A stable intelligent fault diagnosis method with strong robustness to changes in equipment operating conditions

Abstract

The present invention discloses a stable intelligent fault diagnosis method with strong robustness to the change of equipment operating conditions, which relates to the technical field of equipment online detection. It includes the following detection steps: constructing a density clustering subnetwork to pre-label unlabeled online monitoring data, and establishing a feature extraction main network with an incremental sparse pseudo-label matrix; from the perspective of network fast calculation mode and network fast update architecture, respectively constructing a binary neural network and a width network adapted to the fault diagnosis task; extracting equipment operating condition characteristics; reasonably designing fault components and rotor system fault tests, and continuously verifying and improving the proposed new theoretical method with the actual measured signal data of the rotor system, based on theories and new methods such as width learning, adversarial generative networks, and transfer learning, aiming at the requirements of online monitoring of equipment health status, finally forming an online fault diagnosis method that can be stably diagnosed and updated in real time, and obtaining higher agility and stability of intelligent fault diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment online detection, and in particular to a stable intelligent fault diagnosis method with strong robustness for equipment operating condition changes. Background Art

[0002] With the rapid development of modern industry, equipment widely used in engineering vehicles, rail transit, aerospace and other fields, relying on advanced science and technology and manufacturing processes, is also rapidly developing towards automation and integration. The improvement of equipment integration and automation has led to more prominent cooperative operation and coupling between components, and more sensitivity to early and weak faults of components, which has also put forward higher requirements for the prediction and diagnosis of key transmission component faults. However, early fault signals are often submerged by complex noise and vibration environments. Traditional fault diagnosis methods rely largely on expert experience and manual analysis, which can no longer meet the requirements of timely analysis and extraction of fault features. Therefore, the early detection of fault sensitive features has become a key issue at present. If the change trend of sensitive features of some key faults can be effectively detected online in the early stage of fault derivation, the changes in the health status of the equipment can be captured and warned early, and then a more detailed analysis of fault type and location can be carried out to achieve timely fault diagnosis and prediction.

[0003] Online fault diagnosis methods are one of the topics that must be studied in order for fault diagnosis systems to mature in the future. In order to achieve real-time extraction and judgment of fault-sensitive information, whether this information can accurately, timely and stably reflect the evolution of faults is an important issue. However, since the pre-measured data cannot fully reflect all possible fault information and operating environment information of the monitoring equipment, it is impossible to build a complete model for fault-sensitive feature extraction in advance. In addition, once the existing deep network-based diagnostic model is determined, it cannot achieve rapid incremental expansion, resulting in the need for the original network to be re-optimized at a very high time and computational cost when a new fault occurs in the online equipment. Therefore, the real-time and effective update of the online fault diagnosis network is one of the problems that need to be solved urgently.

[0004] At the same time, the sensor collects monitoring data of the rotor system, and after preprocessing and fault feature extraction and identification by the diagnostic network, a judgment result that can reflect the type, location and even degree of the fault is obtained, completing the fault diagnosis task under big data. Although this type of intelligent diagnosis method is effective in some application scenarios, its applicability is greatly limited to the premise of sufficient labeled data. The total amount of labeled fault data is often difficult to meet the optimization requirements of existing high-performance deep diagnostic networks, while the unlabeled online monitoring data is massive. What's more serious is that the variability of the operating conditions and noise environment of rotor systems such as vehicle gearboxes far exceeds the scope that can be covered by labeled data sets. The distribution deviation of fault data caused by the complex and changeable operating conditions of online equipment can easily cause the performance degradation of conventional intelligent diagnosis methods. Therefore, it is very necessary to develop a stable intelligent fault diagnosis method that is highly robust to changes in equipment operating conditions.

[0005] Currently, no effective solution has been proposed for the problems in the related technologies. Summary of the invention

[0006] 1. Technical issues to be resolved

[0007] In view of the shortcomings of the prior art, the present invention provides a stable intelligent fault diagnosis method with strong robustness to changes in equipment operating conditions. Based on theories and new methods such as width learning, generative adversarial networks and transfer learning, and targeting the requirements for online monitoring of equipment health status, an unsupervised incremental extraction method for fault-sensitive features, a binary width incremental network that can be updated in real time, and a generative dynamic migration network are established. Ultimately, an online fault diagnosis method that can be stably diagnosed and updated in real time is formed, thereby obtaining higher agility and stability of intelligent fault diagnosis.

[0008] (II) Technical solution

[0009] In order to achieve the above advantages, the specific technical solution adopted by the present invention includes the following steps:

[0010] Step 1: Construct a density clustering sub-network to pre-label the unlabeled online monitoring data, establish a feature extraction main network with an incremental sparse pseudo-label matrix, and use the above high-confidence samples for training to gradually achieve unsupervised extraction of fault-sensitive features;

[0011] Step 2: Starting from the perspective of network fast calculation mode and network fast update architecture, we construct binary neural network and wide network that are suitable for fault diagnosis tasks respectively, and then organically integrate them to form a binary wide network for online real-time diagnosis, so as to realize real-time update of accurate diagnosis network.

[0012] Step 3: Extract the operating condition characteristics of the equipment and use them as input to construct an adversarial network that can generate rich and realistic target domains. Establish a high-mobility fault feature extraction network with dynamic adaptation of cross-domain feature distribution, and use the above generated samples for optimization to achieve stable extraction of high-mobility features.

[0013] Step 4: Rationally design faulty components and rotor system fault tests, carry out test data collection and analysis, and continuously verify and improve the proposed new theoretical methods using the actual measured signal data of the rotor system.

[0014] Furthermore, the feature extraction method of the incremental sparse pseudo-label matrix comprises the following steps:

[0015] The first step is to analyze the high-dimensional data of the unsupervised density clustering algorithm to achieve dimensionality reduction and shift-variant noise removal of high-dimensional input;

[0016] The second step is to integrate the Convolution 2000 pre-processing module and the unsupervised clustering algorithm to initially establish a convolutional unsupervised clustering subnetwork, cluster and extract high-confidence samples, and combine the health status membership conditional probability distribution vector of the extracted samples;

[0017] The third step is to arrange the extracted samples according to the probability distribution of health status membership, comprehensively consider the sample size, total feature dimension, and number of sample categories, establish a sparse pseudo-label matrix, parameterize it, and design its incremental expansion form;

[0018] The fourth step is to embed it into the constructed feature extraction main network, and combine it with feature sparsity optimization to promote the network to extract and regularly arrange sensitive features of various faults;

[0019] The fifth step is to cycle through the sub-network sample distillation, pseudo-label matrix reconstruction, and main feature extraction network training process until the overall network optimization converges;

[0020] In the sixth step, the main network is combined with a small number of labeled samples to establish a feedback adjustment mechanism for the clustering sub-network, to optimize the clustering performance of the sub-network, and to analyze and verify the network diagnostic performance through multiple data sets to achieve overall fine-tuning of the network and summarize key hyperparameter adjustment strategies.

[0021] Furthermore, the binary width network construction method for online real-time diagnosis comprises the following steps:

[0022] The first step is to construct a binary neural network that fits the monitoring signal, and design the optimal binarization strategy of the fault diagnosis network through multiple existing fault data sets.

[0023] The second step is to design a convolutional layer with adaptive step size, taking into account the requirements of fast convolution and signal translation invariance feature extraction, and integrate it into the binary neural network established above.

[0024] The third step is to establish a width network that matches the monitoring signal, and study the basic structure and optimization model of the width network, focusing on designing the vertical stacking expansion model of the width network.

[0025] The fourth step is to design a feature incremental learning mechanism for the wide network in combination with the incremental sparse pseudo-label matrix, including incremental expansion of the feature dimension to achieve fault identification accuracy and to accommodate sensitive features of new faults, and analyze its effectiveness and stability.

[0026] The fifth step is to conduct diagnostic task tests on existing typical simulated and measured fault signals to preliminarily determine the structure of the incremental width diagnostic network.

[0027] In the sixth step, the binary neural network and the width network are integrated, and the optimal binary width network structure that meets the online update conditions is determined in combination with the online update time threshold and storage cost limit of the diagnostic network.

[0028] Furthermore, the extraction of equipment operating condition characteristics comprises the following steps:

[0029] In the first step, a domain generalization strategy is generated to deal with the problem of invisible target domain data in actual online monitoring, and the specific role of this strategy in fault diagnosis tasks under variable equipment operating conditions is analyzed.

[0030] The second step is to study the changes in monitoring signals under different operating conditions of the equipment, extract the mapping relationship between the operating conditions and the monitoring signals, and realize the extraction of the operating condition style information contained in the monitoring data.

[0031] The third step is to construct a wide adversarial generative network to generate pseudo-variable operating condition data, and integrate fault category distinction, component independence, shared feature accommodation, and original sample approximation to achieve the generation of realistic pseudo samples.

[0032] The fourth step is to adjust the parameter values ​​in the working condition signal feature mapping relationship to form a rich working condition style, and use the obtained working condition style and known source domain data as the input of the wide adversarial generation network to generate pseudo target domain data under various working condition styles.

[0033] The fifth step is to address the problem of cross-domain feature distribution adaptation in migration networks. Starting from the perspective of information entropy, a self-balancing high-order KL divergence entropy is proposed. Based on this, marginal distribution adaptation with self-adjustment of matching weights and conditional distribution adaptation with smooth labels are constructed to form a joint distribution dynamic adaptation method that fits the dynamic learning process of the network.

[0034] The sixth step is to construct a width transfer network based on the Siamese structure. Through two branches, the source domain data and the pseudo target domain data under the new working condition style generated by the adversarial network are processed respectively. The above-mentioned joint distribution dynamic adaptation method is embedded to dynamically adapt the feature distribution of the two branches to achieve high transferability fault-sensitive feature extraction.

[0035] In the seventh step, the binary width diagnosis network and its incremental learning mechanism are integrated to accelerate the training of the generative width transfer network and enhance its ability to extract features of new fault types.

[0036] In the eighth step, the existing measured fault data set under variable working conditions is used as the test object to obtain feedback on the diagnosis effect, and the hyperparameters of the generative migration network are fine-tuned to finally form an online transferable fault diagnosis network.

[0037] Furthermore, the fault component and rotor system fault data collection and analysis comprises the following steps:

[0038] In the first step, the engine rotor system model is selected as the test carrier, and tests are carried out under different system operating conditions (stable operating conditions, speed fluctuations, load changes, etc.) to obtain a large amount of test data under the healthy state of the test bench.

[0039] The second step is to select and arrange sensors according to the fault data collection requirements, and collect corresponding signals for method verification, including setting vibration sensors for vibration signal collection, speed sensors and torque sensors for speed and load adjustment. It should be noted that in the target rotating component test, the speed and load data corresponding to the vibration signal should be collected at the same time for subsequent operation condition style extraction research.

[0040] The third step is to consider the typical failure modes of the target rotating parts (bearings, gears, shafts, etc.), and set the corresponding wear failures, crack failures and installation process failures such as loose support and misalignment in various parts of the rotating parts.

[0041] The fourth step is to consider the different derivative situations of the fault, set different fault levels of weak, mild and moderate for each fault, and study the relationship between the fault characteristics of these three in combination with the fault physical model of the corresponding components.

[0042] The fifth step is to set different stable speed and load conditions to form a fault data set under stable conditions; during the collection, the speed and load are adjusted quickly and widely through the inverter and magnetic powder brake to simulate the variable conditions and collect the corresponding data. Due to the high cost of the test and the preciousness of the data, it is also necessary to reduce the key parts through multi-sensor redundancy, especially the consequences caused by the looseness of the sensor magnetic base and cable connection.

[0043] The sixth step is to organize multiple sets of data for each research content based on the large amount of data obtained from the above experimental research, which are used to repeatedly test the diagnosis and prediction methods proposed in the above three aspects and improve the proposed methods, including verifying the accuracy, stability and real-time performance of fault diagnosis.

[0044] (III) Beneficial effects

[0045] Compared with the prior art, the present invention provides a stable intelligent fault diagnosis method with strong robustness to changes in equipment operating conditions, and has the following beneficial effects: based on theories and new methods such as width learning, generative adversarial networks, and transfer learning, and targeting the requirements for online monitoring of equipment health status, an unsupervised incremental extraction method for fault-sensitive features, a binary width incremental network that can be updated in real time, and a generative dynamic migration network are established, ultimately forming an online fault diagnosis method that can be stably diagnosed and updated in real time, thereby obtaining higher agility and stability of intelligent fault diagnosis. DETAILED DESCRIPTION

[0046] To further illustrate each embodiment, according to an embodiment of the present invention, a stable intelligent fault diagnosis method with strong robustness to changes in equipment operating conditions is provided, comprising the following steps:

[0047] Step 1: Construct a density clustering sub-network to pre-label the unlabeled online monitoring data, establish a feature extraction main network with an incremental sparse pseudo-label matrix, and use the above high-confidence samples for training to gradually achieve unsupervised extraction of fault-sensitive features;

[0048] Step 2: Starting from the perspective of network fast calculation mode and network fast update architecture, we construct binary neural network and wide network that are suitable for fault diagnosis tasks respectively, and then organically integrate them to form a binary wide network for online real-time diagnosis, so as to realize real-time update of accurate diagnosis network.

[0049] Step 3: Extract the operating condition characteristics of the equipment and use them as input to construct an adversarial network that can generate rich and realistic target domains. Establish a high-mobility fault feature extraction network with dynamic adaptation of cross-domain feature distribution, and use the above generated samples for optimization to achieve stable extraction of high-mobility features.

[0050] Step 4: Rationally design faulty components and rotor system fault tests, carry out test data collection and analysis, and continuously verify and improve the proposed new theoretical methods using the actual measured signal data of the rotor system.

[0051] In one embodiment, the feature extraction method of the incremental sparse pseudo-label matrix comprises the following steps:

[0052] The first step is to analyze the high-dimensional data of the unsupervised density clustering algorithm to achieve dimensionality reduction and shift-variant noise removal of high-dimensional input;

[0053] The second step is to integrate the Convolution 2000 pre-processing module and the unsupervised clustering algorithm to initially establish a convolutional unsupervised clustering subnetwork, cluster and extract high-confidence samples, and combine the health status membership conditional probability distribution vector of the extracted samples;

[0054] The third step is to arrange the extracted samples according to the probability distribution of health status membership, comprehensively consider factors such as the number of extracted samples, the total dimension of features, the number of sample categories, etc., establish a sparse pseudo-label matrix, parameterize it, and design its incremental expansion form;

[0055] The fourth step is to embed it into the constructed feature extraction main network, and combine it with feature sparsity optimization to promote the network to extract and regularly arrange sensitive features of various faults;

[0056] The fifth step is to cycle through the sub-network sample distillation, pseudo-label matrix reconstruction, and main feature extraction network training process until the overall network optimization converges;

[0057] In the sixth step, the main network is combined with a small number of labeled samples to establish a feedback adjustment mechanism for the clustering sub-network, to optimize the clustering performance of the sub-network, and to analyze and verify the network diagnostic performance through multiple data sets to achieve overall fine-tuning of the network and summarize key hyperparameter adjustment strategies.

[0058] In one embodiment, the binary width network construction method for online real-time diagnosis comprises the following steps:

[0059] The first step is to construct a binary neural network that fits the monitoring signal, and design the optimal binarization strategy of the fault diagnosis network through multiple existing fault data sets.

[0060] The second step is to design a convolutional layer with adaptive step size, taking into account the requirements of fast convolution and signal translation invariance feature extraction, and integrate it into the binary neural network established above.

[0061] The third step is to establish a width network that matches the monitoring signal, and study the basic structure and optimization model of the width network, focusing on designing the vertical stacking expansion model of the width network.

[0062] The fourth step is to design a feature incremental learning mechanism for the wide network in combination with the incremental sparse pseudo-label matrix, including incremental expansion of the feature dimension to achieve fault identification accuracy and to accommodate sensitive features of new faults, and analyze its effectiveness and stability.

[0063] The fifth step is to conduct diagnostic task tests on existing typical simulated and measured fault signals to preliminarily determine the structure of the incremental width diagnostic network.

[0064] In the sixth step, the binary neural network and the width network are integrated, and the optimal binary width network structure that meets the online update conditions is determined in combination with the online update time threshold and storage cost limit of the diagnostic network.

[0065] In one embodiment, the extracting equipment operating condition characteristics comprises the following steps:

[0066] In the first step, a domain generalization strategy is generated to deal with the problem of invisible target domain data in actual online monitoring, and the specific role of this strategy in fault diagnosis tasks under variable equipment operating conditions is analyzed.

[0067] The second step is to study the changes in monitoring signals under different operating conditions of the equipment, extract the mapping relationship between operating conditions (speed, load, etc.) and monitoring signals, and realize the extraction of operating condition style information contained in the monitoring data.

[0068] In the third step, a wide adversarial generative network is constructed to generate pseudo-variable operating condition data, and the optimization objectives such as fault category distinction, component independence, shared feature accommodation, and original sample approximation are integrated to achieve the generation of realistic pseudo samples.

[0069] The fourth step is to adjust the parameter values ​​in the working condition signal feature mapping relationship to form a rich working condition style, and use the obtained working condition style and known source domain data as the input of the wide adversarial generation network to generate pseudo target domain data under various working condition styles.

[0070] The fifth step is to address the problem of cross-domain feature distribution adaptation in migration networks. Starting from the perspective of information entropy, a self-balancing high-order KL divergence entropy is proposed. Based on this, marginal distribution adaptation with self-adjustment of matching weights and conditional distribution adaptation with smooth labels are constructed to form a joint distribution dynamic adaptation method that fits the dynamic learning process of the network.

[0071] The sixth step is to construct a width transfer network based on the Siamese structure. Through two branches, the source domain data and the pseudo target domain data under the new working condition style generated by the adversarial network are processed respectively. The above-mentioned joint distribution dynamic adaptation method is embedded to dynamically adapt the feature distribution of the two branches to achieve high transferability fault-sensitive feature extraction.

[0072] In the seventh step, the binary width diagnosis network and its incremental learning mechanism are integrated to accelerate the training of the generative width transfer network and enhance its ability to extract features of new fault types.

[0073] In the eighth step, the existing measured fault data set under variable working conditions is used as the test object to obtain feedback on the diagnosis effect, and the hyperparameters of the generative migration network are fine-tuned to finally form an online transferable fault diagnosis network.

[0074] In one embodiment, the fault component and rotor system fault data collection and analysis comprises the following steps:

[0075] In the first step, the engine rotor system model is selected as the test carrier, and tests are carried out under different system operating conditions (stable operating conditions, speed fluctuations, load changes, etc.) to obtain a large amount of test data under the healthy state of the test bench.

[0076] The second step is to select and arrange sensors according to the fault data collection requirements, and collect corresponding signals for method verification, including setting vibration sensors for vibration signal collection, speed sensors and torque sensors for speed and load adjustment. It should be noted that in the target rotating component test, the speed and load data corresponding to the vibration signal should be collected at the same time for subsequent operation condition style extraction research.

[0077] The third step is to consider the typical failure modes of the target rotating parts (bearings, gears, shafts, etc.), and set the corresponding wear failures, crack failures and installation process failures such as loose support and misalignment in various parts of the rotating parts.

[0078] The fourth step is to consider the different derivative situations of the fault, set different fault levels of weak, mild and moderate for each fault, and study the relationship between the fault characteristics of these three in combination with the fault physical model of the corresponding components.

[0079] The fifth step is to set different stable speed and load conditions to form a fault data set under stable conditions; during the collection, the speed and load are adjusted quickly and widely through the inverter and magnetic powder brake to simulate the variable conditions and collect the corresponding data. Due to the high cost of the test and the preciousness of the data, it is also necessary to reduce the key parts through multi-sensor redundancy, especially the consequences caused by the looseness of the sensor magnetic base and cable connection.

[0080] The sixth step is to organize multiple sets of data for each research content based on the large amount of data obtained from the above experimental research and improve the proposed method, including verifying the accuracy, stability and real-time performance of its fault diagnosis.

[0081] To sum up, with the help of the above-mentioned technical scheme of the present invention, the present invention is based on theories and new methods such as width learning, adversarial generative networks and transfer learning, and aims at the requirements of online monitoring of equipment health status. An unsupervised incremental extraction method of fault-sensitive features, a binary width incremental network that can be updated in real time, and a generative dynamic migration network are established, and finally an online fault diagnosis method that can be stably diagnosed and updated in real time is formed, thereby obtaining higher agility and stability of intelligent fault diagnosis.

[0082] In the present invention, unless otherwise clearly stipulated and limited, the terms such as "installation", "setting", "connection", "fixation" and "screw-on" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral one; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be the internal connection of two elements or the interaction relationship between two elements. Unless otherwise clearly defined, ordinary technicians in this field can understand the specific meanings of the above terms in the present invention according to the specific circumstances.

[0083] 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, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A stable intelligent fault diagnosis method with strong robustness to changes in equipment operating conditions, characterized by It includes the following detection steps: Step 1: Construct a density clustering sub-network to pre-label the unlabeled online monitoring data, and establish a feature extraction main network with an incremental sparse pseudo-label matrix. Further, the feature extraction method of the incremental sparse pseudo-label matrix includes the following steps: The first step is to analyze the high-dimensional data of the unsupervised density clustering algorithm to achieve dimensionality reduction and shift-variant noise removal of high-dimensional input; The second step is to integrate the Convolution 2000 pre-processing module and the unsupervised clustering algorithm to initially establish a convolutional unsupervised clustering subnetwork, cluster and extract high-confidence samples, and combine the health status membership conditional probability distribution vector of the extracted samples; The third step is to arrange the extracted samples according to the probability distribution of health status membership, comprehensively consider the sample size, total feature dimension, and number of sample categories, establish a sparse pseudo-label matrix, parameterize it, and design its incremental expansion form; The fourth step is to embed it into the constructed feature extraction main network, and combine it with feature sparsity optimization to promote the network to extract and regularly arrange sensitive features of various faults; The fifth step is to cycle through the sub-network sample distillation, pseudo-label matrix reconstruction, and main feature extraction network training process until the overall network optimization converges; In the sixth step, the main network is combined with a small number of labeled samples to establish a feedback adjustment mechanism for the clustering subnetwork, to optimize the clustering performance of the subnetwork, and to analyze and verify the network diagnostic performance through multiple data sets, to achieve overall network fine-tuning, to summarize key hyperparameter adjustment strategies, and to use the above high-confidence samples for training, to gradually achieve unsupervised extraction of fault-sensitive features; Step 2: Starting from the perspective of network fast calculation mode and network fast update architecture, we construct binary neural network and wide network that are suitable for fault diagnosis tasks respectively, and then organically integrate them to form a binary wide network for online real-time diagnosis, so as to realize real-time update of accurate diagnosis network. The binary width network construction method for online real-time diagnosis comprises the following steps: The first step is to construct a binary neural network that fits the monitoring signal; The second step is to design a convolutional layer with adaptive step size, taking into account the requirements of fast convolution and signal translation invariance feature extraction, and integrate it into the binary neural network established above; The third step is to establish a width network that matches the monitoring signal, and study the basic structure and optimization mode of the width network, focusing on designing the vertical stacking expansion mode of the width network; The fourth step is to design a feature incremental learning mechanism for the width network by combining the incremental sparse pseudo-label matrix; Step 5: Preliminarily determine the structure of the incremental width diagnosis network; Step 6: The binary neural network and the width network are integrated, and the optimal binary width network structure that meets the online update conditions is determined by combining the online update time threshold and storage cost limit of the diagnosis network. Step 3: Extract the equipment operating condition characteristics, use them as input to construct an adversarial network that can generate a rich and realistic target domain, and establish a high-mobility fault feature extraction network with dynamic adaptation of cross-domain feature distribution. Further, the extraction of equipment operating condition characteristics includes the following steps: The first step is to generate a domain generalization strategy to deal with the problem of invisible target domain data in actual online monitoring, and analyze the specific role of this strategy in fault diagnosis tasks under variable equipment operating conditions; The second step is to study the changes in monitoring signals under different operating conditions of the equipment, extract the mapping relationship between the operating conditions and the monitoring signals, and realize the extraction of operating condition style information contained in the monitoring data; The third step is to construct a wide adversarial generative network to generate pseudo-variable working condition data, and integrate fault category distinction, component independence, shared feature accommodation, and original sample approximation to achieve the generation of realistic pseudo samples; The fourth step is to adjust the parameter values ​​in the working condition signal feature mapping relationship to form a rich working condition style, and use the obtained working condition style and known source domain data as the input of the wide adversarial generation network to generate pseudo target domain data under various working condition styles; The fifth step is to solve the problem of cross-domain feature distribution adaptation in the migration network. Starting from the perspective of information entropy, a self-balancing high-order KL divergence entropy is proposed. Based on this, edge distribution adaptation with self-adjustment of matching weights and conditional distribution adaptation with smooth labels are constructed to form a joint distribution dynamic adaptation method that fits the dynamic learning process of the network. The sixth step is to construct a width transfer network based on the Siamese structure, and process the source domain data and the pseudo target domain data under the new working condition style generated by the adversarial network through two branches respectively, and embed the above-mentioned joint distribution dynamic adaptation method to dynamically adapt the feature distribution of the two branches to achieve high transferability fault sensitive feature extraction; Step 7: Integrate the binary width diagnosis network and its incremental learning mechanism to accelerate the training of the generative width migration network and enhance its ability to extract features of new fault types. In the eighth step, the existing measured fault data set under variable working conditions is used as the test object to obtain feedback on the diagnosis effect, and the hyperparameters of the generative migration network are fine-tuned to finally form an online transferable fault diagnosis network; The above generated samples are used for optimization to achieve stable extraction of high mobility features; Step 4: Rationally design faulty components and rotor system fault tests, carry out test data collection and analysis, and continuously verify and improve the proposed new theoretical methods using the actual measured signal data of the rotor system.

2. The stable intelligent fault diagnosis method with strong robustness to changes in equipment operating conditions according to claim 1 is characterized in that: The fault component and rotor system fault data collection and analysis comprises the following steps: In the first step, the engine rotor system model is selected as the test carrier to obtain a large amount of test data under the healthy state of the test bench; The second step is to select and arrange sensors according to the fault data collection requirements and collect corresponding signals for method verification; The third step is to consider the typical failure modes of the target rotating parts and set the corresponding wear failures, crack failures, and installation process failures such as loose support and misalignment in various parts of the rotating parts; The fourth step is to set different fault levels of weak, mild and moderate for each fault according to the different fault derivatives, and combine the fault physical model of the corresponding component to determine the fault characteristic relationship between the three. Step 5: Set different stable speed and load conditions to form a fault data set under stable conditions; The sixth step is to organize multiple sets of data for each research content based on the large amount of data obtained above, including verifying the accuracy, stability and real-time performance of its fault diagnosis.