Cross-working-condition fault diagnosis method and structure capable of explaining three-feature extractor network, and storage medium

Abstract

The invention discloses a cross-working-condition fault diagnosis method capable of explaining a three-feature extractor network, and the method comprises the following steps: inputting an obtained vibration signal to be diagnosed into a pre-trained explaining three-feature extractor migration network model, the vibration signals are subjected to feature extraction through an explainable feature extractor in the explainable three-feature extractor migration network model, and feature data are enabled to extract explainable shared features through a classifier of the explainable three-feature extractor migration network model by using a loss function; and performing cross-working-condition fault diagnosis processing according to the operation state information in the interpretable shared features to obtain a cross-working-condition fault diagnosis result. According to the cross-working-condition fault diagnosis method and structure capable of explaining the three-feature extractor network and the storage medium disclosed by the invention, an explaining migration task is realized by constructing a three-feature extractor strategy and a loss function.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical fault diagnosis, and particularly relates to a cross-condition fault diagnosis method, structure, and storage medium of an interpretable three-feature extractor network. Background Art

[0002] Fault diagnosis is a technology that collects, analyzes, and processes data of systems and equipment to achieve accurate diagnosis of potential faults. Conducting research on mechanical equipment fault diagnosis is of great significance. With the rapid development of artificial intelligence, fault diagnosis technology based on deep learning has been widely applied due to its excellent non-linear feature representation learning and fault diagnosis capabilities. Currently, in mechanical fault diagnosis, widely used representative deep neural networks include artificial neural networks, convolutional neural networks, and residual neural networks, etc. At the same time, many scholars have made improvements and innovations on specific fault diagnosis tasks based on representative methods to improve the accuracy of fault diagnosis. For example, for the fault diagnosis of unbalanced data of rotating mechanical equipment, a deep generalization learning system based on graph embedding has been proposed; for the fault incremental diagnosis task, a dual-branch adaptive aggregation residual network based on continuous learning has been proposed.

[0003] The deep neural network model can achieve intelligent diagnosis through its powerful learning ability. However, its own algorithm structure lacks reliable theoretical support, and there is no clear theoretical mapping relationship between internal layers. These traditional deep neural network models often lack interpretability and are called black-box models. To improve the interpretability of the deep neural network model, researchers have carried out research on constructing pre-explainable network models, including physical knowledge embedding models, functional framework embedding models, and algorithm structure equivalent models.

[0004] Compared with the physical knowledge embedding model and the functional framework embedding model, the algorithm structure equivalent model provides a more excellent scheme for constructing an interpretable network model. It uses a neural network to equivalent the analytical structure in traditional signal processing methods, is a completely interpretable network model, and has higher application potential.

[0005] Among them, algorithm unfolding is one of the most commonly used methods in constructing the algorithm structure equivalent model. Approximating and equating a traditional iterative algorithm with prior knowledge to a deep neural network can construct an efficient and interpretable deep network architecture. As a completely interpretable deep model, the algorithm structure equivalent model has been applied to multiple research fields, including mechanical fault diagnosis, image denoising, and structural damage detection, etc.

[0006] Although the structure equivalent model based on algorithms has achieved good results in mechanical fault diagnosis, its cross-domain transfer learning ability is relatively lacking and it cannot be applied to mechanical fault diagnosis under variable working conditions. Specifically, most of the existing interpretable network structures based on algorithms can only complete the fault diagnosis task under a single working condition. When the working condition of the mechanical system changes, the fault diagnosis ability of the model will decline and it cannot complete the transfer learning task. Summary of the Invention

[0007] The present invention overcomes the deficiencies of the prior art and provides a cross-condition fault diagnosis method, structure, and storage medium for an interpretable three-feature extractor network. By constructing a three-feature extractor strategy for separately extracting the shared features and respective private features of the source domain and the target domain, and designing a loss function for the transfer diagnosis task for effective training of the interpretable three-feature extractor transfer network, an interpretable transfer task is realized.

[0008] To achieve the above object, the technical solution adopted by the present invention is: a cross-condition fault diagnosis method for an interpretable three-feature extractor network, comprising the following steps: Obtain the vibration signal to be diagnosed; Input the obtained vibration signal to be diagnosed into a pre-trained interpretable three-feature extractor transfer network model. The vibration signal passes through the interpretable feature extractor in the interpretable three-feature extractor transfer network model. The interpretable feature extractor extracts feature data from the vibration signal through a three-feature extractor strategy established for the cross-domain transfer diagnosis problem. The extracted feature data passes through the feature adversarial module and classifier of the interpretable three-feature extractor transfer network model, and the constructed network can extract interpretable shared features by using a loss function. Cross-condition fault diagnosis processing is performed according to the operating state information in the interpretable shared features to obtain a cross-condition fault diagnosis result.

[0009] In a preferred embodiment of the present invention, the interpretable three-feature extractor transfer network model includes: Several groups of interpretable feature extractors for collecting vibration signals, one group for obtaining the private feature data of the target domain from different working conditions in the vibration signal, another group for obtaining the private feature data of the source domain from different working conditions in the vibration signal, and another group for obtaining the shared feature data of the target domain and the source domain; the private feature data of the target domain, the private feature data of the source domain, and the shared feature data of the target domain and the source domain pass through the feature adversarial module and classifier to obtain interpretable shared features by using a loss function.

[0010] In a preferred embodiment of the present invention, the three-feature extractor strategy of the interpretable feature extractor includes: Step S1, for the extraction of the deep state features x in the signal, a multi-layer sparse coding model is established. The multi-layer sparse coding model includes: ; Among them, and are respectively the learning dictionary and the matching sparse coding in the i-th layer, is the sparse regularization parameter in the i-th layer, and the state feature x can be transformed into the cascade structure of the L-layer dictionary; Step S2, using the alternating direction multiplier method and the fast iterative soft threshold algorithm to derive the iterative optimization solution algorithm of the multi-layer sparse coding model, and obtaining the iterative optimization solution algorithm of the multi-layer sparse coding model; Step S3, under the expansion theoretical framework of the iterative optimization solution algorithm, expand the iterative optimization solution algorithm along the number of times K and the number of layers L in two directions, construct the equivalent network form of the iterative optimization solution algorithm, and use it as an interpretable feature extractor.

[0011] In a preferred embodiment of the present invention, the state features in the data are extracted based on the iterative optimization solution algorithm of the multi-layer sparse coding model.

[0012] In a preferred embodiment of the present invention, the loss function includes a prediction loss and a feature adversarial loss. The algorithm of the loss function includes: ; where α is the weight ratio of the prediction loss function to the feature adversarial loss function; is the prediction loss function; is the feature adversarial loss function.

[0013] In a preferred embodiment of the present invention, The algorithm of includes: ; Among them, is the loss function generated by the confrontation between the shared features extracted from the source domain and the target domain, is the loss function generated by the confrontation between the shared features and the private features extracted from the source domain, is the loss function generated by the confrontation between the shared features and the private features extracted from the target domain, and β is the weight ratio of the dual-domain shared feature adversarial loss function to the adversarial loss function between the shared and private features within the domain.

[0014] In a preferred embodiment of the present invention, the training method of the interpretable three-feature extractor migration network model includes: Install a vibration sensor on the device to be tested, use a data acquisition card to collect vibration signals, and perform truncation, mean removal, and normalization processing on the vibration signals; obtain a vibration signal dataset; The data from different working conditions in the vibration signal dataset are divided into a source domain dataset and a target domain dataset. The source domain dataset and the target domain dataset are input into an interpretable three - feature extractor transfer network model for training. Among them, the source domain dataset has labels, while the target domain dataset has no labels. After training is completed in the interpretable three - feature extractor transfer network model, the target domain dataset is further input into the network, and the classifier can then determine the equipment operating status corresponding to the samples in the target domain dataset.

[0015] In a preferred embodiment of the present invention, an interpretable three - feature extractor network structure for implementing a cross - working - condition fault diagnosis method of an interpretable three - feature extractor network includes: An interpretable three - feature extractor transfer network model, where the interpretable three - feature extractor transfer network model includes several groups of interpretable feature extractors for collecting vibration signals. One group is used to obtain private feature data of the target domain from different working conditions in the vibration signal, another group is used to obtain private feature data of the source domain from different working conditions in the vibration signal, and there is also a group used to obtain shared feature data of the target domain and the source domain; The private feature data of the target domain, the private feature data of the source domain, and the shared feature data of the target domain and the source domain respectively pass through several feature adversarial modules and at least one classifier through corresponding 、 、 ,and to obtain interpretable shared features.

[0016] In a preferred embodiment of the present invention, the vibration signals in the collection of the vibration signal dataset are obtained from a device under test. The device under test includes a wheel - set bearing test bench, and the wheel - set bearing test bench includes: a motor, the rotating shaft of the motor is drivingly connected to a variable - load nut through a gearbox, the test end of the variable - load nut is connected to a dynamometer, the gearbox is drivingly connected to a rotating shaft, both ends of the rotating shaft are assembled with a carrier through bearings, and an accelerometer is installed on the bearings.

[0017] In a preferred embodiment of the present invention, a computer - readable storage medium stores computer programs / instructions, and when the computer programs / instructions are executed by a processor, they are used to implement the cross - working - condition fault diagnosis method of the interpretable three - feature extractor network structure described in claims 1 - 7.

[0018] The present invention solves the defects existing in the technical background, and the beneficial technical effects of the present invention are: In view of the problem that deep neural networks in the prior art cannot achieve interpretable cross-domain intelligent diagnosis, the present invention discloses a cross-condition fault diagnosis method, structure, and storage medium for an interpretable three-feature extractor network. In view of the interpretability problem, the present invention establishes a multi-layer sparse coding model, derives an iterative solution algorithm for the multi-layer sparse coding model, and obtains an equivalent network form of the solution algorithm through algorithm unfolding, which is used as the feature extractor of the interpretable three-feature extractor migration network. In addition, in view of the cross-domain migration diagnosis problem, a three-feature extractor strategy is constructed to respectively extract the shared features of the source domain and the target domain, as well as their respective private features, and a loss function for the migration diagnosis task is designed for the effective training of the interpretable three-feature extractor migration network, so as to achieve an interpretable migration task. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The present invention will be further described below in conjunction with the drawings and embodiments.

[0020] Figure 1 It is a schematic structural diagram of an interpretable feature extractor according to a preferred embodiment of the present invention; Figure 2 It is a schematic structural diagram of an interpretable three-feature extractor migration network according to a preferred embodiment of the present invention; Figure 3 It is a schematic flow diagram according to a preferred embodiment of the present invention; Figure 4 It is a migration task table of the wheel pair bearing dataset according to a preferred embodiment of the present invention; Figure 5 It is a data table of diagnostic accuracy rates of different methods in the migration task of the wheel pair bearing dataset according to a preferred embodiment of the present invention; Figure 6 It is a data table of diagnostic accuracy rates of the LBP-Net method in the migration task of the wheel pair bearing dataset; Figure 7 It is a data table of diagnostic accuracy rates of the ML-LISTA method in the migration task of the wheel pair bearing dataset; Figure 8 It is a data table of diagnostic accuracy rates of the DANN method in the migration task of the wheel pair bearing dataset; Figure 9 It is a data table of diagnostic accuracy rates of the Coral method in the migration task of the wheel pair bearing dataset; Figure 10 It is a data table of diagnostic accuracy rates of the MMD method in the migration task of the wheel pair bearing dataset; Figure 11 It is a data table of diagnostic accuracy rates of the method of the present invention in the migration task of the wheel pair bearing dataset; Figure 12 It is a schematic structural diagram of the wheel pair bearing test bench in the present invention; Figure 13 This is a photo of the wheel set bearing test bench in the present invention; Among them, 1 - motor, 2 - drive shaft, 3 - gearbox, 4 - variable load nut, 5 - dynamometer, 6 - rotating shaft, 7 - bearing, 8 - load platform, 9 - accelerometer. Specific embodiments

[0021] Now, the present invention will be further described in detail with reference to the accompanying drawings and embodiments. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0022] It should be noted that if there are directional indications (such as up, down, bottom, top, etc.) involved in the embodiments of the present invention, then such directional indications are only used to explain the relative positional relationship and movement conditions between components in a specific posture. If this specific posture changes, then such directional indications will also change accordingly. The terms "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of such features. Unless otherwise clearly specified and limited, the terms "set", "connected", and "connected to" should be understood in a broad sense. For example, it may be a fixed connection, a detachable connection, or an integral connection; it may be directly connected, or indirectly connected through an intermediate medium, and it may be the internal communication of two components. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances. Embodiment 1

[0023] A cross - operating condition fault diagnosis method for an interpretable three - feature extractor network, comprising the following steps: Obtain the vibration signal to be diagnosed; Input the obtained vibration signal to be diagnosed into a pre - trained interpretable three - feature extractor transfer network model. The vibration signal passes through the interpretable feature extractor in the interpretable three - feature extractor transfer network model. The interpretable feature extractor extracts feature data from the vibration signal through a three - feature extractor strategy established for cross - domain transfer diagnosis problems. The extracted feature data is used by the feature adversarial module and classifier of the interpretable three - feature extractor transfer network model to enable the constructed network to extract interpretable shared features, and cross - operating condition fault diagnosis processing is performed according to the operating state information in the interpretable shared features to obtain the cross - operating condition fault diagnosis result. Embodiment 2

[0024] A cross - operating condition fault diagnosis method for an interpretable three - feature extractor network, comprising the following steps: Obtain the vibration signal to be diagnosed; Input the obtained vibration signal to be diagnosed into a pre-trained interpretable three-feature extractor transfer network model. The vibration signal passes through the interpretable feature extractor in the interpretable three-feature extractor transfer network model. The interpretable feature extractor extracts feature data from the vibration signal through a three-feature extractor strategy established for cross-domain transfer diagnosis problems. The extracted feature data is passed through the feature adversarial module and classifier of the interpretable three-feature extractor transfer network model, and the loss function is used to enable the constructed network to extract interpretable shared features. Cross-condition fault diagnosis processing is performed based on the operating state information in the interpretable shared features to obtain a cross-condition fault diagnosis result.

[0025] Among them, the interpretable three-feature extractor transfer network model includes: Several groups of interpretable feature extractors for collecting vibration signals, where one group is used to obtain private feature data of the target domain from different conditions in the vibration signal, another group is used to obtain private feature data of the source domain from different conditions in the vibration signal, and there is also a group used to obtain shared feature data of the target domain and the source domain; the private feature data of the target domain, the private feature data of the source domain, and the shared feature data of the target domain and the source domain obtain interpretable shared features through the feature adversarial module and classifier using the loss function. Embodiment III

[0026] A cross-condition fault diagnosis method for an interpretable three-feature extractor network includes the following steps: Obtain the vibration signal to be diagnosed; Input the obtained vibration signal to be diagnosed into a pre-trained interpretable three-feature extractor transfer network model. The vibration signal passes through the interpretable feature extractor in the interpretable three-feature extractor transfer network model. The interpretable feature extractor extracts feature data from the vibration signal through a three-feature extractor strategy established for cross-domain transfer diagnosis problems. The extracted feature data is passed through the feature adversarial module and classifier of the interpretable three-feature extractor transfer network model, and the loss function is used to enable the constructed network to extract interpretable shared features. Cross-condition fault diagnosis processing is performed based on the operating state information in the interpretable shared features to obtain a cross-condition fault diagnosis result.

[0027] Among them, the interpretable three-feature extractor transfer network model includes: Several groups of interpretable feature extractors for collecting vibration signals, where one group is used to obtain private feature data of the target domain from different working conditions in the vibration signal, another group is used to obtain private feature data of the source domain from different working conditions in the vibration signal, and there is another group used to obtain shared feature data of the target domain and the source domain; the private feature data of the target domain, the private feature data of the source domain, and the shared feature data of the target domain and the source domain obtain interpretable shared features through the feature adversarial module and the classifier using the loss function.

[0028] Among them, the three-feature extractor strategy of the interpretable feature extractor includes: Step S1, for the extraction of the deep state feature x in the signal, a multi-layer sparse coding model is established, and the multi-layer sparse coding model includes: ; Among them, and are the learning dictionary and the matching sparse coding in the i-th layer respectively, is the sparse regularization parameter in the i-th layer, and the state feature x can be transformed into a cascaded structure of L-layer dictionaries; specifically, while ensuring the signal fidelity, the multi-layer sparse coding model ensures the global sparsity of the multi-layer sparse coding model by adding sparsity constraints in each layer.

[0029] Step S2, using the alternating direction multiplier method and the fast iterative soft threshold algorithm to derive the iterative optimization solution algorithm of the multi-layer sparse coding model, and obtaining the iterative optimization solution algorithm of the multi-layer sparse coding model; Step S3, under the expansion theoretical framework of the iterative optimization solution algorithm, expand the iterative optimization solution algorithm along the number K and the number of layers L two directions, construct an equivalent network form of the iterative optimization solution algorithm, and use it as the interpretable feature extractor. As Figure 1 shown, in this example, the number K and the number of layers L are set to 4 and 8 respectively. However, it is not limited to this, and in other embodiments, the parameter settings of the number K and the number of layers L can be adjusted according to actual usage requirements.

[0030] Specifically, the state features in the data are extracted based on the iterative optimization solution algorithm of the multi-layer sparse coding model.

[0031] Specifically, the iterative optimization solution algorithm of the multi-layer sparse coding model includes: .

[0032] In the algorithm, represents a vector containing N elements, N+ represents the set of positive integers, represent the learning dictionary, the penalty parameter, and the bias parameter in the i-th layer respectively.

[0033] Specifically, as Figure 2 shown, the loss function includes a prediction loss and a feature adversarial loss, and the algorithm of the loss function includes: ; where α is the weight ratio of the prediction loss function to the feature adversarial loss function; is the prediction loss function, is to reduce the distance between the predicted label and the true label according to the cross-entropy loss, thereby improving the ability of the network to extract specific fault features; is the prediction loss function, is the feature adversarial loss function, is to reduce the distribution difference between the shared features through the binary cross-entropy loss and the corresponding discriminator, and increase the distribution difference between the extracted shared and private features.

[0034] Specifically, the algorithm of includes: Among them, is the loss function generated by the confrontation between the shared features extracted from the source domain and the target domain, is the loss function generated by the confrontation between the shared features and private features extracted from the source domain, is the loss function generated by the confrontation between the shared features and private features extracted from the target domain, and β is the weight ratio of the dual-domain shared feature adversarial loss function to the adversarial loss function between the shared and private features within the domain. In this example, the weight parameter is set to .

[0035] Referring to Figure 5 shown, 5 comparison methods (LBP-Net, ML-LISTA, DANN, Coral, and MMD) and the state recognition accuracy of the method of this embodiment in 12 migration tasks are listed. It can be seen that the accuracy of the method of this embodiment is lower than 95% in only one task, and the accuracy of multiple tasks exceeds 98%. Its average accuracy is improved by 3.53% compared with the best-performing comparison method LBP-Net. For the five comparison methods, the accuracy of individual tasks is lower than 90%, and even lower than 85%. Their performance is far inferior to the method of the present invention.

[0036] In addition, Figures 6 - 11 gives the confusion matrix of the interpretable cross-domain diagnosis task result of the wheel pair bearing dataset migration task Q9. It can be seen that the probability of misdiagnosis of the method of this embodiment in different state recognitions is less than that of other comparison methods. Example 4

[0037] Based on Example 3, the training method of the interpretable three-feature extractor migration network model includes: Install a vibration sensor on the device under test, collect vibration signals using a data acquisition card, and perform truncation, mean removal, and normalization on the vibration signals; obtain a vibration signal dataset. Divide the data from different working conditions in the vibration signal dataset into a source domain dataset and a target domain dataset, and input the source domain dataset and the target domain dataset into an interpretable three-feature extractor transfer network model for training, where the source domain dataset has labels and the target domain dataset has no labels; after training is completed in the interpretable three-feature extractor transfer network model, input the target domain dataset further into the network, and the classifier can then determine the device operating state corresponding to the samples in the target domain dataset.

[0038] In this example, as Figure 4 shown, for the transfer task Q1: A®B, it means training the network using the source domain dataset A and the target domain dataset B simultaneously, where the source domain dataset A has labels and the target domain dataset B has no labels. After the network training is completed, input the target domain dataset B further into the network, and the classifier can then determine the device operating state corresponding to each sample in the target domain dataset B. The transfer tasks Q2 to Q12 are as Figure 4 shown, and the description of transfer task Q1 is referred to in turn by analogy. Example Five

[0039] An interpretable three-feature extractor network structure for implementing the cross-working-condition fault diagnosis method of the interpretable three-feature extractor network in Example Three, including: An interpretable three-feature extractor transfer network model, where the interpretable three-feature extractor transfer network model includes several groups of interpretable feature extractors for collecting vibration signals, one group for obtaining private feature data of the target domain from different working conditions in the vibration signal, another group for obtaining private feature data of the source domain from different working conditions in the vibration signal, and another group for obtaining shared feature data of the target domain and the source domain; The private feature data of the target domain, the private feature data of the source domain, and the shared feature data of the target domain and the source domain are respectively passed through several feature adversarial modules and at least one classifier through the corresponding , , , and to obtain interpretable shared features. Example Six

[0040] An interpretable three-feature extractor network structure for implementing the cross-working-condition fault diagnosis method of the interpretable three-feature extractor network in Example Three, including: An interpretable three-feature extractor migration network model, the interpretable three-feature extractor migration network model comprising a plurality of groups of interpretable feature extractors for collecting vibration signals, one group for obtaining private feature data of a target domain from different working conditions in the vibration signal, another group for obtaining private feature data of a source domain from different working conditions in the vibration signal, and another group for obtaining shared feature data of the target domain and the source domain; The private feature data of the target domain, the private feature data of the source domain, and the shared feature data of the target domain and the source domain are respectively processed by a plurality of feature adversarial modules and at least one classifier through corresponding , , ,as well as Obtain interpretable shared features.

[0041] Among them, the vibration signal in the vibration signal data set collection is obtained from the tested equipment, and the tested equipment includes a wheelset bearing test bench, and the wheelset bearing test bench includes: a motor 1, a driving shaft 2 of the motor 1 is connected to a variable load nut 4 through a gear box 3, and a test end of the variable load nut 4 is connected to a dynamometer 5. The gear box 3 is driven and connected with a rotating shaft 6, and both ends of the rotating shaft 6 are assembled with a carrier 8 through bearings 7, and an accelerometer 9 is installed on the bearing 7.

[0042] More specifically, in this example, the data is from Figure 12 , Figure 13 The data was collected from the wheelset bearing test bench shown in the figure. The bearing type used in the wheelset bearing test bench is model NJ204ET NSK. The applied load size is controlled by adjusting the nut of the variable load nut. The tested bearings contain two types of faults: inner ring fault and outer ring fault, and the size of each fault is set to 0.2mm, 0.4mm and 0.6mm respectively. Therefore, together with the fault-free state, this data set contains a total of 7 health states. For each bearing in each health state, it is tested under four loads of 0kN, 0.8 kN, 1.6 kN and 2.4 kN. Different loads have different distribution forms in the collected data. The data collected under each load is taken as a domain. Therefore, the data set contains 4 domains, and 12 migration tasks are set to verify the effectiveness of the proposed method, such as Figure 4 Specifically, the data under four loads of 0 kN, 0.8 kN, 1.6 kN and 2.4 kN are defined as domains A, B, C and D respectively. For the data under each working condition, they are truncated, de-meaned and normalized, and the truncation length is set to 1024. Embodiment 7

[0043] A computer-readable storage medium stores a computer program / instructions thereon. When the computer program / instructions are executed by a processor, they are used to implement the cross-condition fault diagnosis method of the interpretable three-feature extractor network structure in Embodiment III. Embodiment VIII

[0044] The method of Embodiment III can be designed as mechanical fault diagnosis software and installed in a host computer. The host computer is connected to a signal acquisition device. The host computer processes the signal in real time according to the above steps to diagnose the fault in time and determine the fault type, ensuring the operation safety of the entire mechanical system.

[0045] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, and the combination of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0046] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device, and the instruction device implements the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0047] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process. Thus, the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0048] Working principle: The present invention discloses a cross-condition fault diagnosis method, structure, and storage medium for an interpretable three-feature extractor network. In view of the interpretability problem, the present invention establishes a multi-layer sparse coding model, derives an iterative solution algorithm for the multi-layer sparse coding model, unfolds the algorithm to obtain an equivalent network form of the solution algorithm, and uses it as the feature extractor of the interpretable three-feature extractor migration network; in addition, in view of the cross-domain migration diagnosis problem, a three-feature extractor strategy is constructed to separately extract the shared features of the source domain and the target domain, as well as their respective private features, and a loss function for the migration diagnosis task is designed for the effective training of the interpretable three-feature extractor migration network, so as to achieve an interpretable migration task.

[0049] The cross-condition fault diagnosis method, structure, and storage medium for the interpretable three-feature extractor network provided by the present invention. The interpretable three-feature extractor migration network constructed by the present invention, on the one hand, embeds the inherent physical knowledge in the multi-layer sparse coding model into the network design through algorithm unfolding, and on the other hand, establishes a three-feature extractor strategy for the cross-domain migration diagnosis problem, so that the constructed network can extract interpretable shared features, thereby realizing interpretable mechanical equipment migration diagnosis under variable working conditions.

[0050] The above specific implementation manners are specific supports for the solution idea proposed by the present invention, and the protection scope of the present invention cannot be limited thereby. Any equivalent change or equivalent modification made on the basis of the technical solution of the present invention according to the technical idea proposed by the present invention still belongs to the protection scope of the technical solution of the present invention.

Claims

1. A cross-condition fault diagnosis method for an interpretable three-feature extractor network, characterized in that Including the following steps: Obtain the vibration signal to be diagnosed; Input the obtained vibration signal to be diagnosed into a pre-trained interpretable three-feature extraction transfer network model. The vibration signal passes through the interpretable feature extractor in the interpretable three-feature extraction transfer network model. The interpretable feature extractor extracts feature data from the vibration signal through a three-feature extraction strategy established for cross-domain transfer diagnosis problems. The extracted feature data enables the constructed network to extract interpretable shared features through the feature adversarial module and classifier of the interpretable three-feature extraction transfer network model using a loss function, and performs cross-condition fault diagnosis processing based on the operating state information in the interpretable shared features to obtain a cross-condition fault diagnosis result.

2. The cross-condition fault diagnosis method of the interpretable three-feature extractor network according to claim 1, characterized in that: The interpretable three-feature extraction transfer network model includes: Several groups of interpretable feature extractors for collecting vibration signals, where one group is used to obtain private feature data of the target domain from different conditions in the vibration signal, another group is used to obtain private feature data of the source domain from different conditions in the vibration signal, and there is also a group used to obtain shared feature data of the target domain and the source domain; the private feature data of the target domain, the private feature data of the source domain, and the shared feature data of the target domain and the source domain obtain interpretable shared features through the feature adversarial module and classifier using a loss function.

3. The cross-condition fault diagnosis method of the interpretable three-feature extractor network according to claim 2, characterized in that: The three-feature extraction strategy of the interpretable feature extractor includes: Step S1, for the extraction of deep state features x in the signal, a multi-layer sparse coding model is established, and the multi-layer sparse coding model includes: ; Among them, and are the learning dictionary and the matching sparse coding in the i-th layer respectively, is the sparse regularization parameter in the i-th layer, and the state feature x can be transformed into the cascaded structure of the L-layer dictionary; Step S2: Derive an iterative optimization solution algorithm for the multi-layer sparse coding model using the alternating direction method of multipliers and the fast iterative soft threshold algorithm to obtain the iterative optimization solution algorithm for the multi-layer sparse coding model; Step S3, under the expansion theoretical framework of the iterative optimization solution algorithm, expand the iterative optimization solution algorithm along the number of times K and the number of layers L in two directions, construct an equivalent network form of the iterative optimization solution algorithm, and use it as an interpretable feature extractor.

4. The cross-condition fault diagnosis method of the interpretable three-feature extractor network according to claim 3, characterized in that: Extract the state features in the data based on the iterative optimization solution algorithm of the multi-layer sparse coding model.

5. The cross-condition fault diagnosis method of the interpretable three-feature extractor network according to claim 2, wherein: The loss function includes a prediction loss and a feature adversarial loss, and the algorithm of the loss function includes: where α is the weight ratio of the prediction loss function to the feature adversarial loss function; is the prediction loss function; is the feature adversarial loss function.

6. The cross-condition fault diagnosis method of the interpretable three-feature extractor network according to claim 5, characterized in that: The algorithm includes: ; Among them, is a loss function generated by the confrontation between the shared features extracted from the source domain and the target domain, is a loss function generated by the confrontation between the shared features and the private features extracted from the source domain, is a loss function generated by the confrontation between the shared features and the private features extracted from the target domain, and β is the weight ratio of the dual-domain shared feature adversarial loss function to the adversarial loss function between the shared and private features within the domain.

7. The cross-condition fault diagnosis method of the interpretable three-feature extractor network according to claim 6, characterized in that: The training method of the interpretable three-feature extraction transfer network model includes: Install a vibration sensor on the device to be tested, use a data acquisition card to collect vibration signals, and perform truncation, mean removal, and normalization processing on the vibration signals; obtain a vibration signal dataset; Divide the data from different conditions in the vibration signal dataset into a source domain dataset and a target domain dataset, and input the source domain dataset and the target domain dataset into the interpretable three-feature extraction transfer network model for training, where the source domain dataset has labels and the target domain dataset has no labels; after training is completed in the interpretable three-feature extraction transfer network model, input the target domain dataset further into the network, and the classifier can determine the device operating state corresponding to the samples in the target domain dataset.

8. An interpretable three-feature extractor network structure, characterized in that, A cross-condition fault diagnosis method for implementing the interpretable three-feature extraction network described in any one of claims 1-7, including: An interpretable three-feature extraction transfer network model, which includes several groups of interpretable feature extractors for collecting vibration signals, where one group is used to obtain private feature data of the target domain from different conditions in the vibration signal, another group is used to obtain private feature data of the source domain from different conditions in the vibration signal, and there is also a group used to obtain shared feature data of the target domain and the source domain; The private feature data of the target domain, the private feature data of the source domain, and the shared feature data of the target domain and the source domain respectively pass through a number of feature adversarial modules and at least one classifier through the corresponding , , , and to obtain interpretable shared features.

9. The interpretable three-feature extractor network structure according to claim 8, characterized in that: The vibration signal in the vibration signal data set collection is obtained from the tested device, and the tested device includes a wheelset bearing test bench, and the wheelset bearing test bench includes: a motor, a driving shaft of the motor is transmission-connected to a variable load nut through a gear box, a test end of the variable load nut is connected to a dynamometer, a rotating shaft is drivingly connected to the gear box, two ends of the rotating shaft are assembled with a carrier through bearings, and an accelerometer is installed on the bearing.

10. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instruction is executed by a processor, it is used to implement the cross-operating condition fault diagnosis method of the interpretable three-feature extractor network described in claims 1-7.

Images