invention, the feature fusion module in S2 is used to fuse the vibration features and the acoustic features from the same source domain to obtain same-domain connection features, and to fuse the vibration features and the acoustic features from different source domains to obtain cross-domain connection features;

[0014] Correspondingly, the feature mapping module is used to extract the same-domain unified feature from the same-domain connection feature and to extract the cross-domain unified feature from the cross-domain connection feature.

[0015] According to the intelligent fault diagnosis method for rotating machinery provided by the present invention, the same domain connection feature z id Specifically:

[0016]

[0017] The cross-domain connection feature z cd Specifically:

[0018]

[0019] in, is the vibration feature belonging to the mth category of the ith source domain; is the acoustic feature belonging to the mth class in the jth source domain.

[0020] According to the intelligent fault diagnosis method for rotating machinery provided by the present invention, S3 includes: training the feature mapping module with the goal of maximizing the mutual information between the same-domain unified features and the cross-domain unified features.

[0021] According to the intelligent fault diagnosis method for rotating machinery provided by the present invention, the feature mapping module is trained with the goal of maximizing the mutual information between the same-domain unified feature and the cross-domain unified feature, specifically comprising:

[0022] The same domain unified features s id and the cross-domain unified features cd The mutual information between them is shown as follows:

[0023]

[0024] Where I represents mutual information; is the expectation; W and N are the batch sizes of samples; f is the evaluation function; is the unified feature of the i-th domain; is the jth cross-domain unified feature;

[0025] Mutual information maximization is achieved by maximizing the lower bound of mutual information, which uses the supervised contrast loss shown in the following formula The feature mapping module is trained:

[0026]

[0027] Where k∈K={1,2,…,2W}; A(k)≡K / k; P(k) is the feature set of positive samples with the same label as sample k; s=s id ∪s cd ;s i is the i-th feature; s p is the pth feature; s a is the negative sample feature; τ is the scaling parameter.

[0028] According to the intelligent fault diagnosis method for rotating machinery provided by the present invention, the classification module in S2 includes a vibration state classifier, an acoustic state classifier and a combined state classifier, the vibration state classifier is used to classify fault modes according to the vibration characteristics, the acoustic state classifier is used to classify fault modes according to the acoustic characteristics, and the combined state classifier is used to classify fault modules according to the unified characteristics, and the output result of the classification module is obtained by applying the majority voting method to the three state classifiers.

[0029] According to the intelligent fault diagnosis method for rotating machinery provided by the present invention, S3 also includes:

[0030] Obtain classification losses respectively according to the classification results of the vibration state classifier, the acoustic state classifier and the combined state classifier, and then obtain a total classification loss;

[0031] The vibration state classifier, the acoustic state classifier, and the combined state classifier are trained based on the total classification loss.

[0032] According to the intelligent fault diagnosis method for rotating machinery provided by the present invention, the fault diagnosis model in S2 further includes a modal converter, and the modal converter is used to generate a false acoustic feature according to the vibration feature, and to generate a false vibration feature according to the acoustic feature;

[0033] Correspondingly, S4 includes: acquiring a vibration signal and / or an acoustic signal of the rotating machinery, and performing online fault diagnosis based on the vibration signal and / or the acoustic signal using a trained fault diagnosis model.

[0034] According to the intelligent fault diagnosis method for rotating machinery provided by the present invention, S3 also includes: using the semantic supervision loss shown in the following formula Train the modality converter:

[0035]

[0036] Among them, L CE is the cross entropy function; is the class probability distribution of the vibration state classifier based on the false acoustic feature output; is the class probability distribution of the acoustic state classifier based on the false vibration feature output; For labels; and The parameters representing the vibration state classifier and the acoustic state classifier are not updated during the training process.

[0037] According to another aspect of the present invention, there is provided an intelligent fault diagnosis system for rotating machinery, the system comprising a memory and a processor, the memory storing a computer program, and the processor executing any one of the above-mentioned intelligent fault diagnosis methods for rotating machinery when executing the computer program.

[0038] In general, compared with the prior art, the above technical solutions conceived by the present invention provide an intelligent fault diagnosis method and system for rotating machinery:

[0039] 1. Using vibration signals and acoustic signals simultaneously in the model training phase is conducive to improving the resistance of the fault diagnosis model to data distribution differences. Using multiple sensors to collect signals of different modes (vibration and acoustics) in the test phase is conducive to comprehensive monitoring of the equipment. The feature fusion module is designed to connect the vibration features and acoustic features so that the connected features cover more information, overcoming the limitations of single-modal information. The features are then mapped into unified features by the feature mapping module, which is conducive to better learning the invariant features. The potential commonalities in multi-modal and multi-domain data are learned from a decoupling perspective, and an invariant unified feature space is explored to capture robust features from cross-domain data of different modes, thereby improving the generalization ability of the model and maintaining a high diagnostic accuracy in cross-domain diagnosis.

[0040] 2. The design of the feature fusion module to generate intra-domain connection features and cross-domain connection features increases the diversity of features and improves the utilization rate of useful information;

[0041] 3. The unified features of the same domain after the same-domain connection feature mapping maintain consistent potential domain-specific information, making them similar to each other and closer in the unified feature space. On the contrary, the cross-domain unified features after the cross-domain connection feature mapping contain different domain-specific information, resulting in relatively scattered positions in the space. By maximizing the mutual information between the same-domain and cross-domain unified features, the domain-specific information is effectively removed and the intra-class differences are reduced. The feature mapping module projects the connected multimodal features onto the invariant unified features by maximizing the mutual information between the same-domain and cross-domain unified features. The invariant unified feature space fully maintains the features related to the fault and robust to domain migration, and removes redundant information that may interfere with the model decision, thereby improving the model generalization ability and diagnostic accuracy.

[0042] 4. Set up three state classifiers and propose a hybrid integrated diagnosis strategy, which is conducive to improving the flexibility and adaptability of the diagnosis model to various tasks and environments;

[0043] 5. Equipped with a modal converter to complete the conversion between modalities when certain modal signals are missing, the modal converter can generate false data of the missing modality based on the available monitoring data. Through semantic supervision, the generated synthetic data can be correctly identified by its respective state classifier, which improves the robustness and applicability of the diagnostic model in the face of various external restrictions. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a model structure diagram of the intelligent fault diagnosis method for rotating machinery provided by the present invention;

[0045] Figure 2 is a schematic diagram of the information enhancement connection method provided by the present invention;

[0046] Figure 3 is a schematic diagram of mutual information maximization provided by the present invention;

[0047] Figure 4 It is a schematic diagram of characteristic mode conversion provided by the present invention. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0049] See also Figure 1 , this embodiment provides an intelligent fault diagnosis method for rotating machinery, the method comprising:

[0050] Offline data processing and training phase:

[0051] S1, constructing a training sample set, wherein the training sample set includes sample signals of multiple source domains, each sample signal includes a vibration signal, an acoustic signal and a fault label;

[0052] S2, constructing a fault diagnosis model including: a vibration feature extractor, an acoustic feature extractor, a feature fusion module, a feature mapping module and a classification module, wherein the vibration feature extractor is used to extract vibration features from the vibration signal, the acoustic feature extractor is used to extract acoustic features from the acoustic signal, the feature fusion module is used to fuse the vibration features and the acoustic features to obtain connection features, the feature mapping module is used to extract unified features from the connection features, and the classification module is used to classify fault modes according to the vibration features, the acoustic features and the unified features;

[0053] S3, training the fault diagnosis model based on the training sample set to obtain a trained fault diagnosis model;

[0054] Online fault diagnosis stage:

[0055] S4, using the trained fault diagnosis model to perform online fault diagnosis on rotating machinery.

[0056] In this embodiment, the training sample set is constructed for rotating machinery, which can be, for example, bearings, gearboxes, motors, etc.; different source domains correspond to different working conditions and / or different devices used for rotating machinery, so different source domain data can be obtained by changing the working conditions on one device, or by changing the device to collect and obtain source domain data on different devices, as long as it is sample data for the same type of rotating machinery. Therefore, when the data of some equipment is not easy to obtain, the sample set can be obtained from other equipment with the same type of rotating machinery, and the model obtained by training is applicable to the same type of rotating machinery, regardless of the specific equipment and working conditions.

[0057] Specifically, the feature fusion module in S2 is used to fuse the vibration feature and the acoustic feature from the same source domain to obtain a same-domain connection feature, and to fuse the vibration feature and the acoustic feature from different source domains to obtain a cross-domain connection feature;

[0058] Correspondingly, the feature mapping module is used to extract the same-domain unified feature from the same-domain connection feature and to extract the cross-domain unified feature from the cross-domain connection feature.

[0059] The classification module in S2 includes a vibration state classifier, an acoustic state classifier and a combined state classifier. The vibration state classifier is used to classify fault modes according to the vibration characteristics, the acoustic state classifier is used to classify fault modes according to the acoustic characteristics, and the combined state classifier is used to classify fault modules according to the unified characteristics. The output result of the classification module is obtained by applying the majority voting method to the three state classifiers.

[0060] In order to improve the flexibility and adaptability of the diagnostic model to various tasks and environments, this embodiment proposes a hybrid integrated diagnostic strategy equipped with three state classifiers: vibration state classifier, acoustic state classifier and combined state classifier. First, the invariant unified features output by the mapper are fed into the combined state classifier to capture the relationship between multimodal monitoring data and the health status of the machine. The vibration state classifier and the acoustic state classifier also mine the potential knowledge in their respective modal data. The result of the hybrid integrated diagnosis is obtained by applying the majority voting method to the three different state classifiers.

[0061] The fault diagnosis model in S2 further includes a modal converter, wherein the modal converter is used to generate a false acoustic feature according to the vibration feature, and to generate a false vibration feature according to the acoustic feature;

[0062] Correspondingly, S4 includes: acquiring a vibration signal and / or an acoustic signal of the rotating machinery, and performing online fault diagnosis based on the vibration signal and / or the acoustic signal using a trained fault diagnosis model.

[0063] This embodiment specifically proposes a generalization and conversion method of sound and vibration integration for intelligent fault diagnosis of rotating machinery. Figure 1 As shown, the fault diagnosis model includes a vibration feature extractor F v , acoustic feature extractor F a , vibration state classifier C v , acoustic state classifier C a , combined state classifier C c , modal converter T av and T va , feature fusion module, the feature mapping module is specifically an integrated feature mapper P n and P u .

[0064] During the model training phase, the vibration feature extractor F v and acoustic feature extractor F a The original vibration signal x v and the original acoustic signal x a As input for feature extraction, the feature extraction is shown as follows:

[0065] z v =F v (x v );

[0066] z a =F a (x a );

[0067] Among them, z v is the high-dimensional vibration characteristic; z a The high dimension is the acoustic features.

[0068] Furthermore, information enhancement connection and mutual information maximization are designed to collaboratively learn invariant unified features. Information enhancement connection is to generate same-domain connection features and cross-domain connection features through feature fusion module. In order to increase the diversity of features and improve the utilization of useful information, same-domain connection features of different modal features from the same field and cross-domain connection features of different modal features from different fields are generated.

[0069] The feature fusion module is used to splice the vibration feature and the acoustic feature for fusion. Splicing together high-dimensional features from different modalities is a simple and effective method to overcome the limitations of single-modal information, thereby improving the performance of the diagnostic model. In order to increase the diversity of features and enhance the utilization of useful information, in addition to generating same-domain connection features of different modal features from the same field, cross-domain connection features of different modal features from different fields are also generated. Ultimately, through the synergy of information enhancement connection and mutual information maximization, an invariant unified feature is obtained, which enhances the resistance of the diagnostic model to data distribution differences.

[0070] Specifically, Figure 2 As shown, the same domain connection feature z after fusion id Specifically:

[0071]

[0072] The cross-domain connection feature z cd Specifically:

[0073]

[0074] in, is the vibration feature belonging to the mth category of the ith source domain; is the acoustic feature belonging to the mth class in the jth source domain.

[0075] Furthermore, the integrated feature mapper P n and P u Connect the features z in the same domain id and cross-domain connection feature z cd The mapping to a unified feature is as follows:

[0076] s id =P n (z id );

[0077] s cd =P u (z cd );

[0078] Among them, s id is the same domain connection feature z id Features in a unified feature space after being mapped by the integrated feature mapper. cd is the cross-domain connection feature z cd Features in a unified feature space after being mapped by the all-in-one feature mapper.

[0079] Furthermore, fault-related, domain-specific, and modality-specific information is disentangled by maximizing the mutual information between same-domain and cross-domain connected features in a unified feature space.

[0080] That is, S3 includes: training the feature mapping module with the goal of maximizing the mutual information between the same-domain unified features and the cross-domain unified features.

[0081] Specifically, the feature mapping module is trained with the goal of maximizing the mutual information between the same-domain unified feature and the cross-domain unified feature, specifically including:

[0082] The same domain unified features s id and the cross-domain unified features cd The mutual information between them is shown as follows:

[0083]

[0084] Since it is not feasible to directly calculate the mutual information, maximizing the mutual information is usually done by increasing its lower bound instead. Where I represents the mutual information; is the expectation; W and N are the batch sizes of samples; f is the evaluation function; is the unified feature of the i-th domain; is the jth cross-domain unified feature.

[0085] Mutual information maximization is achieved by maximizing the lower bound of mutual information. In order to maximize the lower bound of mutual information, the supervised contrast loss is used as shown in the following formula To increase the mutual information between features, such as Figure 3 As shown, based on supervised contrast loss The feature mapping module is trained:

[0086]

[0087] Where k∈K={1,2,…,2W}; A(k)≡K / k; P(k) is the feature set of positive samples with the same label as sample k; s=s id ∪s cd ;s i is the i-th feature; s p is the pth feature; s a is the negative sample feature; τ is the scaling parameter.

[0088] That is, the feature mapping module is trained with the goal of maximizing the mutual information between the same-domain unified features and the cross-domain unified features, specifically including: using supervised contrast learning to train the feature mapping module, and iteratively optimizing the parameters of the feature mapping module based on supervised contrast loss.

[0089] Furthermore, S3 also includes:

[0090] Obtain classification losses respectively according to the classification results of the vibration state classifier, the acoustic state classifier and the combined state classifier, and then obtain a total classification loss;

[0091] The vibration state classifier, the acoustic state classifier, and the combined state classifier are trained based on the total classification loss.

[0092] Specifically, first, the mapper P n and P u The output unified invariant features are input to the combined state classifier C c to capture the relationship between multimodal monitoring data and machine health status:

[0093]

[0094] in, is the loss function of the combined state classifier; L CE is the cross entropy loss function; C c (S) is the prediction result of the combined state classifier; is the label; S is s id and cd A collection of .

[0095] Vibration state classifier C v and acoustic state classifier C a The potential knowledge in their respective modal data is also mined:

[0096]

[0097]

[0098] in, is the loss function of the vibration state classifier; C v (F v (X v )) is the prediction result of the vibration state classifier; is the loss function of the acoustic state classifier; C a (F a (X a )) is the prediction result of the acoustic state classifier;

[0099] Total classification loss Defined as:

[0100]

[0101] The result of the hybrid ensemble diagnosis is obtained by majority voting of three different state classifiers and is defined as follows:

[0102]

[0103] Among them, H(x) is the final prediction result of sample x; m is the mth dimension in the label space; is the probability of the mth dimension of the vibration state classifier based on the vibration feature output; is the probability of the mth dimension output by the acoustic state classifier based on the acoustic features; is the probability of the combined state classifier for the mth dimension of the unified feature.

[0104] like Figure 4 As shown, this embodiment is equipped with a modal converter T in the model av and T va , to achieve inter-modal conversion when some modes are missing. Real vibration characteristics False vibration characteristics Real acoustic characteristics and false acoustic features It can be calculated as follows:

[0105]

[0106] in, It is a high-dimensional vibration feature; is a high-dimensional acoustic feature; It is a false acoustic feature, which is converted from high-dimensional vibration signals by the modal converter; It is a false vibration feature, which is converted from high-dimensional acoustic signals by the modal converter.

[0107] S3 also includes: using the semantic supervision loss as shown below Train the modality converter:

[0108]

[0109] Among them, L CE is the cross entropy function; is the class probability distribution of the vibration state classifier based on the false acoustic feature output; is the class probability distribution of the acoustic state classifier based on the false vibration feature output; For labels; and The parameters representing the vibration state classifier and the acoustic state classifier are not updated during the training process. In this way, only the modality conversion translator is penalized during the training process to establish the connection between different modalities.

[0110] Furthermore, this embodiment provides an intelligent fault diagnosis system for rotating machinery, the system comprising a memory and a processor, the memory storing a computer program, and the processor executing any one of the above-mentioned intelligent fault diagnosis methods for rotating machinery when executing the computer program.

[0111] The present invention discloses a method for generalizing and converting sound and vibration for intelligent fault diagnosis of rotating machinery, which comprises two signal input modules (a vibration feature extractor and an acoustic feature extractor), a feature fusion module, a feature extraction module, a modal converter and a state recognition module (a vibration state classifier, an acoustic state classifier and a combined state classifier). The two signal input modules extract features from vibration signals and acoustic signals respectively. The feature fusion module and the feature extraction module aggregate vibration and acoustic features into unified features based on information enhancement connection and mutual information maximization. The modal converter converts vibration and acoustic features into each other. The state recognition module integrates multiple output results to determine the health status of the equipment. This method effectively utilizes the vibration and acoustic information in rotating equipment monitoring, and is expected to improve the accuracy of equipment diagnosis.

[0112] The technical problem to be solved by this method is the fault diagnosis problem based on multimodal domain generalization, that is, the vibration signal and the acoustic signal are used simultaneously in the training stage to improve the resistance of the diagnosis model to distribution differences and the robustness to sensor failures; its purpose is to use multi-source domain vibration and acoustic signals to learn an invariant unified fault representation from them, thereby improving the diagnostic accuracy of the model. At the same time, the model training data usually uses data from different operating conditions or other machines of the same type. The fault diagnosis model based on domain generalization is trained offline on multimodal multi-source data (vibration and acoustic signals from different working conditions or equipment). After the training is completed, the model is applied to online equipment monitoring. The method of the present invention has low requirements for training data.

[0113] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An intelligent fault diagnosis method for rotating machinery, characterized in that: include: Offline data processing and training phase: S1, constructing a training sample set, wherein the training sample set includes sample signals of multiple source domains, each sample signal includes a vibration signal, an acoustic signal and a fault label; S2, constructing a fault diagnosis model including: a vibration feature extractor, an acoustic feature extractor, a feature fusion module, a feature mapping module and a classification module, wherein the vibration feature extractor is used to extract vibration features from the vibration signal, the acoustic feature extractor is used to extract acoustic features from the acoustic signal, the feature fusion module is used to fuse the vibration features and the acoustic features to obtain connection features, the feature mapping module is used to extract unified features from the connection features, and the classification module is used to classify fault modes according to the vibration features, the acoustic features and the unified features; S3, training the fault diagnosis model based on the training sample set to obtain a trained fault diagnosis model; Online fault diagnosis stage: S4, using the trained fault diagnosis model to perform online fault diagnosis on rotating machinery.

2. The intelligent fault diagnosis method for rotating machinery according to claim 1, characterized in that: The feature fusion module in S2 is used to fuse the vibration feature and the acoustic feature from the same source domain to obtain a same-domain connection feature, and to fuse the vibration feature and the acoustic feature from different source domains to obtain a cross-domain connection feature; Correspondingly, the feature mapping module is used to extract the same-domain unified feature from the same-domain connection feature and to extract the cross-domain unified feature from the cross-domain connection feature.

3. The intelligent fault diagnosis method for rotating machinery according to claim 2, characterized in that: The same domain connection feature z id Specifically: The cross-domain connection feature z cd Specifically: in, is the vibration feature belonging to the mth category of the ith source domain; is the acoustic feature belonging to the mth class in the jth source domain.

4. The intelligent fault diagnosis method for rotating machinery according to claim 2, characterized in that: S3 includes: training the feature mapping module with the goal of maximizing the mutual information between the same-domain unified features and the cross-domain unified features.

5. The intelligent fault diagnosis method for rotating machinery according to claim 4, characterized in that: The feature mapping module is trained with the goal of maximizing the mutual information between the same-domain unified features and the cross-domain unified features, specifically including: The same domain unified features s id and the cross-domain unified features cd The mutual information between them is shown as follows: Where I represents mutual information; is the expectation; W and N are the batch sizes of samples; f is the evaluation function; is the unified feature of the i-th domain; is the jth cross-domain unified feature; Mutual information maximization is achieved by maximizing the lower bound of mutual information, which uses the supervised contrast loss shown in the following formula The feature mapping module is trained: Where k∈K={1,2,…,2W}; A(k)≡K / k; P(k) is the feature set of positive samples with the same label as sample k; s=s id ∪s cd ;s i is the i-th feature; s p is the pth feature; s a is the negative sample feature; τ is the scaling parameter.

6. The intelligent fault diagnosis method for rotating machinery according to any one of claims 1 to 5, characterized in that: The classification module in S2 includes a vibration state classifier, an acoustic state classifier and a combined state classifier. The vibration state classifier is used to classify fault modes according to the vibration characteristics, the acoustic state classifier is used to classify fault modes according to the acoustic characteristics, and the combined state classifier is used to classify fault modules according to the unified characteristics. The output result of the classification module is obtained by applying the majority voting method to the three state classifiers.

7. The intelligent fault diagnosis method for rotating machinery according to claim 6, characterized in that: S3 also includes: Obtain classification losses respectively according to the classification results of the vibration state classifier, the acoustic state classifier and the combined state classifier, and then obtain a total classification loss; The vibration state classifier, the acoustic state classifier, and the combined state classifier are trained based on the total classification loss.

8. The intelligent fault diagnosis method for rotating machinery according to claim 6, characterized in that: The fault diagnosis model in S2 further includes a modal converter, wherein the modal converter is used to generate a false acoustic feature according to the vibration feature, and to generate a false vibration feature according to the acoustic feature; Correspondingly, S4 includes: acquiring a vibration signal and / or an acoustic signal of the rotating machinery, and performing online fault diagnosis based on the vibration signal and / or the acoustic signal using a trained fault diagnosis model.

9. The intelligent fault diagnosis method for rotating machinery according to claim 8, characterized in that: S3 also includes: using the semantic supervision loss as shown below Train the modality converter: Among them, L CE is the cross entropy function; is the class probability distribution of the vibration state classifier based on the false acoustic feature output; is the class probability distribution of the acoustic state classifier based on the false vibration feature output; For labels; and The parameters representing the vibration state classifier and the acoustic state classifier are not updated during the training process.

10. An intelligent fault diagnosis system for rotating machinery, characterized in that: The system includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the intelligent fault diagnosis method for rotating machinery described in any one of claims 1 to 9 when executing the computer program.

Images